Machine Learning For Planetary Science Ebook PDF

Machine Learning for Planetary Science
- eBook PDF
Visit to download the full and correct content document:
https://ebooksecure.com/download/machine-learning-for-planetary-science-ebook-pdf
/
Machine Learning for Planetary Science
- eBook PDF
Visit to download the full and correct content document:
https://ebooksecure.com/download/machine-learning-for-planetary-science-ebook-pdf
/
MACHINE
LEARNING FOR
PLANETARY
SCIENCE
This page intentionally left blank
MACHINE
LEARNING FOR
PLANETARY
SCIENCE
Edited by
JOERN HELBERT
MARIO D’AMORE
MICHAEL AYE
HANNAH KERNER
Elsevier
Radarweg 29, PO Box 211, 1000 AE Amsterdam, Netherlands
The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom
50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States
Copyright © 2022 Elsevier Inc. All rights reserved.
MATLAB® is a trademark of The MathWorks, Inc. and is used with permission.

The MathWorks does not warrant the accuracy of the text or exercises in this book.
This book’s use or discussion of MATLAB® software or related products does not constitute
endorsement or sponsorship by The MathWorks of a particular pedagogical approach or particular
use of the MATLAB® software.
No part of this publication may be reproduced or transmitted in any form or by any means,
electronic or mechanical, including photocopying, recording, or any information storage and
retrieval system, without permission in writing from the publisher. Details on how to seek
permission, further information about the Publisher’s permissions policies and our arrangements
with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency,
can be found at our website: www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by the
Publisher (other than as may be noted herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and experience
broaden our understanding, changes in research methods, professional practices, or medical
treatment may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in
evaluating and using any information, methods, compounds, or experiments described herein. In
using such information or methods they should be mindful of their own safety and the safety of
others, including parties for whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors,
assume any liability for any injury and/or damage to persons or property as a matter of products
liability, negligence or otherwise, or from any use or operation of any methods, products,
instructions, or ideas contained in the material herein.
ISBN: 978-0-12-818721-0
For information on all Elsevier publications

visit our website at https://www.elsevier.com/books-and-journals
Publisher: Candice Janco

Acquisitions Editor: Peter J Llewellyn
Editorial Project Manager: Andrea R. Dulberger
Production Project Manager: R. Vijay Bharath
Designer: Victoria Pearson
Typeset by VTeX
Contents
Contributors ix
Foreword xv
References xvi
1. Introduction to machine learning 1

Hannah Kerner, Joseph Campbell, and Mark Strickland
1.1. Overview of machine learning methods 1
1.2. Supervised learning 3
1.3. Unsupervised learning 6
1.4. Semisupervised learning 9
1.5. Active learning 12
1.6. Popular machine learning methods 13
1.7. Data set preparation 21
References 22
2. The new and unique challenges of planetary missions 25

Jörn Helbert and Mario D’Amore
2.1. Introduction 25
References 35
3. Finding and reading planetary data 39

Michael Aye
3.1. Data acquisition in planetary science 39
4. Introduction to the Python Hyperspectral Analysis Tool (PyHAT) 55

J.R. Laura, L.R. Gaddis, R.B. Anderson, and I.P. Aneece
4.2. PyHAT library architecture 57
4.3. PyHAT orbital 60
4.4. PyHAT in-situ 72
4.5. Conclusion 84
Acronyms 86
Acknowledgments 87
References 87
5. Tutorial: how to access, process, and label PDS image data for
machine learning 91
v
vi Contents
Steven Lu, Kiri L. Wagstaff, Rafael Alanis, Gary Doran, Kevin Grimes, and
Jordan Padams
5.2. Access to PDS data products 92
5.3. Preprocessing PDS data products into standard image formats 97
5.4. Labeling image data 101
5.5. Example PDS image classifier results 106
5.6. Summary 109
Acknowledgments 109
References 109
6. Planetary image inpainting by learning mode-specific regression

models 111
Hiya Roy, Subhajit Chaudhury, Toshihiko Yamasaki, and Tatsuaki Hashimoto
6.2. Related works 113
6.3. Experimental data 114
6.4. Proposed method 115
6.5. Network architecture 116
6.6. Experimental results 120
6.7. Conclusion 127
References 128
7. Automated surface mapping via unsupervised learning and

classification of Mercury Visible–Near-Infrared reflectance spectra 131
Mario D’Amore and Sebastiano Padovan
7.2. Mercury and the MASCS instrument 133
7.3. Data preparation 134
7.4. Learning from multivariate data 136
Acknowledgment 147
References 147
8. Mapping storms on Saturn 151

Ingo P. Waldmann
8.2. Exploratory principal component analysis 152
8.3. A deep learning approach 155
8.4. Saturn’s feature map 163
References 166
9. Machine learning for planetary rovers 169

Contents vii
Masahiro Ono, Brandon Rothrock, Yumi Iwashita, Shoya Higa,

Virisha Timmaraju, Sami Sahnoune, Dicong Qiu, Tanvir Islam, Annie Didier,
Christopher Laporte, Deegan Atha, Vivian Sun, Kyohei Otsu, Mike Paton,
Olivier Lamarre, Shreyansh Daftry, R. Michael Swan, Adam Stambouli,
Flynn Chen, Bhavin Shah, Kathryn Stack, and Chris Mattman
9.2. Risk- and resource-aware AutoNav 173
9.3. Drive-by-science 181
9.4. Demonstration on a test rover 187
9.5. Conclusion and future work 188
References 189
10. Combining machine-learned regression models with Bayesian

inference to interpret remote sensing data 193
Saverio Cambioni, Erik Asphaug, and Roberto Furfaro
10.1. The need for accurate fast-forward functions 193
10.2. Bayesian approach to inverse problems 193
10.3. Machine-learning-based surrogate models 195
10.4. Case study: constraining the thermal properties of asteroids with
surrogate models 196
10.5. Future perspectives for data fusion 200
References 204
Index 209
Contributors
Rafael Alanis
Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA,
United States
R.B. Anderson
U.S. Geological Survey, Astrogeology Science Center, Flagstaff, AZ, United States
I.P. Aneece
U.S. Geological Survey, Western Geographic Science Center, Flagstaff, AZ, United
States
Erik Asphaug
Lunar and Planetary Laboratory, University of Arizona, Tucson, AZ, United States
Deegan Atha
United States
Michael Aye
Laboratory for Atmospheric and Space Physics, University of Colorado, Boulder,
CO, United States
Saverio Cambioni
Department of Earth, Atmospheric & Planetary Sciences, Massachusetts Institute
of Technology, Cambridge, MA, United States
Joseph Campbell
Carnegie Mellon University, Pittsburgh, PA, United States
Subhajit Chaudhury
The University of Tokyo, Department of Information and Communication
Engineering, Tokyo, Japan
Flynn Chen
Yale University, New Haven, CT, United States
United States
ix
x Contributors
Shreyansh Daftry
United States
Mario D’Amore
German Aerospace Center (DLR), Berlin, Germany
Annie Didier
United States
Gary Doran
United States
Roberto Furfaro
Systems and Industrial Engineering Department, University of Arizona, Tucson,
AZ, United States
L.R. Gaddis
Universities Space Research Association, Lunar and Planetary Institute, Houston,
TX, United States
Kevin Grimes
United States
Tatsuaki Hashimoto
The University of Tokyo, Department of Electrical Engineering and Information
Systems, Tokyo, Japan
Jörn Helbert
Shoya Higa
United States
Tanvir Islam
United States
Contributors xi
Yumi Iwashita
United States
Hannah Kerner
University of Maryland, College Park, MD, United States
Olivier Lamarre
University of Toronto, Toronto, ON, Canada
United States
Christopher Laporte
United States
J.R. Laura
Steven Lu
United States
Chris Mattman
United States
R. Michael Swan
United States
Masahiro Ono
United States
Kyohei Otsu
United States
Jordan Padams
United States
xii Contributors
Sebastiano Padovan
EUMETSAT, Darmstadt, Germany
WGS, Darmstadt, Germany
Mike Paton
United States
Dicong Qiu
Carnegie Mellon University, Pittsburgh, PA, United States
United States
Brandon Rothrock
United States
Hiya Roy
The University of Tokyo, Department of Electrical Engineering and Information
Systems, Tokyo, Japan
Sami Sahnoune
United States
Bhavin Shah
United States
Kathryn Stack
United States
Adam Stambouli
United States
Mark Strickland
Arizona State University, Tempe, AZ, United States
Contributors xiii
Vivian Sun
United States
Virisha Timmaraju
United States
Kiri L. Wagstaff
United States
Ingo P. Waldmann
University College London, London, United Kingdom
Toshihiko Yamasaki
The University of Tokyo, Department of Information and Communication
Engineering, Tokyo, Japan
Foreword
Humans have used robots to explore other planets in our solar system since
the beginning of planetary exploration. In the Apollo era, robotic space-
craft were sent to the Moon ahead of humans to collect critical information
about lunar surface properties and landing sites. Today, scientists perform
detailed scientific analyses and field geology on the surface of Mars us-
ing the Perseverance and Curiosity rovers, robotic vehicles equipped with
cameras, spectrometers, and other instruments. Today, humans are actively
exploring planets and other bodies in the solar system with more than 20
robotic spacecraft, with more to come in the next decade [1]. As humans
send more robotic explorers into the solar system carrying increasingly so-
phisticated instruments, each observation contains even more information
for scientists to analyze and increases the volume of archived data from
planetary exploration missions. This is evidenced in Fig. 0.1, which shows
the first close-up image of Mars ever taken in 1965 alongside an image of
Mars taken nearly 50 years later in 2014. Analyzing the enormous volumes
of data returned by past, present, and future planetary exploration missions
will require scientists to adopt a different kind of robot—machine learning,
a subfield of artificial intelligence which learns patterns, perceptions, and
predictions from data in an automated way.
Machine learning is a topic that spans a broad range of methods, mod-
els, learning types, and machine behaviors. In planetary science, machine
learning can be used to facilitate scientific discovery and analysis in multi-
ple ways: by uncovering patterns or features of interest in large, complex
datasets that are difficult for humans to analyze; by inspiring new hypotheses
based on structure and patterns revealed in data; or by automating tedious
or time-consuming tasks. The goal of this book is to provide a bridge be-
tween the communities of machine learning and planetary science to enable
increased uptake of machine learning methods in planetary science, and
improve the accessibility of planetary science data for the machine learn-
ing community. In the first chapter, we will cover the basics of machine
learning, special considerations for applying machine learning to planetary
science datasets, guidelines for implementing machine learning models, and
resources for the reader to find a deeper understanding of machine learning
methods if desired. In the second chapter, we will cover the types of data
and challenges that are encountered in planetary science. The third chapter
xv
xvi Foreword
Figure 0.1 Left: first close-up image of Mars acquired by the Mariner 4 spacecraft on
July 15, 1965. Credit: NASA/JPL-Caltech. Right: HiRISE image of impact crater on Mars
acquired on January 9, 2014. Credit: NASA/JPL-Caltech/UA.
provides tutorials for accessing and preparing planetary science data sets for
machine learning applications. The final chapter presents several case stud-
ies detailing how machine learning has been implemented for a variety of
planetary science applications and data types.
References
[1] E. Lakdawalla, Where we are, The Planetary Report (2019) 22–23.
CHAPTER 1
Introduction to machine learning

Hannah Kernera , Joseph Campbellb , and Mark Stricklandc
a University of Maryland, College Park, MD, United States
b Carnegie Mellon University, Pittsburgh, PA, United States
c Arizona State University, Tempe, AZ, United States
1.1 Overview of machine learning methods

A widely used definition of machine learning (ML) comes from Carnegie
Mellon University professor Tom Mitchell: “A computer program is said
to learn from experience E with respect to some class of tasks T and per-
formance measure P, if its performance at tasks in T, as measured by P,
improves with experience E,” or more simply, “machine learning is the
study of computer algorithms that improve automatically through expe-
rience” [21]. In other words, machine learning involves using computer
programs to learn some task based on experience (in the form of data or
observations) where how well that task is learned is measured by some per-
formance metric. For example, your email spam filter is a machine learning
system: a computer program learns to classify emails as spam or not spam by
observing millions of example emails that were known to be either spam or
not spam, and how well it has learned to do this is measured by the percent-
age of correctly classified emails (the accuracy). Much of machine learning
is considered in the optimization context, where the learning task is de-
fined as learning the parameters θ (weights or coefficients) of a function
y = f (x, θ ) (where x is the input feature or features) by optimizing an ob-
jective function C (y, y ) where y is the correct value or label (e.g., spam or
not spam) and y is the predicted value or label [21]. The objective function
C (y, y ) (also called a loss function) compares y to y and quantifies how
close the prediction is to the correct value.
Machine learning describes a subset of methods within the broader field
of artificial intelligence (AI). Computer vision is also a subarea of AI that
describes methods for perception (including recognition and understand-
ing) of images, videos, or other visual content. There is substantial overlap
between computer vision and machine learning since machine learning
methods (such as convolutional neural networks) are often employed for
computer vision tasks (such as detecting craters in a satellite image of Mars),
Machine Learning for Planetary Science Copyright © 2022 Elsevier Inc.
https://doi.org/10.1016/B978-0-12-818721-0.00007-0 All rights reserved. 1
2 Machine Learning for Planetary Science
though computer vision may also draw from nonML methods (e.g., SIFT,
or scale-invariant feature transform [20]). The focus of this book is on the
broader category of machine learning methods applied to planetary science,
which includes but is not limited to computer vision tasks.
Machine learning methods are typically divided into the following sub-
types:
(1) Supervised learning
(2) Unsupervised learning
(3) Semisupervised learning
Each of these methods represents a different type of learning based on
what data is available as input to the learning model. In supervised learn-
ing, the data provided includes both a “feature vector” and a “label.” The
feature vector is a value (or set of values) intended to represent the observ-
able feature(s), for a particular data sample. The label, in turn, is intended
to represent a correct output for that sample. For example, one sample may
consist of a feature vector in the form of an image (represented as a vector
of pixel values) along with a label in the form of an integer that represents
the category or “class” that describes the image. Fig. 1.1 illustrates an ex-
ample input for classifying whether an image contains a crater or not: the
input features are the pixels from the image and the class label is 1 if the
image contains a crater and 0 if it does not. This is an example of a type
of supervised learning called classification, in which the machine learn-
ing model maps input features to discrete output classes. The second type
supervised learning is called regression, in which the model maps input
features to a continuous output value (i.e., a real number). As an example,
a model could be trained to predict the eccentricity of a crater in an im-
age: the input features are the pixel values and the output is the real-valued
eccentricity of the crater (Fig. 1.1).
In unsupervised learning, the input data is similar but no labels are
available for training. In the following subsections, we will discuss methods
for supervised and unsupervised learning in more detail. In semisuper-
vised learning, models make use of both labeled and unlabeled data
samples during training—in many applications there are far more unlabeled
examples available than labeled examples. This is especially true in planetary
science applications, partly because labels are more difficult to acquire than
traditional machine learning applications because labeling may require do-
main knowledge or may not be clearly defined. For many planetary science
applications, there may be very little data–labeled or unlabeled–requiring
Introduction to machine learning 3
Figure 1.1 Example input and output pair for supervised classification (top) and regres-
sion (bottom).
specialized methods. Section 1.4 will cover methods for dealing with small
labeled datasets, including semisupervised learning.
1.2 Supervised learning

Supervised learning refers to the set of machine learning techniques in
which the model learns from input examples that contain an associated label
for every example. There are two main categories of supervised learning
techniques: classification, in which the model prediction is a discrete class
or label, and regression, in which the model prediction is continuous (i.e.,
a real number).
1.2.1 Classification
Though supervised classification problems have inspired the majority of
machine learning solutions, many real-world applications require significant
effort to compile a training dataset since the data samples are usually labeled
by humans. For example, in order to train a model to classify whether an
image contains a crater or not (Fig. 1.1), one must first assemble a dataset
containing hundreds or thousands of examples of images and an associated
label of “crater” or “noncrater.” Techniques for labeling planetary science
datasets are discussed in more detail in Chapter 5 while techniques for
Figure 1.2 Classifier as function.
making use of a small number of labeled examples are discussed in Section

1.4. Classification involves predicting a class that corresponds to a given
input. The learning model is essentially a function that maps the input
feature values to the output class value, as illustrated in Fig. 1.2.
One of the simplest classifiers is a “linear classifier,” in which the func-
tion that maps input to output is a linear function. For example, a simple
linear function of the input that distinguishes two classes is:
y(x) = W T x + w0 (1.1)
where x is the input feature vector, W is called the weight matrix, and
w0 is called the bias. For this function, a “decision rule” can be used to
distinguish the classes:

C1 if y(x) ≥ 0
y= (1.2)
C2 otherwise
The “decision boundary” is therefore defined by the linear relation y(x) =

0, which is a hyperplane of dimension N − 1 where N is the dimensionality
of the input features. One way to illustrate a linear classifier is shown in
Fig. 1.3a for the case of a 2-dimensional feature vector. The blue (dark
gray in print version) circles represent data samples from class C1 while the
red (mid gray in print version) squares are data samples from class C2 . The
magenta (light gray in print version) line represents the decision boundary
at y(x) = 0 that separates the classes.
Although the boundary shown in Fig. 1.3a is a simple line, the data
samples that correspond to each class may not be linearly separable (e.g.,
Fig. 1.3b). In such cases, models that are capable of finding nonlinear de-
cision boundaries (e.g., support vector machines and other kernel-based
methods, or neural networks) or different feature extraction methods that
enable the classes to be linearly separable are required.
A common obstacle in developing supervised classifiers is overfitting,
which occurs when the classification model parameters are designed to
Figure 1.3 Linear decision boundary separating data samples from two classes.
Figure 1.4 Example of overfitting.
perform particularly well for the training data, but the model performs rel-
atively poorly when tested with data unseen during training (e.g., testing
or validation data). Fig. 1.4 shows an example of a decision boundary for
which overfitting has occurred. This problem can occur when the model
parameters are specifically tailored to the training samples used rather than
the underlying distribution of the samples. In this example, a higher-order
polynomial was chosen to separate the classes C1 and C2 (magenta (light
gray in print version) decision boundary), but the underlying distribution
likely could be approximated using a simpler (e.g., linear) decision bound-

ary and fewer model parameters.
1.2.2 Regression
Regression involves predicting a real number from a continuous range of
possible values that corresponds to the input features. Regression can be
considered analogous to curve-fitting: given a particular input, a regression
model predicts a corresponding output that best fits the data used to train
the model. One example of regression is the prediction of a weather-related
parameter such as barometric pressure given a feature vector that includes
values for temperature, humidity, and wind velocity. Another example of
regression is prediction of stock or bond prices over time.
The regression model can take the form of a linear or nonlinear func-
tion of the input feature vector. The process of selecting a regression model
includes choosing the order of the model. Following the curve-fitting anal-
ogy, selecting a regression model is analogous to choosing the order of a
polynomial that fits the training data samples. If a polynomial of order one
is chosen, the result is a linear regression model. If a quadratic or higher-
order polynomial is chosen, the result is a nonlinear regression model.
For example, if we use linear regression to model a training dataset con-
sisting of a number of data samples with corresponding continuous labels,
we are essentially fitting the following linear model to the training data
samples:
f (x) = w1 x + w0
where w1 and w0 are model parameters, also called regression coefficients.
Here, x is the input feature vector and y(x) is the output predicted value
(or label). For a particular dataset with training data samples, the model
parameters or regression coefficients can be determined using commonly-
used methods, e.g., least squares, which attempt to find coefficients that
minimize error between predicted value and the given label value.
1.3 Unsupervised learning

Unsupervised Learning refers to ML techniques in which the dataset con-
sists of unlabeled data samples. An unsupervised learning model may be
used to group similar data samples together (clustering) or map the input
feature vector to an alternate (i.e., simplified or more compact) representa-
tion (dimensionality reduction). These applications are described in detail
in this section. Other types of unsupervised learning applications include

anomaly detection and probability density estimation.
In general, different types of learning models are used for unsupervised
learning than those described in the previous section for supervised learn-
ing. This is because inputs for unsupervised learning models do not have
corresponding labels, so the model must optimize a different target during
learning—e.g., the likelihood of samples under a probability distribution.
Unsupervised learning models can provide insights into the underlying
structure of a dataset (e.g., identifying classes of minerals in a dataset of ther-
mal infrared spectra), transform high-dimensional data samples into salient
features that can be used as input for supervised techniques, and inspire new
hypotheses based on patterns detected in planetary datasets.
1.3.1 Clustering
Clustering involves organizing data samples in a way that groups simi-
lar samples together. For example, a clustering algorithm could be used
to group similar spectra observed by a spacecraft instrument and visualize
the patterns of mineral formation on the surface. Chapter 7 describes an
example of this for Mercury. These resulting groups or “clusters” often
demonstrate patterns or commonalities that may not be immediately ap-
parent without more extensive analysis of the samples. Because the number
of groups is a finite number (often determined beforehand), the output of
a clustering learning model is an identification of each group, along with
an assignment of each data sample to one of the groups. Popular clustering
methods include k-means clustering and spectral clustering.
In k-means clustering, a distance metric is chosen (e.g., Euclidean
distance) and used to determine the distance between each pair of data
samples. The Euclidean distances are used to partition the data samples
into k clusters, in which the within-cluster sum of squared distances is
minimized while also ensuring that each cluster contains at least one data
sample. Although this type of clustering can be computationally expensive,
approximate algorithms have been developed that iterate to local minima
with lower computational cost.
Spectral clustering involves the generation of a similarity matrix relat-
ing each data sample to each other data sample (e.g., the inverse of the
distance between samples). The eigenvalues and eigenvectors of the Lapla-
cian of the similarity matrix are computed and then a clustering algorithm
(e.g., K-means) is applied to the eigenvectors. Spectral clustering can be
more computationally efficient than other clustering methods depending
Figure 1.5 Example of Spectral Clustering.
on the sparseness of the similarity matrix. It can also result in improved

cluster definitions for contiguous groups of data samples. See Fig. 1.5 for
an example.
1.3.2 Dimensionality reduction

Many methods do not perform well when data samples are high-
dimensional. This is referred to as the “curse of dimensionality.” In these
cases, methods to automatically reduce the dimensionality of data can be
used to transform high-dimensional data samples into more compact rep-
resentations that capture the most salient features in the data. This reduced
representation can then be used as the input to supervised or unsupervised
algorithms or to visualize high-dimensional data samples. At the core of di-
mensionality reduction is the manifold hypothesis, which is the theory that
high-dimensional data actually lie on a lower-dimension manifold which
can be represented by the sample’s latent representation. For example, sup-
pose we have a dataset containing 100 × 100-pixel images of faces, where
each pixel can have an integer value between 0-255. The dimensionality of
samples in this dataset is 100 × 100 = 10,000. However, the intrinsic dimen-
sionality of these data is not 10,000. We know this because if we were to
randomly sample from the distribution of all possible 100 × 100 pixel values,
we would not get a face. Thus, the data must lie on a lower-dimensional
manifold. Dimensionality reduction methods learn a mapping from high-

dimensional data to lower-dimensional latent representations. Common
algorithms for dimensionality reduction include principal component anal-
ysis (PCA), t-distributed Stochastic Neighbor Embedding (t-SNE), and
autoencoder neural networks.
1.4 Semisupervised learning

The previous sections introduced both supervised and unsupervised learn-
ing paradigms, which exploit labeled data and unlabeled data respectively, in
order to learn underlying latent information about the dataset. However, in
planetary science it is often the case that labeled data is not plentiful enough
to train a sufficiently accurate supervised learning model. This can be due
to many factors, for example,
• Obtaining data for bodies other than Earth is extremely difficult and
requires planning multiyear missions with tight constraints on what
sensors can be used and where they can be employed.
• Creating accurate labels, such as image annotations for another plane-
tary body, is time consuming and requires a high degree of expertise
(making it difficult to utilize crowd-sourcing and citizen science ap-
proaches).
• Labels cannot be physically validated with on-site field surveys (bar-
ring exceptional circumstances), so low-confidence annotations must
be treated with care.
• Communication is bandwidth- and power-constrained for bodies fur-
ther away from Earth, limiting the amount of data that can be physically
transmitted back. For example, the Cassini orbiter transmitted data at a
rate between 14-166 kb/s once beyond the inner solar system.
While some of these issues might result in a small amount of labeled
and unlabeled data, for now we concern ourselves with the more common
case that we have a small set of labeled data and a larger set of unlabeled
data. One way to approach this problem is to combine elements of both
supervised and unsupervised learning in a complementary fashion, resulting
in the aptly-named semisupervised learning.
Suppose we have a dataset consisting of surface images of a planetary
body and we want to classify whether or not each image contains a re-
curring slope lineae. This is a binary classification problem and so a subset
of our images have been annotated by experts as either y = 1 (contains an
Figure 1.6 The labeled data points as produced by self-training: (left) initially, (middle)
after half of the points have been processed, (right) after all of the points have been
processed.
anomaly) or y = 0 (does not). However, given the time and expertise re-
quired to perform annotation, our labeled set of images is a small subset
of the total image set available to us. Semisupervised learning provides an
approach for exploiting what we know about the unlabeled data in order
to use it to improve our classifier.
1.4.1 Self-training
In the simplest case, we train a classifier over the labeled set of images and
use that to predict labels for the remaining unlabeled images. We then take
the most confident predictions and add them to the labeled data set. The
classifier is re-trained and we repeat this process until all images have been
assigned labels. This process is visualized in Fig. 1.6. By repeatedly building
up the labeled data set from its own predictions, the classifier is in a sense
training itself – hence this is known as self-training [33].
The choice of which predicted label/image pairs to add to the training
set in each iteration is typically based on confidence (thus requiring a prob-
abilistic interpretation of the output label), although other criteria such as
selection metrics based on distance have also been employed [27].
This is a simple, yet flexible method which allows the user to specify
their own classification model. The caveat is that we assume the classifier’s
predictions are actually correct. If this is not the case and the classifier makes
an incorrect prediction which is then added to the labeled data set, the
algorithm can diverge from the true class labels by “doubling down” on its
own mistakes, leading to worse classification results [35].
1.4.2 Self-training with Expectation Maximization

Alternatively, we can attempt to estimate the parameters for a generative
model over the labeled and unlabeled data, which models the joint proba-
bility distribution p(x, y) over input variables x and target variables y. This
generative model is assumed to be a mixture distribution [25] or a classifica-
tion model based on one, e.g., naive Bayes [23]. While it is straightforward
to estimate the parameters for such a distribution with a maximum likeli-
hood estimation approach [5] given the labeled data set, this is only a noisy
estimate of the true distribution. We would like to use the large amount
of unlabeled data that may be available in addition to the labeled data to
improve our estimate of the model parameters, however, the unlabeled data
on its own only allows us to estimate the marginal probability p(x) over all
class labels. In order to integrate the unlabeled data into our estimate of the
joint distribution p(x, y), we adopt an iterative two-step process in which
we 1) estimate class label probabilities for each unlabeled sample given our
current best estimate of the distribution parameters, and 2) improve our
estimate of the distribution parameters given the new labels. This process
repeats until the parameters converge and is known as Expectation Max-
imization [10], for which we treat the unknown labels as latent variables
that are to be estimated along with the parameters.
As in regular self-training, the EM-based version iteratively trains itself
over time; the difference lies in how unlabeled data is used in training. In
self-training only a subset of the predictions (the most confident) are in-
corporated into the labeled data set at each step. However, in self-training
with EM, all predictions are weighted by their expectation and immediately
incorporated into training. Thus, self-training can be thought of as an incre-
mental algorithm and self-training with EM as an iterative algorithm. While
EM is well-studied and has been applied to many different applications, it
requires strong assumptions to be made about the underlying distribution of
the data that cause performance degradation when violated; namely, that it
is generated by the proposed mixture model and that there is a relationship
between the model components and the data classes [22].
1.4.3 Cotraining
In a slightly more complex strategy, we can instead opt to subdivide the data
into two distinct feature sets in a procedure called cotraining [6]. Intuitively,
these feature sets represent two different ways – or views – of looking at
the same data. Consider the anomaly detection classification scenario from
before and suppose that we have two input features associated with our
images: pixel color and topographic elevation denoted as xc and xe respec-
tively. One way of looking at the data is that pixel color alone is indicative
of whether there is an anomaly or not, and so we can train a classifier Cc to
predict if there is an anomaly based solely on xc with the labeled data set. A
different, complementary view is that topographic elevation is sufficient to
indicate the presence of an anomaly, and we train a second classifier Ce over
this view using the input feature xe . The fundamental idea behind cotrain-
ing is that each classifier can identify complementary information about the
unlabeled data set that can be used to build up the other, in a fashion sim-
ilar to self-training. The color classifier Cc is used to predict labels for the
unlabeled data set and a subset of these predictions is selected—often the
most confident predictions—and added to the labeled data set. Similarly,
Ce predicts labels for the unlabeled instances and also selects a subset to be
added. Then both classifiers are re-trained over the new labeled data and
this iterative process repeats. Without diving into formal theory, the reason
that this works is that both classifiers must agree on the predicted labels for
the unlabeled data. When considering the hypothesis space of all models,
the net effect is that we prune models that do not agree, thus reducing the
overall size of the space [35]. As in self-training, this method allows flexibil-
ity in the choice of the classification model, however, there are assumptions
which must hold in order to provably learn [22]: a) each view must be suf-
ficient to predict the class label on its own and the two views must agree
over most of the labels, and b) the views must be conditionally independent
given the class label.
1.5 Active learning

Semisupervised learning seeks to exploit what we know about the un-
labeled data in order to improve our model performance from a limited
set of labeled samples. There is an opposite, yet complementary approach
known as active learning which seeks to exploit what we don’t know about
the unlabeled data [29]. A subset of unlabeled samples is queried for the
class label; intuitively, if we assume that the act of labeling samples is a
time-consuming task then if we can carefully select which samples would
be most informative then we can limit the number of labeled samples that
are required.
1.5.1 Uncertainty sampling

A simple strategy for selecting which samples to label is to query the labels
for the instances in which the classifier is least confident about its pre-
diction. This strategy is referred to as uncertainty sampling [19]. This is an
iterative method in which a classifier is first trained on the labeled data set,
predicts labels for the unlabeled samples, then queries labels for the sam-
ples in which it is least confident (e.g., whose predictions are closest to
the decision boundary). These newly labeled samples are then labeled and
added to the labeled data set and the process is repeated until all unlabeled
samples can be confidently predicted. Uncertainty sampling can be seen
as the active learning counterpart to self-training—in uncertainty sampling
we are querying labels for the least confident predictions rather than adding
the most confident as in self-training. The choice of classification model is
constrained to those that are capable of providing a probabilistic output or
a similar measure of uncertainty.
1.5.2 Query-by-committee
A more complex approach called query-by-committee [30] is to partition the
model hypothesis space by training multiple classification models over the
labeled data set. Each classifier then predicts labels for the unlabeled sam-
ples and we query labels for the samples on which the predictions do not
agree. Two assumptions are made by this method: a) that we can partition
the hypothesis space with multiple classification models, and b) that these
models disagree on a subset of unlabeled samples. As in cotraining, we seek
to reduce the size of the hypothesis space to narrow down the list of can-
didate models. However, while cotraining gradually builds up the models
such that they always agree, here we achieve the same effect by querying
for labels in regions where they disagree.
1.6 Popular machine learning methods

While there are myriad models and algorithms to choose from when im-
plementing the approaches discussed in this chapter, there is a much smaller
subset of more commonly used algorithms. We will summarize common
methods in this section to provide a starting point for those looking to ap-
ply machine learning and data science techniques for planetary science. We
illustrate these methods using an example dataset of HiRISE grayscale or-
bital images of Mars labeled as crater, bright dune, dark dune, slope streak,
impact ejecta, swiss cheese, spider, or other [12]. This dataset is also used in
Chapter 5 and Wagstaff et al. [32]—see Fig. 5.7 in Chapter 5 for an example
image in each class. We excluded the ‘other’ class to simplify our visualiza-
tions since this class has substantially higher intra-class variance than other
classes. The python code used to create the visualizations in this chapter is
available at https://github.com/hannah-rae/planetary-ml-book.
In general, the choice of which model to use will depend on a variety of
factors specific to the application, such as number of training labels, dimen-
sionality of data samples, computational speed or storage requirements, or
interpretability. Scikit-learn, a popular machine learning library for python,
provides a flowchart to help users choose the right algorithm for their ap-
plication: https://scikit-learn.org/stable/tutorial/machine_learning_map/
index.html. Readers may also find Raghu & Schmidt [24] helpful for
understanding if a neural network approach, and which type of neural net-
work, is best suited for a scientific application.
1.6.1 Principal component analysis

Principal component analysis (PCA) is a dimensionality reduction method
used to project data to a lower-dimensional space. PCA is widely used in
planetary science—for example, Chapter 8 uses PCA for exploratory data
analysis of hyperspectral image observations of Saturn from the Cassini mis-
sion. PCA defines a linear projection of the data onto a principal subspace
that retains maximal variance in the data. The principal components are the
eigenvectors of the data covariance matrix, which can be computed using
singular value decomposition (SVD). PCA is commonly used to reduce the
dimensionality of high-dimensional data sets for feature extraction, visual-
ization, or other purposes. The number of principal components to retain
(often referred to as the variable k) is a hyperparameter that can be cho-
sen or tuned. A common approach to choosing k is to determine when the
fraction of total variance explained by the k components starts to plateau for
higher values of k. Fig. 1.7 shows a plot of the percentage of total variance
explained by k principal components in the HiRISE dataset for k ∈ [0, 20]
and the dataset projected onto the 2-dimensional principal subspace.
1.6.2 K-means clustering

K-means is a popular unsupervised clustering algorithm that tries to as-
sign data points to k clusters such that each sample belongs to the cluster
whose mean is closest to the sample value. The standard algorithm begins
Figure 1.7 Left: percentage of total variance explained by k principal components for
k ∈ [0, 20]. Right: HiRISE dataset projected onto principal subspace (k = 2).
by randomly initializing k cluster centroids and iteratively performing two

steps: 1) assigning points to the centroid with the nearest cluster mean,
and 2) re-computing each cluster mean (i.e., the mean of all samples as-
signed to the cluster). The algorithm converges and stops updating cluster
assignments when the cluster assignment is no longer changing with each
update. The Euclidean distance is often used as the distance metric for
finding the nearest centroid. The number of clusters k is a tunable hy-
perparameter, which could be chosen based on domain knowledge (e.g.,
number of minerals likely to be present if clustering spectra) or such that
within-cluster variance is minimized. Because the Euclidean distance tends
to be an ineffective metric for high-dimensional data sets, K-means is fre-
quently combined with PCA or other dimensionality reduction algorithms
so that clustering is performed in the low-dimensional space. Chapter 7
uses K-means clustering with dimensionality reduction and manifold learn-
ing techniques to map similar classes of surface composition on the surface
of Mercury using data from the MESSENGER mission. Fig. 1.8 shows the
clusters identified using the k-means algorithm with k = 7. While some
clusters appear to correlate with a class label when comparing to Fig. 1.7
(e.g., dark dune and cluster 2), overall the cluster assignments do not map
directly to the class labels, as is often the case for high-dimensional (e.g.,
image) data sets.
1.6.3 Support vector machines

Support vector machines (SVMs) [5] are a supervised learning method that
finds the hyperplane (or set of hyperplanes) in the n-dimensional feature
space (where n is the number of input features) which maximizes the
distance to the nearest training samples from each class. Maximizing this
Figure 1.8 K-means clusters of PCA-reduced HiRISE data set.
Figure 1.9 SVM decision boundary (solid gray line) for two classes in the HiRISE data
set with two different values of hyperparameter C. The dashed lines show the margin
and the circled points are the support vectors.
margin ensures that the decision boundary will be as generalizable as pos-

sible to new, unseen data points. SVMs can be used for both classification
and regression. The main hyperparameter to choose is the regularization
parameter C, which represents the strength of the penalty incurred during
training for allowing samples to be closer to the margin boundary (since a
perfect decision boundary is not attainable for most problems). SVM also
uses a kernel function K to map samples to a higher dimensional space (this
is referred to as the “kernel trick” [5]); the radial basis function (RBF) is a
commonly chosen kernel function. Fig. 1.9 visualizes the decision bound-
ary found by SVM on the HiRISE data set with PCA using a linear kernel
(we reduced the data set to two classes to simplify the visualization). The
Figure 1.10 Decision tree with maximum depth of 2 learned from two classes of the
HiRISE data set. Box colors represent the majority class and the Gini impurity (lighter
colors represent higher impurity) for each node.
two plots illustrate the effect of smaller (C = 0.01) and larger (C = 100)
values of C on the decision boundary.
1.6.4 Decision trees and random forests

Decision trees [8] are a supervised learning method used for both classi-
fication and regression that predict a target variable by learning a set of
rules that progressively narrow down the decision space. These rules are
inferred from the training data and have a simple “if-then-else” structure.
For classification, the model learns rules that split the data in such a way
that the “impurity” (e.g., measured by Gini impurity) of training samples
is minimized. The impurity refers to the degree to which each branch of a
split contains more than one class. For regression, a scalar loss metric such
as Mean Squared Error or Mean Absolute Error is minimized. The maxi-
mum depth of the tree is a hyperparameter that can be tuned: the higher
the depth of the tree, the more complex and susceptible to overfitting the
model is. Fig. 1.10 shows the decision tree learned from the HiRISE data
set maximum depth set to 3. We used only two classes and statistics fea-
tures (mean, standard deviation, minimum, maximum, skew, and kurtosis)

instead of PCA to simplify the visualization and provide interpretability.
The random forest method [7] is an extension of decision trees meant to
combat overfitting that is typically experienced with decision trees in prac-
tice. A random forest is essentially an ensemble of decision trees in which
each tree is learned from a random subset of samples from the training set.
Each sample is drawn with replacement. This is referred to as “bootstrap-
ping.” Since the parameters of each decision tree are very sensitive to the
training samples, sampling with replacement allows many different samples
to be drawn from the same training data set to create many different trees
that are ensembled in the final prediction. The final random forest predic-
tion is obtained by averaging or taking a vote of the predictions from all
of the trees. Tunable hyperparameters include the number of trees, max-
imum tree depth, and number of samples drawn for each tree. Random
forests often show good performance in practice and are commonly used
in scientific applications such as remote sensing [3].
1.6.5 Neural networks

A neural network consists of layers of composable nonlinear modules
(“neurons”) that each transform the data representation from the previous
layer into a new representation that is more abstract than at the previous
layer. Every neural network contains an input layer, one or more hidden
layers, and an output layer; the network is called “deep” if it contains two
or more hidden layers. Inspired by neurons firing in human brains, the
output from each neuron is also called the “activation” and computes the
following function:
a = h(W T X + b) (1.3)
where h(·) is a nonlinear function (called the activation function), X is a
tensor containing outputs from neurons in the previous layer (or the net-
work input, in the case of the first hidden layer), W is a tensor containing
weights wijk representing the kth weight connecting layer i to layer j (pur-
ple arrows in Fig. 1.11), and b is a tensor of biases. During training, the
parameters W and b are continually updated using a standard procedure
called backpropagation such that a cost function C (·), also called the loss
function, is minimized.
By learning many layers of hierarchical representations from input data,
deep neural networks are able to discover complex structures in high-
dimensional data (e.g., images) that ignore irrelevant variations while fo-
Figure 1.11 Example feed-forward neural network with neurons shown in black and
weights in purple.
cusing on subtle but important variations [18]. The primary difference be-
tween deep learning or neural network techniques compared to traditional
machine learning approaches is that traditional methods typically require
manual specification of how to transform raw data (e.g., pixel values) into
feature representations suitable for the learning algorithm, requiring signif-
icant engineering effort and usually domain expertise. In neural networks,
the representation does not need to be designed by human engineers; in-
stead, it is learned automatically from the data using a general-purpose
procedure. Another important difference is that traditional algorithms, such
as support vector machines and random forests, typically require 1D vectors
as input, which requires that multidimensional samples (such as multispec-
tral images) be flattened into 1D vectors or transformed into a feature
vector, thus removing spatial or local context. On the other hand, methods
that can ingest multidimensional tensors and preserve the spatial topology
of inputs such as convolutional neural networks can learn 2D or higher
dimensional patterns in the data.
The primary disadvantage of deep learning compared to traditional ma-
chine learning techniques is that they require large labeled datasets (thou-
sands or even millions of examples, depending on the architecture) for
tuning the network parameters and their complex architectures can make it
difficult to interpret the representations and decision functions learned by
the network. A common approach for applications with smaller data sets is
to pretrain a neural network on a large generic data set such as ImageNet
[11] and then fine-tune the network on the smaller domain-specific data
Figure 1.12 Relative rankings of popular classification models on the basis of accuracy
and interpretability.
set. Fine-tuning consists of initializing a network with weights from the

pretraining data set and further tuning the weights, often in the final or
penultimate layers only, based on a different data set with different output
categories. The effectiveness of fine-tuning relies on the assumption shown
in prior work that the representations learned by a neural network trained
on a large dataset are generic and can be useful for classification tasks that
may be very different from the original task [26]. Chapter 5 provides a tu-
torial for fine-tuning a neural network on the HiRISE data set that was
pretrained on the ImageNet data set.
Another important limitation of neural networks is that they are less
interpretable than traditional ML methods. Fig. 1.12 shows the relative
trade-offs between popular classification methods on the axes of accuracy
and interpretability. While neural networks tend to have higher accu-
racy than traditional methods, they are the least interpretable. Conversely,
if/then rules are highly interpretable but not very accurate. These trade-
offs are important to consider because in many scientific and real-world
applications, it may be desirable to sacrifice some accuracy for more in-
terpretability. Developing methods for interpreting neural networks is an
active area of research that is summarized well by this tutorial: https://

explainml-tutorial.github.io/.
Popular deep learning methods include feed-forward neural networks,
recurrent neural networks, convolutional neural networks, autoencoder
networks, and generative adversarial networks [13,14,18]—several of these
are applied in the case study chapter of this book. Using these deep learning
methods, scientists have achieved unprecedented results using deep learn-
ing in image processing in recent years, including high-resolution image
generation [1,2,17], image-to-image [15,34] and sequence-to-sequence [9]
translation, and visual question answering [28]. We refer the reader to Le-
Cun et al. [18] for a general overview of deep learning, Raghu & Schmidt
[24] for an overview of deep learning for scientific discovery, Karpatne
et al. [16] and Bergen et al. [4] for surveys of deep learning in the geo-
sciences, and Zhu et al. [36] for an overview of deep learning in remote
sensing. Chapters 6, 8, 9, and 10 present case studies using a variety of deep
learning methods for both classification and regression in planetary science
applications.
1.7 Data set preparation

A typical machine learning workflow starts with a data preparation stage.
In addition to dataset-specific preprocessing, e.g., removing clouds or data
artifacts, coregistering observations, or extracting spectral indices or other
features, three datasets are usually created from the available data: the train-
ing, validation, and test datasets. The training set is used for training the
model, i.e., iteratively tuning model parameters or weights to optimize
some cost or objective function. The validation set is used for select-
ing model hyperparameters. Hyperparameters are variables not tuned during
training but that determine the configuration of the model and influence
the learning process, such as the number of trees in a random forest. Finally,
the test set is used for evaluating model performance on unseen data.
The examples in each of these datasets should be independent and
nonoverlapping. If the source data set contains independent and identically-
distributed examples, it is common to split the source data set into train-
ing, validation, and test sets using a 80%/10%/10%, 70%/15%/15%, or
60%/20%/20% split. Using a 80%/10%/10% split, a random set of 80% of
the available examples will be assigned to the training set while 10% will be
assigned to the validation and test sets each. However, in many real-world
(including planetary science) data sets, there exist correlations between ex-
amples in the source data set that should be considered when creating these
three data sets. This is important because correlations between examples in
the training and validation or test sets can result in overestimation of model
performance. This is especially important for geospatial data since, as stated
succinctly by Tobler [31], “all things are related, but nearby things are more
related than distant things.” For example, in Chapter 5, the HiRISE data set
contains 227 × 227-pixel image tiles cropped from HiRISE images with a
much larger size and footprint. Since tiles cropped from the same HiRISE
images would have higher correlation than tiles from different HiRISE im-
ages, Lu et al. divided the tiles into train, validation, and test sets in such a
way that all landmarks that originated from the same HiRISE source image
were confined to the same subset.
In some data sets, especially small data sets, model performance might
be sensitive to the particular split of the data and thus the validation or
test performance is not an accurate measure of the model performance
across the entire data set. A popular technique to address this is k-fold
cross validation. In k-fold cross validation, the data set is split into k folds
or partitions each of size N /k where N is the number of samples in the
data set. The model is then evaluated on each of the k folds, using the
remaining folds as training data. Leave-one-out cross validation (LOOCV)
is a special case of k-fold cross validation in which each fold contains only
one example. Performance metrics are averaged across the k iterations.
References
[1] G. Antipov, M. Baccouche, J.-L. Dugelay, Face aging with conditional generative ad-
versarial networks, eprint, arXiv:1702.01983, 2017.
[2] M. Arjovsky, S. Chintala, L. Bottou, Wasserstein GAN, eprint, arXiv:1701.07875,
2017.
[3] M. Belgiu, L. Drăguţ, Random forest in remote sensing: a review of applications and
future directions, ISPRS Journal of Photogrammetry and Remote Sensing 114 (2016)
24–31.
[4] K.J. Bergen, P.A. Johnson, V. Maarten, G.C. Beroza, Machine learning for data-driven
discovery in solid Earth geoscience, Science 363 (2019).
[5] C.M. Bishop, in: Pattern Recognition and Machine Learning, Springer, 2006, Chap-
ter 7.
[6] A. Blum, T. Mitchell, Combining labeled and unlabeled data with co-training, in:
Proceedings of the Eleventh Annual Conference on Computational Learning Theory,
1998, pp. 92–100.
[7] L. Breiman, Random forests, Machine Learning 45 (2001) 5–32.
[8] L. Breiman, J.H. Friedman, R.A. Olshen, C.J. Stone, Classification and regression
trees. Belmont, CA: Wadsworth, International Group 432 (1984) 151–166.
[9] C. Chan, S. Ginosar, T. Zhou, A.A. Efros, Everybody dance now, in: Proceedings of
the IEEE International Conference on Computer Vision, 2019, pp. 5933–5942.
[10] A.P. Dempster, N.M. Laird, D.B. Rubin, Maximum likelihood from incomplete data
via the em algorithm, Journal of the Royal Statistical Society: Series B (Methodologi-
cal) 39 (1977) 1–22.
[11] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, ImageNet: a large-scale
hierarchical image database, in: CVPR09, 2009.
[12] G. Doran, E. Dunkel, S. Lu, K. Wagstaff, Mars orbital image (HiRISE) labeled data
set version 3.2, https://doi.org/10.5281/zenodo.4002935, 2020.
[13] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A.
Courville, Y. Bengio, Generative adversarial nets, in: Advances in Neural Information
Processing Systems (NIPS), 2014, pp. 2672–2680.
[14] G.E. Hinton, R.R. Salakhutdinov, Reducing the dimensionality of data with neural
networks, Science (New York, N. Y.) 313 (2006) 504–507, https://doi.org/10.1126/
science.1127647.
[15] P. Isola, J.-Y. Zhu, T. Zhou, A.A. Efros, Image-to-image translation with conditional
adversarial networks, in: Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2017, pp. 1125–1134.
[16] A. Karpatne, I. Ebert-Uphoff, S. Ravela, H.A. Babaie, V. Kumar, Machine learning
for the geosciences: challenges and opportunities, IEEE Transactions on Knowledge
and Data Engineering 31 (2018) 1544–1554.
[17] T. Karras, S. Laine, T. Aila, A style-based generator architecture for generative adver-
sarial networks, in: Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, 2019, pp. 4401–4410.
[18] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (2015) 436–444, https://
doi.org/10.1038/nature14539.
[19] D.D. Lewis, W.A. Gale, A sequential algorithm for training text classifiers, in: SI-
GIR’94, Springer, 1994, pp. 3–12.
[20] D.G. Lowe, Distinctive image features from scale-invariant keypoints, International
Journal of Computer Vision 60 (2004) 91–110.
[21] T. Mitchell, Machine Learning, 1st ed., McGraw Hill, 1997.
[22] K. Nigam, R. Ghani, Analyzing the effectiveness and applicability of co-training, in:
Proceedings of the Ninth International Conference on Information and Knowledge
Management, 2000, pp. 86–93.
[23] K. Nigam, A.K. McCallum, S. Thrun, T. Mitchell, Text classification from labeled and
unlabeled documents using em, Machine Learning 39 (2000) 103–134.
[24] M. Raghu, E. Schmidt, A survey of deep learning for scientific discovery, arXiv
preprint, arXiv:2003.11755, 2020.
[25] J. Ratsaby, S.S. Venkatesh, Learning from a mixture of labeled and unlabeled examples
with parametric side information, in: Proceedings of the Eighth Annual Conference
on Computational Learning Theory, 1995, pp. 412–417.
[26] A. Razavian, H. Azizpour, J. Sullivan, S. Carlsson, Cnn features off-the-shelf: an
astounding baseline for recognition, in: Proceedings of the IEEE Conference on Com-
puter Vision and Pattern Recognition Workshops, 2014, pp. 806–813.
[27] C. Rosenberg, M. Hebert, H. Schneiderman, Semi-supervised self-training of object
detection models, in: WACV/MOTION, vol. 2, 2005.
[28] A. Santoro, D. Raposo, D.G.T. Barrett, M. Malinowski, R. Pascanu, P. Battaglia, T.
Lillicrap, D. London, A simple neural network module for relational reasoning, in:
Proceedings of the 30th International Conference on Neural Information Processing
Systems, 2017, pp. 4967–4976.
[29] B. Settles, Active learning literature survey, Technical Report, University of
Wisconsin-Madison, Department of Computer Sciences, 2009.
[30] H.S. Seung, M. Opper, H. Sompolinsky, Query by committee, in: Proceedings of the
Fifth Annual Workshop on Computational Learning Theory, 1992, pp. 287–294.
[31] W.R. Tobler, A computer movie simulating urban growth in the Detroit region, Eco-
nomic Geography 46 (1970) 234–240.
[32] K. Wagstaff, Y. Lu, A. Stanboli, K. Grimes, T. Gowda, J. Padams, Deep Mars: Cnn
classification of Mars imagery for the pds imaging atlas, in: Proceedings of the AAAI
Conference on Artificial Intelligence, vol. 32, 2018.
[33] D. Yarowsky, Unsupervised word sense disambiguation rivaling supervised methods,
in: 33rd Annual Meeting of the Association for Computational Linguistics, 1995,
pp. 189–196.
[34] J.-Y. Zhu, T. Park, P. Isola, A. Efros, Unpaired image-to-image translation using
cycle-consistent adversarial networks, in: IEEE International Conference on Com-
puter Vision (ICCV), 2017.
[35] X. Zhu, A.B. Goldberg, Introduction to semi-supervised learning, Synthesis Lectures
on Artificial Intelligence and Machine Learning 3 (2009) 1–130.
[36] X.X. Zhu, D. Tuia, L. Mou, G.-S. Xia, L. Zhang, F. Xu, F. Fraundorfer, Deep learning
in remote sensing: a comprehensive review and list of resources, IEEE Geoscience and
Remote Sensing Magazine 5 (2017) 8–36.
CHAPTER 2
The new and unique challenges of

planetary missions
Jörn Helbert and Mario D’Amore
2.1 Introduction
2.1.1 50 years of Mercury exploration
A good example of the trend to larger and more complex datasets is a com-
parison between three missions to Mercury: the NASA missions Mariner
10 and MESSENGER and the upcoming ESA-JAXA BepiColombo mis-
sion. Those three missions span a time period of almost 50 years encompass-
ing most of planetary exploration by missions. At the same time Mercury
has surprised planetary scientists with every mission.
Mariner 10 was the seventh successful launch in the Mariner series and
the first spacecraft to visit Mercury. It was also the first spacecraft to use the
gravitational pull of one planet (Venus) to reach another (Mercury), and
the first spacecraft mission to visit two planets. The spacecraft flew by Mer-
cury three times in a retrograde heliocentric orbit and returned images and
data on the planet. Mariner 10 returned the first-ever close-up images of
Venus and Mercury [1,2]. The primary scientific objectives of the mission
were to measure Mercury’s environment, atmosphere, surface, and body
characteristics and to make similar investigations of Venus. That spacecraft
was able to map between 40 and 50% of the planet due to the orientation
of the flybys. This revealed a surface covered in craters that looked at first
glance similar to the Moon. As the first major surprise of Mercury Mariner
10 revealed the presence of a significant internal magnetic field, measured
constituents of the exosphere and confirmed Mercury’s high uncompressed
density.
The NASA MESSENGER mission [3] performed three flybys of Mer-
cury before entering orbit around Mercury at 00:45 UTC on March 18,
2011. MESSENGER’s primary mission began on April 4, 2011, followed
by two extended missions starting on March 17, 2012. By March 6, 2013
the mission had imaged 100% of the surface [4]. The mission ended with a
Machine Learning for Planetary Science Copyright © 2022 Elsevier Inc.
https://doi.org/10.1016/B978-0-12-818721-0.00009-4 All rights reserved. 25
crash of the spacecraft on Mercury on April 30, 2015. MESSENGER was

the first mission to enter orbit around Mercury. MESSENGER revealed
the surface to be surprisingly rich in volatile elements like sulfur [5], pro-
vided the first detailed views of the so-called hollow structures indicating
currently ongoing surface modification [6], provided visual evidence of past
volcanic activity on the surface of Mercury [7] and strong evidence for wa-
ter ice deposit in the permanently shadowed craters close to the poles of
Mercury [8].
BepiColombo is an interdisciplinary mission to explore Mercury car-
ried out jointly between the European Space Agency and the Japanese
Aerospace Exploration Agency [9]. It consists of two spacecraft, the ESA
provided Mercury Planetary Orbiter (MPO) and the JAXA provided Mer-
cury Magnetospheric Orbiter (MMO) which fly in a stack configuration
to Mercury propelled by the Mercury Transfer Module. Upon arrival at
Mercury the modules will separate and enter their respective science orbits.
The mission launched successfully in October 2018. After 6 Mercury fly-
bys (as well as one Earth and two Venus flybys) the mission will enter orbit
around Mercury in December 2025.
The following table shows the payload of each of the missions. When
possible an attempt was made to align similar types of instruments on each
of the missions.
Obviously the instruments on the three missions spanning almost 50
years can not be compared 1-to-1. The comparison in Table 2.1 is based
on similar functionality in the widest terms. Focusing on the first three
lines, Table 2.1 highlights exactly the growing complexity that requires
new methods for data analysis. On first glance the Television Photography
on Mariner 10 [1], MDIS on MESSENGER [3] and the SIMBIO-SYS
on BepiColombo [10] are similar instruments that returned (or will return)
images of the surface. However looking in details we see the increase in
spatial and spectral resolution as well as in complexity of the data acquisi-
tion.
The Television Photography experiment on the NASA Mariner 10
consisted of two spherical (150 mm diameter) Cassegrain telescopes with
eight filters, each attached to a GEC 1 inch vidicon tube camera (TV)
(1500 mm focal length and 0.5 degree field of view) for narrow-angle pho-
tography. An auxiliary optical system mounted on each camera provided
wide-angle (62 mm focal length, 11 x 14 degree field of view) photog-
raphy by moving a mirror on a filter wheel to a position in the optical
path. The TV picture consisted of 700 scan lines with 832 picture elements
The new and unique challenges of planetary missions 27
Table 2.1 Comparison of the payloads of the NASA missions Mariner 10 and MESSEN-
GER and the ESA BepiColombo Planetary Orbiter based on their functionality.
Mariner 10 MESSENGER BepiColombo
Television Mercury Dual Imaging Spectrometer AND
Photography System (MDIS) Imagers for MPO
BepiColombo Integrated
Observatory SYStem
(SIMBIO-SYS)
- Mercury Atmospheric SIMBIO-SYS
and Surface Composition
Spectrometer (MASCS)
Extreme Ultraviolet MASCS Probing of Hermean
Spectrometer Exosphere by Ultraviolet
Spectroscopy (PHEBUS)
Triaxial Fluxgate Magnetometer (MAG) MPO-MAGnetometer
Magnetometer (MPO-MAG)
Celestial Mechanics Radio Science (RS) Mercury Orbiter
and Radio Science Radio-science Experiment
(MORE)
Scanning Search for Exospheric
Electrostatic Analyzer Refilling And Emitted
and Electron Natural Abundances
Spectrometer (SERENA)
Energetic Particles Energetic Particle and SERENA
Experiment Plasma Spectrometer
(EPPS)
- Mercury Laser Altimeter BepiColombo LasEr
(MLA) Altimeter (BELA)
Two-Channel - MErcury Radiometer and
Infrared Radiometer Thermal infared Imaging
Spectrometer (MERTIS)
- Gamma-Ray Mercury Gamma ray and
Spectrometer (GRS) Neutron Spectrometer
(MGNS)
- Neutron Spectrometer MGNS
(NS)
- X-Ray Spectrometer Mercury Imaging X-ray
(XRS) Spectrometer (MIXS)
- - Italian Spring
Accelerometer (ISA)
- - Solar Intensity X-ray and
particles Spectrometer
(SIXS)
per line, which were digitally coded into eight-bit words for transmission.
Exposure time ranged from 3 ms to 12 s, and each camera took a picture
every 42 s. The average bit rate from this experiment was 22 kbps. The
system had eight filter wheel positions: (1) wide-angle image relay mirror;
(2) blue bandpass; (3) UV polarizing; (4) minus UV high pass; (5) clear; (6)
UV bandpass; (7) defocusing lens (for calibration); and, (8) yellow band-
pass. About 7000 photographs were obtained of Venus and Mercury, with
a maximum resolution of 100 m for Mercury.
The Mercury Dual Imaging System (MDIS) on the NASA MESSEN-
GER consists of a wide angle refractive optic imager with a 10.5 degree
field of view and a narrow angle reflective optic imager with a field of view
of 1.5 degrees. The wide angle imager has a modified achromatic Cooke-
triplet lens which is 30 mm in diameter and a 12 position filter wheel with
two clear filters centered at 750 nm, one with a 100 nm width, the other at
600 nm. The ten color filters are centered at 415 nm (40 nm width), 480
(30), 560 (10), 650 (10), 750 (10), 830 (10), 900 (10), 950 (20), 1000 (30),
and 1020 (40). The light then passes through a small field flattening lens
and strikes the 1024 x 1024 frame transfer CCD. The narrow angle imager
uses an off-axis section of a Ritchie-Chretien reflective telescope to achieve
a focal length of 550 mm with mirrors to correct spherical aberration and
coma. The focal ratio is 18. A single band-limiting filter is used. As with
the wide angle imager, a 1024 x 1024, 14 µm/pixel, frame transfer CCD
is used. Each CCD records 12 bits/pixel and has manual and automatic
exposure control over a range of 1 ms to 10 s.
The MASCS experiment consists of two instruments, a UV/Visible
Spectrometer (UVVS) and a Visible/IR Spectrograph (VIRS) [11]. A baf-
fled 250 mm Cassegrain f/5 telescope focuses light through a common
boresight to both instruments. The UVVS consists of an Ebert-Fastie
diffraction grating spectrometer. An 1800 groove/mm grating gives an av-
erage spectral resolution of 1.0 nm (0.5 nm in the far ultraviolet). The
grating is rotated in 0.25 nm steps for scanning. Three photomultiplier
tubes are situated behind separate slits, one covers the far ultraviolet (115-
190 nm), one the middle ultraviolet (160-320 nm), and one the visible
(250-600 nm). The VIRS is designed to measure surface reflectance in
the 0.3-1.45 µm band with a spatial resolution of 100 m to 7.5 km. The
field of view is 0.023 x 0.023 degrees. Light reaches the detector through
a fused silica fiber optic bundle. A concave holographic diffraction grating
with 120 lines/mm and a dichroic beam splitter which separates the visible
(0.30-1.025 µm) and infrared (0.95-1.45 µm) parts of the spectrum are used
Another random document with
no related content on Scribd:
The Project Gutenberg eBook of Sunnuntai
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.
Title: Sunnuntai
Romaani
Author: Ester Ståhlberg
Release date: October 8, 2023 [eBook #71834]
Language: Finnish
Original publication: Porvoo: WSOY, 1922
Credits: Juhani Kärkkäinen and Tapio Riikonen
*** START OF THE PROJECT GUTENBERG EBOOK SUNNUNTAI

***
SUNNUNTAI
Romaani
Kirj.
Ester Ståhlberg
Porvoossa, Werner Söderström Osakeyhtiö, 1922.
Äidilleni.
1.
Ilta hämärtää. Koivut ikkunani ulkopuolella ovat paksussa

huurteessa, ja tuulen henki huojuttaa hiljaa niiden valkoisia oksia. On
syvä rauha ja hiljaisuus. Viereisestä huoneesta kuuluu yksinäisen
keinutuolin natinaa, muuten ei napsahdustakaan koko talossa.
Siellä istuu isä, hitaasti kiikuttaen keinutuolia, pää kätensä
varassa. On sunnuntai-ilta. Koko päivä on ollut samaa äänetöntä
sunnuntaita, samaa pyhien pyhää. Olemme yhdessä katselleet
valokuviasi, yhdessä entisiä kirjeitäsi. Talo on ollut sinua täynnä, ja
jokainen soppi ja jokainen esine on kertonut sinusta.
Olen joskus tarkannut isää, kun hän siinä kumarassa istuu sinun
kirjeesi hyppysissään. Olen tarkannut hänen raukeata ilmettänsä,
hänen laihtuneita kasvojansa. Hän ei saa öisin unta. Hänen kätensä
ovat entistä kalpeammat, hänen huulensa väräjävät, kun hän
kuvaasi katselee. Ja kun luen kirjeistäsi jonkun kuvauksen entisestä
kesäelämästämme tai mistä tahansa, mihin mielesi on ollut kiintynyt,
sulkee hän silmänsä ja hänen rinnastaan nousee raskas huokaus
ikäänkuin jostakin syvältä, hänen olemuksensa pohjimmaisesta
pohjukasta.
Nyt on ilta. Kuuntelen tuota keinutuolin natinaa ja tiedän varsin

hyvin missä hänen mielensä vaeltaa. Hän soudattaa siinä esille
kuvan toisensa jälkeen murheellisen sydämensä lohdutukseksi. Hän
muistaa niitä äskeisiä päiviämme, jolloin sinä muutamalla
harppauksella juosten portaat saavuit koulusta ja kerroit illan suussa
hänelle juttujasi. Silloin oli vielä viihtymisemme häiriintymätön
yhteisen kotilieden ympärillä. Ja karhuntaljalla isän jalkojen juuressa
nukahti joka ilta pieni liinatukkainen tyttölapsi.
Kaikki on siitä muuttunut. Nyt on maailma äänetön ja elämämme

kulku on pysähtynyt. Arki touhuineen on poissa, ja meillä on
sunnuntai. Yhtämittainen, salaperäinen sunnuntai. Tuulen suhina on
toisenlainen kuin ennen, junan vihellys kuuluu kuin toisesta, meille
vieraasta maailmasta. Huomaan usein isän silmistä, että hän
katselee jotakin, jota eivät muut näe, ja kuuntelee jotakin, jota eivät
muut kuule. Meille on avattu salattu tie, jota kulkiessamme entisten
näköalojen takaa avautuvat uudet, jonka varsilla oudot kukkaset
loistavat ja vieraat linnut meille visertävät. Kuitenkin se on oman
elämämme polku. Astumme tätä polkua poispäin nykyisyydestämme,
mutta on kuin emme olisi siitä ennen mitään aavistaneet.
Pelkään pahoin, ettei isän silmä koskaan enää käänny siitä

maailmasta pois edes sen verran, että hän jaksaisi nykypäivän
kutsua seurata. Pelkään, että hän on menneen ajan lumoissa niin
kokonaan, ettei hän tule mitään tulevaisuutta enää tavoitelleeksi.
Uskon oi jospa voisin olla sitä uskomatta! — mutta uskon joskus,
ettei hänen katseensa enää koskaan palaja siltä veräjältä, jonka
taakse sinä hävisit.
Menen hiljaa hänen huoneeseensa. Hän ei nosta päätään. Lasken

käteni hellävaroin hänen olkapäällensä. Hän ei näy huomaavan
tuloani. Otsa on uurtunut, suupielissä on syvät vaot. Annan katseeni
liukua ulos ikkunasta huurteisten oksien lomitse yli hämärtävien
lumikenttien. Miten kutsun hänet takaisin elämään?
Kaareutuuhan täällä vielä taivas ja paistaahan joskus taas aurinko.

Onhan tässä vielä koti tallella, on pieni tyttönen vuoteessansa, ovat
tuossa ruohokentät lumen alla, ja ensi keväänä ne taas viheriöinä
helottavat. Olethan täällä sinäkin, Yrjö. Olet — vaikka oletkin poissa,
olet kuitenkin luonamme pysyväisemmin, muuttumattomammin kuin
koskaan ennen.
— Onhan hän kuitenkin luonamme — sanon ääneen.
Valkotukkainen mies katsahtaa minuun ja nyökkää päätään. Sitä

katsetta en unohda. Siinä ei ollut sillä hetkellä kaipauksen
katkeruutta. Siinä on varmuutta jostakin. Mistä?
— Muistellaan yhdessä! — sanon taasen ääneen. — Kaivataan
yhdessä! Sillä lailla on helpompi kantaa kuormaa, ja yhteinen kaipuu
muuttuu elämän voimaksi. Se, mikä meillä on yhteistä, ei ole
kenenkään muun omaa kuin meidän kahden, ja meillä on suuret
aarteet. Yrjö on avannut meille portin omaan elämäämme ja
elämään yleensä. Hän tulee itse meitä vastaan kaikkialla. Nyt vasta
näemme silmillämme, tajuamme toisin kuin ennen oman
kohtalomme ja muittenkin kohtaloita. Ja kaiken takana ovat
loppumattomat avaruudet.
Hän sulkee käteni omaansa ja minä istuudun tuolille hänen

viereensä.
Tuossa ovat mahonkituolit molemmin puolin pöytää ja sohva

pöydän takana. Yrjö! Niissä istuimme ennen iltaisin. Tule! Istuudu
tuohon isän tyhjään nojatuoliin kuten muinoin! Jutelkaamme
keskenämme kuin ennen, ja antakaamme tuon väsyneen miehen
lämmetä ääntemme kuulemisesta. Suokaamme hänelle nykyhetken
unohduksen lepo ja menneitten päivien lohdutus. Yrjö, auttakaamme
häntä, rohkaiskaamme häntä, niin että hän entisyyden muistoista
saa voimaa elämään uutta elämää. Aika ei todellisuudessa pysähdy,
ihmisen on riennettävä eteenpäin, sillä yksin kohtalo asettaa hänelle
hänen matkansa päämäärän. Mutta hän jaksaa paremmin kestää
kovat kaudet, kun hän saa rauhassa muistella ja niistä
muisteloistaan koota itsellensä uutta tarmoa. Niinpä tehkäämme
hänelle tämä ilta leppoisaksi ja rauhan täyttä suloa uhkuvaksi.
Kulukoon lohdullisesti tämä pyhäinen sunnuntai loppuansa kohti,
lampun hauskasti valaistessa hänen kirjoituspöytäänsä ja hämärän
kurkistaessa sisään punaviiruisten uutimien lomitse.
*****
Kerro, poikaseni, tulimmeko sinulle koskaan sanoneeksi, miten
paljon me sinua rakastimme? Emme suinkaan. Sillä emmehän
tulleet vakuuttaneeksi sinulle päivänselvää asiaa. Emmekä
kiittäneeksi sinua siitä, että olit olemassa. Emme siitäkään, että olit
meille olemassa, sillä niin oli kohtalomme säätänyt.
Jospa voisin puhua sinulle sanattomasti! Sanat eivät kuitenkaan

ilmaise sitä mitä tunnen. Niillä on ahtaat rajansa, mutta en tule ilman
niitä toimeen. Kun vaivun sinun kanssasi muistelemaan entistä
elämäämme, olen riippuvainen rajoista, muodoista ja väreistä Ja
kuitenkin on kaiken takana jotakin rannatonta, jotakin sanoin
sanomatonta.
2.
Ajattelen sitä hetkeä, jolloin kohtalomme langat yhtyivät. Et voi sitä

muistaa, mutta olen usein kertonut siitä sinulle. Se oli se hetki, jolloin
näimme toisemme ensi kerran. Synnyit silloin toistamiseen, sillä
synnyit silloin meille.
Siinä huoneessa ei kuulunut kiirehtiviä askelia, ei valituksia, ei

huudahduksia. Et tervehtinyt minua kirkunalla enkä minä lähestynyt
sinua tyydyttääkseni alkuperäisintä vaistoasi. Siinä huoneessa
vallitsi äänettömyys. Ja me kohtasimme toisemme ainoastaan
viipyvässä ihmettelevässä katseessa, ikäänkuin olisimme jonkin
kadonneen jälleen nähneet.
Eikä ollut silloin muuta olemassa kuin tämä lapsi, joka oli äitinsä
löytänyt, ja tämä äiti, joka oli löytänyt lapsensa.
Varmaan sinäkin vaistomaisesti tunsit, että se päivä oli ratkaiseva.
Sillä sinä olit siitä asti toisessa suhteessa minuun kuin olit
kehenkään muuhun tutussa ympäristössäsi. Enkä koskaan unohda
sitä onnen hymyä, joka aina levisi kasvoillesi myöhemmin joskus
puhuessamme tästä yhtymisestämme ja lapsuutesi varhaisimmista
muistoista.
Sinä päivänä istuit valkoisen jakkarapöydän ääressä, edessäsi

kulunut kuvakirja. Huoneen muusta sisustuksesta minulla ei ole
aavistustakaan. Sen vain tiedän, että siinä oli pieniä lastenvuoteita ja
niissä nukkuvia pikkulapsia, ja meidät saattoi sinne laitoksen
johtajatar, diakonissanpuku yllään. Hän oli korkeasukuinen nainen,
hienopiirteinen ja komearyhtinen, tottunut käskemään, mutta sinä olit
hänen silmäteränsä, ja hän valvoi tarkasti tulevaisuutesi
kaikinpuolista turvaamista. Meihin hän oli jostakin syystä saanut
ehdottoman luottamuksen. Ja sen takia hän seisoi syrjässä odottaen
kohtalomme kehitystä.
Isä jäi äänetönnä sinua katselemaan. Minä lähestyin sinua ja

kyykistyin viereesi lattialle. Sinä vain suurilla silmilläsi minua katsoit
ja katsoit. Sydämeni oli kuin vellova vesi, toivon ja pelon vaiheilla
läikkyvä. Uskaltaisinko koskettaa noita pehmeitä kämmeniä sinun
siitä säikkymättä? Otin kätesi omaani ja sinä annoit sen silmiä
räpäyttämättä tapahtua. Tunsin silkkisen hipiäsi kosketuksen ja suljin
molemmat kätesi omiini. Silloin irroitit ne otteestani ja kiersit hennot
käsivartesi kaulaani.
Silmistäni valahtivat kyyneleet ja minä nostin sinut käsivarsilleni.

Suljin sinut syliini ja painoin otsani karkean mekkosi poimuihin.
Ja minusta tuntui kuin olisi jossain ajan kello lyönyt, erottaen

ajanjakson edellisestä. Se löi sinulle ja minulle. Ja sinä olit kuitenkin
vasta pieni lapsi, joka sopertelit ensimäisiä sanojasi.
Minä nousin ja käännyin isään päin. Näin, että asia hänenkin

puoleltaan oli ratkaistu. Huoneen kynnyksellä seisoi johtajatar-
diakonissa, eikä hänkään selvityksiä kaivannut. Mitään sanomatta
läksimme huoneesta pois, sillä tavanmukaiset kysymykset ja
vastaukset olivat tässä tarpeettomat.
Tuosta hetkestä tuli elämämme rajapyykki. Monet muut ovat

vähitellen häipyneet tietoisuudestamme, mutta se on aina loistanut
vastaamme yhtä kirkkaana muistojemme kätköistä.
3.
Niin se tapahtui. Mutta kesti kauan ennenkuin saimme ottaa sinut

mukaamme kotiin, sillä olimme vieraassa maassa, jolla oli omat
lakinsa. Odottaessamme asian laillista järjestelyä meidän oli pakko
pysyä kärsivällisinä kauankin, ja tämän ajan kestäessä oli aina
mahdollista, että jotakin arvaamatonta voisi tapahtua.
Vihdoin isä ei enää voinut aikaansa hukata, hänen täytyi, jos kohta
vastahakoisesti, matkustaa kotimaahamme. Minä jäin. Ja olin valmis
odottamaan sinua vaikka vuosia, jos sikseen tulisi.
Päivästä päivään elin vain sitä hetkeä varten, jolloin sain sinut
nähdä. Sinä tähystelit tuloani ikkunasta ja tallustelit minua vastaan.
En tahtonut tavata sinua muitten läsnäollessa, ja siksi vein sinut aina
ulos. Pistin yllesi päällystakkisi, jos oli sateista, muuten sait tulla
kanssani kävelylle hoitolapuvussasi. Sinä astua tepastelit vieressäni
paksut villasukat jalassasi kesähelteestä huolimatta ja punaruutuinen
esiliina peittämässä harmaata mekkoasi. Mutta mitäpä minä
tamineistasi! Olit silmissäni maailman herttaisin lapsi, ja vieläkin olen
varmasti vakuutettu, että todella olitkin. Ja sitäpaitsi, kukapa siinä
suurkaupungin vilinässä olisi ennättänyt kiinnittää huomiotaan
nuoreen naiseen, joka talutti pientä palleroistaan!
Ne kävelyt olivat minulle elämän puhtainta onnea. Pojan pieni käsi

omassani tunsin hänen rajattoman luottamuksensa virtaavan
lävitseni ja oman sydämeni sykkivän rakkautta häntä kohtaan.
Löytäessäni tyhjän istuimen puistossa nostin hänet syliini ja hän
nojasi minuun. Hän istui siinä kuin varmimmassa turvassaan, hän
nukkui siihen kuin lintunen pesäänsä. Ja hänen herätessään laulelin
hänelle outoja lauluja omalla kielelläni, katsoen hänen ihmetteleviin
silmiinsä hänen siinä polvellani kiikkuessaan, rusopintaiset kasvot
minuun käännettyinä.
Silloin tunsin, ettei mikään mahti voisi erottaa minua sinusta. Että
kaikki äidinvaistoni olivat kasvamassa voimaksi, joka voisi siirtää
vuoria sinun tähtesi.
Ja kuitenkin oli vielä mahdollista, että hän, joka oli sinut niin
vähäiseksi arvioinut, että miltei synnyinhetkenäsi oli sinut hylännyt,
voisi sinut minulta riistää.
Mutta en sitä mahdollisuutta muuta kuin ohimennen pelännyt. Sillä

en voinut uskoa hänen vaativan sinua takaisin. Hän oli luopunut
sinusta vapaaehtoisesti, jättäen sinut elämän purteen satunnaisen
kohtalon varaan. Olit kyllä ollut hyvissä käsissä tähän asti, mutta
tällä suojelijallasi oli lupa siirtää sinut eteenpäin muitten huomaan.
Synnyttäjäsi oli enää vain nimellisesti äitisi. Sillä elämän laki on
lahjomaton.
Eikä tässä ollut muusta kysymystäkään kuin hänen valkoisen
kätensä allekirjoituksesta. Uusi sormus välkkyi hänen sormessaan,
tiesin sen. Se kiiltäisi siinä hänen piirtäessään nimensä paperille,
joka ratkaisisi elämäsi kohtalon.
Vapaaksi hän oli tahtonut tulla — siis olkoon hän vapaa ainiaaksi.
Pukekoon hän yllensä nuoruutensa ja kauneutensa hulmuavat
vaatteet ja käyköön elämän karkeloon! Huuhdelkoon kulmiltaan
öitten salaiset murheet ja rientäköön onnen tanhuville! Tulkoon
maailman loisto ja upeus hänen osakseen, jos hän sitä haluaa,
mutta älköön hän koskaan vaipuko muistelemaan varhaisimman
nuoruutensa haaveita! Silloin hiipii ehkä hänen sydämeensä avuton
autiuden tunne, joka hiljaa hänen elinlankaansa jäytää.
Nyt hän ei halunnut menneestä ajasta mitään tietää. Hän tahtoi

unohtaa nuo keväiset yöt, jolloin satakielet lehdoissa lauloivat ja
kastehelmet ruohonkorsissa hohtivat. Hän heitti ne hetket unholan
hämärään, sillä hän ei voinut sitoa itseään siihen, joka oli vain
virvatulen leikkiä. Eikä siihenkään, joka teki tämän haihtuvan kuplan
todelliseksi hänen silmissään. Hän ei ollut niitä, jotka ottavat
lapsensa käsivarrelleen astuaksensa ulos elämään sen kanssa.
Hän siis pakeni, koska hän ei muuta voinut. Hän heitti kaikki
luotansa ja sulki portin takanansa, koska hän ei tahtonut mistään
vastata. Silloin tulin minä kulkien tietä pitkin ja löysin luomakunnan
aarteen hylättynä. Jouduttakoon hän askeleitansa, häntä ei enää
tarvittu. Suljin tämän hyljätyn olennon syliini ja kannoin sen ilosta
laulahtaen oman kotini kynnyksen yli.
Niin. Minä tunsin, että tämän lapsen takia olin valmis tekemään
mahdottomiakin. Nouskoon vaikka koko maailma sanomaan, ettei
hän ole omani! Uhmailisin maailmaa vastaten, että se tuomitsi
harhaan. Kukaan ei tule äidiksi synnyttämällä lapsen, jota hän ei ole
voinut olla synnyttämättä ja jonka hän sitten hylkää. Mutta äidiksi
tulee se, joka lasta rakastaen vaalii sitä omana sielunaan, oman
elämänsä kukkana.
Mutta ainoastaan harvoin ja sivumennen ajattelin niitä, jotka olivat

antaneet sinulle elämän lahjan. Useimmiten unohdin ne tykkänään,
eivätkä ne koskaan tulleetkaan tajunnassani täysin eläviksi. Olin
ihmeteltävän huoleton. Todella tuntui sieltä, kuin olisit syntynyt sinä
hetkenä, jolloin sinut ensi kerran näin. He olivat vain kangastuksia,
jotka häipyivät olemattomiin, he olivat vain harhailevaa sumua, joka
utupilvinä pakenee päivän kehrän ensimmäisiä valoenteitä.
Mutta sinä — sinä olit kastehelminen kukkaiskupu, jonka yksin

meidän silmämme oli nähnyt. Sinä olit kaukaisessa vuoristossa
säteilevä hohtokivi, joka meidän yksin oli sallittu löytää. Tai olitko
linnunpoikanen, joka oli pudonnut pesästään ja jonka piipitystä me
yksin kuulimme? Jonka pelastin käteeni hoitaakseni sydänkäpynäni.
Ja elämä sulki meidät sykähtävään helmaansa, ja tuuli riemuitsi ohi
liitäessään. Sillä itse pyhä kohtalo oli tarttunut käteemme ohjaten
meidät sinne, missä sinä varroit. Se oli meidät vihkinyt löytämään
sinut ja ilolla ja ylpeydellä omistamaan sinut Jumalan ja ihmisten
edessä.
4.
Kävelyiltämme palattuamme tuli joskus diakonissa-täti vastaamme

käytävässä. Ja jos silloin kysäisin ratkaisun päivää tai hetkeä, kohotti
hän vain olkapäitään merkiksi siitä, ettei hän mitään tiennyt.
— Ei syytä pelkoon — sanoi hän. — Kukaan ei teiltä lastanne vie.
Olen kokenut tarpeeksi tietääkseni sen. On vain yhtä ja toista
huolehdittavaa, ennenkuin asiat ovat laillisessa järjestyksessä.
Hän oli oikeassa.
Koitti vihdoin se päivä, jolloin sain noutaa sinut luokseni

viedäkseni sinut mukaani merten ja mantereitten taakse omaan
kotiimme.
Niinpä tulit omaksemme. Ja se päivä on ollut elämämme

kohtalokkaimpia.
Viimeisen kerran tullessani laitokseen juoksit vastaani ilosta

kirkuen ja polviini kiinni tarrautuen. Nostin sinut korkealle ilmaan,
jolloin sinulta pääsi helähtävä nauru, ja olkapäilläni kieppuen ja
käsiäsi huitoen riensit kanssani käytävää pitkin huoneeseesi. Oivalsit
heti, että tänään sinua odotti jokin tavallista hauskempi, ja melkein
vapisit kuumeisesta malttamattomuudesta. Meillä oli molemmilla
tulinen kiire saada päällystakki yllesi ja lakki päähäsi. Ja sitten
kiertää laitosta sanoaksemme hyvästit kilteille tädeille.
Hyvästi ja kiitos kaikesta!
Ja sitten seisoimme yhdessä portaitten alla katukäytävällä, sinä ja

minä! Muistatko sinä sen — sinä ja minä!
Eikä meitä enää mikään eroittaisi!
Tartuit heti käteeni ja aloit epävakaisilla jaloillasi tarmokkaasti

taivaltaa eteenpäin ikäänkuin arvaten, että nyt oli kysymyksessä
pitempiaikainen matka. Sinä pieni veijari olit niin sanomattoman
tyytyväisen näköinen. Lyhyet jalantypykkäsi ponnistelivat
urhoollisesti ja syrjäsilmin kurkistit minuun ikäänkuin urkkien, mihin
maailmanmatkojen päähän nyt oli tarkoitus vaeltaa. Sinä et
rahtuakaan epäröinyt, sinä olit varma, että siellä paratiisin ovi meille
avautuisi. Enkä minäkään sitä epäillyt.
Asfaltti huokui lämpöä allamme, ja raitiovaunut kilisivät. Pikkupojat

nappipelineen unohtivat koko muun olevaisuuden, katukaupustelijat
huusivat hajamielisesti sanomalehtiänsä, banaanejansa,
kengänmustettansa kaupaksi. Avopäälapset tirkistelivät
portinpielistä, härnäsivät toisiansa ja tömistivät sitten yhtenä vilinänä
pakoon rotevaa poikaa, joka silmät vilkkuen ajoi heitä takaa.
Meitä ei mikään liikuttanut emmekä mitään huomanneet. Me

kuljimme käsi kädessä ja katse katseessa kohti yhteistä
tulevaisuuttamme.
*****
Millaisia päiviä sitten seurasikaan! Jos voisi niin varhaisista ajoista

ammentaa esille unholan pimentoon vajonneita kuvia, muistaisit sinä
kuten minäkin ennen kaikkea sen hämyhohteisen iltapäivän, jolloin
yhdessä liikuskelimme pauhaavan kaupungin keskuksessa. Oli
ainoastaan vuorokausi kulunut siitä, kun ensi kerran omana
lapsenani olit käsivarrelleni nukkunut. Seuraavana päivänä
lähdimme ulos heti päivällisen jälkeen, sillä en kärsinyt nähdä sinua
hoitolavaatteissasi. Pois siis kaikki vanha ja kuta pikemmin sitä
parempi! Pojallani tuli olla omat somat vaatteensa eikä rihmaakaan
entisyydestään.
Tuo menneisyys oli kuitattu ja ainiaaksi lopussa, ja mikä sitä

muistutti, oli karkotettava pois tietoisuudestamme. Tule siis,
mennään hakemaan pikkumiehelle sellaiset vaatteet, jotka hänen
arvolleen sopivat!
Tuolla jo raitiovaunu tulla hurisi, ja saimme paikan siinä. Matka

syrjäkaupungilta keskikaupungille kesti kauan, ja minusta tuntui, että
noitten vieraitten vaunussa istujien huomio kiintyi meihin. Enkä sitä
ensinkään ihmetellyt. Sillä kukapa ei olisi sellaista lasta ihaillut ja
hänen äitiään kadehtinut! Tunsin arvoni monin verroin kohonneen.
Olimme siis nyt niin pitkällä, että istuit polvellani kenen tahansa
nähden ja olit omanani kaikkien näitten ihmisten edessä. Se oli
meille suuri-arvoinen ja merkillinen päivä.
Katsokaa te vain! Kyllä kannattaa!
Sinä et hiiskunut halaistua sanaa koko matkalla, mutta silmäsi

seisoivat tarkkaavaisina päässäsi ja korvasi olivat pystyssä. Sinä
sirkeäsilmäinen matkamies! Mistä minä silloin arvasin, että luontaiset
harrastuksesi sinä hetkenä saivat ensimmäisen vaistomaisen
herätteen, että sinussa paloi väkevä intohimo kaikkeen, mikä
puhkuen, säikkyen, sihisten tai ulisten liikkui tai pani liikkeelle. Sitten
sen selvästi huomasin myöhemmin, kun jouduit tavalla tai toisella
tekemiseen koneellisten ilmiöiden kanssa. Nyt katsoin vain hymyillen
vaaksanpituiseen mieheen ja hänen tutkivaan ilmeeseensä, jättäen
hänet omaan rauhaansa maailman ilmiöitä havaitsemaan.
Astuimme alas jonkin lasten vaatetusliikkeen kohdalla, ja aloin

tarkastaa sen varastoja. Seisoit tuolilla lampun alla hiukan ymmällä
kaikesta tuosta kääntämisestä ja koettelemisesta, pyörittelemisestä
ja mittaamisesta, jonka alaiseksi käsittämättömästä syystä olit
joutunut. Mutta koska täällä emme saaneet kaikkia tarpeitamme
tyydytetyksi, jatkoimme retkeämme seuraavaan ja sitä seuraavaan
liikkeeseen, ja kun nämäkin sijaitsivat keskikaupungilla, kannoin
sinut toisesta toiseen niiden välisen lyhyen matkan. Ensimmäiset
valot sytytettiin katujen kulmauksissa, sanomalehtipoikain kiljunta
kävi yhä äänekkäämmäksi, polkupyöräilijäin tiu'ut helisivät yhä
tiheämmin, ja poika painoi yhä enemmän käsivarsillani. Hänen
päänsä lämmitti kosteana olkapäätäni, hänen hengityksensä kävi
tasaiseksi ja syväksi, hänen jäsenensä raukenivat rennoiksi. Hän
nukkui. Jo olin lysähtää kokoon uupumuksesta, ennenkuin tapasin
auton ja pääsin ystäviemme kotiin, jossa asuimme ja jossa suurella
myötätunnolla seurattiin tarinamme kehitystä.
Mutta tämä ensimmäinen uupumukseni sinun tähtesi värisi

ruumiissani niin suloisena raukeutena, että mieltäni vieläkin huumaa
sen muisto. Tunnen vieläkin pääsi painon rinnoillani ja käsivartesi
kiinteän puristuksen kaulani ympäri. Suu on puoliraollaan ja
valkoisten helmirivien lomitse hivelee poskeani hengityksesi keveä
kosketus. Ja minusta tuntuu kuin kaikella tällä minut olisi kultaisin
kirjaimin kirjoitettu onnellisten ihmisten kirjaan, ja kuin iloni
yltäkylläisyydessä tahtoisin jakaa omaa saamaani hyvää koko
maailmalle, siunaten kaikkia ja kaikkea.
Siunattu olkoon tämä hämärikkö ahtaitten katujen melussa!

Siunattu olkoon grammofoonien laulu avatuista kahvila-ikkunoista,
rattaiden räminä ja kellojen kilinä, raitiovaunujen vauhti ja
näyteikkunoitten silmiä hivelevä komeus! Siunattu olkoon tuokin
ihmismeri, joka hyväntahtoisesti sulki meidät hyrskivään helmaansa,
kieltämättä meiltä osuuttamme elämän onneen!
Siunattu olkoon tuo rauhaa henkivä päivänlaskun hetki, jolloin olin

vaipua maahan väsymyksestä sinua hoivatessani! Olisin iloiten
ponnistellut voimiani sinun tähtesi vielä tänäänkin ja joka päivä
tämän päivän maille mentyä. Ja tiedäthän, mikä onni minulle olisi
ollut, jos olisit pysynyt vanhuuteni huolenpidon esineenä ja päivieni
valona myöhäiseen iltaan asti.
*****
Vihdoin olimme valmiit matkustamaan kotiin.
Teit sen matkan kerran myöhemmin edestakaisin. Tiedät,

minkälainen on meri, kun se kirkkaina elokuun päivinä
auringonpaisteessa kimmeltää. Silloin se oli kuin peili, ja kalalokit
kiersivät laivaamme tehden pitkiä, loivia kaaria.
Sinä peräkannella katselit potkurin kiertämää ryöppyävää vaahtoa

tai paistattelit päivää polvellani istuen valkoisessa korituolissa
naisten salongin ikkunan alla. Se oli toinen esittelysi maailmalle,
paljon tärkeämpi kuin ensimmäinen. Mutta hauskinta meistä oli
piiloutua hyttiimme, jossa saimme leikkiä ja jutella mielinmäärin
kahden kesken. Siellä kenkäsi purjehtivat edestakaisin sohvalla, ja
sinä asuit milloin matkalaukussa, milloin sohvatyynyn takana. Ja
samallahan sinun oli opittava ensimmäiset suomalaiset sanasi —
minkä niminen oli oma paitasi, mitä rannukkaat puolisukkasi ja nuo
muutkin tavarat, jotka olimme yhdessä ostaneet tuona
puolihämyisenä iltapäivänä pauhaavassa suurkaupungissa.
Näihin ensimmäisiin tavaroihin on liittynyt niin paljon muistoja,

etten koskaan ole hennonut luopua niistä kaikista. Tiedät, että
minulla on niitä vieläkin joitakuita piironginlaatikossani, ja olet joskus
nauraen levittänyt ne polvellesi tai vetänyt lastensukat suurille
kämmenillesi. Lapselliseksi olet minua silloin sanonut ja olkapäitäsi
kohottaen minulle naurahdellut, mutta uskon sentään, että olet salaa
siitä pitänytkin. Aina, kun sattumalta olemme tulleet niitä yhdessä
katselleeksi tai edes vilaukseltakaan nähneeksi, on edessämme ollut
hohtava meren pinta, kone on kumeasti jyskyttänyt lattian alla ja
olemme häiritsemättä omistaneet toisemme ja aarteemme
ihastuttavassa kammiossamme.
Oli ainoastaan pari tuttavaa laivassa, ja muistan vielä, miten

eräänä päivänä toinen heistä ihmetellen sanoi minulle:
3°
— Sinähän jumaloit häntä!
Totta hän puhui. Minä jumaloin sinua, Yrjö.
*****
Kun kallioinen saaristomme lähestyi, nousimme kannelle. Siellä

näkyi — Suomi! Sitä sanaa olimme hartaasti harjoittaneet pitkin
matkaa, ja sinä lausuit sen jo miltei suomalaisesti ääntäen. Suomi oli
jotakin suunnattomasti hauskaa. Kaikki oli Suomea. Metsäiset
saaret, vaaleaksi maalatut huvilat, veneet laivasiltojen kupeella, uivat
lapset rannoilla, haukkuvat koirat kallioilla. Jo näkyivät Viaporin
linnoitukset ja vallit, kirkot ja sipulinmuotoiset kupukatot. Molemmin
puolin kapeata väylää laivat ja moottoriveneet jyskyttivät ja
täksyttivät, ja vihdoin sukelsi meren helmasta esille valkoisen
hohtava kaupunki torneineen ja huippuineen.
Näin tuon venäläisen lipun Viaporin valleilla ikäänkuin olisin

nähnyt sen ensi kerran. Ja nuo venäläiset sotamiehet, jotka siinä
vahtia pitivät pyssyt olallaan. Selkääni karmi ja melkein häpesin.
Sellaisen orjuutetun kansan jäseneksi hänet tuon! Vapaasta
kansasta sorron alaisuuteen! Mutta olkoon, se on sittenkin kunniakas
maa, tämä maamme. Emme ole koskaan sisäisesti taipuneet, ja kai
Another random document with
no related content on Scribd:
young Cæsar Valentinian, between East and West, a new, if
unsubstantial, cordiality appeared. Italy at least was restored to
prosperity, while in Aetius she possessed a general as great as the
great Stilicho. But if Italy was safe the provinces were in peril and
she herself saw Africa betrayed by Boniface and ravaged by and lost
to the Vandals under Genseric. Nor was the domestic state of her
household and court such as to inspire her with confidence in the
future. If her son Valentinian was a foolish and sensual boy, her
daughter Honoria was discovered in a low intrigue with a
chamberlain of the palace, and when in exile at Constantinople sent,
perhaps longing for the romantic fate of her mother, her ring to the
new and youthful King of the Huns, soon to be famous as Attila,
inviting him to carry her off as Adolphus, the Goth, had carried off
Placidia.
Such was the condition of things in the royal household of the
West. In Constantinople things were not more promising.
Theodosius, the young Emperor, called the Calligrapher, was a
dilettante of the fine arts, not a statesman. Those who surrounded
him were mediocrities intent rather on theological controversies than
on the safety of the State, or sunk in a cynical corruption in which
everything noble was lost. No one East or West seemed able to
grasp or to realise that there was any danger. Had the Imperial
Governments failed altogether to understand the fundamental cause
of the Gothic advance, the Vandal attack, indeed of all their
embarrassments? Had they failed to remember what was there
beyond the Rhine and the Danube? Had they forgotten the Huns?
FOOTNOTES:
[2] See my “Ravenna” (Dent, 1913), pp. 1-10.
II
THE HUNS AND ATTILA
The people called the Huns, “scarcely mentioned in other records,”

are fully described by that Ammianus Marcellinus[3] whom I have
already quoted. He lived at the end of the fourth century, was a
Roman historian born of Greek parents at Antioch, and after fighting
in Gaul, in Germany and the East, settled in Rome and devoted
himself to history. He describes the Huns as “living beyond the Sea
of Azov on the borders of the Frozen Ocean.” And adds that they
were a people “savage beyond all parallel.” He then gives us the
following careful description of them:—
“In their earliest infancy deep incisions are made in the cheeks of
their boys[4] so that when the time comes for the beard to grow the
sprouting hairs may be kept back by the furrowed scars, and
therefore they grow to old age as beardless as eunuchs. At the same
time all have strong and well-built limbs and strong necks; they are
indeed of great size, but so short-legged that you might fancy them
to be two-legged beasts, or the figures which are hewn out in a rude
manner with an axe on the posts at the end of bridges.[5]
“They do, however, just bear the likeness of men (horribly ugly
though they be), but they are so little advanced in civilisation that
they make no use of fire, nor of seasoned food, but live on roots
which they find in the fields, or on the half raw flesh of any animal
which they merely warm a little by placing it between their own thighs
and the backs of their horses.
“They do not live under roofed houses but look upon them as
tombs and will only enter them of necessity. Nor is there to be found
among them so much as a cabin thatched with reed; but they
wander about over the mountains and through the woods training
themselves to bear from their infancy the extremes of frost and
hunger and thirst.
“They wear linen clothes or else the skins of field mice sewn
together, and this both at home and abroad. When once such a tunic
is put on, it is never changed till from long decay it falls to pieces.
Their heads are covered with round caps and their hairy legs with
goat skins and their shoes which are ignorant of any last are so
clumsy as to hinder them in walking.
“For this cause they are not well suited for infantry; but, on the
other hand, they are almost one with their horses, which are poorly
shaped but hardy; often they sit them like women. In truth they can
remain on horseback night and day; on horseback they buy and sell,
they eat and drink, and bowed on the narrow neck of their steeds
they even sleep and dream. On horseback too they discuss and
deliberate. They are not, however, under the authority of a king, but
are content with the loose government of their chiefs.
“When attacked they sometimes engage in regular battle formed in
a solid body and uttering all kinds of terrific yells. More often,
however, they fight irregularly, suddenly dispersing, then reuniting
and after inflicting huge loss upon their enemy will scatter over the
plains hither and thither, avoiding a fortified place or an
entrenchment. It must be confessed that they are very formidable
warriors....
“None of them ploughs or even touches a plough-handle; for they
have no settled abode, but are alike homeless and lawless,
continually wandering with their waggons which indeed are their
homes. They seem to be ever in flight.... Nor if he is asked can any
one tell you where he was born; for he was conceived in one place,
born in another far away, and bred in another still more remote.
“They are treacherous and inconstant and like brute beasts are
utterly ignorant of the distinction between right and wrong. They only
express themselves with difficulty and ambiguously, have no respect
for any religion or superstition, are immoderately covetous of gold,
and are so fickle and cantankerous that many times in a day they will
quarrel with their comrades without cause and be reconciled without
satisfaction.”[6]
Such were the people who according to Ammianus were “the
original cause of all the destruction and manifold calamities” which
descended upon the Roman Empire, in the fifth century of our era.
Fifty-six years before they began directly to menace civilisation
and the Roman Empire, they had, as we have seen, in 376 a.d.,
driven the Goths before them to the first of those famous assaults
upon the frontiers of the Roman world. They themselves, utter
barbarians as they were, attempted then no direct attack upon our
civilisation, though in 396 they crossed the Caucasus, raided
Armenia and as Claudius notes, “laid waste the pleasant fields of
Syria.” In 409, however, Alaric being then intent on Italy, they
crossed the Danube and pushed on into Bulgaria, Uldis, their chief,
boasting in true Barbarian fashion, “All that the sun shines upon I
can conquer if I will.” It was the first claim of the Barbarian, vocal and
explicit, to “a place in the sun”—someone else’s place. Uldis’ boast,
however, had been but the prelude to his flight and fall. Amid the
welter of Barbarians less barbarous than he, Visigoths, Vandals,
Suevi, Alani, the Hun in fact was unable to do much more than drive
them on. When they had passed into the Empire, into Gaul and
Spain and Africa, he, worse than them all, was free at last to
threaten Christendom and its capitals, Constantinople and Rome.
It was not till the two brothers Attila and Bleda ascended the
Hunnish throne, if throne it can be called, in the year 423, that the
Huns really became immediately and directly dangerous to
civilisation.
That civilisation already half bankrupt and in transition had, as we
have seen, been bewildered and wounded by the actual incursion of
Barbarian armies south of the Danube and the Rhine, nay within the
heart of the Empire, within reach of Constantinople, within the very
walls of Rome. It was now to be assaulted by a savage horde, wholly
heathen, intent on murder and rape, loot and destruction.
The contrast between the two attacks, the attack of Alaric and that
of Attila, is very striking. To admire Alaric, even to defend him, is
obviously not impossible, since so many historians have been found
ready to do both. No voice unless it be Kaiser Wilhelm’s has ever
been raised in behalf of Attila. Here was the Empire, Christendom;
he fell upon it like a wild beast. At least the Goths were Christian—
though Arian—the Huns were pagan heathen. At least Alaric had
revered the Roman name and sought to assume it; Attila despised
and hated it and would have destroyed it utterly. But if there is this
moral contrast between the Gothic and the Hunnish attacks upon the
Empire, militarily they are alike in this above all that both were
directed first upon the East and were only turned upon the West after
a sort of failure. Happily for us the attacks of Attila, while infinitely
more damaging, were not nearly so dangerous as those of Alaric.
The Empire was assaulted by an assassin; it was delivered.
The Roman system with regard to the Barbarians had long been
established when Theodosius II ascended the Eastern throne. It
consisted not only in employing Barbarians as auxiliaries—thus Uldis
and his Huns had fought under Stilicho against Radagaisus at the
battle of Fiesole; but in setting the different Barbarian tribes and
races one against another. The Huns especially had been favoured
by the Empire in this way, Stilicho knew them well and Aetius who
was at last to defeat them upon the Catalaunian plains owed them
perhaps his life in the crisis that followed the death of his rival
Boniface in 433. But that policy, always dangerous, and the more so
if it were inevitable, was already bankrupt. The dispersal through the
provinces of the Goths, the Vandals, the Alani, Suevi and other tribes
left the Empire face to face upon its northern frontier with the real
force which had driven them on. In 432 we find Roua, King of the
Huns, in receipt of an annual subsidy, scarcely to be distinguished
from a tribute, of 350 pounds’ weight of gold. He it was who perhaps
first broke the old Roman policy. When the Empire, according to its
custom, made alliances with certain Barbarian tribes his neighbours,
he claimed them as his subjects and immediately swore that he
would denounce all his treaties with the Empire unless the Emperor
broke off these alliances. Moreover, he demanded that all those of
his subjects then within the Empire should be restored to him; for
many had entered the Roman service to escape his harsh rule.
These demands could not be ignored or refused. In 433 Theodosius
was on the point of sending an embassy to treat with Roua, when he
heard that he was dead and that his two nephews, still young men,
Attila and Bleda, had succeeded him. It was they who received the
Imperial ambassadors.
The conference met on the right bank of the Danube within the
Empire, that is near the Roman town of Margus or Margum, a city of
Moesia, where the Danube and the Morava meet. The place was
known as the Margum planum on account of the character of the
country, and was famous as the spot where Diocletian had defeated
Carinus.[7]
The Byzantine historian Priscus has left us an account of this
strange meeting. The Huns it seems came on horseback and as they
refused to dismount the Roman ambassadors also remained on their
horses. It was thus they heard the arrogant demands of the Hunnish
kings: the denunciation by Theodosius of his alliance with the
Barbarians of the Danube, the expulsion of all the Huns serving in
the Imperial armies or settled within the Empire, an undertaking not
to assist any Barbarian people at war with the Huns, and the
payment by the Empire as tribute, tributi nomine, of seven hundred
pounds’ weight of gold instead of the three hundred and fifty given
hitherto. To all these demands the ambassadors were forced to
agree as Attila insisted either upon their acceptance or upon war,
and Theodosius preferred any humiliation to war. The famous
conference of Margus was thus a complete victory for the Huns, a
victory Attila never forgot.
That Theodosius was ready to accept any terms which Attila might
insist upon is proved by the fact that he immediately delivered up to
him his two guests, young princes of the Huns, and made no protest
when Attila crucified them before the eyes of his ambassadors.
This act seems to symbolise at the outset the character of Attila
and his reign. He was then, we may suppose, between thirty and
forty years old, and although the younger always the master of his
brother Bleda, whom he was soon to murder. Of the place of his birth
we know nothing,[8] but he grew up on the Danube and there learned
the use of arms, perhaps in the company of the young Aetius, who
had been a Roman hostage of Roua and who was one day to
conquer Attila. If we look for a portrait of him we shall unhappily not
find it in any contemporary writer; but Jornandes, probably repeating
a lost passage of some earlier writer, perhaps Priscus himself, tells
us that he was short, with a mighty chest, a large head, eyes little
and deep-set, a scant beard, flat nose and dark complexion. He
thrust his head forward as he went and darted his glances all about,
going proudly withal, like one destined to terrify the nations and
shake the earth. Hasty and quarrelsome, his words, like his acts,
were sudden and brutal, but though in war he only destroyed, and
left the dead unburied in their thousands for a warning; to those who
submitted to him he was merciful, or at least he spared them. He
dressed simply and cleanly, ate as simply as he dressed, his food
being served on wooden dishes; indeed his personal temperance
contrasted with the barbaric extravagance he had about him.
Nevertheless he was a Barbarian with the instincts of a savage.
Constantly drunk he devoured women with a ferocious passion,
every day having its victim, and his bastards formed indeed a
people. He knew no religion but surrounded himself with sorcerers,
for he was intensely superstitious.[9] As a general he was seldom in
the field, he commanded rather than led and ever preferred
diplomacy to battle.[10] His greatest weapon was prevarication. He
would debate a matter for years and the continual embassies of
Theodosius amused without exhausting him and his patience. He
played with his victims as a cat does with a mouse and would always
rather buy a victory than win it. He found his threat more potent than
his deed, and in fact played with the Empire which had so much to
lose, very much as Bismarck played with Europe. Like Bismarck too
his business was the creation of an Empire. His idea, an idea that
perhaps even Roua had not failed to understand, was the creation of
an Empire of the North, a Hunnish Empire, in counterpoise against
the Roman Empire of the South, to the south that is of the Rhine and
the Danube. For this cause he wished to unite the various Barbarian
tribes and nations under his sceptre, as Bismarck wished to unite the
tribes of the Germans under the Prussian sword. He was to be the
Emperor of the North as the Roman Emperors were Emperors of the
South. Had he lived in our day he would have understood that
famous telegram of the Kaiser to the Tsar of Russia—“the Admiral of
the Atlantic....”
It was the business of Theodosius to prevent the realisation of this
scheme, nor did he hesitate to break the treaty of Margus to achieve
this. His emissaries attempted to attach to the Empire the Acatziri, a
Hunnish tribe that had replaced the Alani on the Don. Their chief,
however, fearing for his independence, or stupidly handled, sent
word to Attila of the Roman plot. The Hun came down at the head of
a great army, and though he spared the Acatziri, for their chief was
both wily and a flatterer, he brought all the Barbarians of that part
within his suzerainty and, returning, soon found himself master of an
Empire which stretched from the North Sea to the Caucasus, and
from the Baltic to the Danube and the Rhine, an Empire certainly in
extent comparable with that of Rome.
It was in achieving this truly mighty purpose that Attila exhibits two
of his chief characteristics, his superstition and his cruelty.
It seems that the ancient Scythians on the plain to the east of the
Carpathians had for idol and perhaps for God a naked sword, its hilt
buried in the earth, its blade pointed skyward. To this relic the
Romans had given the name of the sword of Mars. In the course of
ages the thing had been utterly forgotten, till a Hunnish peasant
seeing his mule go lame, and finding it wounded in the foot, on
seeking for the cause, guided by the blood, found this sword amid
the undergrowth and brought it to Attila who recovered it joyfully as a
gift from heaven and a sign of his destined sovereignty over all the
peoples of the earth. So at least Jornandes relates.[11]
The other episode exhibits his cruelty. In founding his empire Attila
had certainly made many enemies and aroused the jealousy of those
of his own house. At any rate he could not remember without
impatience that he shared his royalty with Bleda. To one of his
subtlety such impatience was never without a remedy. Bleda was
accused of treason, perhaps of plotting with Theodosius, and Attila
slew his brother or had him assassinated; and alone turned to enjoy
his Barbary and to face Rome.
FOOTNOTES:
[3] In the thirty-first book of his History of Rome: see Appendix
I.
[4] The Prussian student is even to-day famous for the scars on
his face inflicted in the duels at the Universities.
[5] Cf. the physique of the ordinary Prussian at its most
characteristic in Von Hindenberg, who really seems to have been
hewn out of wood.
[6] It was a modern and famous German who not long since
declared that the Prussians were such quarrelsome and
disagreeable brutes that it was only their propensity to drink beer
and that continually that mollified them sufficiently to be regarded
as human beings.
[7] It is curious to remember that this first encounter of Attila
with the Imperial power took place in what is now Servia only fifty
miles further down the Danube than Belgrade.
[8] It has been suggested that his name Attila is that of the
Volga in the fifth century and that therefore he was born upon its
banks; but as well might one say that Roua was born there
because one of the ancient names of that river was Rha.
[9] For all this see Appendix: Jornandes, R. Get., 35 and
especially for his dress and food, Priscus, infra.
[10] Cf. Jorn., R. Get., 36: “Homo subtilis antequam arma
gereret , arte pugnabat....”
[11] See Appendix, Jornandes, R. Get., 35.
III
ATTILA AND THE EASTERN EMPIRE
When Attila had achieved the hegemony of the North he turned his
attention upon the Empire; and it is curious for us at this moment to
note the coincidence that this first attack upon civilisation was
delivered at the very spot upon the Danube where the Germanic
powers in August, 1914, began their offensive. Attila directed his
armies upon the frontiers of modern Servia at the point where the
Save joins the Danube, where the city of Singidunum rose then and
where to-day Belgrade stands.
The pretext for this assault was almost as artificial and
manufactured as that which Austria put forward for her attack upon
Servia. Attila asserted that the Bishop of that same frontier town of
Margus, on the Morava, where he had made treaty with the Empire,
had crossed the Danube, and having secretly obtained access to the
sepulchre of the Hunnish kings had stolen away its treasures. The
Bishop, of course, eagerly denied this strange accusation, and it
seemed indeed so unlikely that he was guilty that Theodosius was
exceedingly reluctant to sacrifice him. The people of Moesia
clamoured for a decision; if the Bishop were guilty then he must be
delivered to Attila, but if not Theodosius must protect both him and
them. For Attila had waited for nothing; he had crossed the Danube
before making his accusation and had occupied Viminacium, one of
the greater towns upon the frontier.
Meanwhile the Bishop, seeing the hesitation of Theodosius and
expecting to be sacrificed, made his way to the camp of the Huns
and promised in return for his life to deliver Margus to them, and this
he did upon the following night. Then, dividing his forces into two
armies, Attila began his real attack upon the Empire.
The first of these armies was directed upon Singidunum, the
modern Belgrade, which was taken and ruined, and when that was
achieved it proceeded up the Save to Sirmium, the ancient capital of
Pannonia, which soon fell into its hands. The second crossed the
Danube further eastward and besieged Ratiaria, a considerable
town, the head-quarters of a Roman Legion and the station of the
fleet of the Danube.
THE ATTACK OF ATTILA UPON THE EAST.

Having thus, with this second army, secured the flank, Attila
marched his first army from Singidunum up the Morava to Naissus
(Nisch), precisely as the Austrians tried to do but yesterday. They
failed, but he succeeded and Naissus fell. Thence he passed on to
Sardica where he was met by his second army which had taken
Ratiaria. Sardica was pillaged and burnt.
Attila thus possessed himself in the year 441 of the gateways of
the Balkans, almost without a protest from Theodosius. Five years
later, in 446, he was ready to advance again. In that year and the
next he destroyed two Roman armies, took and pillaged some
seventy towns, and pushed south as far as Thermopylae, and
eastward even to Gallipoli; only the walls of Constantinople saved
the capital. Theodosius was forced to buy a disgraceful peace at the
price of an immediate payment of 6000 pounds’ weight of gold, an
annual tribute, no longer even disguised, of 2000 pounds, and an
undertaking that the Empire would never employ or give refuge to
any of those whom Attila claimed as his subjects.
It was easier to agree to such terms than to fulfil them. The
provinces were ruined, the whole fiscal system of the East in
confusion, and even what wealth remained was, as Priscus tells us,
“spent not in national purposes, but on absurd shows and gaudy
pageants, and all the pleasures and excesses of a licentious society
such as would not have been permitted in any properly governed
State, even in the midst of the greatest prosperity.” Attila, who
marked the decay and the embarrassment of the Imperial
Government, forewent nothing of his advantage. He became more
and more rapacious. When he did not obtain all he desired he sent
an embassy to Constantinople to intimidate the government, and this
became a regular means of blackmail with him, a means more
humiliating than war and not less successful.
The first of these embassies arrived in Constantinople immediately
after the terms of peace had been agreed upon. It made further
demands, and was treated with the most extravagant hospitality.
Three times within a single year other embassies arrived; they were
a means of blackmail and were assured of an ever-increasing
success.
The most famous and the most important of these embassies was
that which arrived in Constantinople in 449. The ambassadors then
employed by Attila are worthy of notice, for in them we see not only
the condition of things at that time, but also the naive cunning of the
Hun. The two chief legates whom Attila dispatched to Constantinople
upon this occasion were Edecon and Orestes. Edecon was a
Scythian or Hun by birth, a heathen of course, and a Barbarian, the
commander of the guard of Attila, and the father of Odoacer, later to
be so famous. Orestes, on the other hand, who was one of Attila’s
chief ministers, was a Roman provincial of Pannonia, born at
Petavium (probably Pettau on the Drave), who had made a fortunate
marriage as a young man when he allied himself with Romulus, a
considerable Roman personage of that province. He had, however,
deserted the Imperial service, certainly open to him, for that of the
Barbarians, and had made his fortune. Nor was his part in history to
be played out in the service of Attila, for his son Romulus was to be
the last of the Western Emperors, contemptuously known to history
as Romulus Augustulus.
Orestes was then an adventurer pure and simple, but in sending
him with the Barbarian Edecon, we see the system of Attila in his
blackmail of the Empire. The employment of a Roman provincial was
a check upon the Barbarian envoy. A bitter jealousy subsisted
between them, each spied on the other, and thus Attila was well
served. The fact that the Hun was able to command the services of
such as Orestes is a sufficient comment upon the condition of the
frontier provinces.
It was these two jealous envoys that, in the early months of 449,
appeared in Constantinople bringing, of course, new demands. Their
mission, indeed, was the most insolent that Attila had so far dared to
send. It demanded three main things; first, that all the country to the
south of the Danube as far as Naissus should be regarded as a part
of the Hunnish Empire; second, that in future Theodosius should
send to the Hunnish court only the most illustrious ambassadors, but
if this were done Attila for his part would consent to meet them on
the frontier at Sardica; third, that the refugees should be delivered
up. This last demand was a repetition of many that had gone before
it. As before Attila threatened if his requests were not granted he
would make war.
The ambassadors Edecon and Orestes came to Constantinople
where a “Roman” named Vigilas acted as their guide and interpreter,
an indiscreet and vulgar fellow of whom we shall hear more
presently. Received in audience by Theodosius in the famous palace
on the Bosphorus, the ambassadors with the interpreter later visited
the chief minister, the eunuch Chrysaphius. On their way they
passed through the noble halls of Constantine decorated with gold
and built of marble, the whole a vast palace, perhaps as great as the
Vatican. Edecon, the Hun, was stupefied by so much splendour, he
could not forbear to express his amazement; Vigilas was not slow to
mark this naive astonishment nor to describe it to Chrysaphius, who
presently proposed to put it to good use. Taking Edecon apart from
Orestes as he talked he suggested to him that he also might enjoy
such splendour if he would leave the Huns and enter the service of
the Emperor. After all it was not more than Orestes had done. But
Edecon answered that it would be despicable to leave one’s master
without his consent. Chrysaphius then asked what position he held
at the court of Attila, and if he was so much in the confidence of his
master as to have access freely to him. To which Edecon answered
that he approached him when he would, that he was indeed the chief
of his captains and kept watch over his person by night. And when
Chrysaphius heard this he was content and told Edecon that if he
were capable of discretion he would show him a way to grow rich
without trouble, but that he must speak with him more at leisure,
which he would do presently if he would come and sup with him that
evening alone without Orestes or any following. Already in the mind
of the eunuch a plan was forming by which he hoped to rid the
Empire once for all of the formidable Hun.
Edecon accepted the invitation. Awaiting him he found Vigilas with
Chrysaphius, and after supper heard apparently without
astonishment the following amazing proposal. After swearing him to
secrecy, Chrysaphius explained that he proposed to him the
assassination of Attila. “If you but succeed in this and gain our
frontiers,” said he, “there will be no limit to our gratitude, you shall be
loaded with honours and riches.”
The Hun was ready in appearance at least to agree, but he
insisted that he would need money for bribery, not much, but at least
fifty pounds’ weight of gold. This he explained he could not carry
back with him as Attila was wont upon the return of his ambassadors
to exact a most strict account of the presents they had received, and
so great a weight of gold could not escape the notice of his own
companion and servants. He suggested then that Vigilas should
accompany him home under the pretext of returning the fugitives and
that at the right moment he should find the money necessary for the
project. Needless to say, Chrysaphius readily agreed to all that
Edecon proposed. He does not seem either to have been ashamed
to make so Hunnish a proposal or to have suspected for a moment
that Edecon was deceiving him. He laid all before Theodosius, won
his consent and the approval of Martial his minister.
Together they decided to send an embassy to Attila, to which the
better to mask their intentions Vigilas should be attached as
interpreter. This embassy they proposed to make as imposing as
possible, and to this end they appointed as its chief a man of a high,
but not of consular rank, and of the best reputation. In this they
showed a certain ability, for as it seemed to them if their plot failed
they could escape suspicion by means of the reputation of their
ambassador. The man they chose was called Maximin, and he
fortunately chose as his secretary Priscus, the Sophist, to whose pen
we are indebted for an account of all these things. He asserts, and
probably with truth, that neither Maximin nor he himself was aware of
the plot of assassination. They conceived themselves to be engaged
in a serious mission and were the more impressed by its importance
in that its terms were far less subservient to the Hun than had been
the custom in recent times. Attila was told that henceforth he must
not evade the obligations of his treaties nor invade at all the Imperial
territories. And with regard to the fugitives he was informed that
beside those already surrendered seventeen were now sent but that
there were no more. So ran the letter. But Maximin was also to say
that the Hun must look for no ambassador of higher rank than
himself since it was not the Imperial custom towards the Barbarians;
on the contrary, Rome was used to send to the North any soldier or
messenger who happened to be available. And since he had now
destroyed Sardica his proposal to meet there any ambassador of
consular rank was merely insolent. If indeed the Hun wished to
remove the differences between Theodosius and himself he should
send Onegesius as ambassador. Onegesius was the chief minister
of Attila.
Such were the two missions, the one official, the other secret,
which set out together from Constantinople.
The great journey seems to have been almost wholly uneventful
as far as Sardica, 350 miles from Constantinople, which was
reached after a fortnight of travel. They found that town terribly
pillaged but not destroyed, and the Imperial embassy bought sheep
and oxen, and having prepared dinner invited Edecon and his
colleagues to share it with them, for they were still officially within the
Empire. But within those ruins, even among the ambassadors, peace
was impossible. Priscus records the ridiculous quarrel which
followed. The Huns began to magnify the power of Attila,—was not
his work around them? The Romans knowing the contents of the
letter they bore sang the praises of the Emperor. Suddenly Vigilas,
perhaps already drunk, asserted that it was not right to compare men
with the gods, nor Attila with Theodosius, since Attila was but a man.
Only the intervention of Maximin and Priscus prevented bloodshed,
nor was harmony restored till Orestes and Edecon had received
presents of silk and jewels. Even these gifts were not made
altogether without an untoward incident. For Orestes in thanking
Maximin exclaimed that he, Maximin, was not like those insolent
courtiers of Constantinople “who gave presents and invitations to
Edecon, but none to me.” And when Maximin, ignorant of the
Chrysaphian plot, demanded explanations, Orestes angrily left him.
Already the plan of assassination was beginning to fester.
The ambassadors went on from ruined Sardica to desolate
Naissus (Nisch) utterly devoid of inhabitants, full only of horror and
ruins. They crossed a plain sown with human bones whitening in the
sun, and saw the only witness to the Hunnish massacre of the
inhabitants—a vast cemetery. “We found,” Priscus tells us, “a clean
place above the river where we camped and slept.”
Close to this ruined town was the Imperial army, commanded by
Agintheus, under whose eagles five of the seventeen refugees to be
surrendered had taken refuge. The Roman general, however, was
obliged to give them up. Their terror as they went on in the
ambassadorial train towards the Danube may well be imagined.
The great river at length came in sight; its approaches lined and
crowded with Huns, the passages served by the Barbarians in dug-
outs, boats formed out of the hollowed trunks of trees. With these
boats the whole Barbarian shore was littered as though in readiness
for the advance of an army. Indeed, as it appeared Attila was in
camp close by, and intent on hunting within the Roman confines to
the south of the river, a means certainly of reconnaissance as
habitually used by the Huns as commerce has been for the same
end by the Germans.
We do not know with what feelings Maximin and Priscus saw all
this and crossed the great river frontier at last and passed into
Barbary. To their great chagrin, for they had made the way easy for
the Hunnish ambassadors on the road through the Imperial
provinces, Edecon and Orestes now left them brusquely enough. For
several days they went on alone but for the guides Edecon had left
them, till one afternoon they were met by two horsemen who
informed them that they were close to the camp of Attila who awaited
them. And indeed upon the morrow they beheld from a hill-top the
Barbarian tents spread out innumerable at their feet, and among
them that of the King. They decided to camp there on the hill; but a
troop of Huns at once rode up and ordered them to establish
themselves in the plain. “What,” cried they, “will you dare to pitch
your tents on the heights when that of Attila is below?”
They were scarce established in their appointed place when to
their amazement Edecon and Orestes and others appeared and
asked their business, the object of their embassy. The astonished
ambassadors looked at one another in amaze. When the question
was repeated Maximin announced that he could not disclose his
mission to any other than Attila to whom he was accredited. Scotta,
the brother of Onegesius, then announced angrily that Attila had sent
them and they must have an answer. When Maximin again refused
the Huns galloped away.
The Romans, however, were not left long in doubt of the reception
they were to get. Scotta and his friends soon returned without
Edecon, and to the further amazement of Maximin repeated word for
word the contents of the Imperial letter to Attila. “Such,” said they, “is
your commission. If this be all depart at once.” Maximin protested in
vain. Nothing remained but to prepare for departure. Vigilas who
knew what Chrysaphius expected was particularly furious; better
have lied than to return without achieving anything, said he. What to
do? It was already night. They were in the midst of Barbary, between
them and the Danube lay leagues of wild unfriendly country.
Suddenly as their servants loaded the beasts for their miserable
journey other messengers arrived from the Hun. They might remain
in their camp till dawn. In that uneasy night, had Vigilas been less of
a fool, he must have guessed that Edecon had betrayed him.
It was not the barbarous Vigilas, however, who found a way out of
the difficulty, for at dawn the command to depart was repeated, but
that Priscus who has left us so vivid an account of this miserable
affair. He it was who, seeing the disgrace of his patron, sought out
Scotta, the brother of Onegesius, the chief minister of Attila, in the
Hunnish camp. With him went Vigilas as interpreter, and so cleverly
did the Sophist work upon the ambition of Scotta, pointing out to him
not only the advantages of peace between the Huns and the
Romans, but also the personal advantage Scotta would gain thereby
in honour and presents, and at last feigning to doubt Scotta’s ability
to achieve even so small a matter as the reception of the embassy
that he had his way. Scotta rode off to see Attila, Priscus returned to
his patron, and soon after Scotta returned to escort them to the royal
tent.
The reception must have been a strange spectacle. The tent of
Attila was quite surrounded by a multitude of guards; within, upon a
stool of wood, was seated the great Hun. Priscus, Vigilas and the
servants who attended them bearing the presents remained upon
the threshold. Maximin alone went forward and gave into Attila’s
hands the letter of Theodosius saying: “The Emperor wishes Attila
and all that are his health and length of days.” “May the Romans
receive all they desire for me,” replied the instructed Barbarian. And
turning angrily to Vigilas he said: “Shameless beast, why hast thou
dared to come hither knowing as thou dost the terms of peace I
made with thee and Anatolius. Did I not then tell thee that I would
receive no more ambassadors till all the refugees had been
surrendered!” Vigilas replied that they brought seventeen fugitives
with them and that now there remained no more within the Empire.
This only made Attila more furious: “I would crucify thee and give
thee as food for the vultures but for the laws regarding envoys,” cried
he. As for the refugees, he declared there were many still within the
Empire, and bade his people read out their names, and this done he
told Vigilas to depart with Eslas, one of his officers, to inform
Theodosius that he must forthwith return all the fugitives who had
entered the Empire from the time of Carpilio, son of Aetius, who had
been his hostage. “I will never suffer,” said he, “that my slaves shall
bear arms against me, useless though they be to aid those with
whom they have found refuge.... What city or what fortress have they
been able to defend when I have determined to take it?” When he
had said these words he grew calmer; informed Maximin that the
order of departure only concerned Vigilas, and prayed the
ambassador to remain and await the reply to the letter of the
Emperor. The audience closed with the presentation and acceptance
of the Roman presents.
Vigilas must surely have guessed now what his dismissal meant.
Perhaps, however, he was too conceited and too stupid to notice it.
At any rate he did not enlighten his companions but professed
himself stupefied by the change of Attila’s demeanour towards him.
The whole affair was eagerly discussed in the Roman camp. Priscus
suggested that Vigilas’ unfortunate indiscretion at Sardica had been
reported to Attila and had enraged him. Maximin did not know what
to think. While they were still debating Edecon appeared and took
Vigilas apart. The Hun may well have thought he needed
reassurance. He declared that he was still true to the plan of
Chrysaphius. Moreover, seeing what a fool Vigilas was, he told him
that his dismissal was a contrivance of his own to enable the
interpreter to return to Constantinople and fetch the money
promised, which could be introduced as necessary to the embassy
for the purchase of goods. Vigilas, however, can scarcely have
believed him, at any rate for long; a few hours later Attila sent word
that none of the Romans were to be allowed to buy anything but the
bare necessities of life from the Huns, neither horses, nor other
beasts, nor slaves, nor to redeem captives. Vigilas departed with the
order ringing in his ears, upon a mission he must have known to be
hopeless.
Two days later Attila broke camp and set out for his capital, the
Roman ambassadors following in his train under the direction of
guides appointed by the Hun. They had not gone far on their way
northward when they were directed to leave the train of Attila and to
follow another route, because, they were told, the King was about to
add one more to his innumerable wives, Escam, the daughter of a
chief in a neighbouring village.
Very curious is Priscus’ description of the way followed by the
patron and his embassy. They journeyed across the Hungarian plain,
across horrible marshes and lakes which had to be traversed
sometimes on rafts; they crossed three great rivers, the Drave, the
Temes, and the Theiss in dug-outs, boats such as they had seen on
the Danube hollowed out of the trunks of trees. They lived for the
most part on millet which their guides brought or took from the
wretched inhabitants, they drank mead and beer, and were utterly at
the mercy of the weather, which was extremely bad. On one
occasion, indeed, their camp was entirely destroyed by tempest, and
had it not been for the hospitality of the widow of Bleda they would
perhaps have perished.
For seven days they made their way into the heart of Hungary till
they came to a village where their way joined the greater route by
which Attila was coming. There they were forced to await the King,
since they must follow and not precede him. It was in this place that
they met another Roman embassy, that of the Emperor in the West,
Valentinian III, who was quarrelling with Attila about the holy vessels
of Sirmium. It seems that the Bishop of Sirmium in 441, seeing his
city invested, had gathered his chalices and patens and plate,
sacred vessels of his church, and had sent them secretly to a certain
Constantius, a Gaul, at that time Attila’s minister. In case the city fell
they were to be used as ransom, first of the Bishop, and in case of
his death of any other captives. Constantius was, however, untrue to
the trust placed in him by the Bishop, and sold or pawned the plate
to a silversmith in Rome. Attila hearing of it when Constantius was
beyond his reach claimed the booty as his own. It was upon this
miserable business that Valentinian had sent an embassy to Attila
from Ravenna.
It is certainly a shameful and an amazing spectacle we have here.
In that little village of Barbary the ambassadors of the Emperors,

Machine Learning For Planetary Science Ebook PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Machine Learning For Planetary Science Ebook PDF

Uploaded by

Copyright:

Available Formats

Machine Learning for Planetary Science

MATLAB® is a trademark of The MathWorks, Inc. and is used with permission.

For information on all Elsevier publications

Publisher: Candice Janco

1. Introduction to machine learning 1

2. The new and unique challenges of planetary missions 25

3. Finding and reading planetary data 39

4. Introduction to the Python Hyperspectral Analysis Tool (PyHAT) 55

6. Planetary image inpainting by learning mode-speciﬁc regression

7. Automated surface mapping via unsupervised learning and

8. Mapping storms on Saturn 151

9. Machine learning for planetary rovers 169

Masahiro Ono, Brandon Rothrock, Yumi Iwashita, Shoya Higa,

10. Combining machine-learned regression models with Bayesian

Introduction to machine learning

1.1 Overview of machine learning methods

1.2 Supervised learning

Figure 1.2 Classiﬁer as function.

making use of a small number of labeled examples are discussed in Section

The “decision boundary” is therefore defined by the linear relation y(x) =

Figure 1.4 Example of overﬁtting.

likely could be approximated using a simpler (e.g., linear) decision bound-

1.3 Unsupervised learning

in this section. Other types of unsupervised learning applications include

Figure 1.5 Example of Spectral Clustering.

on the sparseness of the similarity matrix. It can also result in improved

1.3.2 Dimensionality reduction

manifold. Dimensionality reduction methods learn a mapping from high-

1.4 Semisupervised learning

1.4.2 Self-training with Expectation Maximization

1.5 Active learning

1.5.1 Uncertainty sampling

1.6 Popular machine learning methods

1.6.1 Principal component analysis

1.6.2 K-means clustering

by randomly initializing k cluster centroids and iteratively performing two

1.6.3 Support vector machines

Figure 1.8 K-means clusters of PCA-reduced HiRISE data set.

margin ensures that the decision boundary will be as generalizable as pos-

1.6.4 Decision trees and random forests

tures (mean, standard deviation, minimum, maximum, skew, and kurtosis)

1.6.5 Neural networks

set. Fine-tuning consists of initializing a network with weights from the

active area of research that is summarized well by this tutorial: https://

1.7 Data set preparation

The new and unique challenges of

crash of the spacecraft on Mercury on April 30, 2015. MESSENGER was

Author: Ester Ståhlberg

Release date: October 8, 2023 [eBook #71834]

Original publication: Porvoo: WSOY, 1922

Credits: Juhani Kärkkäinen and Tapio Riikonen

*** START OF THE PROJECT GUTENBERG EBOOK SUNNUNTAI

Porvoossa, Werner Söderström Osakeyhtiö, 1922.

Ilta hämärtää. Koivut ikkunani ulkopuolella ovat paksussa

Nyt on ilta. Kuuntelen tuota keinutuolin natinaa ja tiedän varsin

Kaikki on siitä muuttunut. Nyt on maailma äänetön ja elämämme

Pelkään pahoin, ettei isän silmä koskaan enää käänny siitä

Menen hiljaa hänen huoneeseensa. Hän ei nosta päätään. Lasken

Kaareutuuhan täällä vielä taivas ja paistaahan joskus taas aurinko.

— Onhan hän kuitenkin luonamme — sanon ääneen.

Valkotukkainen mies katsahtaa minuun ja nyökkää päätään. Sitä

Hän sulkee käteni omaansa ja minä istuudun tuolille hänen

Tuossa ovat mahonkituolit molemmin puolin pöytää ja sohva