Professional Documents
Culture Documents
CERTIFICATE
————————-
————————-
(Signature of the Head of College)
DECLARATION
I declare that this written submission represents our ideas in my own words and where
others’ ideas or words have been included, I have adequately cited and referenced the
original sources. I also declare that I have adhered to all principles of academic honesty
and integrity and have not misrepresented, fabricated, or falsified any idea, data, fact, or
source in my submission. I understand that any violation of the above will be cause for
disciplinary action by the Institute and can also evoke penal action from the sources which
have thus not been properly cited or from whom proper permission has not been taken
when needed.
—————
(Signature)
Karan upare
(Name of the student)
C-55
(Institute Roll Number)
ACKNOWLEDGEMENT
I have taken efforts in this project. However, it would not have been possible without the
kind support and help of many individuals and organizations and I would like to extend
I am highly indebted to Prof. Ghodichor sir, for his guidance and constant super-
vision as well as for providing necessary information regarding the project also for his
support in helping me work on the project. I would like to express my gratitude towards
my parents for their kind co-operation and encouragement. I would also like to thank all
people who have willingly helped me out with their abilities.
4
Contents
1 Introduction1
2 Literature Survey3
3 Data Engineering4
List of Figures
3.1 Band1 of Radar signa ................................................................................. 5
Abstract
model.
1
Chapter 1
Introduction
methods are not feasible, and the only viable monitoring option is via satellite.Considering this
need to bring in the advancements of Machine Learning and Deep Learning truly in
real-life problems.
1.2 Objective
The main objective is to build an algorithm that automatically identifies if a remotely sensed
target is a ship or iceberg using machine learning and deep learning.
currently available datasets. We extensively talk about what work have been done in this
project. Chapter 3 covers the data collection for our thesis project and mentions the
2
resources used for creating the dataset until now. Also it talks about the preprocessing
Chapter 1. Introduction
conducted on the dataset. Chapter 4 covers the Feature Extraction methods viz. Baseline
CNN, Pre-Train model etc and talks about the process of obtaining it. It also talks about
the algorithms to train and test the dataset for Iceberg Ship classification. Chapter 5 talks
about the results obtained.
3
Chapter 2
Literature Survey
Several research papers have been published with different features and models for classi-
fying Satellite Radar images into ship and Icerberg classes. In [2] Cheng Zhan presented
a convolutional neural network (CNN) designed to work with a limited training data and
features, while demonstrating its effectiveness in this problem. Results showed that trans-
fer learning resulted in a significant boost in accuracy. The augmentation used for feature
engineering is accomplished by a variety of image transformations,smoothing, first and
second derivatives, gradient and Laplacian.In [3] Ankita Rane and Vadivel Sangili pre-
sented a Semisupervised Approach in which they labelled the test data using Pseudola-
beling learning and helped in data augmentation and increase in accuracy.
Chapter 3
Data Engineering
companies such as C-CORE to provide Dataset at Kaggle. C-CORE has been using satel-
lite data for over 30 years and has been developing surveillance network based on com-
puter vision.On November 2017, St. John’s-based applied RD organization C-CORE and
international energy company Equinor provided the dataset on kaggle[1].
Following were the main steps undertaken to ensure that the data provided for feature
extraction is clean:
• Normalization: Horizontal and Vertical band are a flat vector so we have to reshape
them into 75x75 matrix to work an image.
• Filters: Bilateral and Wavelet denoising filters are using to remove noise from both
reducing filter. It averages pixels based on their spatial closeness and radiometric
Chapter 3. Data Engineering
similarity.A wavelet denoising filter relies on the wavelet representation of the im- age. The
noise is represented by small values in the wavelet domain which are set to 0.
• Smothing: Resultant image is De-noised again with smoothing with Gaussian filter
• Third Band: Normalized HH and HV band and add then to create an extra feature as
Chapter 4
Networks
To Extract Features ,for example edge detection, from an image conventionally Haar
filter or Canny filter are convoluted with the image with fixed weight.Similarly in an Con-
volution Neural Network we use convolution of filter to extract features from an image
but the filters weight are not pre-determined.Those weight are determined through back-
propogation through the last layer.
In Convolution Neural Network(CNN) the first few layers are used to extract basic
features like edge,line,shape etc .Then these edge enhanced images are passed through
another Convolution layer to extract more important features and so on.
A CNN model was trained having 2 CNN layers with kernel size of 3x3 with "relu"
activation function followed by Batch Normalization and MaxPooling and Dropout of 0.2
even small-scale data collection can be incredibly costly or even almost impossible in
9
certain real-world use cases (e.g. medical imaging). To improve the accuracy of model
and collect more data images from training set are mirror flipped left-right,up-down and
rotated.
Chapter 4. Feature Extraction and the Training Networks 10
starting point for computer vision and natural language processing tasks due to the vast
computational and time resources required to build neural network models on these issues
and the huge skill leaps they provide on related issues.
The Residual Network, or ResNet for short, is a model that makes use of the residual module
involving shortcut connections.
It was developed by researchers at Microsoft and described in the 2015 paper titled
and then new architecture was used.Convolution layer with 512 features was removed and
the output of fourth 128 layer was Concatenate with incidence angle and passed through
neural network with dense layer of 256 and 64 neurons.It’s output was passed through
sigmoid activation layer to give a probability of belonging to each class.
VGG16 is a convolution neural net (CNN ) architecture which was used to win
till date. Most unique thing about VGG16 is that instead of having a large number of
hyper-parameter they focused on having convolution layers of 3x3 filter with a stride 1
and always used same padding and maxpool layer of 2x2 filter of stride 2. It follows this
arrangement of convolution and max pool layers consistently throughout the whole archi-
tecture. In the end it has 2 FC(fully connected layers) followed by a softmax for output.
The 16 in VGG16 refers to it has 16 layers that have weights. This network is a pretty
large network and it has about 138 million (approx) parameters.
Features were extracted from VGG16 and mobileNet model and an extra global max-
pooling layer was applied.Then those features were concatenate with incidence
angle.Resultant was passed through neural network followed by sigmoid activation layer.Due
to less GPU model was trained on small dataset. Following architecture was used :
predictions. It gives us insight not only into the errors being made by a classifier but more
importantly the types of errors that are being made.
16
Chapter 5
Following are the results that were obtained in case of Resnet convolution neural network
model.
Chapter 5. Experiments and Results 19
Following are the results that were obtained when extracting features from convolution
neural network classified with Random Forest as classifier.
Following are the results that were obtained when extracting features from convolution
neural network classified with Kneighbors algorithm.
Chapter 5. Experiments and Results 23
There are 64 filter in each convolution neural network and these are 6 filters of size
3 X 3 .These filter are applied to an iceberg or ship images to detect edges and shape
in an radar image .These filters perform convolution with image matrix. After
Chapter 5. Experiments and Results 24
Convolution the resultant image have more sharp detected shape.This edge detected
Bibliography
[1] Statoil/c-core radar images of iceberg and ship dataset. https://www.kaggle.com/
c/statoil-iceberg-classifier-challenge.
[2] Zhan Cheng, Zhang Licheng, Zhong Zhenzhen, Lin Sher Didi-Ooi, and Youzuo,
and Zhang andShujiao Huang; Changchun Wang Yunxi. Deep learning approach in
au- tomatic iceberg – ship detection with sar remote sensing data. University of
Bristol, Dec 2018.