You are on page 1of 17

RESEARCH ARTICLE How Machine Learning and High‐Resolution Imagery

10.1029/2019JC015569
Can Improve Melt Pond Retrieval From MODIS Over
Key Points:
• Current methods of estimating pond
Current Spectral Unmixing Techniques
coverage on sea ice from MODIS
have high error and bias when
Nicholas C. Wright1 and Chris M. Polashenski1,2
compared to new high‐resolution 1
data sets
Thayer School of Engineering, Dartmouth College, Hanover, NH, USA, 2U.S. Army Cold Regions Research and
• Spectral unmixing used to observe Engineering Laboratories, Hanover, NH, USA
sea ice features fails to meet key
assumptions that reduce its potential
accuracy Abstract Meltwater that pools on the surface of Arctic sea ice enhances solar absorption and accelerates
• A new machine learning model that
further ice melt. The impact of melt ponds on energy absorption is controlled primarily through their
improves subpixel surface
estimations from MODIS is influence on ice albedo, which is, in turn, governed in large part by the ponds' spatial coverage. This work
presented seeks to observe the spatial coverage of melt ponds across the Arctic basin with sufficient accuracy to
investigate pond‐albedo feedback and presents an improved technique to achieve this goal. We approach the
problem by using the Open Source Sea Ice Processing algorithm to classify surface features in submeter
Correspondence to: resolution optical satellite imagery over select sites where such imagery is available. These data establish
N. C. Wright, “true” estimates of pond coverage and the ponds' spectral reflectance. This information is then used to
ncwright.th@dartmouth.edu
inform, improve, and test spectral unmixing and machine learning techniques that seek to determine melt
pond coverage from more widely available, but lower resolution, optical satellite imagery (e.g., Moderate
Citation: Resolution Imaging Spectroradiometer). The new machine learning approach improves accuracy from prior
Wright, N. C., & Polashenski, C. M.
(2020). How machine learning and
work and can contribute to improved efforts to validate melt pond models or understand trends in pond
high‐resolution imagery can improve coverage. Nevertheless, we encounter and carefully document significant challenges to retrieving melt pond
melt pond retrieval from MODIS over fractions from low‐resolution optical imagery. These limit accuracy to levels below that necessary for
current spectral unmixing techniques.
Journal of Geophysical Research:
resolving climatologically important trends. We conclude that greatly expanding the collection of high‐
Oceans, 125, e2019JC015569. https:// resolution satellite imagery over sea ice is necessary to monitor melt pond coverage with the accuracy
doi.org/10.1029/2019JC015569 needed by the scientific community.

Received 15 AUG 2019 Plain Language Summary As temperatures in the Arctic rise during summer months, meltwater
Accepted 22 JAN 2020 pools on the surface of sea ice. These ponds are darker than snow/ice and cause more sunlight to be
Accepted article online 29 JAN 2020
absorbed. This creates a feedback cycle that accelerates ice melt. A primary factor in this cycle is the
percentage of the ice surface that is covered by ponds. It is therefore important to observe the spatial
coverage of ponds across the Arctic to better understand their impact sea ice loss. Due to the remoteness of
the Arctic Ocean, it is necessary to utilize satellite imagery to investigate the properties of sea ice. In
satellite imagery there is a trade‐off between resolution and spatial coverage: Images can be either “sharp,”
seeing many details of the surface but covering a small region, or “blurry” covering a large region but where
each pixel contains many ponds. This work seeks to improve techniques (called spectral unmixing) for
determining the fraction of each low‐resolution pixel that is a pond. We investigate the theoretical limits of
spectral unmixing and conclude that it has several limitations that negatively impact its usefulness. Finally,
we trial a new machine learning approach to the same problem that shows promising results.

1. Introduction
The albedo of summer Arctic sea ice is greatly determined by the presence of melt ponds (Fetterer &
Untersteiner, 1998; T. C. Grenfell & Maykut, 1977; T. Grenfell & Perovich, 1984, 2004; D. K. Perovich
et al., 2002). These pools of meltwater darken the ice surface and play a critical role in the solar energy bal-
ance that controls further melt. They form quickly; an ice cover can transition from a premelt state with no
pond coverage to fully developed ponds with 50% coverage or more in as little as two days (Polashenski et al.,
2012). The ponds can disappear equally fast; once a critical threshold of melt is reached the ponds can
quickly drain to the ocean below, again revealing the relatively high‐albedo ice surface. As such, the most
important metric describing melt ponds' impact on albedo is their areal coverage. The result is that melt
©2020. American Geophysical Union. pond fractional coverage is one of the strongest effects on seasonal albedo change of the Arctic Ocean and
All Rights Reserved. a major player in the ice‐albedo feedback that govern annual ice retreat (Flocco et al., 2010; Perovich,

WRIGHT AND POLASHENSKI 1 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

2005; Perovich et al., 2002; Perovich & Polashenski, 2012). It is particularly notable that pond formation
causes changes in albedo early in summer, initiating ice‐albedo feedbacks that cascade through the
summer season (Schröder et al., 2014). The coverage of ponds may also be changing over the long term; it
has been widely theorized that melt ponds cover a larger fraction of first year ice than of multiyear ice
because flat, undeformed pans of first year ice allow melt water to spread horizontally, while the rough
surface of multiyear ice contains melt water in well‐defined ponds (Eicken et al., 2004; Fetterer &
Untersteiner, 1998). As the Arctic sea ice cover transitions from one dominated by thick multiyear ice
toward a seasonal ice cover (Maslanik et al., 2007; Rigor & Wallace, 2004), the coverage of melt ponds and
their role in albedo feedback is likely changing. The importance of melt ponds, and the changes they are
likely experiencing, demand we observe seasonal and long‐term trends in their prevalence.
Many efforts have been undertaken to assess the coverage of melt ponds, identify the physical mechanisms
governing how pond coverage changes, and to incorporate these mechanisms into models (Flocco et al.,
2010; Hunke et al., 2013; Lüthje et al., 2006; Zhang et al., 2018). These efforts can be divided into three broad
categories based on their spatial and temporal scales: (1) those quantifying pond coverage very precisely over
small domains and short time periods using ground‐based studies (e.g., Eicken et al., 2004; Landy et al., 2014;
Perovich et al., 2003; Polashenski et al., 2017, 2012); (2) those using aerial photography (Miao et al., 2015;
Perovich, 2002; Webster et al., 2015; Wright & Polashenski, 2018), high‐resolution optical satellite imagery
(Wright & Polashenski, 2018, and other citations therein), and C‐band SAR imagery (Scharien et al., 2014)
to quantify pond coverage explicitly over multi‐kilometer domains at snapshots in time; and (3) those
attempting to determine pond coverage over extensive spatial and temporal scales from lower resolution
remote sensing by de‐convolving the pond‐related signals from other contributions to the surface signature
(Rösel et al., 2012; Rösel & Kaleschke, 2011; Tschudi et al., 2008). To date, these efforts have not met the need
for a spatially and temporally continuous data set that would serve as a climatological data record suitable
for examining trends in pond coverage or as a basis against which to validate model predictions.
Recent advances in processing high‐resolution optical imagery of sea ice provide high‐accuracy measures of
the state of the sea ice surface, but only at snapshots in space and time (Miao et al., 2015; Miao et al., 2016;
Webster et al., 2015; Wright & Polashenski, 2018). Submeter resolution imagery excels at capturing the state
of sea ice because it can explicitly resolve surface features, providing a great deal of information for proces-
sing algorithms. As technological advances increase the availability of high‐resolution optical imagery, these
methods will play an increasingly important role in the study of Arctic sea ice. However, despite the benefits
of high‐resolution imagery, image coverage remains very limited. The entire annual collection of cloud free
high‐resolution satellite imagery over the Arctic Ocean in recent years has totaled 1–3% of the Arctic by area
and few locations are imaged repeatedly to observe seasonal changes. Lower resolution imaging sensors are
not constrained by this lack of data availability, collecting data over nearly the entire Arctic Ocean on a daily
basis, clouds permitting. Furthermore, lower resolution optical data have been collected over the past several
decades and provide the opportunity to create historical metrics of melt ponds by retrospectively analyzing
prior observations.
At lower resolutions, individual surface features like melt ponds cannot be resolved, meaning techniques for
extracting subpixel features must be used to determine melt pond fraction. Rösel et al. (2012) and Tschudi
et al. (2008) used a spectral unmixing technique to determine melt pond coverage as a subpixel phenomenon
from Moderate Resolution Imaging Spectroradiometer (MODIS) data, resulting in an Arctic basin‐wide
eight‐day pond coverage product from 2000 to 2011 (Rösel et al., 2012). Although the data set was validated
against data available at the time, the limited availability of high‐resolution imagery and field observations
restricted the scope of the validation. Recent advances in high‐resolution image collection and processing
now allow for more extensive validation and allow for tuning or training low‐resolution processing algo-
rithms using the more accurate results of high‐resolution image snapshots.
This work focuses on testing current tools for observing melt ponds in low‐resolution optical imagery and
trialing new approaches. The goal is to create a high‐accuracy metric of melt pond coverage that spans large
spatial and temporal domains. Such a data set could be used to validate recent improvements to the parame-
terization of melt ponds in sea ice models (e.g., Hunke et al., 2013) and to test theories about how the role of
ponds in governing ice melt is changing. Newly available libraries of processed high‐resolution imagery are
utilized to assess the accuracy of existing MODIS processing techniques, and modern machine learning

WRIGHT AND POLASHENSKI 2 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

techniques are leveraged to improve the analyses of low‐resolution data sets. The level of accuracy needed to
understand trends in melt pond coverage and their impacts on Arctic energy balance is also considered, as is
the potential for further improvement.

2. Background
2.1. Uses of Melt Pond Fraction Data
The observation of melt ponds is primarily motivated by the need to better understand the processes and
changes in Arctic climate. We anticipate that the core uses of basin‐wide melt pond fraction data sets will
be to validate large‐scale models of melt pond evolution (e.g., as in Hunke et al., 2013; Zhang et al., 2018)
and to understand trends in the energy balance of the Arctic Ocean (e.g., Perovich et al., 2008). These uses
dictate a level of accuracy that any data set of melt pond coverage must seek to match, which we will define
specifically in terms of root‐mean‐square error (RMSE) and overall bias.
First, we consider the demands of data being used for validation of climate models. Recent advances in glo-
bal climate models have implemented explicit parameterization of melt pond formation and evolution pro-
cesses as a component of calculating sea ice surface albedo, with the goal of increasing physical realism,
accuracy, and model resiliency under changing conditions (e.g., Hunke et al., 2013). In order to assess and
improve the parameterization of albedo in these models it is necessary to validate the modeled pond cover-
age, which is an intermediary in calculating the albedo. This requires basin‐wide, time series measurements
of melt pond coverage having accuracy at least as high as current modeling ability. We quantify this accuracy
level by considering validation conducted to date. When model estimates of melt ponds on sea ice are com-
pared to high‐resolution observations, the modeled and observed melt pond coverage agrees to within ~0.10
RMSE areal fraction (Hunke et al., 2013). This provides a baseline threshold for the required accuracy of
future observational data sets—the observational data sets must have accuracy substantially better than
0.10 RMSE areal fraction in order for a disagreement to be confidently attributed to model error rather than
observational error.
Second, we should question whether such further model improvements are really necessary by estimating
the consequence of misrepresenting melt pond fraction in terms of changes in net energy balance and
impacts on sea ice. Following the approach to solar partitioning used by Perovich (2005) we calculated the
net absorption of incident shortwave energy under two scenarios: (1) observed surface fractions and (2) sur-
face fraction with a +0.05 bias in melt pond fraction. The first case follows the results presented in Perovich
(2005) with a total incident solar energy absorption of 831 MJ m−2 from 1 April through 5 October. Shifting a
constant 0.05 surface fraction from the ice category to the melt pond category yields an additional energy
absorption of ~83 MJ m−2. This additional energy is enough to melt ~28 cm of ice which is a larger effect
on sea ice volume than is typically caused by other parameters in model sensitivity studies (Urrego‐
Blanco et al., 2016). Note that this calculation does not take into account any ice‐albedo feedback, which
would likely serve to increase the amount of energy absorbed over the course of the melt season (Curry
et al., 1995). We conclude, therefore, that model validation and energy balance calculations demand a melt
pond coverage product that has an RMS error over large areas of better than 0.10 melt pond fraction, and an
absolute accuracy (bias) of better than 0.05, and set these values as the target for our efforts herein.

2.2. Types of Available Imagery


The work herein focuses on the use of optical imagery of the sea ice. Such imagery comes in a wide range of
spectral and spatial resolutions. For the purpose of the discussions herein we define high‐ and low‐resolution
imagery with respect to the dimension of sea ice surface features. High spatial resolution refers to imagery
where the area represented by a single pixel is less than the typical horizontal scale of surface features of
interest. For example, the pixel size of 0.3–0.5 m of WorldView imagery is considerably smaller than the
average several meter horizontal dimensions of a melt pond. In contrast, a low‐resolution pixel observes
an area much larger than the features of interest, such as the 250–500‐m pixel size of MODIS or the
750 m of Visible Infrared Imaging Radiometer Suite. High‐resolution imagery of Arctic sea ice enables direct
observations of melt ponds because their size and shape are explicitly resolved, and most pixels represent
only one surface type. The fraction of a scene covered by melt ponds can be determined by classifying which
surface each pixel represents and counting those that represent pond‐covered areas. While there is no hard

WRIGHT AND POLASHENSKI 3 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

cutoff between “low” and “high” resolution, as pixel size grows beyond ~2 m, pixels begin to commonly
represent mixed surfaces. By 10 m, a substantial majority of pixels represent more than one surface type
and fractional coverage can only be determined by calculating subpixel categories. The imagery used in this
paper does not require us to confront determining an exact cutoff. We use only imagery that is of sufficiently
high resolution (submeter pixel size) or sufficiently low resolution (250 m+ pixel size) to clearly fall into one
category or the other.

2.3. Spectral Unmixing: Theoretical Basis


Melt pond fraction can be estimated from imagery where the resolution is significantly lower than the aver-
age size of a melt pond by treating each pixel signal as a mixture of several fixed surface types with varying
spatial coverage. These spectral unmixing techniques are well‐established in terrestrial remote sensing (e.g.,
Asner & Heidebrecht, 2002; Keshava & Mustard, 2002), and have been demonstrated for calculating sea ice
melt pond fractions from MODIS imagery (Rösel et al., 2012; Tschudi et al., 2008). In the sea ice case,
the observed surface reflectance is interpreted to be the mixed contributions of the reflectance from the three
predominant surface types: snow/ice, melt ponds, and open water, which will be referred to as the
“component surfaces.” As described by Fetterer and Untersteiner (1998), the bulk reflectance of a region
is then represented using the following relationship:

α ¼ αsi *f si þ αmp *f mp þ αow *f ow (1)

where α is the regional reflectance; αsi, αmp, and αow are the reflectances of each open water, melt ponds, and
snow/ice, respectively; and f is the fractional coverage of those same surface types, such that

f si þ f mp þ f ow ¼ 1 (2)

Note particularly that equation (1) assumes that the reflectance of each of the three component surfaces is
uniform and unchanging, a simplifying assumption that will be revisited later. Surface fractions are then
determined by equating the observed reflectance at specific wavelengths to the combined contributions of
each surface, governed by its fractional coverage and reflectance in that wavelength. These equations form
the basis of a system of linear equations (equations (2)–(5)) presented in Rösel et al. (2012) that can be solved
by linear decomposition:

r ðλ1 Þ ¼ αsiðλ1 Þ *f si þ αmpðλ1 Þ *f mp þ αowðλ1 Þ *f ow (3)

r ðλ2 Þ ¼ αsiðλ2 Þ *f si þ αmpðλ2 Þ *f mp þ αowðλ2 Þ *f ow (4)

r ðλ3 Þ ¼ αsiðλ3 Þ *f si þ αmpðλ3 Þ *f mp þ αowðλ3 Þ *f ow (5)

1 ¼ f si þ f mp þ f ow Equation 2 ðrepeatedÞ

where r (λb) is the observed surface reflectance in spectral band b and αðλb Þ is the spectral reflectance of a sur-
face type in spectral band b. Converting this system of linear equations to matrix form yields

r⃑ ¼ A f⃑ (6)

where
2 3 2 3
r ½λ1  αsi½λ1  αmp½λ1  αow½λ1  2 3
6 7 6 α f si
⃑ 6 r ½λ2  7 6 si½λ2  αmp½λ2  αow½λ2  7
7 ⃑ 6 7
r¼6
6
7
7 A¼6 7 f ¼ 4 f mp 5
4 r ½λ3  5 4 αsi½λ3  αmp½λ3  αow½λ3  5
f ow
1 1 1 1

This system can be solved for an observed aggregate surface reflectance (r⃑) using a known set of surface
reflectance values (A) to determine the areal fraction of each component surface (f⃑). For the remainder of
this study A will be referred to as the “reflectance matrix.” Following prior work, we will reference

WRIGHT AND POLASHENSKI 4 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

ground‐based albedo observations (bihemispherical reflectance) to determine the values for this matrix. This
system has four equations with three unknowns, yielding an overdetermined problem with the possibility of
no exact solution. However, f⃑ can be estimated by a value bf which is the “best fit” to the observed data by
solving for the areal fraction of each surface type with a linear least squares approximation.

2.4. Limitations of Spectral Unmixing


The spectral unmixing technique (equation (6)) relies on a core principle that the reflectance of each com-
ponent surface within a pixel is both unique and constant. However, it is well established from numerous
field studies that the albedo of common sea ice surfaces in the Arctic evolves over the melt season
(Grenfell & Perovich, 1984, 2004; Perovich, 2005; Perovich & Polashenski, 2012). As the ice transitions from
winter dry‐snow‐covered ice to melting summer ice pack and subsequently evolves through the summer, the
albedo properties of both the ponded and the unponded ice change dramatically. Unponded ice evolves as
dry snow has a much higher albedo than wet snow, which in turn has a higher albedo than bare ice. Even
bare ice has an albedo weakly dependent on thickness. The albedo of pond covered ice is governed by the
thickness of ice beneath the ponds and the inclusion of bubbles in the ice itself. Melt ponds that form on first
year sea ice are generally darker than their multiyear ice counterparts, as first year ice tends to be thinner
and may have lower bubble content (Grenfell & Maykut, 1977). Unlike snow, ice, and melt ponds, the albedo
of open water is relatively stable.
A spectral unmixing method that uses only a single reflectance to encompass every variety of melt pond
cannot detect the true fraction of melt ponds at all points in the Arctic basin. Critically, using a single
reflectance is likely to lead to a systemic flaw where ponds that do not match the initially prescribed
reflectance are not accurately quantified in a biased manner. Of particular concern, as the sea ice shifts
toward an ice cover dominated by first year ice with ponds spectrally distinct from those typical on multi-
year ice, spectral unmixing algorithms calibrated on multiyear ice will likely fail to accurately observe
these changes. Changes in pond spectral albedo may instead be interpreted as changes in pond area.
Similarly, as ponds evolve and darken during the melt season, changes in pond albedo could be mistaken
for changes in area. The key question that arises is how severe these biases are and how sensitive the
algorithm is to small variations in pond color. With large increases in the availability of processed
high‐resolution image snapshots to use as validation across seasonal and ice type gradients we can explore
these concerns.

3. Methods
3.1. Data Sources
Two sources of imagery were used in this study: (1) the MODIS daily surface reflectance product from
the Aqua satellite (MYD09GA collection 6) and (2) high‐resolution optical satellite imagery from
DigitalGlobe's WorldView (WV) constellation in both panchromatic and eight‐band multispectral formats.
We used MODIS bands 1 through 7 at a spatial resolution of 500 m to test the algorithms described herein.
We use the MOD09GA cloud flag to ignore pixels that were flagged as containing cloud.
High‐resolution panchromatic and multispectral imagery were used to create validation sites for comparison
to coincident MODIS imagery (section 3.2.2). Multispectral imagery from the WV 2 and 3 satellites was used
to test the theoretical limits of spectral unmixing in controlled experiments (section 3.2.3). The high‐
resolution satellite images were selected to encompass a range of possible sea ice surface conditions, includ-
ing a wide variety of possible melt pond colors (spectral reflectance signatures) as well as areas of both high
and low ice concentrations, and areas in early and late stages of melt. The panchromatic bands of WV 1 and 2
both have a spatial resolution of 0.46 m at nadir and WV 3 has a spatial resolution of 0.31 m at nadir. The WV
1/2/3 satellite panchromatic bands sample the visible spectrum between 400/450/450 and 900/850/800 nm,
respectively (DigitalGlobe, 2014). In addition, WV 2 and 3 have eight multispectral bands at 1.84‐ and 1.24‐m
nadir resolution, respectively, within the range of 400 to 1,040 nm. Each satellite image was orthorectified to
mean sea level and the multispectral images were pansharpened to the resolution of the panchromatic ima-
gery following the process described in section 3.1 of Wright and Polashenski (2018).

WRIGHT AND POLASHENSKI 5 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

3.2. Spectral Unmixing


3.2.1. Spectral Unmixing Simplification
We use the approach presented in Rösel et al. (2012), and described in section 2.3 to determine surface frac-
tions from MODIS imagery. We modify the implementation slightly by solving the system of equations with
a simplified technique which produces equivalent results. In place of adding a nonlinear cost function to
equation (6) and solving the system as an optimization problem (as in Rösel et al., 2012) we use a modified
linear least squares approximation. The cost function ensures the physical realism of the solution: The area
fraction of each component surface, bf, represents a physical parameter that must be positive for each element
bf . This constraint can instead be directly incorporated into a least squares formulation. The standard expres-
i
sion for an ordinary least squares approximation with this constraint is
 
⃑ b2
minr−A f  s:t: bf ≥0 (7)
f

also known as a nonnegative least squares approximation. While bf is also subject to the constraint bf i ≤1, this is
handled implicitly by the requirement that bf i ≥0 and by equation (4) of the system of linear equations ∑bf i ¼
1. This novel method solves the same problem and yields equivalent results. We chose the new method for
ease of reproduction and improved computational time.
3.2.2. Expanded Unmixing Validation
Recent increases in image acquisition and developments in mass image processing allow for validation of
unmixing results against a larger range of possible sea ice conditions. Spectral unmixing estimates derived
from MODIS observations were compared to collocated observations from high‐resolution imagery at 175
sites, covering a total area of over 90,000 km2 (Wright & Polashenski, 2019). These scenes cover a wide range
of possible ice surface conditions including both premelt and postmelt onset images.
Acquisition times did not match exactly for the coincident WV‐MODIS pairs. Due to sea ice drift, two images
of the same geographic coordinates at different acquisition times may not see the same ice surface. The dif-
ference in acquisition time between the MODIS and WV imagery can be up to 24 hr depending on weather
and satellite orbital paths but is typically less than 6 hr. We account for image acquisition mismatch and sea
ice drift by finding the set of contiguous MODIS pixels—with area equal to the WV scene—that best matches
the WV surface fractions. The search area has a square radius of 12 km from the WV image center point lati-
tude and longitude. This 24‐km window allows for a mean ice drift rate of up to 12 km/day in an unknown
direction for up to a maximum difference in acquisition time of 24 hr. This is slightly above typical drift
speeds of 6–10 km/day (Rampal et al., 2009). While this does not ensure a perfect overlap as might be
achieved with explicit drift information, it does find the best possible match between WV and MODIS within
realistic constraints. It is therefore a slightly optimistic estimate of error—the technique could marginally
improve the match but cannot worsen it.
3.2.3. Unmixing Validation in an Ideal Data Set Case
A perfect match between MODIS pixels and a WorldView image in both space and time will still contain dif-
ferences in observed reflectance due to satellite viewing geometries, atmospheric correction models, and
image sensor differences. In order to eliminate these variables and assess the theoretical limits of spectral
unmixing, we created an experiment to test this method in an ideal manner. A 0.5‐m resolution panshar-
pened WorldView 2 image was transformed into a “pseudo‐MODIS” image by downsampling it to a resolu-
tion of 500 m by pixel averaging (see Figure 1 for an example pseudo‐MODIS image). This yielded a high‐
and low‐resolution image pair of exactly the same sea ice, effectively collected through exactly the same
viewing geometry and atmospheric conditions, removing several sources of error in MODIS‐WorldView
comparisons. The high‐resolution image was processed using the Open Source Sea‐ice Processing (OSSP;
Wright & Polashenski, 2018) algorithm to label each pixel as ice, melt pond, or ocean. The low‐resolution
image was processed in two different ways to assess spectral unmixing, henceforth referred to as “stage 1”
and “stage 2” unmixing.
In stage 1, a custom‐tailored reflectance matrix (A in equation (6)) was created for each 500‐m pseudo‐
MODIS pixel. The reflectance values for each surface in matrix A were determined by calculating the mean
of all high‐resolution WV pixels assigned to that surface by the OSSP algorithm. The three WV bands, 7, 5,
and 2 (770–895, 630–690, and 450–510 nm, respectively), were used to best match the MODIS band

WRIGHT AND POLASHENSKI 6 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

wavelengths chosen in Rösel et al. (2012) (MODIS bands 2, 1, and 3 at 841–


876, 620–670, and 459–479 nm). This reflectance matrix was then used to
solve equation (7) using the mean total reflectance of the 500 × 500‐m
region as the input reflectance, r⃑. Spectral unmixing was run on every
pseudo‐MODIS pixel in the same way, using a reflectance matrix that indi-
cates the exact mean reflectance values for each surface type in that 500 ×
500‐m pseudo‐MODIS pixel. The results from unmixing each pseudo‐
MODIS pixel were then compared to the “true” surface fractions deter-
mined by OSSP from the high‐resolution image. In stage 2, all pseudo‐
MODIS‐pixel‐specific reflectance matrices from stage 1 were averaged
(mean) to create a single matrix for the whole ~15 × 15‐km downsampled
image. The unmixing algorithm was run again on every pseudo‐MODIS
pixel in the image, this time using only the single reflectance matrix. The
results were again compared to the OSSP‐determined surface fractions.
3.2.4. Error Propagation by Resampling
We conducted an error analysis of the spectral unmixing system to better
understand its sensitivity to uncertainty in the input parameter values.
The uncertainties in predicted surface fractions (bf ) were estimated using
a Monte Carlo method for error propagation; the spectral unmixing sys-
tem was solved for many different values of a given input and the variance
in the output was compared to the variance in the input. Two tests were
Figure 1. (top) Example WorldView 2 high‐resolution image at 0.5‐m reso- run: one with varying surface reflectance values ( r⃑ ) and constant
lution and (bottom) the resulting pseudo‐MODIS image at 500‐m resolution.
surface‐type reflectance matrix (A) and the other with the opposite.
This image was acquired at 72°N. 128°W on 13 June 2014. Imagery © 2016
DigitalGlobe, Inc. Each value in r⃑ was sampled from a normal distribution with a mean at
the observed reflectance and with a standard deviation set equal to the
published uncertainty in the atmospheric correction of the MODIS surface reflectance product; 0:005 þ 0:05
*r⃑(Vermote & Ray, 2015). Values in A were also sampled from a normal distribution with a mean at the true
albedo being used in our unmixing based on field observations and standard deviation of 0.05. The standard
deviation of bf calculated from each sample was then compared to the standard deviation of the input para-
meters to understand how uncertainty propagates through this system.

3.3. Developing a New Technique


In an effort to improve processing of low‐resolution imagery, we developed a new technique for determin-
ing subpixel surface fractions using a random forest machine learning model (Breiman, 2001). We imple-
mented a multi‐output random forest regressor (implemented in the scikit‐learn Python library
(Pedregosa et al., 2011)) with the goal of predicting a pixel's surface distribution (i.e., f⃑ from equation (6))
based on that pixel's observed reflectance (i.e., r⃑). The random forest regressor was built with 100 trees
(the remaining parameters were assigned the default value as implemented in scikit‐learn version 0.20).
The technique is not constrained by the need for analytical solving, and so more bands and more surface
types may be considered. In this model, r⃑ is a 6 × 1 vector with reflectance values from MODIS spectral
bands 1 through 6. Band 7 was not included as the reflectance of melting sea ice surfaces is too low in this
wavelength to be a useful predictor in the random forest (note its indistinguishability from band 6 in
Table 1). The model was trained on a data set of 64,000 simulated “observations.” Each data point in
the training set was a feature vector of 6 simulated reflectance values (r⃑s ) paired to a simulated ice‐
pond‐ocean surface fraction array (f⃑s). Generating a simulated data set allowed us to create training data
that included members with all possible permutations of surface‐type fractions. Limited quantities of
available true observations do not encompass the full range of surface fraction possibilities and would
leave the algorithm untrained on some parts of the solution space.

Simulated reflectance values (r⃑s Þ were calculated from equation (1) using field‐observed albedos (Table 1)
⃑ Þ. The sum of fractional albedos can be rewritten in matrix form
and randomly generated surface fractions ðf s
(equation (6) with altered notation) as

WRIGHT AND POLASHENSKI 7 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

Table 1
Field Observed Albedos for Common Sea Ice Surfaces, as Recorded by Grenfell and Perovich (2004), Plus the
Addition of Bright Melt Ponds and Open Water
Band wavelength Cold Melting Determined Dirty Bright melt Early melt Late melt Open
[average] (nm) snow snow melting ice ice ponds pond pond water

620–670 [645] 0.98 0.76 0.57 0.41 0.30 0.26 0.16 0.06
841–876 [858.5] 0.95 0.64 0.37 0.40 0.08 0.10 0.07 0.04
459–479 [469] 0.97 0.77 0.63 0.36 0.56 0.42 0.19 0.05
545–565 [555] 0.98 0.77 0.64 0.40 0.50 0.41 0.22 0.07
1230–1250 [1240] 0.54 0.17 0.04 0.24 0.06 0.06 0.05 0.03
1628–1652 [1640] 0.17 0.02 0.01 0.16 0.04 0.04 0.04 0.03
2105–2155 [2130] 0.11 0.01 0.01 0.15 0.04 0.04 0.03 0.03

r⃑s ¼ A f⃑s (8)

where
2 3 2 3 2 3
r ðλ1 Þ αsi;λ1 αmp;λ1 αow;λ1 f si
6 7 6 7 6 7
r⃑s ¼ 4 ⋮ 5 A¼4 ⋮ ⋮ ⋮ 5 f⃑s ¼ 4 f mp 5
r ðλ6 Þ αsi;λ6 αmp;λ6 αow;λ6 f ow

This equation is solved for r⃑s 64,000 times, each with a separately generated f⃑s, where ∑f⃑s ¼ 1. The result is
64,000 pairs of a feature vector (f⃑s) and the target prediction (r⃑s) that can be used to train the random forest. A
three‐surface model was created with training data generated from albedos of three surfaces only; cold snow
for αsi;λn , early melt ponds for αmp;λn , and ocean/leads for αow;λn (values listed in Table 1) to match the albedos
in the spectral unmixing matrix presented in Rösel et al. (2012). An eight‐surface model was created with
training data where the albedo values for snow/ice (αsi;λn ) and melt ponds (αmp;λn ) were created from their
own sum of fractional albedos that represent subcategories of those surface types:
2 3
2 3 f
αcs;λ1 αms;λ1 αdmi;λ1 αdi;λ1 6 cs 7
6 76 f ms 7
⃑ 6
α si ¼ 4 ⋮ ⋮ ⋮ ⋮ 7 6 7 (9)
56 7
4 f dmi 5
αcs;λ6 αms;λ6 αdmi;λ6 αdi;λ6
f di

2 32 3
αbmp;λ1 αemp;λ1 αlmp;λ1 f bmp
6 76 f 7
α⃑mp ¼ 4 ⋮ ⋮ ⋮ 54 emp 5 (10)
αbmp;λ6 αemp;λ6 αlmp;λ6 f lmp

where again each element of f⃑ is randomly generated and the sum of f⃑ is 1. The subcategories of surface
types are chosen to mirror those observed in Grenfell and Perovich (2004), plus the addition of a “bright
melt pond” type from field observations made during the SHEBA campaign (Perovich et al., 1999).
Snow/ice was divided into cold snow (cs), melting snow (ms), deteriorated melting ice (dmi), and dirty
ice (di). The melt pond category was divided into bright, early, and late melt ponds (bmp, emp, and
lmp, respectively). The ocean albedo is a constant based on field observations of leads (Perovich et al.,
1999). All field observed values (α) used here are shown in Table 1. The out‐of‐bag error was 0.99 for
the three‐surface model and 0.90 for the eight‐surface model (OOB error is equivalent to cross‐validation
with a 50/50 test/train split in random forests (Breiman, 1996)). A better OOB score for the three‐surface
model is expected because there is more separability in the chosen three surface reflectance signatures
than when using all eight. However, the use of eight surface albedos can better capture the variation that
exists in the real‐world sea ice system, and we will later see that the eight‐surface model is better able to
predict surface distributions from real MODIS observations. The machine learning method was tested on
true MODIS data and compared to high‐resolution observations (mimicking methods in section 3.2.2) and
on pseudo‐MODIS data (mimicking methods in section 3.2.3).

WRIGHT AND POLASHENSKI 8 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

4. Results
4.1. Validation of Existing Spectral Unmixing Techniques
The Rösel et al. (2012) unmixing algorithm performed well during pre-
vious validation efforts when compared against aerial imagery, show-
ing an RMSE of ~4–11% (Rösel et al., 2012). However, these
validations did not span the full range of seasonal and spatial variabil-
ity in melt pond morphology. Figure 2 compares true melt pond frac-
tion calculated from newly processed high‐resolution satellite imagery
with pond fraction derived from spectrally unmixing coincident
MODIS pixels at 200 sites spread across the 2011–2015 melt seasons.
The spectral unmixing results from MODIS show considerable scatter
in individual point values relative to the Worldview estimate and show
a positive bias in melt pond fraction of 0.14. While the correlation
(0.79) between MODIS estimates and WorldView observations is
encouraging, the MODIS algorithm fails to achieve our standard of
accuracy by a significant margin: The RMS error between MODIS
Figure 2. Comparison of melt pond fraction estimated from spectral unmix-
and WV surface fraction estimates is 0.18 for this data set. The accu-
ing applied to real MODIS data to that observed in coincident
high‐resolution WorldView scenes. The MODIS estimates have a racy of the OSSP high‐resolution image processing algorithm was
positive bias of 0.14 and an RMS error of 0.18. shown to be much higher than our stated requirements—96%—and
comparable to the agreement between two human classifiers (Wright
& Polashenski, 2018). Note that this error is pixel‐by‐pixel accuracy
in high‐resolution imagery and thus a different measure than RMSE. The RMSE of high‐resolution
observations is difficult to quantify, but likely much lower than 0.04 due to errors in object classification
cancelling when aggregated into surface fractions (e.g., a pond segment misclassified as ice will cancel
an ice segment misclassified as pond).
The comparison between high‐resolution data sets and MODIS spectral unmixing agrees with a similar
study done in Zhang et al. (2018) which also found considerable scatter in the relationship between
unmixing‐determined melt ponds and true pond coverage. (Note that the MODIS‐determined melt pond
fraction data set (version 2 of data set from Rösel et al. (2012)) used in Zhang et al. (2018) already has a
0.08 bias subtracted from melt pond fraction estimates.) It is important to note that each data point in
Figure 2 is the result of averaging many contiguous MODIS pixels. The difference between each MODIS esti-
mate and the WorldView observation represents an incorrect unmixing for that region, and not just an error
for a single MODIS pixel. These data show that using a single albedo
matrix with spectral unmixing techniques over the entire Arctic basin
Table 2 does not generate observations of sufficient accuracy to address process‐
Stage 1 and Stage 2 RMSE for a Set of 18 WorldView Images based questions about how melt ponds drive albedo feedback or improve
Date Stage 1 RMSE Stage 2 RMSE model validations beyond the current state of the art.
12/07/2011 0.044 0.165
13/07/2013 0.008 0.035 4.2. Controlled Spectral Unmixing
13/06/2014 0.006 0.058 4.2.1. Stage 1 Unmixing
28/06/2014 0.013 0.099 The following tests were designed to understand if regionally tuning the
28/06/2014 0.014 0.098 reflectance matrix could improve accuracy. The first case seeks to
01/07/2014 0.022 0.115
understand how well spectral unmixing performs when the reflectance
06/07/2014 0.017 0.071
08/07/2014 0.024 0.083 matrix is perfectly determined. The results from spectrally unmixing
11/07/2014 0.029 0.128 18 downsampled WorldView images (i.e., pseudo‐MODIS images) in
15/07/2014 0.034 0.149 this ideal case are shown in Table 2. Recall from section 3.2 that every
15/07/2014 0.010 0.093 pseudo‐MODIS pixel in these 18 images was unmixed with its own
26/07/2014 0.079 0.107
27/07/2014 0.011 0.061
unique reflectance matrix which exactly matches the properties of the
30/07/2014 0.027 0.083 surface types within that image pixel. Further, recall that the pseudo‐
30/07/2014 0.012 0.189 MODIS image is created from the WorldView image so that the images
01/06/2015 0.001 0.059 are precisely coincident (no drift) and there is no mismatch in viewing
14/06/2015 0.001 0.044 angle, illumination geometry, or atmospheric correction. Results from
07/07/2015 0.003 0.064
two of these images are plotted in Figure 3, which shows the

WRIGHT AND POLASHENSKI 9 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

Figure 3. Examples of (top) stage 1 and (bottom) stage 2 unmixing from two WorldView scenes, showing the fraction
of each component surface. In the 08 July scene (left) the stage 2 accuracy remains within acceptable limits. In the
15 July scene, however, the stage 2 unmixing is not successful. The error bars are the result of Monte Carlo
simulations adding 0.05 random error to the input. These show that even when the surface can correctly unmixed, the
system is very sensitive to errors in the input. Note that each unmixed pixel yields three points—one for each surface.

estimates of each component surface compared to their true value. There is a strong correlation
(correlation coefficient: 0.99) between the surface fractions determined from spectral unmixing and the
true surface distributions from high‐resolution imagery. Figure 4a shows the melt pond area as a
fraction of total ice area for stage 1 unmixing of a 30 July 2014 WorldView image. The melt pond
fraction is accurately resolved in this stage 1 case. In the ideal case, with precisely matched data and
reflectance matrix, it appears that there is sufficient information available in the system. Spectral
unmixing can differentiate these surface types and retrieve highly accurate surface distributions of sea
ice from low‐resolution optical imagery.
4.2.2. Stage 2 Unmixing
We next tested how accuracy changes when using a “regionally representative” reflectance matrix to
derive surface fractions. Rather than a precisely determined reflectance matrix for each component
surface in each pseudo‐MODIS pixel, we used one that represents the average properties for the entire
Worldview image. Since it is not possible to custom‐tailor the reflectance matrix for every true MODIS
pixel, this tested the idea that high‐resolution imagery could be used to provide snapshots of relevant
surface reflectance matrix values for unmixing in the surrounding region, tailoring the reflectance
matrix regionally. The results of this tuned unmixing, compared to the OSSP‐derived surface distribu-
tions, are shown in Table 2, and the estimates of each component surface for two sample images are
plotted in Figure 3. Surface fractions are estimated to acceptable accuracy in stage 2 for the 08 July
image, but the estimation fails for the 15 July image. Melt ponds as a fraction of total ice area for a

WRIGHT AND POLASHENSKI 10 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

30 July scene are shown in Figure 4b, where results have degraded sig-
nificantly from the stage 1 case. These tests indicate that in many stage
2 cases the algorithm fails to accurately unmix the component surfaces.
However, in some the method produces reasonable results, but ones
that are lower accuracy than stage 1.
4.3. Propagation of Errors in Inputs
The cause for algorithm failure or success in the stage 2 unmixing test can
be better understood by assessing how small changes in the spectral
unmixing system inputs alter the resulting estimates of surface fractions.
As described in section 3.2.4, the Monte Carlo method for error propaga-
tion was used to assess the sensitivity of the spectral unmixing system to
error in either observed surface reflectance or reflectance matrix values.
The sensitivity of the unmixing result to a 0.05 standard deviation in input
values (representing reasonably likely observational errors) is presented
in the form of error bars added to Figures 2, 3, and 5. Table 3 contains a
detailed report on the Monte Carlo errors from unmixing each pseudo‐
MODIS image. In most images, the magnitude of error in the unmixed
surface fractions is several times higher than the error in the input para-
meters. This is indicative of an unstable system where the results are
not robust to low‐precision inputs: a small error in the observations leads
to a large uncertainty in the result. Importantly, the majority of error
shows up in the melt pond category, while the snow/ice category is more
stable (error in output roughly equals error in input). The images that
show accurate unmixing results, however, also show a smaller sensitivity
to error in input values. These images share a common feature; the melt
ponds are much bluer and lighter in color than areas of open water in
the same image—in other words, they are distinct in spectra and bright-
ness from ocean, providing a spectral unmixing problem less sensitive to
small errors in input.
4.4. Next‐Generation MODIS Processing
Figure 4c presents the results of the machine learning method alongside
the results of spectral unmixing (Figures 4a and 4b). In this figure, a
single pseudo‐MODIS image of 220 pixels is analyzed using three tech-
niques: frame (a) stage 1 unmixing, (b) stage 2 unmixing, and (c) a ran-
dom forest machine learning model. The accuracy of the random forest
method is better than that of stage 2 unmixing—and still has little bias
despite having no a priori knowledge of the properties of that image.
The machine learning method is also significantly more feasible to
implement on a large scale because it does not require the processing
of coincident high‐resolution data sets to obtain regionally representa-
tive component reflectances.
Testing the method at large scale shows further encouraging results.
Figure 5 is equivalent to Figure 2, except that it uses the new random for-
est method instead of spectral unmixing. Real MODIS data are processed
and compared against near‐coincident WorldView data. Here we see
Figure 4. Estimated melt pond fraction compared to true melt pond fraction
using (a) stage 1 unmixing, (b) stage 2 unmixing, and (c) a random forest marked improvements over the spectral unmixing method. The mean
model. RMS error is 0.07, with a positive bias of 0.02 and a correlation coefficient
of 0.87 as compared to an RMSE of 0.18 and bias of +0.14 in the untuned
spectral unmixing case. Recall that the simulated training data set
includes surface reflectance data from eight surface states. Repeating the above test with a simulated train-
ing set generated from only three ice states (melting snow, bright melt ponds, and ocean) yielded an RMS
error to 0.16 (not plotted). These tests indicate that the random forest technique may even meet our

WRIGHT AND POLASHENSKI 11 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

standard of accuracy. Given that the first‐order drift correction


employed here could artificially improve this comparison more
extensive tests would be required to conclusively determine if the
machine learning method passes our criteria in all cases. The
machine learning approach also has the potential for further refine-
ments in areas that are not accessible to spectral unmixing, for exam-
ple, including the time of year and location of pixels or synthesizing
data sets from other sensors to create a complex multimodal setup
(e.g., from radar remote sensing).

5. Discussion
This study began by validating published spectral unmixing techni-
ques using resolved observations of melt pond fractions. The compar-
ison (section 4.1 and Figure 2) revealed several issues with the
accuracy of spectral unmixing. These limit the ability to produce a
robust melt pond fraction data set suitable for validation of melt pond
Figure 5. Comparison of melt pond fraction estimated from the new random models or for describing long‐term trends in pond coverage. In
forest model applied to real MODIS data to that observed in coincident high‐
section 2.4 we discussed well‐known issues with this technique based
resolution WorldView scenes. The MODIS estimates have a positive bias of 0.02
and an RMS error of 0.07. on the physical properties of the sea ice system. Specifically, there are
two key assumptions underlying the equations that define spectral
unmixing: (1) the mean spectral signatures of each surface type must
not change over space or time (2) inter‐surface differences in reflectance must be large, particularly if input
data being analyzed are noisy. In essence, the categories of sea ice surfaces to be separated must be homo-
geneous, unchanging, and distinct from one another. It is well known that these assumptions are violated
to varying degrees, with a negative impact on accuracy of unmixing. We posited that improvements to the
accuracy of this method could be accrued by reducing violations of these assumptions. We were less success-
ful in achieving improvements in accuracy than we expected, and so we devote a significant portion of our
discussion to exploring why, to investigating what limits accuracy.

5.1. Failure of Reflectance Matrix Tuning


Perhaps the most important and easiest to address violation of assumptions is the changing of melt pond
properties in space and time. Since sea ice thickness and melt state, which govern pond reflectance, vary over
regional scales, we hypothesized that replacing the single, unchanging, reflectance matrix used to describe
the properties of each surface type in prior work with one that is tuned to regional sea ice state would
improve the accuracy of spectral unmixing—and make it possible to meet accuracy requirements. High‐
resolution imagery collected in select locations around the Arctic and processed with the OSSP code
(Wright & Polashenski, 2018) could provide a source for these regionally tuned reflectance matrices.
Unfortunately, while results of tests did indicate improvement, regionally tuning the unmixing reflectance
matrix did not improve results to a level meeting the accuracy requirements. Exploring why leads us to
examine the imagery in the context of assumptions implicit in spectral
unmixing. We come to the conclusion that it is not possible to consistently
Table 3
Errors in Output Surface Fractions After Adding Error to the Input
derive component surface fractions of sea ice from low‐resolution imagery
Parameters Using the Monte Carlo Method using spectral unmixing techniques at an accuracy suitable for validating
melt pond models or establishing unambiguous long‐term trends.
Surface type Stage 1 Stage 2
Our most idealized tests showed that highly accurate spectral unmixing
r A r A
is possible under highly controlled circumstances (section 4.2.1). Using
Snow/ice 0.070 0.057 0.068 0.055 a unique custom reflectance matrix that exactly matches the mean
Melt pond 0.164 0.141 0.157 0.134
component surface reflectance for every pixel in a low‐resolution
Ocean 0.110 0.093 0.105 0.088
image yields results with perfect accuracy (as is mathematically
Note. The amount of error for each test was selected from a Gaussian dis- required). Unfortunately, this level of tuning is unreasonable in real
tribution centered at 0 with a standard deviation of 0.05. Tests were per-
formed for both stage 1 and stage 2 unmixing, and in each case errors conditions. Developing perfectly tuned reflectance matrices for each
were either added to the input surface reflectance (r) or the reflectance MODIS pixel necessitates already having a categorized high‐resolution
matrix (A), but not both. image that resolves the melt ponds explicitly, from which to derive

WRIGHT AND POLASHENSKI 12 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

the component surface properties—in other words, already having


superior overlapping data.
The study presented in section 4.2.2 tested the next level of abstrac-
tion from the ideal case. A “locally tuned” reflectance matrix was
used to unmix many low‐resolution pixels from a several hundred
square kilometer area. The results of these tests showed a degradation
of unmixing accuracy to below our required threshold (discussed in
section 2.1) in some, but not all, circumstances. Note that this test
is still somewhat idealized. It uses a reflectance matrix that is better
(that is, more accurately tuned and applied to represent a smaller
region) than would be possible in large‐scale processing of
MODIS data.

5.2. Limits to Tuning: A Closer Look


Understanding why our effort to improve unmixing by tuning the
Figure 6. RMSE of spectral unmixing for 17 pseudo‐MODIS scenes compared to
the difference in color between ocean and ponds within each scene. As the color reflectance matrices was not adequate requires further consideration
difference between ponds and ocean increases (typically because the ponds are of the major assumptions implicit to spectral unmixing. First, we
bluer and brighter), the unmixing method is more accurate. examine the degree to which melt pond and ocean surfaces are dis-
tinct from one another (assumption #2). Then we evaluate the degree
to which component surface reflectance varies in ways that are not captured by our tuning, and the impor-
tance of this variation (addressing assumption 1).
To evaluate the importance of separation between categories, we employed a quantitative metric to deter-
mine the difference between the spectral signature of melt ponds and that of open water. This metric is
the Euclidian distance in RGB space between the reflectance of the pixels classified as melt pond and those
classified as open ocean in a high‐resolution image. Note that in WorldView imagery bands 5, 3, and 2 are
the red, green, and blue bands, respectively. Typically, the bluer and the brighter melt ponds appear, the
greater their difference from open water. Figure 6 plots the accuracy of stage 2 spectral unmixing when
applied to pseudo‐MODIS pixels (method from section 3.2.3) as a function of the ocean‐pond color difference
in the area which that pixel represents in the source WorldView image. Unsurprisingly, as the color of ponds
and open water become more similar, determining their respective contributions to the reflectance observed
in a low‐resolution pixel becomes increasingly inaccurate. No similar issue exists in stage 1 unmixing, when
the reflectance matrix is perfectly tuned. The correlation between accuracy and ocean‐pond color difference
in the stage 2 case indicates that violation of assumption #2 is a major contributor to remaining error. When
melt ponds and open water are insufficiently distinct, the solution becomes highly sensitive to error from
using a slightly imperfect reflectance matrix. Dark‐colored ponds are,
therefore, likely to always be challenging to differentiate from ocean in
spectral unmixing in real‐world cases.
We next considered the homogeneity of each surface type's reflective
properties. We find that variation in component surface properties occurs
regionally, as expected, but also over spatial scales too small to be
addressed by plausible local or regional tuning. These heterogeneous
characteristics are particularly large in the reflectance of melt ponds.
Bright and dark melt ponds often coexist in single high‐resolution images
and even within the area covered by single MODIS pixels. Figure 7 shows
a pixel brightness histogram derived from a 2.5 × 2.5‐km region of a
WorldView scene acquired 30 July 2014, labeled by the surface category
of the pixels. Pond pixels have a relatively broad distribution of brightness
compared to the sharp peak of ocean pixels. The variation in pond color
over scales not addressed by regional tuning remains a violation of the
assumptions inherent to spectral unmixing in stage 2.
Figure 7. Histogram of pixel reflectance values for band 3 of Worldview,
colored by the surface to which they belong. Black corresponds to ocean In order to explore the impact that this intrasurface variation in reflec-
pixels, blue to melt pond pixels, and gray to snow and ice pixels. tance has on errors in unmixing surfaces we (1) quantify the magnitude

WRIGHT AND POLASHENSKI 13 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

of the variation in reflectance at different scales and (2) determine the sen-
sitivity of the unmixing to a mismatch between anticipated and actual
reflectance. We evaluate the “local” variation in melt pond color by calcu-
lating the mean and standard deviation of the Euclidian distance between
three‐band pond reflectance in each pseudo‐MODIS pixel and the mean
pond color of the complete WorldView scene. Combining these statistics
across 15 images yields a mean of 0.03 and standard deviation of 0.02.
This is the variation in pond color that feasible tuning will not address.
We also calculated the mean and standard deviation of mismatch between
a single standard basin‐wide reflectance and the pixel‐by‐pixel mean
reflectance. For this calculation we use the standard reflectance matrix
previously published (matrix A; Rösel et al., 2012), and a new global
reflectance matrix that represents the mean of all regional (stage 2)
matrices (matrix B; the ideal basin‐wide reflectance matrix). The mean
and standard deviation of pond color difference between pixel‐by‐pixel
Figure 8. Error in pond fraction estimation that results from a mismatch values and the global matrices is 0.12 ± 0.03 and 0.07 ± 0.03 for matrices
between the ideal (stage 1) spectral reflectance matrix and the one used A and B, respectively. The difference between the global pond color var-
to calculate pond fraction. Each line is the result of altering a single
component of the spectral reflectance matrix in either the positive or iance (0.07) and the local pond color variance (0.03) is the maximum ben-
negative direction. In other words, this plot is the result of unmixing one efit of regionally tuning: 0.04. We conclude that regionally tuning reduces
pseudo‐MODIS pixel with many varying reflectance matrices. the mismatch between the anticipated and actual surface reflectance
matrices slightly more than half.
Placing the degree of mismatch into context requires understanding the sensitivity of the unmixing method
to a reflectance matrix mismatch. Figure 8 shows the error in the pond fraction estimate as a function of the
mismatch between pond color in the reflectance matrix used and the true pond color in the pixel. Data for
Figure 8 is the result of a Monte Carlo simulation unmixing a constant observed reflectance with a changing
reflectance matrix. The error in pond fraction rises rapidly as pond color mismatch increases. A mismatch of
as little as 0.01 results in an absolute error in melt pond fraction of 0.1, which violates our stated accuracy
goal. Achieving our desired accuracy with spectral unmixing would, therefore, require a reflectance matrix
that matches the actual surface reflectance to within 0.01. It is not possible to tune the reflectance to this
level. Nonaddressable variance will reduce pixel‐by‐pixel accuracy to below the desired threshold, exactly
as the data indicate.
5.3. Sensitivity to Input Errors
So far, we have primarily investigated the response of spectral unmixing to
the controllable parameters (e.g., selecting the spectral reflectance matrix
values to best match the true surface). In satellite observations of sea ice
there are a number of sources of error that introduce uncontrollable var-
iance in the observed reflectance. These errors in input image data might
come from atmospheric clarity, instrument response, look angle, illumi-
nation angle, or any number of other factors. For example,
the accuracy of the MOD09 atmospheric correction is lower at extreme
off‐nadir look angles and high solar zenith angles, which are typical
in polar regions (E. F. Vermote & Vermeulen, 1999). Error in reflectance
on the order of 0.01–0.1 can be reasonably expected from these
sources. Tests that introduced this expected random error degraded the
unmixing accuracy by twofold to threefold when using a tuned reflectance
matrix (section 4.3). This is a serious issue for the potential success of a
spectral unmixing method—the method is highly sensitive to errors over
which we have no control. Further exploring this propagation of error
Figure 9. Sensitivity of unmixing pseudo‐MODIS pixels as a function of the showed a notable feature. Plotting the Monte Carlo error as a function
fraction of each pixel that is open water. Results from a single
pseudo‐MODIS scene (30 July 2014). Sensitivity is largest in pixels where
of true ocean fraction shows that sensitivity is lowest when ocean is
there are three component surfaces (where ocean fraction is less than 1, but near zero or 100%, shown in Figure 9. In essence, when there are only
larger than 0). two (zero ocean) or one (complete ocean) component surfaces,

WRIGHT AND POLASHENSKI 14 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

unmixing sensitivity is low. When all three surface types (pond, ocean, ice) are present in significant quan-
tities—exactly the time when melt pond determination is required—unmixing becomes less reliable.
5.4. Machine Learning as a Superior Approach
Upon discovering the limits of regional tuning and recognizing the sensitivity of the system to input errors,
we concluded that achieving the goals set out in section 2.1 by refining the tuning technique was not feasible.
Instead, we turned to a machine learning model. We theorized that machine learning would be able to incor-
porate extra implicit information, such as the likelihood of common surface distributions rather than solely
the mathematical feasibility. The approach of creating a training data set based on field observations and
simulated surface fractions easily accommodated more surface categories (eight versus three) in order to
reduce the amount of intra‐surface difference. The impact of increasing the number of component categories
in the machine learning training data set was to reduce RMS error by a factor of ~2, from 0.16 to 0.07. A key
flaw in the spectral unmixing tuning approach was the increase in sensitivity to input data error. The
machine learning approach shows less sensitivity here; Monte Carlo error in the random forest predictions
(shown as error bars in Figure 5) is on the same order of magnitude as those in the untuned spectral unmix-
ing case (Figure 2). That is to say, the error in output is roughly equal to the error in input.
Overall results of the machine learning model presented here are markedly better than the prior state of the
art technique for deriving subpixel surface fractions. They show less bias and less scatter when compared to
real‐world observations (Figure 5). Absolute accuracy remains right at the edge of being adequate for long‐
term processing, but we acknowledge that the simple drift correction implemented here to match
WorldView and MODIS observations will have a small tendency to cause the MODIS results to appear better
than reality. Future iterations of the machine learning method could incorporate additional information
other than optical reflectance. These could be simple factors like time of year and location, or more complex
multimodal sensing. Incorporating additional observations from, for example, radar remote sensing may
provide additional information to help determine surface fractions. We recommend reprocessing historic
MODIS data with machine learning methods to extract an improved historical record, but also emphasize
that accuracy limitations remain. Due to the similarity of certain pond modalities and open water, especially
given atmospheric distortions common in the Arctic, subpixel estimations of surface fraction will always
contain ambiguity. In contrast, high‐resolution imagery can observe surface properties such as shape or
roughness, which allow melt pond fraction to be determined with considerably less ambiguity.

6. Conclusion
This study explored the strengths and weaknesses of retrieving sea ice surface fractions with the spectral
unmixing techniques presented by Tschudi et al. (2008) and refined by Rösel et al. (2012). By validating these
techniques against high‐resolution image processing we discovered significant uncertainty in determining
surface fraction coverage from low‐resolution optical imagery. Unmixing accuracy is highest when the
reflectance of component surface types match their prescribed properties and when pond reflectance signa-
tures are significantly different from those of open water. Unfortunately, with the large temporal and spatial
variation in reflectance of sea ice surface types, these conditions are often not met. Therefore, spectral
unmixing does not produce high‐accuracy results in all sea ice conditions.
We tested spectral unmixing in controlled cases and used these to explore possible avenues to improve its
performance. Unsurprisingly, accuracy can be improved by tuning the reflectance matrix to better match
the properties of the sea ice surfaces being unmixed. This indicates a potential to improve results by system-
atically tuning the reflectance matrix to better replicate spatial and temporal changes in melt pond spectral
reflectance. Unfortunately, we find that these improvements will not go far enough to satisfy the accuracy
requirements needed for the expected usage of this data set. Our theoretical tests show that the spectral
unmixing solution is highly sensitive to mismatch between anticipated and actual reflectance of the compo-
nent surface types and is also highly sensitive to error in the input surface reflectance data, as could be
expected from atmospheric distortion and look angle dependencies. A single albedo matrix is often not ade-
quate even for regions that are only tens to hundreds of square kilometers. An exploration of both sensitivity
to error and scale dependence in reflectance matrix variation showed little hope of reducing mismatch
between anticipated and actual reflectance of component surface types to a level that would meet
accuracy demands.

WRIGHT AND POLASHENSKI 15 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

Once discovering the limitations to spectral unmixing, we trialed a new machine learning technique for
deriving subpixel surface fractions from low‐resolution imagery. Our tests show the accuracy of the
machine learning method to be better than regionally tuning spectral unmixing and it is significantly
more feasible to implement. Using machine learning and an artificial training data set we have improved
the RMS error in melt pond determination from the 0.18 we found in existing published spectral unmix-
ing techniques to 0.07. These tests span large geographic range and the full seasonal cycle, indicating that
they are likely robust error magnitudes. The apparent success of this method supports its consideration by
the community as an improved technique for determining pond coverage from low‐resolution imagery
and we suggest reprocessing the MODIS archives to produce an improved melt pond product using
these techniques.
Note that while these improvements are promising, the remaining error in the improved machine learn-
ing method is still large, and considering the errors in input spectra, likely near our cutoff. In order to
further our understanding of the complexity of melt pond evolutionary pathways, more precise observa-
tions of melt ponds are needed. This suggests a need going forward for superior sensing methods as
could be achieved with collecting larger quantities of high‐resolution imagery. Semi‐high‐resolution sen-
sors, such as Sentinel‐2 or Landsat‐8 (10 and 15 m, respectively), are a possible solution to the low accu-
racy of low‐resolution unmixing. At 10 m, pixels would still have to be separated into subpixel
categories, but they are significantly more likely to contain only one to two surfaces rather than more
than three surfaces that are common within a 500‐m pixel (recall that Figure 9 shows that accuracy
increases when only one to two surfaces comprise a pixel). These platforms could provide better aerial
coverage over submeter resolution sensors while maintaining high accuracy—future work should focus
on establishing reliable ways of deriving surface fractions from those data. In addition, we also suggest
that the community begin systematically collecting submeter resolution imagery data that, at a mini-
mum, sample the regional variability in melt ponds over scales on the order of 100 km. This will pro-
vide the data required by future studies to continue to refine basin‐scale observations of melt
pond behaviors.

Acknowledgments References
This work was funded by the Office of
Naval Research (grant Asner, G. P., & Heidebrecht, K. B. (2002). Spectral unmixing of vegetation, soil and dry carbon cover in arid regions: Comparing multi-
N0001413MP20144) and the National spectral and hyperspectral observations. International Journal of Remote Sensing, 23(19), 3939–3958. https://doi.org/10.1080/
Science Foundation, Office of Polar 01431160110115960
Programs (grant 1417436). WorldView Breiman, L. (1996). Out‐of‐Bag Estimation.
images processed with the OSSP Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/10.1023/A:1010933404324
technique are available at the Arctic Curry, J. A., Schramm, J. L., & Ebert, E. E. (1995). Sea ice‐albedo climate feedback mechanism. Journal of Climate, 8, 240–247. https://doi.
Data Center (Wright & Polashenski, org/10.1175/1520‐0442(1995)008<0240:SIACFM>2.0.CO;2
2019), and the OSSP code is archived on DigitalGlobe. (2014). WorldView‐3. Retrieved January 4, 2019, from https://www.spaceimagingme.com/downloads/sensors/datasheets/
Zenodo (DOI: 10.5281/ DG_WorldView3_DS_2014.pdf
zenodo.3066324). WorldView imagery Eicken, H., Grenfell, T. C., Perovich, D. K., Richter‐Menge, J. A., & Frey, K. (2004). Hydraulic controls of summer Arctic pack ice albedo.
was provided with the NextView license Journal of Geophysical Research, 109, C08007. https://doi.org/10.1029/2003JC001989
through the Polar Geospatial Center Fetterer, F., & Untersteiner, N. (1998). Observations of melt ponds on Arctic sea ice. Journal of Geophysical Research, 103(C11),
(PGC) at the University of Minnesota. 24,821–24,835. https://doi.org/10.1029/98JC02034
WorldView imagery are subject to Flocco, D., Feltham, D. L., & Turner, A. K. (2010). Incorporation of a physically based melt pond scheme into the sea ice component of a
copyright and can be obtained directly climate model. Journal of Geophysical Research, 115, C08012. https://doi.org/10.1029/2009JC005568
from DigitalGlobe or through the PGC. Grenfell, T., & Perovich, D. (1984). Spectral albedos of sea ice and incident solar irradiance in the southern Beaufort Sea. Journal of
The MODIS data used in this work are Geophysical Research, 89(C3), 3573. https://doi.org/10.1029/JC089iC03p03573
available from NASA (https://doi.org/ Grenfell, T., & Perovich, D. (2004). Seasonal and spatial evolution of albedo in a snow‐ice‐land‐ocean environment. Journal of Geophysical
10.5067/MODIS/MYD09A1.006). Code Research, 109, C01001. https://doi.org/10.1029/2003JC001866
for creating and analyzing the pseuo‐ Grenfell, T. C., & Maykut, G. A. (1977). The optical properties of ice and snow in the Arctic basin. Journal of Glaciology, 18(80), 445–463.
MODIS imagery, creating the simulated https://doi.org/10.3189/S0022143000021122
training data sets and machine learning Hunke, E. C., Hebert, D. A., & Lecomte, O. (2013). Level‐ice melt ponds in the Los Alamos sea ice model, CICE. Ocean Modelling, 71, 26–42.
model, and analyzing real MODIS https://doi.org/10.1016/j.ocemod.2012.11.008
imagery is archived at Zenodo (DOI: Keshava, N., & Mustard, J. F. (2002). Spectral unmixing. IEEE Signal Processing Magazine, 19(1), 44–57. https://doi.org/10.1109/79.974727
10.5281/zenodo.3608344). The training Landy, J., Ehn, J., Shields, M., & Barber, D. (2014). Surface and melt pond evolution on landfast first‐year sea ice in the Canadian Arctic
data sets used herein are archived at the Archipelago. Journal of Geophysical Research: Oceans, 119, 3054–3075. https://doi.org/10.1002/2013JC009617
same location as the code. The authors Lüthje, M., Feltham, D. L., Taylor, P. D., & Worster, M. G. (2006). Modeling the summertime evolution of sea‐ice melt ponds. Journal of
declare that they have no conflicts of Geophysical Research, 111, C02001. https://doi.org/10.1029/2004JC002818
interest. Maslanik, J. A., Fowler, C., Stroeve, J., Drobot, S., Zwally, J., Yi, D., & Emery, W. (2007). A younger, thinner Arctic ice cover: Increased
potential for rapid, extensive sea‐ice loss. Geophysical Research Letters, 34, L24501. https://doi.org/10.1029/2007GL032043
Miao, X., Xie, H., Ackley, S., Perovich, D., & Ke, C. (2015). Object‐based detection of Arctic sea ice and melt ponds using high spatial
resolution aerial photographs. Cold Regions Science and Technology, 119, 211–222. https://doi.org/10.1016/j.coldregions.2015.06.014

WRIGHT AND POLASHENSKI 16 of 17


Journal of Geophysical Research: Oceans 10.1029/2019JC015569

Miao, X., Xie, H., Ackley, S. F., & Zheng, S. (2016). Object‐based Arctic sea ice ridge detection from high‐spatial‐resolution imagery. IEEE
Geoscience and Remote Sensing Letters, 13(6), 787–791. https://doi.org/10.1109/LGRS.2016.2544861
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., et al. (2011). Scikit‐learn: Machine learning in Python.
Journal of Machine Learning Research, 12(Oct), 2825–2830. Retrieved from. http://jmlr.csail.mit.edu/papers/v12/pedregosa11a.html
Perovich, D. (2005). On the aggregate‐scale partitioning of solar radiation in Arctic sea ice during the Surface Heat Budget of the Arctic
Ocean (SHEBA) field experiment. Journal of Geophysical Research, 110, C03002. https://doi.org/10.1029/2004JC002512
Perovich, D., Grenfell, T., Richter‐Menge, J., Light, B., Tucker, W., & Eicken, H. (2003). Thin and thinner: Sea ice mass balance mea-
surements during SHEBA. Journal of Geophysical Research: Oceans, 108(C3), 8050. https://doi.org/10.1029/2001JC001079
Perovich, D. K., Grenfell, T. C., Light, B., Richter‐Menge, J. A., Sturm, M., Tucker, W. B. III, et al., SHEBA: Snow and Ice Studies [CD‐ROM],
U.S. Army Cold Reg. Res. and Eng. Lab., 1999.
Perovich, D., & Polashenski, C. (2012). Albedo evolution of seasonal Arctic sea ice. Geophysical Research Letters, 39, L08501. https://doi.org/
10.1029/2012GL051432
Perovich, D. K. (2002). Aerial observations of the evolution of ice surface conditions during summer. Journal of Geophysical Research,
107(C10), 8048. https://doi.org/10.1029/2000JC000449
Perovich, D. K., Grenfell, T. C., Light, B., & Hobbs, P. V. (2002). Seasonal evolution of the albedo of multiyear Arctic sea ice. Journal of
Geophysical Research, 107(C10), 8044. https://doi.org/10.1029/2000JC000438
Perovich, D. K., Richter‐Menge, J. A., Jones, K. F., & Light, B. (2008). Sunlight, water, and ice: Extreme Arctic sea ice melt during the
summer of 2007. Geophysical Research Letters, 35, L11501. https://doi.org/10.1029/2008GL034007
Polashenski, C., Golden, K., Perovich, D., Skyllingstad, E., Arnsten, A., Stwertka, C., & Wright, N. (2017). Percolation blockage: A process
that enables melt pond formation on first year Arctic sea ice. Journal of Geophysical Research: Oceans, 122, 413–440. https://doi.org/
10.1002/2016JC011994
Polashenski, C., Perovich, D., & Courville, Z. (2012). The mechanisms of sea ice melt pond formation and evolution. Journal of Geophysical
Research, 117, C01001. https://doi.org/10.1029/2011JC007231
Rampal, P., Weiss, J., & Marsan, D. (2009). Positive trend in the mean speed and deformation rate of Arctic sea ice, 1979–2007. Journal of
Geophysical Research, 114, C05013. https://doi.org/10.1029/2008JC005066
Rigor, I. G., & Wallace, J. M. (2004). Variations in the age of Arctic sea‐ice and summer sea‐ice extent. Geophysical Research Letters, 31,
L09401. https://doi.org/10.1029/2004GL019492
Rösel, A., & Kaleschke, L. (2011). Comparison of different retrieval techniques for melt ponds on Arctic sea ice from Landsat and MODIS
satellite data. Annals of Glaciology, 52(57), 185–191. https://doi.org/10.3189/172756411795931606
Rösel, A., Kaleschke, L., & Birnbaum, G. (2012). Melt ponds on Arctic sea ice determined from MODIS satellite data using an artificial
neural network. The Cryosphere, 6(2), 431–446. https://doi.org/10.5194/tc‐6‐431‐2012
Scharien, R. K., Hochheim, K., Landy, J., & Barber, D. G. (2014). First‐year sea ice melt pond fraction estimation from dual‐polarisation C‐
band SAR: Part 2—Scaling in situ to Radarsat‐2. The Cryosphere, 8(6), 2163–2176. https://doi.org/10.5194/tc‐8‐2163‐2014
Schröder, D., Feltham, D. L., Flocco, D., & Tsamados, M. (2014). September Arctic sea‐ice minimum predicted by spring melt‐pond frac-
tion. Nature Climate Change, 4(5), 353–357. https://doi.org/10.1038/nclimate2203
Tschudi, M. A., Maslanik, J. A., & Perovich, D. K. (2008). Derivation of melt pond coverage on Arctic sea ice using MODIS observations.
Remote Sensing of Environment, 112(5), 2605–2614. https://doi.org/10.1016/j.rse.2007.12.009
Urrego‐Blanco, J. R., Urban, N. M., Hunke, E. C., Turner, A. K., & Jeffery, N. (2016). Uncertainty quantification and global sensitivity
analysis of the Los Alamos sea ice model. Journal of Geophysical Research: Oceans, 121, 2709–2732. https://doi.org/10.1002/
2015JC011558
Vermote, E. F., & Vermeulen, A. (1999). Atmospheric correction algorithm: Spectral reflectances (MOD09). ATBD Version.
Vermote, J. C., & Ray, J. P. (2015). MODIS Surface Reflectance User's Guide collection 6. MODIS Land Surface Reflectance Science
Computing Facility.
Webster, M. A., Rigor, I. G., Perovich, D. K., Richter‐menge, J. A., Polashenski, C. M., & Light, B. (2015). Seasonal evolution of melt ponds
on Arctic sea ice. Journal of Geophysical Research: Oceans, 120, 1–15. https://doi.org/10.1002/2015JC011030
Wright, N., & Polashenski, C. (2018). Open‐source algorithm for detecting sea ice surface features in high‐resolution optical imagery.
Cryosphere, 12(4), 1307–1329. https://doi.org/10.5194/tc‐12‐1307‐2018
Wright, N., & Polashenski, C. (2019). Surface classifications of Arctic sea ice from WorldView Satellite Imagery. https://doi.org/10.18739/
A22Z12P4J.
Zhang, J., Schweiger, A., Webster, M., Light, B., Steele, M., Ashjian, C., et al. (2018). Melt pond conditions on declining Arctic sea ice over
1979–2016: Model development, validation, and results. Journal of Geophysical Research: Oceans, 123, 7983–8003. https://doi.org/
10.1029/2018JC014298

WRIGHT AND POLASHENSKI 17 of 17

You might also like