Remote Sensing of Environment 300 (2024) 113895

Contents lists available at ScienceDirect

Remote Sensing of Environment

journal homepage: www.elsevier.com/locate/rse

Mapping surface water dynamics (1985–2021) in the Hudson Bay

Lowlands, Canada using sub-pixel Landsat analysis
Ian Olthof a, *, Robert H. Fraser b
a
National Wildlife Research Centre, Environment and Climate Change Canada, 1125 Colonel By Drive, Ottawa K1A 0H3, Canada
b
Canada Centre for Mapping and Earth Observation, Natural Resources Canada, 560 Rochester St, Ottawa K1S 5K2, Canada

A R T I C L E I N F O A B S T R A C T

Edited by Menghua Wang The fate of the carbon stored in frozen peatlands in Canada’s Hudson Bay Lowlands (HBL) depends strongly on
water, with wetter conditions favoring long-term storage. Dynamic surface water products generated from his­
Keywords: torical Landsat data exist to inform surface water trends in the HBL based on binary classifications of land vs
Unmixing water at 30 m spatial resolution. However, the HBL contains many water features smaller than 30 m, including
Surface water
streams and patterned fens that require a sub-pixel mapping approach to fully capture their dynamics. In this
Peatland
paper, we leverage an existing binary dynamic surface water product to generate a spatially and temporally
Climate change
Landsat comprehensive Landsat sub-pixel surface water time-series over the HBL. We first generate Landsat Red, NIR and
Time-series SWIR composites from 1985 to 2021 and evaluate data-driven machine learning and physical models to derive
sub-30 m water fractions across the entire mapping domain. Data-driven models were calibrated on 30 m water
fractions obtained from 16 of 17 WorldView scene water fraction maps and assessed on the 17th held-out scene
in a jackknife validation. Data-driven models were also calibrated by relating scaled Landsat reflectance to water
fractions obtained from binary water masks at corresponding scales from 90 to 270 m. Physical models involved
simple linear unmixing using a new approach that samples land endmembers in the vicinity of each pixel to
represent land reflectance, and a global water endmember sampled beneath permanent water in each image. The
assumption of linear scaling was verified at sub-30 m to 270 m resolution for NIR, SWIR and NDVI, which were
all highly correlated with water fraction and had no significant differences among water fraction – spectral
feature relationships across the full range of scales. Each of 11 different methods and input spectral features were
evaluated against WorldView water fractions and ranked based on nine criteria. We found that machine learning
using all spectral features including local NIR land endmembers performed best among data-driven approaches,
and linear unmixing of the NIR band using local land endmembers performed best among physical. Both best-
performing methods were applied to the entire time-series and surface water area trends were calculated and
compared with trends obtained from binary water masks. All three methods agree on the direction of surface
water change towards wetting, with sub-pixel inter-annual variation agreeing with stream discharge noted in
other studies.

1. Introduction carbon storage and emissions from peatlands during climate warming.
Drying peatlands are projected to accelerate warming due to carbon
The Hudson Bay Lowlands (HBL) is the wettest ecozone in Canada dioxide emissions that likely outweigh the reduction in methane emis­
with 80% of its area covered by wetlands (NWWG, 1988). It forms the sions under wetter conditions (Morris, 2021).
third largest wetland in the world (Fraser and Keddy, 2005) and is Continuous permafrost forms a band within the HBL ecozone along
composed almost entirely of peatlands that store more carbon in the first the northern coast of Hudson Bay while the southernmost portion con­
2 m of soil than the total carbon stored in any other ecozone in Canada tains no permafrost (Fig. 1). This transition from permafrost-free peat­
(Sothe et al., 2022). The future of HBL’s peatlands as a carbon store is lands and soils in the south, to discontinuous and then continuous
uncertain under projected development and climate warming. Water, permafrost northward makes water in this ecozone’s wetlands, lakes and
both in permafrost and at the surface, plays a key role in changes to rivers particularly susceptible to climate change. While global

* Corresponding author.
E-mail addresses: [email protected] (I. Olthof), [email protected] (R.H. Fraser).

https://doi.org/10.1016/j.rse.2023.113895
Received 10 July 2023; Received in revised form 27 October 2023; Accepted 1 November 2023
Available online 9 November 2023
0034-4257/Crown Copyright © 2023 Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

temperature increases are approaching 1.5 degrees C since the start of national and global products using binary classes of water versus land
the industrial revolution (IPCC, 2023), warming in the HBL was (Olthof and Rainville, 2022; Pekel et al., 2016, Pickens et al., 2020). This
moderated by sea ice until the early 1990s. However, temperatures have thematic level is sufficient to depict surface water spatially and
been rising by 0.5–1 degrees C per decade since (Hadley et al., 2019), temporally in most locations where the majority is stored in large
and are projected to continue to increase by up to eight degrees by the waterbodies for applications including freshwater availability (Rodell
end of the century (Gagnon and Gough, 2005). Understanding surface et al., 2018), flooding (Olthof and Tolszczuk-Leclerc, 2018) and wetland
water dynamics in frozen peatlands such as the HBL is key to identifying mapping (Bourgeau-Chavez et al., 2015). The minimum mapping unit
the source of past carbon emissions and projecting future ones under (MMU) of 30 m Landsat is generally accepted as being in the range of 1
different climate scenarios (Fewster et al., 2022). Water in peatlands ha, or approximately 3 × 3 pixels. At this scale, pure pixels in the centre
influences rates of carbon sequestration and whether emissions are CO2 of the 3 × 3 window are composed of a single cover type that can reliably
or methane under dryer or wetter conditions (McLaughlin and Webster, be detected and mapped, reducing confusion between adjacent cover
2014; Hopple et al., 2020). types. In the HBL, a large proportion of water is contained in marshes,
The series of Landsat satellites have been used in numerous studies to patterned fens and streams, many of which are narrower than the MMU
map surface water dynamics at national to global scales (Olthof and of Landsat and thus mixed with land at that scale. In regions where many
Rainville, 2022; Pekel et al., 2016; Pickens et al., 2020). Landsat rep­ water features are smaller than the MMU of Landsat, a binary approach
resents the most detailed, continuous satellite data record available, to water mapping will miss a significant amount of water area (DeVries
with consistent global acquisitions at 30 m resolution since 1982 with et al., 2017), rendering maps and changes based on those maps inac­
the launch of Landsat 4, to present with the launch of Landsat 9 in 2021. curate for many applications. Higher resolution data exist from a range
Enormous effort has been put into the mission since its launch, focussing of newer satellite sensors (Cooley et al., 2019), but these lack the tem­
on both geometric and radiometric data quality as well as making the poral coverage needed to document multi-decadal changes to surface
entire record available to the public. Because of these efforts, Landsat is water. Alternative methods exist to extract sub-pixel surface water in­
the state of the art for long-term global change studies (Wulder et al., formation to generate maps that portray the percentage of water in each
2019). For surface water specifically, Landsat NIR and SWIR bands Landsat pixel. Mapping the fractional area of water in each pixel should
provide well-calibrated surface reflectance in parts of the electromag­ provide a more detailed depiction of the location of water and a more
netic spectrum that are relatively transparent to haze while absorbing accurate estimate of surface water area.
most incident radiation (Frazier and Page, 2000; Mueller et al., 2016), Sub-pixel mapping can be broadly divided into two approaches; one
providing good contrast between water and land under a range of im­ physical and the other data-driven. Physical approaches involve
aging conditions. unmixing the spectral contribution of multiple cover types to the overall
Previously, Landsat has been used to map dynamic surface water in pixel reflectance. It uses spectral endmembers that represent the

Fig. 1. Location of the Hudson Bay Lowlands, Wapusk National Park and permafrost zones in Northern Ontario and Manitoba, Canada.

2
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

reflectance of each cover type contained within the pixel footprint to mapping domain to ensure that model extension is reliable everywhere.
determine their fractions (Settle and Drake, 1993). Linear unmixing is As such, more data and effort are needed to generate reference data for
often assumed since it facilitates inversion of the forward mixture model model calibration than for physical approaches.
to estimate fractions. Linear mixing is reasonable for mixtures involving An efficient alternative for generating data to calibrate data-driven
single scattering from a relatively flat surface, as is the case of water models that does not require fine-resolution reference data involves
mixed with soil or short vegetation. Where scattering is more complex aggregating native resolution (30 m) water from a binary water mask
due to height differences among multiple cover types causing mutual along with reflectance to calibrate a water fraction versus reflectance
shadowing that depends on illumination and viewing geometry, mixing data-driven model (Rover et al., 2010; DeVries et al., 2017). This
may be nonlinear (Du, 2018). However, the assumption of linearity can approach has the advantage of depending only on Landsat data to
be checked against reference 30 m fractions obtained by aggregating a generate a model that can be applied to map water fractions. It assumes
finer-resolution classification. that reflectance – water fraction relations are scale-independent such
In the case of physical unmixing of land versus water, two spectral that a model calibrated at scales from windows of multiple Landsat
endmembers are needed to determine a pixel’s water fraction with the pixels can be applied to unmix at the scale of a single Landsat pixel. A
constraint that land and water fractions sum to 100%. This approach has disadvantage of this approach is that it only provides a limited number
been shown to perform well using linear unmixing and histogram of water fraction bins depending on the scale at which calibration is
breakpoint analysis to determine land and water endmembers in the performed. For example, calibration using a 3 × 3 window will provide
Arctic where thermokarst lake dynamics were examined using Landsat nine bins that can be mapped with that model. A regression model can
time-series over the Tuktoyaktuk Peninsula in NWT, Canada (Olthof interpolate between bins or alternatively, the inclusion of larger win­
et al., 2015). However, in that environment, the land endmember dows can also provide a greater number of bins. Several scales may be
reflectance was relatively stable across the study domain due to the combined to generate more precise water fraction estimates as long as
presence of limited vegetation functional types above treeline, consist­ the assumption of scale-independence of reflectance – water fraction
ing mainly of vascular graminoid and shrub. In the HBL, vegetation relations holds across the range of scales.
reflectance is more diverse, ranging from bright lichen mats and sedge The objectives of this paper are to 1. Evaluate and compare data-
fens to dark, dense riparian shrub and sparse conifer trees that cast driven and physical sub-pixel Landsat water mapping methods over
shadows. the HBL, 2. Apply the best method to a multi-decadal Landsat time-series
Endmember variability has been accounted for by considering mul­ to map inundation trends, and 3. Examine broad-scale patterns of wet­
tiple candidate endmembers and their fractions in forward mixture ting and drying and discuss potential greenhouse gas implications of
models in a widely applied technique called Multiple Endmember wetness trends in a large northern peatland.
Spectral Mixture Analysis (MESMA) (Roberts et al., 1998). From
candidate models, this technique selects the set of endmember fractions 2. Methods
that produce a sum of fractions between 0 and 1 with the smallest RMS
error and residuals between modeled and observed spectra. Endmember High-resolution water masks are first generated from 17 WorldView
variability has also been accounted for in coastal mapping, where sub­ scenes acquired over the HBL and scaled to produce 30 m water fractions
pixel waterlines were determined by first mapping pixels to land and for Landsat data-driven and physical unmixing model calibration and
water before unmixing along the land / water interface using adjacent validation. The best data-driven and physical models are applied next to
land and water pixels to represent their respective endmembers (Bishop- a 1985–2021 Landsat time-series over the HBL and annual surface water
Taylor et al., 2019). The relatively high variability of land reflectance in area is calculated and compared to annual water area from binary land /
the HBL necessitates a similar approach developed and tested here that water maps (Olthof and Rainville, 2022). Spatial patterns of per-pixel
uses a land endmember sampled in the local vicinity of each pixel to water fraction trends are investigated and potential implications of
represent its reflectance. The basic equation for pixel reflectance from a temporal trends on peatland GHG emissions are discussed. Surface water
linear mixture consisting of two (water and land) endmembers where information from an existing binary dynamic surface water dataset
Ewaterλ and Elandλ represent water and land spectral endmembers of (Olthof and Rainville, 2022) is leveraged throughout the analysis.
wavelength λ, fwater and fland represent water and land fractions that sum
to 1, and pλ represents pixel reflectance in wavelength λ is as follows (Eq. 2.1. Data
(1)):
2.1.1. Dynamic surface water
Ewaterλ fwater + Elandλ fland = pλ
A dynamic surface water dataset generated over Canada from his­
fwater + fland = 1 torical Landsat data between 1984 and 2021 was used in the current
study (Olthof and Rainville, 2022). The dataset includes binary annual
pλ − Elandλ water masks at national-scale, and derived inundation frequency and
= fwater (1) trends. The inundation frequency product depicts the percent frequency
Ewaterλ − Elandλ
that water was historically mapped on the surface, from 0% frequency
A second sub-pixel mapping approach is data-driven and relies on a representing permanent land, 100% frequency permanent water, and
relationship between reflectance and reference water fractions that are frequencies between 0% and 100% ephemeral water where inundation
usually obtained from finer resolution surface water maps (Olthof et al., has occurred occasionally. This product was overlaid on WorldView data
2015; Liang and Liu, 2021; Li et al., 2022). Reference surface water maps to sample land and water pixels for generating high-resolution water
generated from high resolution satellite or airborne data are first masks using machine learning, and on Landsat data to sample image-
aggregated to the coarser satellite resolution aligned with its grid to specific reflectance endmembers representing land and water for input
provide precise sub-pixel water fractions. Data-driven statistical or into linear unmixing.
machine-learning models are calibrated where finer resolution reference
surface water maps exist, and applied to satellite data to map other 2.1.2. WorldView
areas. Data-driven models can be simple linear regression between WorldView-2 imagery obtained over the HBL include four spectral
reflectance and water fraction (Li et al., 2022), or more complex models bands in blue (450–510 nm), green (510–580 nm), red (630–690 nm)
that use multispectral data and indices input into multiple regression, and near infrared (770–895 nm). 17 orthorectified WorldView scenes
machine learning regression or classification models (Li et al., 2018). projected to UTM zones 15–17 with a reported two-dimensional RMSE
This approach requires reference data that is representative of the entire of 6.18 m were acquired between years 2018 and 2021, with 13 of 17 in

3
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

the month of June and the remaining 4 in August (Table 1). Scenes were regression to obtain gain and offset calibration coefficients. For each
selected for minimal cloud cover, with 15 of 17 being completely cloud- year and band, this was repeated ten times to produce a set of co­
free and the remaining two having <2.3% cloud cover. All but one of the efficients that were used to calculate final coefficients from the mean of
scenes was pan-sharpened to a resolution of 0.5 m from their 2 m native 30 sets obtained from 10 trials and three years (Table 2). Final
resolution. Scenes were located primarily in the southeastern portion of 2012–2013 composites included data from Landsat 5/7 and cross-
the HBL below treeline, while coverage in the northeastern part in and calibrated Landsat 8, while subsequent composites included only
around Wapusk National Park was sparse, consisting of a single scene Landsat 7 and cross-calibrated Landsat 8 due to Landsat 5 failure in
(Fig. 2). 2013.

2.1.3. Landsat 2.2. Generating WorldView water masks

Rolling 3-year median Red, NIR and SWIR Landsat composites were
generated over the HBL from 1985 to 2021 in NAD83 Canada Atlas Reference water fractions were needed to calibrate machine learning
Lambert (EPSG 3979) projection. Based on Olthof and Rainville (2022), models and validate both physical and data-driven water fraction re­
the mean annual number of per-pixel observations after screening for trievals. These were generated from WorldView imagery at its native
cloud, shadow and poor classification performance during the snow-free resolution before calculating 30 m pixel water fractions expressed as a
season was between 5 and 8 in the HBL, which proved to be an insuf­ percentage of inundated area in each pixel aligned with the Landsat grid.
ficient number to generate high-quality, single-year annual composites High-resolution water masks were produced from 17 WorldView
for sub-pixel water body detection based on our visual assessment. scenes using fully-automated methods adapted from Olthof (2019).
Therefore, Landsat 5, 7 and 8 Collection 2 Tier 1 Level 2 orthorectified WorldView scenes were almost cloud-free, and individual tiles with
scenes measuring surface reflectance between June 1 to September 30 cloud cover were removed prior to water extraction to minimize errors
anniversary dates in three consecutive years were input to create com­ due to cloud and shadow. Landsat inundation frequency maps (Olthof
posites representing the middle year. Three years were included to and Rainville, 2022) were resampled to match the resolution and foot­
ensure sufficient data for high-quality, ice-free clear composite pro­ print of each WorldView tile in a scene, and WorldView land pixels were
duction for small water feature detection at these latitudes. All available randomly sampled where 30 m inundation was 0%, and water where
scenes in Google Earth Engine with <50% cloud cover were selected and inundation was >50%. A 50% inundation threshold was used to ensure a
cloud-screened prior to median compositing on a band-by-band basis. sufficiently large water sample size in subsequent machine learning but
Composites representing years 1985–2011 were produced from may introduce some water commission in training data where World­
Landsat 5 and 7 only, which are considered to be well inter-calibrated View pixels were not inundated at their time of acquisition. Because
(Roy et al., 2016). Known inconsistencies between the Thematic Map­ Landsat fails to capture water bodies smaller than its MMU, some 0%
per (TM and ETM+) onboard Landsat 5 and 7, and Operational Land Landsat inundation land samples may also miss water at the WorldView
imagery (OLI) onboard Landsat 8 due to differences between spectral scale. The impact of this was reduced by reclassifying the 5% darkest
response functions (Roy et al., 2016) were accounted for through cross- NIR land pixels in training data to water. In Olthof and Tolszczuk-Leclerc
calibration of coincident 2012, 2013 and 2014 TM and OLI composites. (2018), we showed that binary land / water classification with machine
Landsat 8 was launched on February 11, 2013 and therefore the 3-year learning is robust to errors that were deliberately introduced into
2012 composite included Landsat 8 data from 2013 only, while the 2013 training data to a level of roughly 30% error, and therefore errors in
composite included 2013–2014 Landsat 8, and the 2014 composite years WorldView training data should not significantly affect classification
2013–2015. While calibration coefficients generated from nearly coin­ performance. One hundred thousand land and one hundred thousand
cident ETM+ and OLI scenes over the conterminous US are provided in water pixels were sampled from each tile in a scene in four spectral
Roy et al. (2016), we felt it necessary to verify whether these apply to bands and input into the random forest machine learning algorithm to
composites generated some 700 km north over the HBL. generate a single model to classify all tiles in that scene.
Landsat 8 OLI was calibrated to match Landsat 5 and 7 TM / ETM+ The initial random forest classifications were inspected and com­
by regressing Landsat 8 composite reflectance on Landsat 5/7 compos­ mission errors were noted where shrubs and sparse trees cast shadow.
ites on an annual, band-wise basis. For three years from 2012 to 2014 Water tends to be more spectrally homogeneous over large areas than
inclusive and Red, NIR and SWIR bands, 10,000 pixels from Landsat 8 vegetation shadow, particularly in regions with shrubs and sparse, small
composites (x) and spatially coincident pixels from Landsat 5/7 com­ trees such as the HBL. A filter was applied that considered the local NIR
posites (y) were randomly sampled and input into a robust Thiel-Sen homogeneity within a 5 × 5 window surrounding each water pixel, and
water pixels with greater than a 100 digital number contrast between
minimum and maximum values within the window were re-classified as
Table 1 land. Commission errors due to remaining shadow were further reduced
WorldView scenes used for calibration / validation of 30 m water fractions. by subsequently applying a 3 × 3 mode filter. These processing steps
scene date # tiles resolution (m) Cloud removed the majority of shadow, but also eliminated some narrow water
bodies as well as water near shore.
013920478040_01_P001 2019-06-21 1 2 0
015011295160_01_P001 2018-06-10 16 0.5 0 After initial classification and filtering, the majority of remaining
015011295170_01_P001 2019-06-19 44 0.5 0 pixels represent portions of water bodies excluding nearshore water,
015011295180_01_P001 2018-06-14 9 0.5 0 narrow channels and tiny individual water features. Remaining water
015011372280_01_P001 2018-08-18 16 0.5 0 pixels represented seeds for region growing to the full extent of water
015011372280_01_P002 2020-08-31 21 0.5 0
015011372290_01_P001 2019-06-20 15 0.5 1.3%
features using a dark NIR threshold. Scene-based thresholds were
015011372290_01_P002 2019-06-20 1 0.5 0 calculated from the darkest 5% land pixels in NIR from the initial filtered
015011372290_01_P003 2019-06-20 1 0.5 0 classification, and pixels adjacent to water seeds and darker than this
015011372290_01_P004 2019-06-09 17 0.5 2.3% threshold were classified as water. WorldView water masks were
015011372300_01_P001 2021-08-01 27 0.5 0
inspected against input WorldView and high-resolution imagery in
015011372300_01_P002 2021-08-01 21 0.5 0
015011372310_01_P001 2021-06-11 2 0.5 0 Google Earth to optimize parameters and thresholds prior to settling on
015011372310_01_P002 2021-06-11 22 0.5 0 the final set. Water omission errors remained in certain tiles where
015011372310_01_P004 2018-06-07 23 0.5 0 specular reflectance from large water bodies occurred due to wind,
015011372310_01_P005 2021-06-11 34 0.5 0 causing a bright return and high local contrast in open water. These tiles
015011372310_01_P006 2021-06-11 30 0.5 0
were removed from final masks.

4
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

Fig. 2. Location of 17 WorldView scenes used for calibration / validation of 30 m water fractions in the Hudson Bay Lowlands.

machine learning. In order to select a pure land pixel in the vicinity of

Table 2
each pixel that will be unmixed, we used 1984–2021 Landsat inundation
Calibration coefficients to match Landsat OLI to Landsat TM/ETM+ from Roy
frequency from the dynamic surface water dataset generated over
et al. (2016), and over the HBL in the current study.
Canada (Olthof and Rainville, 2022) to remove pixels representing
Roy et al., 2016 Current study candidate endmembers where inundation was observed historically at
Red TM/ETM+ = 0.0123 + 0.9372 OLI TM/ETM+ = 0.0111 + 0.8080 OLI any point in time because water may be present, while remaining pixels
NIR TM/ETM+ = 0.0448 + 0.8339 OLI TM/ETM+ = 0.00193 + 0.8674 OLI represent permanent land where inundation frequency is equal to 0%
SWIR TM/ETM+ = 0.0306 + 0.8639 OLI TM/ETM+ = 0.0074 + 0.9339 OLI
and water is unlikely to occur. For each pixel, the algorithm first samples
all pixels in a 5 × 5 surrounding window as candidate land endmembers.
Tiles that formed a scene were mosaiced and re-projected to NAD83 From these, pixels are removed if they were inundated at any time in the
Canada Atlas Lambert (EPSG 3979) before generating 30 m percent past based on an inundation frequency > 0%, and the local land end­
water fraction maps aligned with the Landsat grid (Fig. 3). Windows member reflectance is calculated as the median of remaining pixels.
surrounding WorldView pixels located at the centre of each Landsat Where no land endmember is found, e.g. for a pixel located in open
pixel were sampled and percent water was calculated as the number of water, a global land endmember is used, calculated as the mean
water pixels divided by the total number of pixels in the window. For reflectance for all land and periodically flooded pixels beneath historical
pan-sharpened 0.5 m WorldView imagery, the window size was 61 × 61 inundation <80%.
= 3721 pixels while for the 2 m unsharpened scene, the window size was
15 × 15 = 225 pixels.
2.4. Data-driven water fraction retrieval

2.3. Sub-pixel water mapping from Landsat 2.4.1. WorldView-based calibration and validation
Machine learning random forest models relating WorldView water
2.3.1. Local land endmember selection fractions from 17 scenes and Landsat spectral data were calibrated,
In the current study, we assume a pixel in the vicinity of a mixed applied and evaluated using a scene-wise jackknife. For each WorldView
water pixel most likely contains the landcover whose reflectance rep­ scene, coincident rolling 3-year median Red, NIR and SWIR reflectance
resents the subpixel land endmember. A local land endmember was used was cropped to the extent of the scene. Both random forest classification
for physical linear unmixing and also tested as a predictor variable in and regression models were calibrated and tested with spectral bands

5
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

Fig. 3. Example of an area with lakes and patterned fens in the HBL where a 0.5 m resolution WorldView (top left) water mask was generated (top right) before
calculating 30 m water fractions (bottom left) aligned with the Landsat grid (bottom right) with NIR, SWIR, Red displayed as RGB. (For interpretation of the ref­
erences to colour in this figure legend, the reader is referred to the web version of this article.)

only, spectral bands and indices (NDVI and NDWI), and bands, indices performed at 10000 random locations at scales from 90 m to 270 m in
and local NIR land endmembers. For each test, water fractions were 60 m increments (from 3 × 3 to 9 × 9 windows in 2 × 2 window in­
estimated and validated for each scene in a jackknife hold-out in turn, crements). By aggregating over a range of scales, the assumption of
with calibration performed on the 16 remaining scenes. linear scaling can be checked and verified. In addition to the use of
random forest regression that generates continuous water fraction pre­
2.4.2. Calibration by Landsat aggregation with WorldView validation dictions necessary in DeVries et al. (2017) to interpolate 4% water
The relation between water fraction and reflectance can be estab­ fraction bins, we also tested random forest classification to predict the
lished by aggregating binary water masks and reflectance data at coarser 93 bins resulting from aggregation over a range of scales. Models were
scales. In DeVries et al. (2017), water fractions were calculated using an applied to Landsat data over WorldView footprints and evaluated
initial 30 m water mask that was aggregated to 150 m by summing the against mapped water fractions from all WorldView scenes.
number of water pixels and dividing by 25, representing the total
number of pixels in the 5 × 5 aggregation window. Note that aggre­
gating using a 5 × 5 window produces 25 possible water fraction bins in 2.5. Physical linear unmixing
4% ranges. Reflectance values consisting of six Landsat reflectance
channels, and derived spectral indices were similarly aggregated by The linear slope predicting water fraction from pixel reflectance is
averaging over the same 5 × 5 window. DeVries et al. (2017) then mainly governed by the subpixel landcover type and its reflectance.
calibrated a random forest regression model with water fraction as the Using conventional linear unmixing, water fractions were predicted for
dependent variable, and spectral values and indices as independent each pixel using eq. (1) between local land endmembers, and a global
variables, and applied this model to the original spectral bands and water endmember assuming water reflectance is stable across the image
indices to estimate 30 m water fractions. (Fig. 4). The global water endmember was calculated from the image as
The same approach was adapted for evaluation in the current study the average of the darkest 1% of all pixels where water was permanent
with a few notable differences. First, the binary water mask cropped to based on a historical inundation of 100%. After linear unmixing of the
the HBL was obtained for year 2020 from Olthof and Rainville (2022), pixel was performed, a series of operators were applied to constrain and
and 2020 Red, NIR and SWIR rolling 3-year median reflectance from the refine the prediction.
current study. NDVI and NDWI, as well as a local NIR land endmember Initial water fraction predictions sometimes fell beyond the limit of
in the surrounding 5 × 5 window were included as independent vari­ 0–100%, and these were truncated to either 0 or 100%. Pixels with
ables. Instead of aggregating to a fixed scale of 150 m, aggregation was reflectance greater than the 10% darkest land pixels based on an inun­
dation threshold <80% were assumed to contain no water and assigned

6
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

Fig. 4. Physical linear unmixing of a shoreline pixel (red) using a local land pixel (green) representing the land endmember, and a global water pixel (blue) to
represent water over Landsat (top left) and high-resolution imagery from Google Earth (top right). Note that in the current study, the global water endmember is
derived from the darkest 1% of all water pixels in the image. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)

a water fraction of 0%. Both pixels with reflectance less than the global coefficient and mean absolute error (MAE) were calculated between
water endmember, and those beneath historical inundation >95%, were predicted and reference water fractions for each WorldView validation
assumed to be complete water and therefore assigned a water fraction of scene and then averaged for all scenes. The standard deviation of each
100%. Finally, linear predictions were rounded to the nearest integer. measure among scenes was also included as an indicator of performance
consistency, measuring how well a model extended in space across
2.6. Evaluation scenes. Water fractions were also binarized in both predicted and
reference maps where pixels were assigned a value of 1 if any amount of
The performance of each unmixing approach (Table 3) was evalu­ water was present (water fraction >0), and 0 where no water was
ated using several metrics. The percent water area difference ((Area observed (water fraction = 0). Traditional classification assessment
metrics of omission, commission and overall accuracy (OAc) based on
predicted – AreaWorldView) / AreaWorldView x 100%), Pearson correlation

Table 3
Data-driven and physical water fraction modelling tests.
Method Name Calibration Validation Classifier

Data- WorldView 30 m water fractions from 16 WorldView scenes 30 m water fractions from 17th WorldView
driven scene (jack knife)
rf.img Spectral bands only RandomForest
classification
rf.img.vi Spectral bands + NDVI + NDWI RandomForest
classification
rf.img.reg Spectral bands only RandomForest
regression
rf.img.reg.vi Spectral bands + NDVI + NDWI RandomForest
regression
rf.img.vi.landEM Spectral bands + NDVI + NDWI + local land NIR EM RandomForest
classification
rf.img.reg.vi.landEM Spectral bands + NDVI + NDWI + local land NIR EM RandomForest
regression
Data- Landsat aggregation Landsat 90 m - 270 m water fractions and reflectance in 3 × 3 30 m water fractions from 17 WorldView
driven to 9 × 9 windows scenes
rf.img.scale.vi. Spectral bands + local land NIR EM RandomForest
landEM classification
rf.img.scale.reg.vi. Spectral bands + NDVI + NDWI + local land NIR EM RandomForest
landEM regression
Physical Local land and global water endmembers 30 m water fractions from 17 WorldView
scenes
lin.nir Linear unmixing NIR NA
lin.swir Linear unmixing SWIR NA
lin.ndvi Linear unmixing NDVI NA

7
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

confusion matrices were calculated from binary water fraction masks. on regression coefficients using scale as separate regressors in combined
For each of the 11 methods, individual metrics and their standard de­ models, which indicated that for the three spectral features, no signifi­
viations were assigned a rank from best (1) to worst (11) and ranks were cant differences at p < 0.05 were found between any pair-wise slopes
summed to give a relative measure of overall performance. representing different scales. For Red, turbidity causing high red
reflectance and median compositing that selected turbid pixels in lakes
2.7. Methods application and trends and off the coast in the HBL beneath WorldView scene footprints led to a
number of pure water pixels with high red reflectance. These outliers
Based on the evaluation, two methods were applied to the Landsat affect the slope of NDVI versus water fraction and cause it to be different
time-series and annual surface water area was calculated by summing (although not significantly) than the slope at other scales. Linearity
the water area in all pixels to determine the overall direction and suggested by high correlation coefficients is confirmed in plots, while
statistically similar regression lines from 30 m to 270 m confirms
magnitude of surface water change in the HBL for the 1985–2021
period. The two methods that were applied included the best physical scalability.
method based on linear unmixing, and the best data-driven method
based on random forest classification or regression. These two ap­ 3.1. Binary water masks
proaches were applied and compared due to the fact that results from the
physical method rely less on Landsat sensor and atmospheric calibration Water area and per-pixel water presence / absence omission, com­
than data-driven methods since global water and local land endmembers mission and OAc were compared between water fractions generated
were sampled from each annual image, whereas data-driven methods from each WorldView scene and annual water masks generated in Olthof
relied on absolute radiometric values and model extension in time. and Rainville (2022) corresponding to the year of each WorldView scene
A post-processing step was applied to the annual water fraction acquisition. The average WorldView water fraction beneath all pixels
products to account for recent wildfire burns that have low NIR reflec­ mapped as water in the binary masks covered by WorldView was 66.3%,
tance and can be assigned a low surface water fraction. Burned areas which suggests that in general, more than half of a Landsat pixel must be
were identified in each annual Landsat composite using a threshold of < inundated to be reliably detected as water in this environment. Using
− 0.02 applied to the Normalized Burn Ratio (Key and Benson, 2006) this average water fraction value for a binary water pixel, binary masks
and then set to 0% water. This spectral index, computed as the underestimated WorldView water area by 31.1%, with high variability
normalized difference between NIR and SWIR reflectance, is highly among scenes indicated by a standard deviation of 22.6%. When
effective in separating burned vegetation and water due to the elevated comparing water presence / absence between binary water masks to
SWIR signal of burns. In addition to these two methods, annual surface reference WorldView water fractions, a high omission error of 62.9%
water area from 1985 to 2021 binary Landsat water masks (Olthof and was noted, indicating that locations where WorldView detected water at
Rainville, 2022) was calculated for comparison. Water area from annual pixel to sub-pixel Landsat scale were missed frequently in binary masks.
binary masks was calculated by first determining the average World­ A relatively low commission error indicated that binary masks seldom
View water fraction in a water pixel for all WorldView scenes, and then map water where none is present in WorldView, leading to a relatively
multiplying the average water fraction by the area of a Landsat pixel high OAc (Table 5).
(900 m2) times the number of water pixels in each annual mask.
Trends were also mapped by applying per-pixel Theil-Sen linear 3.2. Data-driven
regression and Mann-Kendall significance to water fraction map time-
series derived from physical linear unmixing of the NIR band. Pixels 3.2.1. WorldView calibration
with significant increasing and decreasing water fraction trends at p < Table 5 presents performance metrics and their standard deviations
0.05 were mapped corresponding to areas of statistically significant for each of the eleven methods tested, with ranks from best (1) to worst
wetting and drying. To facilitate visualization across the entire HBL, (11) shown in parentheses. For calibration using WorldView water
trends were aggregated to ~1 km spatial resolution by calculating the fractions and random forest classification, a baseline using only reflec­
percentage of significantly wetting and drying 30 m pixels in each 990 m tance channels (rf.img) produced Landsat water fractions that under­
(33 × 33 pixel Landsat window) cell, and calculating the net change as estimated reference WorldView water area by 20.4%. Overall, extension
the difference between percent wetting and drying. performance measured by the standard deviation of metrics was slightly
lower than median values, while a relatively high overall accuracy was
achieved due to low commission error offset by high omission. There­
3. Results
fore, predicted inundated areas tended to be correct, but many areas that
were inundated in WorldView were missed in Landsat classifications. A
Pearson correlation coefficients between spectral features, and 30 m
slightly lower than median correlation and slightly higher than median
water fractions from WorldView, and combined water fractions scaled
MAE suggest average performance in terms of spatial agreement
from 90 m to 270 m Landsat are shown in Table 4. Correlation co­
encompassing both the location of inundation occurrence and their
efficients are similar between scales for all spectral features except for
water fraction magnitudes.
the red band, with high coefficients suggesting linearity.
The inclusion of NDVI and NDWI indices produced worse results than
The linearity and scalability of water fraction versus spectral features
baseline due to model overfitting to the set of calibration scenes in each
are shown in Fig. 5. 30 m and 90 m to 270 m average Landsat NIR, SWIR
hold-out iteration and poor extension performance, indicated by last-
and NDVI are plotted against corresponding 30 m water fractions from
place metric standard deviations. Binary commission was the lowest of
WorldView, and 90 m to 270 m water fractions aggregated from the
all methods, but was offset by the highest omission error, causing the
binary Landsat water mask, respectively. Pair-wise tests were conducted
second worst area underestimation of all methods at 59.6%. By
substituting random forest classification for regression using the same
Table 4
input data as the baseline, errors flipped from high omission to high
Correlations between Landsat spectral features and water fractions (WF) from
commission, indicating that many areas predicted to have water in
WorldView, and aggregated Landsat.
Landsat had none in WorldView, but that little water that was mapped in
Red NIR SWIR NDVI NDWI
WorldView was missed. This over-prediction of inundated areas is seen
WorldView WF (n = 1700) − 0.02 − 0.65 − 0.59 − 0.79 − 0.29 in the percent area difference that, while second lowest, was caused
Landsat aggregated WF (n = − 0.28 − 0.69 − 0.62 − 0.74 − 0.27 more by mis-classification than by overpredicting water fractions in
4000)
inundated areas. Model overfitting caused by including indices in

8
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

Fig. 5. Linear regressions of NIR, SWIR reflectance and NDVI versus water fraction at scales from sub-30 m to 270 m. Plots confirm linearity while no significant
differences (p < 0.05) among regression coefficients at different scales confirms scalability for all three spectral features.

Table 5
Evaluation of data-driven and physical water fraction modelling tests.
Name % area % area correlation correlation MAE MAE inundated pixels inundated pixels inundated sum of
difference difference SD SD SD omission error commission error pixels OAc ranks

binary − 31.1% 22.6% NA NA NA NA 62.9% 6.7% 88.6% NA

Data-driven
RF class − 20.4% (5) 18.5% (6) 0.82 (9) 0.09 (7) 3.89 3.42 56.4% (10) 9.6% (3) 89.4% (2) 55 (7)
Spectral bands (5) (8)
rf.img
RF class − 59.6% 15.4% (4) 0.68 (11) 0.13 (11) 5.09 4.66 84.5% (11) 0.1% (1) 85.4% (7) 75 (10)
Spectral bands (10) (9) (11)
+ VI
rf.img.vi
RF reg 6.7% (2) 27.6% (9) 0.87 (2) 0.08 (4) 4.41 3.75 8.2% (4) 69.8% (9) 61.9% (9) 57 (8)
Spectral bands (8) (10)
rf.img.reg
RF reg 224.8% (11) 248% (11) 0.73 (10) 0.13 (10) 19.5 2.59 0% (1) 82.7% (11) 17.3% (11) 81 (11)
Spectral bands (11) (5)
+ VI
rf.img.reg.vi
RF class − 21% (6) 17% (5) 0.83 (7) 0.08 (5) 3.7 (3) 3.06 56.1% (9) 6.5% (2) 89.7% (1) 45 (4)
Spectral bands (7)
+ VI + NIR EM
rf.img.vi.
landEM
RF reg 6.5% (1) 24.3% (8) 0.88 (1) 0.07 (2) 4.26 3.56 7.2% (3) 68.9% (8) 63.1% (8) 46 (5)
Spectral bands (6) (9)
+ VI + NIR EM
rf.img.reg.vi.
landEM
RF class scale 18.4% (3) 24.2% (7) 0.83 (8) 0.09 (9) 4.37 2.29 38% (5) 30.3% (7) 88.8% (4) 54 (6)
Spectral bands (7) (4)
+ VI + NIR EM
rf.img.scale. vi.
landEM
RF reg scale 45% (9) 38.9% (10) 0.85 (4) 0.08 (6) 5.72 2.96 3% (2) 76.1% (10) 46% (10) 67 (9)
Spectral bands (10) (6)
+ VI + NIR EM
rf.img.scale.
reg.vi.landEM
Physical
Linear unmix NIR − 26.3% (7) 12.8% (1) 0.86 (3) 0.07 (3) 3.42 1.96 52.5% (7) 26.3% (6) 88.0% (5) 35 (1)
lin.nir (1) (2)
Linear unmix − 20.2% (4) 14.2% (3) 0.85 (5) 0.09 (8) 3.48 1.84 53.5% (8) 25.8% (5) 87.9% (6) 42 (3)
SWIR (2) (1)
lin.swir
Linear unmix − 40.1% (8) 12.9% (2) 0.84 (6) 0.06 (1) 3.71 2.25 50.2% (6) 21.5% (4) 88.9% (3) 37 (2)
NDVI (4) (3)
lin.ndvi

9
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

random forest classification was similar for regression only more 3.2.2. Aggregating landsat water fractions
extreme, as inundation was predicted to occur in large areas where Water fraction – spectral feature relations obtained by aggregating
water was not truly present according to WorldView and almost no Landsat input into random forest classification or regression over­
water was missed, causing a net predicted water area more than twice estimated water area by 18.4% and 45%, respectively. Classification
the true area. balanced binary omission and commission errors best of all methods,
The addition of a NIR land endmember improved predictions slightly achieving a relatively high overall accuracy, while regression produced
compared to spectral bands only, and considerably compared to spectral low omission and high commission and a relatively poor accuracy.
bands and indices for both random forest classification and regression. Overall, classification achieved the median sum of ranks of all methods
As seen earlier, higher omission error and underpredicted water area (6), while regression performance was near the bottom (9).
were noted for classification, with higher commission errors and a small
water overpredicted area for regression. Both were among the best of the 3.2.3. Physical linear unmixing
data-driven methods tested, achieving 4th and 5th best sum of ranks, Physical linear unmixing using local NIR, SWIR and NDVI end­
respectively. members achieved the three best overall performances, with NIR being
the best among them. All three underestimated water area by 20% -
40%; however, these underestimations were consistent among

Fig. 6. An RGB overlay of water features including patterned fens at locations 1–4, with Random Forest classification (RF_img) displayed as Red, and linear NIR
unmixing (Lin_NIR) as green and blue (top left). Yellow corresponds to features present in both RF_img and Lin_NIR classifications. The binary surface water map
from Olthof and Rainville (2022) that captures the majority of water area but lacks detail in small features such as patterned fens is shown (top right) with cor­
responding high-resolution natural colour image from Google Earth™ (bottom right), and an enlargement of the patterned fen shown at location (1) (bottom left).
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

10
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

WorldView scenes as indicated by the lowest area difference standard 3.3. Temporal and spatial trends
deviations. This is due to the fact that physical methods do not rely on
calibration that can lead to overfitting to training data seen when Trends from three different methods including binary land / water
extending data-driven models to independent data. Consistency implies masks, calibration of random forest classification models against
that area bias corrections will be most reliable of all methods when WorldView water fractions, and physical linear unmixing of the NIR
applied in time and space, providing the best wall-to-wall estimates of band using local land endmembers suggest an overall wetting trend in
water fractions and their changes. Relatively high correlation and low the HBL since 1985 (Fig. 7). Both random forest and linear unmixing
MAE against reference water fractions suggest good spatial agreement. water fraction products suggest drying from 1985 to 1990, followed by
Close to median binary omission, commission and overall accuracy progressive wetting to 2021. The number of inundated pixels increased
implies good spatial agreement is caused less by predicted water loca­ in all three products but at a higher rate for binary masks and random
tion, and more by agreement in water fraction magnitude (Fig. 6). forest classification than for physical linear unmixing. Physical linear

Fig. 7. Trends in number of inundated pixels (water fraction >0%) and total surface water area for binary land / water maps from Olthof and Rainville (2022), and
linear NIR unmixing and machine learning random forest using all spectral features and local NIR land endmembers from the current study.

11
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

unmixing also had the most inundated pixels among the three products (Mondejar and Tongco, 2019), while comparisons here were made using
in all years due to the lowest omission error. 2020 Landsat composite data generated from both Landsat 7 and 8.
Overall water area suggests that not only do binary water masks miss Narrower NIR and SWIR bands measuring slightly different wavelengths
inundated pixels spatially, but they also generate water area predictions in Landsat 8 OLI compared to previous TM / ETM+ sensors may
that are about 40% less than both water fraction products assuming an contribute to performance differences (Mondejar and Tongco, 2019).
average binary pixel water fraction of 66.3%. Using a different average Better contrast between bright vegetation and water often exists for NIR,
water fraction for binary water pixels will not change the slope of ab­ especially if vegetation is dense or wet causing low SWIR reflectance.
solute water area increase but will change the offset. Both water fraction WorldView data-driven calibration using WorldView water fractions
products predicted similar water area with random forest suggesting as a reference produces comparable or slightly better results overall
approximate 9% more water area, and interannual variability was also compared to binary water maps, but in general does not extend partic­
significantly related between the two products at r = 0.59, p < 0.05. The ularly well especially in the north-south direction (Olthof et al., 2005).
fact that random forest predicted more water area but fewer inundated Compared to physical methods, it also requires more time and effort to
pixels compared to linear unmixing suggests that inundated pixels in generate calibration data and apply machine learning models. Use of
random forest maps had higher percentage water fractions overall. spectral features only produced marginal results due to relatively poor
Broad spatial patterns suggest an overall drying trend in zones con­ model extension, particularly when including indices due to overfitting
taining lichen peat bogs in the interior of Wapusk National Park (Brook to calibration data, while the addition of local land endmembers
and Kenkel, 2002) as well as bogs inland of Hudson Bay, and wetting improved results. If the entire mapping domain is well represented with
trends in coastal fens near the Hudson Bay coast (OMNR, 2014) (Fig. 8). high-resolution image samples, calibration using high-resolution water
masks can be a good option and high-resolution masks are also needed to
4. Discussion validate model performance. Because data-driven or machine learning
models are developed on coincident WorldView – Landsat dates, time-
While other studies using Landsat TM / ETM+ have shown SWIR or series analysis requires temporal extension of models and therefore
NDWI to be optimal for water fraction retrieval (Frazier and Page, 2000; depends on the quality of Landsat radiometric calibration. Note that
Gao, 1996; Olthof et al., 2005), results here suggest the NIR and NDVI validation of all data-driven and physical methods depends on the
performed better. At least one other study showed that the Landsat 8 NIR spatial alignment of WorldView and Landsat; however, only data-driven
band outperformed other bands and indices for surface water extraction calibration using WorldView reference water masks depends on this

Fig. 8. Linear NIR unmixing per-pixel Thiel-Sen spatial trends aggregated to 1 km resolution to facilitate visualization. Each 1 km pixel represents the sum per­
centages of significantly trended pixels (p < 0.05) in positive (wetting) and negative (drying) directions.

12
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

alignment since other methods use only Landsat to map water fractions. with other cover types, particularly dark dense vegetation and soil.
Therefore, the reported 6.18 m accuracy of orthorectified WorldView Results presented in this study for the HBL may represent an upper limit
may have penalized data-driven methods that rely on WorldView to what is achievable in other regions due to limited tree cover and
calibration. relatively consistent water colour compared to coastal regions for
Scaling Landsat water fraction – reflectance can also be a good op­ example, where waves and turbidity can be significant and highly var­
tion, since machine learning classification produced better results iable. One issue that was sometimes noticed with the linear NIR
overall with balanced omission and commission errors and good overall unmixing specifically was in narrow strips of vegetation, for example
accuracy compared to binary water masks. However, processing time dark riparian vegetation corridors that would sometimes be mistaken for
can be an issue since calibration requires extraction of large windows of water due to median sampling of brighter, local land endmembers in the
reflectance / water fraction over the whole mapping domain and vicinity. High spatial heterogeneity of surface cover types may also limit
training on large sample sizes is limited due to processing power. the performance of the method since sampled local endmembers may
Applying random forest models over large spatial and temporal mapping not be representative of the land portion of the unmixed pixel in these
domains can be time consuming as well. Because machine learning environments.
models are calibrated on scaled Landsat and derived binary water masks All three methods (data-driven using WorldView water fraction, and
independent of any external data, date-specific models can be generated Landsat scaled water fraction, as well as physical linear unmixing),
and applied to reduce the dependence of trends on radiometric cali­ leverage binary dynamic surface water maps as an intermediate source
bration, but this increases processing time further. Random forest of information to generate improved water dynamics at a sub-pixel
regression causes large commission error since it interpolates between scale. Historical inundation frequency is used to sample pure land and
water fraction bins input into the regression, while classification reduces water to generate calibration / validation data from high-resolution
commission error by fitting nonlinear, or nonmonotonic relations. WorldView imagery. Landsat scaling also requires binary land / water
Applying a 9 × 9 window and below generates 93 water fraction bins maps to generate water fractions at a range of scales, and physical
with classification, which is an improvement over use of a single win­ unmixing uses inundation frequencies to select local land endmembers
dow size but represents slightly lower precision compared to prediction from candidates in surrounding windows, and water endmembers
over the entire 0–100% range. beneath permanent water. Therefore, production of binary historical
Physical linear unmixing of land / water is relatively efficient and water masks and inundation frequencies can be considered a necessary
requires a single band of imagery with the NIR performing best in the first step towards producing enhanced, sub-pixel water maps that
case of the HBL. The use of local endmembers accounts for the vari­ portray surface water more accurately in environments such as the HBL
ability in land reflectance to adequately represent the land portion of where a relatively large percentage of water features are smaller than 30
mixed water pixels compared to the use of a fixed land endmember. m. We used a dynamic surface water product that is optimized for
Physical unmixing errs on the side of omission and has low commission Canada (Olthof and Rainville, 2022), but global datasets exist that can
error, which enables a more accurate water area bias correction than in be leveraged for this application elsewhere (Pekel et al., 2016; Pickens
the inverse case of high commission. This is because correcting com­ et al., 2020).
mission bias requires the subtraction of the percentage of incorrectly Trends suggest overall wetting in the HBL, with unmixing detecting a
classified water, whereas in the current study the sources of commission larger number of inundated pixels, but a smaller water area overall
error are unknown. This is particularly problematic in change detection compared to binary maps. In the case of unmixing, a decreasing number
since sources of commission error can vary from year to year depending of inundated pixels and water area were observed from 1985 to 1990,
on how falsely classified water targets change. For example, dark conifer followed by a progressive wetting trend from 1990 to 2021 (Fig. 9). This
forest is sometimes incorrectly classified as water; however these com­ is broadly consistent with a 15% decline in Hudson Bay river discharge
mission errors may be greater or less in different years depending on observed between 1964 and 1994 explained largely by variations in the
conifer darkness due to conditions such as forest stress, insects and Arctic Oscillation (Déry and Wood, 2004), followed by a wet period
drought that can affect tree colour. from 1995 to 2008 due to a shift in regional atmospheric circulation,
WorldView suggests approximately 26% water area is missed in then sustained streamflow in 2009–2019 likely due to permafrost thaw
Landsat linear NIR unmixing, but this amount is most consistent of all (Champagne et al., 2021). High interannual variability was typically
methods across WorldView scenes as measured by the area difference noted in shallow lakes at local scales due to variations in precipitation
standard deviation among WorldView scenes. One source of omission (Bouchard et al., 2013). Significant (p < 0.05) wetting / drying trends
error is likely caused by shallow or turbid water, causing the water were mapped for binary water maps using logistic regression in Olthof
endmember to be brighter locally than the global water endmember and Rainville, 2022, and for water fractions using Thiel-Sen. Both show
used across the whole study domain. Adapting the water endmember to similar patterns overall, with water fraction trends producing more
local conditions was attempted but proved challenging due to the far significant wetting and drying area. In few cases, trends can be opposite
proximity of water to many pixels that in most cases, did not represent between products, which suggests more frequent inundation but less
local conditions well. The effect of water endmember variability was water fraction over time, or vice-versa. In Wapusk National Park, binary
examined by determining the 95% NIR water reflectance confidence water mask trends suggest wetting and many wetting and drying fea­
intervals for year 2020 (1.4–5.6% reflectance), and calculating the range tures are confirmed in water fraction trends. Many of these occur around
of mixed water fractions using the land endmember and mixed pixel the margins of lakes as both lake expansion and contraction, with water
reflectance in Fig. 4. This resulted in a range of predicted water fractions fraction trends generally producing wider margins of change. However,
between 32 and 40%, which suggests that in 95% of cases, water frac­ sub-pixel water features that are ubiquitous in bogs appear to be drying
tions are within 2–6% of the 34% value predicted using the global water in areas where no trends are detected in binary masks, producing an
endmember (Fig. 4). A more thorough analysis of the effects of water overall drying trend and highlighting the value of unmixing and com­
endmember variability may reveal wider predicted water fraction con­ plementary information to binary masks.
fidence intervals in some cases; however, this analysis suggests rela­
tively small intervals and an in-depth analysis of the effects of 4.1. Implications
endmember variability is beyond the scope of the current study. Because
this unmixing method is adaptive to local land and scene-specific water All permafrost zones that exist in Canada occur in the HBL, from the
endmembers, trends will also depend less on radiometric calibration unfrozen zone in the southeast through to the continuous permafrost
than data-driven classification methods. zone along the Hudson Bay coast and to the northwest into Wapusk NP.
All water detection from optical data has limitations due to confusion Whether northern peatlands remain as carbon sinks or become a source

13
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

Fig. 9. Linear NIR unmixing inundated area changes corresponding to a period of declining Hudson Bay Lowlands river discharge observed in Déry and Wood
(2004), followed by a 1995–2008 wet period caused by a regional shift in regional atmospheric circulation and sustained streamflow in 2009–2019 due to permafrost
thaw (Champagne et al., 2021).

of atmospheric carbon emissions as temperatures increase is largely thermal and wind erosion are also contributing to peat plateau degra­
governed by hydrology (Sim et al., 2021). Frozen peatlands in perma­ dation and conversion to fen in the HBL (Dyke and Sladen, 2010).
frost zones are especially vulnerable due to climate change (Sim et al., Under saturated conditions peat soils emit methane (Hopple et al.,
2021), while non-frozen peatlands are expected to remain long-term 2020), a greenhouse gas with >20 times the warming potential of CO2.
carbon sinks under most projected climate change scenarios (Qiu However, collapsed peatland and saturation by surface and groundwater
et al., 2020). can also cause vegetation mortality in the short-term, followed by
The net effect of warming on northern peatland greenhouse gas increased primary productivity due to the influence of local hydro­
emissions is uncertain (Harris et al., 2022); however, evidence suggests chemistry, leading to conditions where carbon accumulation is rapid
that increased carbon emissions are likely to occur in the short-term due (Swindles et al., 2016). Recent evidence suggests that increasing surface
to peat desiccation at the surface and enhanced decomposition at depth wetness in unfrozen peatlands can cause them to switch from being net
while in the longer-term, increasing wetness due to permafrost thaw and sources to net sinks of greenhouse gases. Huang et al. (2021) found that
conversion from bog to fen may enhance carbon accumulation. Swindles changes in peat wetness was due to changes in water table depth and
et al., 2016 describe a 5-phase model of permafrost peatland response to was the dominant control on carbon and methane fluxes. Evans et al.
increasing temperature from intact peatland, to collapse due to perma­ (2021) suggest that the reduction in carbon dioxide emissions more than
frost thaw and eventual conversion to artic fen that have higher rates of offsets the accompanying increase in methane emissions from wetting
carbon accumulation compared to bogs. peatlands, causing them to be net sinks. Therefore, the wetting trend
Swindles et al., 2016 model suggests that initially, warming causes observed in Landsat across the HBL may produce a net benefit for carbon
drying of the upper layer of sphagnum, increasing the rate of decay and budgets, since evidence suggests that wetting peatlands act as a carbon
carbon emissions. Warming also increases the active layer depth leading sink and thus form part of a nature-based climate solution.
to decomposition of previously frozen soil organic matter while also
increasing the rate of soil microbial activity, both of which contribute to 5. Conclusions
enhanced GHG emissions (Sim et al., 2021). Spatial trends suggest that
drying is occurring mainly in sub-pixel water features in bogs in the The HBL contains all Canadian permafrost zones from unfrozen to
interior of Wapusk National Park, and inland of the Hudson Bay coast. continuous, making its peatlands particularly vulnerable to climate
As warming continues, permafrost begins to degrade and thaw through change. Water plays a key role in the fate of northern peatlands as a
enhanced thermokarst, leading to increased wetness. In their metanal­ significant carbon sink, or as a large source of atmospheric carbon
ysis, McLaughlin and Webster (2014) found that recent carbon accu­ emissions. Existing binary surface water maps from Landsat, which
mulation in the decades to centuries timeframe is significantly higher in represents the longest and most detailed EO mission, may not capture
wet peatland features compared to dry, while Sim et al. (2021) found no changes in the many peatland water features smaller than its 30 m
clear relationship between peatland wetting and carbon sequestration resolution in the HBL such as patterned fens, thus requiring a sub-pixel
rates in northern Sweden. mapping approach to reliably inform peatland wetting / drying trends
A wetting trend observed in the current study, particularly in coastal and associated carbon emissions. A comparison of data-driven machine
fens accompanied by increasing temperatures observed in the HBL since learning approaches to a new physical linear unmixing approach be­
1990 suggest that permafrost thaw may be ongoing and widespread. tween water and local land endmembers revealed the superior perfor­
Permafrost thaw causes peat domes to collapse or erode at their margins mance of the physical approach using the NIR band. When applied to the
via thermokarst encroachment, eventually converting peat bogs to 1985–2021 time-series, spatial trends suggest that drying is occurring
wetter fens (Dyke and Sladen, 2010; Swindles et al., 2016). Conversion mainly in sub-pixel water features in bogs in the interior of Wapusk
from bog to fen has been observed in palaeoecological records in north- National Park and inland of the Hudson Bay coast, while an overall
central Alberta, Saskatchewan and Manitoba since the end of the little wetting trend is observed in the HBL along coastal fens since 1990 when
ice age (Vitt et al., 1994), and is ongoing northward into the HBL with increasing temperatures began to occur. While considerable work has
increased warming (Dyke and Sladen, 2010; Pironkova, 2017; Kirkwood been done in recent years on the response of peatland carbon to climate
et al., 2021). Numerous shallow bog lakes that are expanding due to warming, significant knowledge gaps still exist. Since water is known to

14
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

strongly influence whether peatlands remain as carbon sinks or become Frazier, P.S., Page, K.J., 2000. Water body detection and delineation with Landsat TM
data. Photogramm. Eng. Remote. Sens. 66, 1461–1468.
a source of carbon emissions as temperature increases, we hope that
Gagnon, A.S., Gough, W.A., 2005. Climate change scenarios for the Hudson Bay region:
detailed surface water time-series such as the one developed in this an intermodal comparison. Clim. Chang. 69, 269–297.
paper can help address remaining uncertainties. Gao, B.-C., 1996. NDWI—A normalized difference water index for remote sensing of
vegetation liquid water from space. Remote Sens. Environ. 58, 257–266. https://doi.
org/10.1016/S0034-4257(96)00067-3.
CRediT authorship contribution statement Hadley, K.R., Paterson, A.M., Rühland, K.M., White, H., Wolfe, B.B., Keller, W., Smol, J.
P., 2019. Biological and geochemical changes in shallow lakes of the Hudson Bay
Ian Olthof: Conceptualization, Methodology, Software, Validation, lowlands: a response to recent warming. J. Paleolimnol. 61, 313–328. https://doi.
org/10.1007/s10933-018-0061-9.
Formal analysis, Investigation, Resources, Data curation, Writing – Harris, L.I., Richardson, K., Bona, K.A., Davidson, S.J., Finkelstein, S.A., Garneau, M.,
original draft, Writing – review & editing, Visualization, Supervision, McLaughlin, J., Nwaishi, F., Olefeldt, D., Packalen, M., Roulet, N.T., Southee, F.M.,
Project administration. Robert H. Fraser: Validation, Formal analysis, Strack, M., Webster, K.L., Wilkinson, S.L., Ray, J.C., 2022. The essential carbon
service provided by northern peatlands. Front. Ecol. Environ. 20, 222–230. https://
Investigation, Writing – review & editing, Visualization. doi.org/10.1002/fee.2437.
Hopple, A.M., Wilson, R.M., Kolton, M., Zalman, C.A., Chanton, J.P., Kstka, J.,
Declaration of Competing Interest Hanson, P.J., Keller, J.K., Bridghsm, S.D., 2020. Massive peatland carbon banks
vulnerable to rising temperatures. Nat. Commun. 11, 2373. https://doi.org/
10.1038/s41467-020-16311-8.
The authors declare that they have no known competing financial Huang, Y., Ciais, P., Luo, Y., Zhu, D., Wang, Y., Qiu, C., Goll, D.S., Guenet, B.,
interests or personal relationships that could have appeared to influence Makowski, D., De Graaf, I., Leifeld, J., Kwon, M.J., Hu, J., Qu, L., 2021. Tradeoff of
CO2 and CH4 emissions from global peatlands under water-table drawdown. Nat.
the work reported in this paper.
Clim. Chang. 11, 618–622. https://doi.org/10.1038/s41558-021-01059-w.
IPCC, 2023. Summary for policymakers. In: Core Writing Team, Lee, H., Romero, J.
Data availability (Eds.), Climate Change 2023: Synthesis Report. A Report of the Intergovernmental
Panel on Climate Change. Contribution of Working Groups I, II and III to the Sixth
Assessment Report of the Intergovernmental Panel on Climate Change. IPCC,
Data will be made available on request. Geneva, Switzerland, 36 pages. (in press).
Key, C.H., Benson, N.C., 2006. Landscape assessment: ground measure of severity, the
Acknowledgements composite burn index; and remote sensing of severity, the normalized burn ratio. In:
Lutes, D.C., Keane, R.E., Caratti, J.F., Key, C.H., Benson, N.C., Sutherland, S.,
Gangi, L.J. (Eds.), FIREMON: Fire Effects Monitoring and Inventory System. USDA
Thanks to Taylor Biccum for creating map layouts. Funding was Forest Service, Rocky Mountain Research Station, Gen. Tech. Rep, Ogden, UT.
provided under Environment Climate Change Canada’s Nature-based Kirkwood, J.A.H., Roy-Léveillée, P., Mykytczuk, N., Packalen, M., McLaughlin, J.,
Laframboise, A., Basiliko, N., 2021. Soil microbial community response to
Climate Solutions Program. permafrost degradation in palsa fields of the Hudson Bay lowlands: implications for
greenhouse gas production in a warming climate. Glob. Biogeochem. Cycles 35.
References https://doi.org/10.1029/2021GB006954 e2021GB006954.
Li, L., Vrieling, A., Skidmore, A., Wang, T., Turak, E., 2018. Monitoring the dynamics of
surface water fraction from MODIS time series in a Mediterranean environment. Int.
Bishop-Taylor, R., Sagar, S., Lymburner, L., Alam, I., Sixsmith, J., 2019. Sub-pixel
J. Appl. Earth Obs. Geoinf. 66, 135–145. https://doi.org/10.1016/j.
waterline extraction: characterising accuracy ands to indices and spectra. Remote
jag.2017.11.007.
Sens. 11, 2984. https://doi.org/10.3390/rs11242984.
Li, X., Jia, X., Yin, Z., Du, Y., Ling, F., 2022. Integrating MODIS and Landsat imagery to
Bouchard, F., Turner, K.W., MacDonald, L.A., Deakin, C., White, H., Farquharson, N.,
monitor the small water area variations of reservoirs. Sci. Remote Sens. 5, 100045.
Medeiros, A.S., Wolfe, B.B., Hall, R.I., Pienitz, R., Edwards, T.W.D., 2013.
ISSN 2666-0172. https://doi.org/10.1016/j.srs.2022.100045.
Vulnerability of shallow subarctic lakes to evaporate and desiccate when snowmelt
Liang, J., Liu, D., 2021. Automated estimation of daily surface water fraction from
runoff is low. Geophys. Res. Lett. 40, 6112–6117. https://doi.org/10.1002/
MODIS and Landsat images using Gaussian process regression. Int. J. Remote Sens.
2013GL058635.
42, 4261–4283. https://doi.org/10.1080/01431161.2021.1892859.
Bourgeau-Chavez, L., Endres, S., Battaglia, M., Miller, M.E., Banda, E., Laubach, Z.,
McLaughlin, J., Webster, K., 2014. Effects of climate change on peatlands in the far north
Higman, P., Chow-Fraser, P., Marcaccio, J., 2015. Development of a bi-national
of Ontario, Canada: a synthesis. Arct. Antarct. Alp. Res. 46, 84–102. https://doi.org/
Great Lakes coastal wetland and land use map using three-season PALSAR and
10.1657/1938-4246-46.1.84.
Landsat imagery. Remote Sens. 7, 8655–8682. https://doi.org/10.3390/
Mondejar, J.P., Tongco, A.F., 2019. Near infrared band of Landsat 8 as water index: a
rs70708655.
case study around Cordova and Lapu-Lapu City, Cebu, Philippines. Sustain. Environ.
Brook, R.K., Kenkel, N.C., 2002. A multivariate approach to vegetation mapping of
Res. 29 https://doi.org/10.1186/s42834-019-0016-5.
Manitoba’s Hudson Bay lowlands. Int. J. Remote Sens. 23, 4761–4776.
Morris, P.J., 2021. Wetter is better for peat carbon. Nat. Clim. Chang. 11, 561–562.
Champagne, O., Altaf Arain, M., Wang, S., Leduc, M., Russell, H., 2021. Interdecadal
https://doi.org/10.1038/s41558-021-01072-z.
variability of streamflow in the Hudson Bay lowlands watersheds driven by
Mueller, N., Lewis, A., Roberts, D., Ring, S., Melrose, R., Sixsmith, J., Lymburner, L.,
atmospheric circulation. J. Hydrol. Reg. Stud. 36, 100868. ISSN 2214-5818. htt
McIntyre, A., Tan, P., Curnow, S., Ip, A., 2016. Water observations from space:
ps://doi.org/10.1016/j.ejrh.2021.100868.
mapping surface water from 25 years of Landsat imagery across Australia. Remote
Cooley, S.W., Smith, L.C., Ryan, J.C., Pitcher, L.H., Pavelsky, T.M., 2019. Arctic-boreal
Sens. Environ. 174, 341–352. https://doi.org/10.1016/j.rse.2015.11.003.
lake dynamics revealed using CubeSat imagery. Geophys. Res. Lett. 46, 2111–2120.
NWWG (National Wetlands Working Group), 1988. Wetlands of Canada. Ecological Land
https://doi.org/10.1029/2018GL081584.
Classification Series No. 24. Environment Canada. Sustainable Development Branch
Déry, S.J., Wood, E.F., 2004. Teleconnection between the Arctic oscillation and Hudson
and Polyscience Publications Inc, Montreal QC, 452pp.
Bay river discharge. Geophys. Res. Lett. 31, L18205. https://doi.org/10.1029/
Olthof, I., 2019. Automated surface water extraction from RapidEye imagery including
2004GL020729.
cloud and cloud shadow detection. Geomat. Can. Open File 52. https://doi.org/
DeVries, B., Huang, C., Lang, M.W., Jones, J.W., Huang, W., Creed, I.F., Carroll, M.L.,
10.4095/315176.
2017. Automated quantification of surface water inundation in wetlands using
Olthof, I., Fraser, R.H, Schmitt, C., 2015. Landsat-based mapping of thermokarst lake
optical satellite imagery. Remote Sens. 9, 807. https://doi.org/10.3390/rs9080807.
dynamics on the Tuktoyaktuk Coastal Plain, Northwest Territories, Canada since
Du, Q., 2018. Pixel unmixing. In: Liang, Shunlin (Ed.), Comprehensive Remote Sensing.
1985. Remote Sens. Environ. 168, 194–204. https://doi.org/10.1016/j.
Volume 2: Data Processing and Analysis Methodology. Elsevier, New York, USA,
rse.2015.07.001.
pp. 186–198.
Olthof, I., Rainville, T., 2022. Dynamic surface water maps of Canada from 1984 to 2019
Dyke, L.D., Sladen, W.E., 2010. Permafrost and peatland evolution in the northern
Landsat satellite imagery. Remote Sens. Environ. 279 https://doi.org/10.1016/j.
Hudson Bay lowland, Manitoba. Arctic 63, 429–441.
rse.2022.113121.
Evans, C.D., Peacock, M., Baird, A.J., Artz, R.R.E., Burden, A., Callaghan, N.,
Olthof, I., Tolszczuk-Leclerc, S., 2018. Comparing Landsat and RADARSAT for current
Chapman, P.J., Cooper, H.M., Coyle, M., Craig, E., Cumming, A., Dixon, S., Gauci, V.,
and historical dynamic flood mapping. Remote Sens. 10, 780. https://doi.org/
Grayson, R.P., Helfter, C., Heppell, C.M., Holden, J., Jones, D.L., Kaduk, J., Levy, P.,
10.3390/rs10050780.
Matthews, R., McNamara, N.P., Misselbrook, T., Oakley, S., Page, S.E., Rayment, M.,
Olthof, I., Butson, C., Fraser, R., 2005. Signature extension through space for northern
Ridley, L.M., Stanley, K.M., Williamson, J.L., Worrall, F., Morrison, R., 2021.
landcover classification: a comparison of radiometric correction methods. Remote
Overriding water table control on managed peatland greenhouse gas emissions.
Sens. Environ. 95 https://doi.org/10.1016/j.rse.2004.12.015.
Nature 593, 548–552. https://doi.org/10.1038/s41586-021-03523-1.
OMNR, 2014. Ontario ministry of natural resources. Far North Land Cover v1.4.
Fewster, R.E., Morris, P.J., Ivanovic, R.F., Swindles, G.T., Peregon, A.M., Smith, C.J.,
https://geohub.lio.gov.on.ca/documents/lio::far-north-land-cover/about.
2022. Research briefing: permafrost peat carbon approaching a climatic tipping
Pekel, J.F., Cottam, A., Gorelick, N., Belward, A., 2016. High-resolution mapping of
point. Nat. Clim. Chang. 12, 320–321. https://doi.org/10.1038/s41558-022-01298-
global surface water and its long-term changes. Nature 540, 418–436. https://doi.
5.
org/10.1038/nature20584.
Fraser, L.H., Keddy, P.A., 2005. The World’s Largest Wetlands: Ecology and
Conservation. Cambridge University Press, Cambridge, UK, 488pp.

15
I. Olthof and R.H. Fraser Remote Sensing of Environment 300 (2024) 113895

Pickens, A.H., Hansen, M.C., Hancher, M., Stehman, S.V., Tyukavina, A., Potapov, P., Settle, J.J., Drake, N.A., 1993. Linear mixing and the estimation of ground cover
Marroquin, B., Sherani, Z., 2020. Mapping and sampling to characterize global proportions. Int. J. Remote Sens. 14, 1159–1177. https://doi.org/10.1080/
inland water dynamics from 1999 to 2018 with full Landsat time-series. Remote 01431169308904402.
Sens. Environ. 243 https://doi.org/10.1016/j.rse.2020.111792. Sim, T.G., Swindles, G.T., Morris, P.J., Baird, A.J., Cooper, C.L., Gallego-Sala, A.V.,
Pironkova, Z., 2017. Mapping Palsa and peat plateau changes in the Hudson Bay Charman, D.J., Roland, T.P., Borken, W., Mullan, D.J., Aquino-López, M.A.,
lowlands, Canada, using historical aerial photography and high-resolution satellite Gałka, M., 2021. Divergent responses of permafrost peatlands to recent climate
imagery. Can. J. Remote. Sens. 43, 455–467. https://doi.org/10.1080/ change. Environ. Res. Lett. 16 https://doi.org/10.1088/1748-9326/abe00b.
07038992.2017.1370366. Sothe, C., Gonsamo, A., Arabian, J., Kurz, W.A., Finkelstein, S.A., Snider, J., 2022. Large
Qiu, C., Zhu, D., Ciais, P., Guenet, B., Peng, S., 2020. The role of northern peatlands in soil carbon storage in terrestrial ecosystems of Canada. Glob. Biogeochem. Cycles 36.
the global carbon cycle for the 21st century. Glob. Ecol. Biogeogr. 29, 956–973. https://doi.org/10.1029/2021GB007213 e2021GB007213.
https://doi.org/10.1111/geb.13081. Swindles, G.T., Morris, P.J., Mullan, D., Watson, E.J., Turner, E., Roland, T.P.,
Roberts, D.A., Gardner, M., Church, R., Ustin, S., Scheer, G., Green, R.O., 1998. Mapping Amesbury, M.J., Kokfelt, U., Schoning, K., Pratte, S., Gallego-Sala, A., Charman, D.J.,
chaparral in the Santa Monica Mountains using multiple endmember spectral Sanderson, N., Garneau, M., Carrivick, J.L., Woulds, C., Holden, J., Parry, L.,
mixture models. Remote Sens. Environ. 65, 267–279. https://doi.org/10.1016/ Galloway, J.M., 2016. The long-term fate of permafrost peatlands under rapid
S0034-4257(98)00037-6. climate warming. Sci. Rep. 5, 17951. https://doi.org/10.1038/srep17951.
Rodell, M., Famiglietti, J.S., Wiese, D.N., Reager, J.T., Beaudoing, H.K., Landerer, F.W., Vitt, D.H., Halsey, L.A., Zoltai, S.C., 1994. The bog landforms of continental western
Lo, M.H., 2018. Emerging trends in global freshwater availability. Nature 557, Canada in relation to climate and permafrost patterns. Arct. Alp. Res. 26, 1–13.
651–659. https://doi.org/10.1038/s41586-018-0123-1. https://doi.org/10.2307/1551870.
Rover, J., Wylie, B.K., Ji, L., 2010. A self-trained classification technique for producing Wulder, M.A., Loveland, T.R., Roy, D.P., Crawford, C.J., Masek, J.G., Woodcock, C.E.,
30 m percent-water maps from Landsat data. Int. J. Remote Sens. 31, 2197–2203. Allen, R.G., Anderson, M.C., Belward, A.S., Cohen, W.B., Dwyer, J., Erb, A., Gao, F.,
Roy, D.P., Kovalskyy, V., Zhang, H.K., Vermote, E.F., Yan, L., Kumar, S.S., Egorov, A., Griffiths, P., Helder, D., Hermosilla, T., Hipple, J.D., Hostert, P., Hughes, M.J.,
2016. Characterization of Landsat-7 to Landsat-8 reflective wavelength and Huntington, J., Johnson, D.M., Kennedy, R., Kilic, A., Li, Z., Lymburner, L.,
normalized difference vegetation index continuity. Remote Sens. Environ. 185, McCorkel, J., Pahlevan, N., Scambos, T.A., Schaaf, C., Schott, J.R., Sheng, Y.,
57–70. https://doi.org/10.1016/j.rse.2015.12.024. Storey, J., Vermote, E., Volgelmann, J., White, J.C., Wynne, R.H., Zhu, Z., 2019.
Current status of Landsat program, science, and applications. Remote Sens. Environ.
225, 127–147. https://doi.org/10.1016/j.rse.2019.02.015.

16