1 s2.0 S0303243421002889 Main

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

International Journal of Applied Earth Observations and Geoinformation 105 (2021) 102581

Contents lists available at ScienceDirect

International Journal of Applied Earth


Observations and Geoinformation
journal homepage: www.elsevier.com/locate/jag

Machine learning with high-resolution aerial imagery and data fusion to


improve and automate the detection of wetlands
Santiago López-Tapia a, Pablo Ruiz a, b, Mitchell Smith c, Jeffrey Matthews d, Bradley Zercher e,
Liliana Sydorenko f, Neelanshi Varia c, Yuanzhe Jin a, Minzi Wang d, Jennifer B. Dunn g, h, i, *,
Aggelos K. Katsaggelos a, h, i
a
Department of Electrical and Computer Engineering, Northwestern University, Evanston, IL, USA
b
OriGen.AI, Brooklyn, NY, USA
c
Department of Computer Science, Northwestern University, Evanston, IL, USA
d
Natural Resources and Environmental Sciences, University of Illinois at Urbana-Champaign, USA
e
Illinois Natural History Survey, Prairie Research Institute, University of Illinois at Urbana-Champaign, USA
f
Department of Environmental Science, Northwestern University, Evanston, IL, USA
g
Department of Chemical and Biological Engineering, Northwestern University, Evanston, IL, USA
h
Center for Engineering Sustainability and Resilience, Northwestern University, Evanston, IL, USA
i
Northwestern-Argonne Institute of Science and Engineering, Evanston, IL, USA

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

Keywords: Wetlands serve many important ecosystem services, yet the United States lacks up-to-date, high-resolution
Machine learning wetland inventories. New, automated techniques for developing wetland segmentation maps from high-
Wetland resolution aerial imagery can improve our understanding of the location and amount of wetlands. We assem­
High-resolution aerial imagery
bled training and testing data sets (patch sizes of 28 × 28 m2 and 56 × 56 m2) of high-resolution aerial imagery
Segmentation
Neural networks
of wetlands using Illinois Natural History Survey wetland location data and National Agricultural Imagery
Project data. Each patch was labeled as wetland or non-wetland. To augment these data sets with additional
information, we incorporated digital surface and digital terrain models and topographic wetness index data in
the same two patch sizes. Subsequently, we evaluated convolutional neural network (CNN) and Gaussian
process-based machine learning methods to produce wetland segmentation maps. We developed the best per­
forming method into a new CNN algorithm, WetSegNet. It exhibited an area under the curve of 98% when used
with 56 × 56 m2 patch sizes. WetSegNet developed reliable wetland segmentation maps in test cases in which
wetlands would have gone undetected using only the National Land Cover Database. The development of
WetSegNet exemplifies the types of data sets and methods that are needed to accelerate the use of high-resolution
aerial imagery towards an improved understanding of wetlands. This algorithm could be used by state and
federal agencies or other groups to identify wetlands with higher accuracy and at a finer scale than previously
possible.

1. Introduction (CDL), the National Land Cover Database (NLCD), and the National
Wetlands Inventory (NWI) (United States Environmental Protection
In the United States, the Renewable Fuel Standard, which aims to Agency, 2018; U.S. Fish and Wildlife Service, 2020; U.S. Geological
spur the adoption of biofuels, has caused controversy regarding the Survey, 2016). These data sets have also been the basis of academic
amount of uncultivated grasslands, wetlands, and forests (“natural studies of agricultural land expansion in the United States. For example,
lands”) that would become corn or soy farms to produce biofuels. To Alemu et al. (2020) used the CDL to assess agricultural expansion in the
evaluate agricultural land expansion in the United States each year, the ecologically important Prairie Pothole Region, which spans portions of
U.S. Environmental Protection Agency relies on several data sets, the Dakotas and Minnesota and reaches into Canada. They estimated
including the U.S. Department of Agriculture’s Cropland Data Layer agricultural expansion into grassland and wetlands in this region.

* Corresponding author at: Department of Chemical and Biological Engineering, Northwestern University, Evanston, IL, USA.
E-mail address: jennifer.dunn1@northwestern.edu (J.B. Dunn).

https://doi.org/10.1016/j.jag.2021.102581
Received 8 February 2021; Received in revised form 30 September 2021; Accepted 9 October 2021
Available online 24 October 2021
1569-8432/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

Similarly, Johnston (2013) integrated data from the NWI, the CDL, and substantially when LiDAR data were combined with multispectral Earth
the NLCD to characterize wetland loss from agricultural expansion in the observation data (86.5%), as opposed to a classifier using solely multi­
Prairie Pothole Region within the Dakotas. Finally, Lark et al. (2015) spectral imagery (29–59%). Using the Google Earth Engine cloud
and Wright et al. (2017) have undertaken CDL-based studies aiming to calculation to classify land cover types, Huang et al. (2017) used Landsat
quantify agricultural expansion in the United States. The coarse scale of imagery and corresponding Normalized Difference Vegetation Index
the CDL and NLCD (30 m) and the age of NWI data (founded on imagery data to understand major land dynamic changes resulting from urban­
collected in the 1980 s) limit the ability to quantify areas of natural land ization and afforestation in Beijing, including wetland loss and conver­
– small wetlands in particular - with certainty (Copenhaver et al., 2021). sion. The classification map they developed had an overall accuracy of
The presence of small wetlands is important to capture because they 86%. By combining multiple high-resolution remote sensing and
offer important ecosystem services (Christensen et al., 2016; Matthews geographic information system data sets, such as NWI, DEM, and the Soil
et al., 2016; Van Meter and Basu, 2015) but will not be captured by data Survey Geographic Database, Kloiber et al. (2015) developed a data
sets such as the NLCD that are at a 30 m resolution. fusion update for an inventory of a Minnesota region’s wetlands.
To improve the accuracy and efficiency of agricultural land expan­ Despite progress in the application of ML methods to high-resolution
sion assessments, new techniques are needed that leverage aerial and imagery analysis for wetland segmentation, important gaps remain. One
satellite imagery to detect the presence of ecologically important lands foundation that makes ML methods possible is high-resolution, wetland-
including wetlands (Maxwell et al., 2019). Wetlands offer many specific testing and training data sets. Such data sets are not widely
ecosystem services including flood abatement, water quality improve­ available. Additionally, no studies have compared the performance of
ment and biodiversity support yet they are globally threatened (Zedler, different ML and data fusion approaches. Furthermore, no previous ML
2003). Imagery data sources for this purpose include the National approach generated segmentation maps at high resolution (1 m), which
Agricultural Imagery Project (NAIP) (1 m resolution), RapidEye (5 m), allows detection of very small wetlands.
Landsat (30 m), and the Moderate Resolution Imaging Spectroradi­ We report herein several advances we have made to address these
ometer (250–1000 m). Previous studies have used these data to develop gaps in data availability and methods development for wetland detec­
algorithms for mapping wetlands and wetland dynamics, but the reso­ tion in high-resolution aerial imagery. First, we developed a unique
lution of these studies is 30 m, which may be too coarse to detect testing and training data set of Illinois wetlands that can be used in ML-
ecologically important small wetlands. With ground-truthed imagery based algorithms for wetland detection. We used this data set to evaluate
from NAIP and Landsat and subsequent image classification and vali­ CNN- and Gaussian Process (GP)-based methods for segmentation of
dation, Xie et al. (2015) mapped wetlands (overall accuracy 89%, 30 m wetlands using high-resolution (1 m) NAIP data. Seeing a need for al­
resolution) and invasive species using Landsat imagery with NAIP im­ gorithms that could improve the performance of existing methods, we
agery as ground reference data. Pekel et al. (2016) used Landsat images developed a new ML architecture to generate wetland segmentation
to quantify changes in wetland area at a 30 m resolution and developed maps. The high resolution of these segmentations is particularly critical
the European Commission’s Joint Research Centre Global Surface Water given the importance of capturing the presence and extent of small
products. Wu et al. (2019) developed an algorithm that analyzes NAIP wetlands. We also explored the effect of variables, such as image patch
and Google Earth Engine imagery to map wetlands (30 m resolution) in size and fusion of NAIP data with geophysical data sets including DEM
the Prairie Pothole Region and evaluated wetland inundation dynamics. and TWI, on the accuracy of wetland segmentation. Our efforts focused
Wu et al. determined that their method often detected small, sub-hectare on the State of Illinois, with the goal of developing a general method that
wetland features that the Joint Research Commission’s Monthly Water could be expanded to other regions, because the Illinois Natural History
History project did not. It is clear that efficient methods for detecting Survey (INHS) has produced a unique wetland location dataset that we
wetlands of this size are important to protect them. used to train and test the ML algorithms we developed.
Several ML-based techniques have been proposed to develop seg­
mentation of land cover (e.g., cropland versus natural lands), in general, 2. Materials and methods
with both aerial imagery and remote sensing products, including the
CDL and NLCD. Deep learning, especially, has shown promise for these The study area included the entire State of Illinois, USA. Climate in
tasks. In their review paper, Ma et al. (2019) mention heavy use of Illinois is continental, with cold winters and warm, humid summers.
recurrent neural networks, deep belief networks, convolutional neural Average annual precipitation varies from 1,220 mm in the south to 890
networks (CNNs), autoencoders, and generative adversarial networks mm in the north (Illinois State Climatologist’s Office, 2021). The
for image fusion, image registration, scene classification, object detec­ topography is relatively flat, with elevation in the state ranging from 85
tion, land use and land cover classification, segmentation, and object- to 376 m. Several types of wetlands exist within Illinois, including fens
based image analysis. Specific examples include the application of and bogs in the northeastern part of the state, and herbaceous-
deep learning such as CNNs on multispectral data to characterize land dominated marshes and wet meadows, ponds with floating-leaved and
cover. Additionally, IrrMapper (Ketchum et al., 2020) is an ML model for submerged vegetation, and forested wetlands bordering streams
mapping irrigated agricultural lands; it provides 90.8% accuracy for throughout the state (Mohlenbrock, 2014). Wetlands are estimated to
land cover mapping when compared to Census of Agriculture irrigation have once covered approximately 23% of Illinois, but up to 90% of this
estimates based on Landsat imagery using Random Forests. Maxwell wetland area has been destroyed due to drainage and land-use change
et al. (2019) used geographic object-based image analysis and Random (Suloway and Hubbell, 1994). The next three subsections describe how
Forest-based classification on NAIP data for general land cover mapping we developed training and testing data sets, trained the model, and then
and achieved an accuracy of 96.7%. used it to segment wetlands in NAIP data. Fig. 1 provides a flow chart of
The effectiveness of ML methods for image analysis can be enhanced our methodology.
through integration of additional data including LiDAR-derived Digital
Elevation Models (DEM), the Normalized Difference Vegetation Index, 2.1. Data sets
the Soil Survey Geographic Database, and the Topographic Wetness
Index (TWI). With such fusion approaches, a robust classifier is devel­ To segment wetlands in Illinois in high-resolution aerial imagery, we
oped by integrating multiple sources of imagery and remote sensing created unique tailored training and testing data sets for wetlands in the
products to inform the classification process. For example, Wu et al.’s state. Prior data sets (Basu et al., 2015) were used with classical machine
wetland inundation dynamics-mapping algorithm (Wu et al., 2019) in­ learning approaches (Deep Belief Networks and RF) and were smaller,
tegrates LiDAR and NWI data. Rapinel et al. (2015) concluded that having few instances, not being suited for training DL models without
wetland habitat mapping and classification accuracies improved overfitting. They also used imagery acquired from across the United

2
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

Fig. 1. Methodology for developing testing and training data sets, training and evaluating the model, and generating segmentation maps. Green rounded rectangles
indicate data. Blue rectangles indicate a process or calculation. Red hexagons indicate output. (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)

States without geotagging. It is therefore not possible to know whether Using the INHS wetland location data, we extracted 28 × 28 m2 and
imagery from Illinois was included. Given regional wetland variation, 56 × 56 m2 patches from images of Illinois at a 1-m ground sample space
inclusion of some regionally specific data, i.e., imagery that is known to resolution from 2010, 2012, 2014, 2015, and 2017, from images that
be from Illinois, is required for generation of accurate results. were taken between June and September of each year (Table 1). This
Our training and testing data sets were based on Illinois wetland was done to ensure that the maximum extent of wetlands was repre­
location data from the Illinois Natural History Survey (INHS), collected sented. Notably, this time period represents the peak of the growing
on-site between 2002 and 2018 from environmental survey corridors season in Illinois, when wetland vegetation is most clearly identifiable
where Illinois Department of Transportation projects were slated to on imagery. The June-September time period also largely overlaps the
occur (Matthews et al., 2016). All wetlands within these corridors are season of INHS field data collection. Hard copy NAIP data were pur­
identified and delineated following U.S. Army Corps of Engineers chased directly from the United States Department of Agriculture Aerial
methodology. At each wetland, the observer walked the edge of the Photography Field Office. While NAIP data is freely available online,
wetland with a GPS unit to capture the extent of the wetland as a artefacts can be introduced during image compression, and we found
polygon. that the hard-copy data had fewer artifacts and visual imperfections. It
To be considered a wetland, positive evidence is required for the should be noted that overall, imagery varied in quality from 2010 to
presence of hydrophytic vegetation, wetland hydrology, and hydric soils 2017 (Fig. 2).
(Environmental Laboratory, 1987; U.S. Army Corps of Engineers, 2010). Digital surface model (DSM), digital terrain model (DTM), and TWI
INHS data span 95 counties and approximately 2,600 ha. Wetlands data were used to supplement the imagery in our analysis. The DSM and
include ponds with submerged or floating vegetation, marshes, wet DTM are types of elevation data generated from LiDAR point cloud data
meadows, wet shrublands, and forested wetlands. While the data also collected via aerial fly-overs. The DSM data capture all surface features,
include information about wetland class (Cowardin et al. 1979), we did both man-made and natural. The DTM data (used to calculate the TWI)
not account for these differences in our analysis because wetland class is were processed to remove these surface features, leaving behind only
not required for evaluation of agricultural expansion to meet the stat­ natural terrain features, similar to a bare-Earth surface. Both DSM and
utory requirements of the Renewable Fuel Standard. The INHS data set DTM can improve the detection of forested wetlands that are not
offers a higher resolution (1 m) compared to the NLCD data set (30 m) distinguishable in the NAIP data. All DTM and DSM data are collected by
and is more recent than the NWI data, which was originally delineated county and publicly available online from the Illinois Natural Resources
manually from 1980 s-era aerial imagery and is only partially updated Geospatial Data Clearinghouse (Illinois State Geological Survey Prairie
by the U.S. Fish and Wildlife Service as it receives data from partner Research Institute, 2020). Temporal resolution varies by county. While a
institutions and state governments. Although NWI data are the most statewide image server is available for accessing the Illinois Clearing­
widely available wetland maps in Illinois and much of the U.S., the data house data, the county-level data are merged, upscaled, and downscaled
contain inaccuracies (Matthews et al., 2016; Stolt and Baker, 1995). In using an unspecified method. To avoid over-processing the data and to
Illinois, many areas mapped as forested wetlands on NWI maps are often ensure that we used the highest resolution data, we downloaded each
not truly wetlands, and many small (~ <0.06 ha) wetlands are omitted county’s data individually instead of using the statewide image server.
(Matthews et al., 2016). TWI data were derived from the Illinois Clearinghouse DTM data
Notably, the INHS data is unique as a ground-truth wetland-delin­ using a public ArcPy script (Wolf and Fricker, 2013). TWI is a measure of
eating data set. While various state environmental or natural resource how ‘wet’ a certain point is, based on water flow direction, flow accu­
departments in the Midwest may possess similar data sets, these often mulation, slope, and other features of surrounding elevation data (Bal­
contain personal identifiable information that prevents data sharing. lerine, 2017). We expect that areas with high wetness indices exhibit
Other wetland delineation projects are conducted privately, often water accumulation and are more likely to be wetlands.
commissioned as part of environmental impact reports associated with From the DSM, DTM, and TWI data, we extracted 28 × 28 m2 and 56
new construction projects, and are thus of limited scale. × 56 m2 patches to correspond with the NAIP patches. The 28 × 28 m2

3
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

Table 1
Raw data sources for creating training and testing data sets.
Data Type Source Dates Bands Resolution pixel/ Total covered
m2 area

Aerial NAIP 2010, 2012, 2014, 2015, 2017(all June- Red, Green, Blue (RGB), Near Infrared 1 149,998 km2
Imagery September) (4)
DSM and Illinois Clearinghouse 2008–2018 (varies by county) 2
DTM
TWI Derived from DEM 1
data

Fig. 2. Example of change in imagery quality from NAIP in 2010, 2014, and 2017 (left to right).

size was selected following previous studies in automatic wetland features of the area included in the patch. Moreover, using only a subset
detection (Basu et al., 2015). We included the 56 × 56 m2 size to test the of the DTM data also significantly reduced overall processing time. After
effect of providing more context information to the algorithm. Table 1 all DSM, DTM, and TWI patches were extracted, some of them did not
summarizes key features of the data sources we used. The next section have reliable information (due, for example, to cloud cover), and were
describes the patch extraction process in detail. therefore removed from the dataset, bringing the number of patches
from 122,383 to 116,380.
To build a data set of non-wetland images, we extracted 120,000
2.2. Patch extraction: Data set generation
random patches from Will County, Illinois from NAIP, DTM, DSM, and
TWI data. The same extraction processes and patch sizes used for the
All of the 28 × 28 and 56 × 56 m2 patches from NAIP and the de­
wetland data were performed. To minimize wetland patches in the non-
rivative DSM, DTM, and TWI data were extracted using various ArcPy
wetland data, the INHS data regions were excluded from the possible
Python scripts that we developed. INHS data were first preprocessed to
extraction regions within Will County. Areas classified as wetlands in the
yield a 28 × 28 m2 cell, whose center was computed. For each center, a
NWI data were also excluded.
28 × 28 m2 and 56 × 56 m2 patch were extracted for each year. Because
The resulting training data set contained 189,117 patches, 93,105 of
of the 28 × 28 m2 cell, 56 × 56 m2 have overlap. The overlap within the
which were wetlands and 96,012 were non-wetlands. The testing data
larger patches did not influence our analysis because we only classified
set contained 47,263 patches, 23,275 of which were wetlands and
the center pixel of each patch. All 28 × 28 m2 NAIP patches were
23,988 were non-wetlands. The split between training and test was
manually reviewed to determine whether they contained wetlands. The
performed randomly, selecting 80% of the samples for training and the
30,222 patches that did not contain wetland pixels in the 28 × 28 m2
remaining 20% for testing.
patches were excluded from both the 28x28 m2 and the 56 × 56 m2
patch data set. This step was important because since the INHS data
were collected between the years 2002 and 2018 while NAIP imagery 2.3. Development of ML algorithms
was collected in 2010, 2012, 2014, 2015 and 2017, man-made con­
struction could have destroyed a fraction of the INHS-recorded wet­ We developed ML algorithms for wetland segmentation by fusing
lands. Across all five years of NAIP data, we collected 122,383 patches of high-resolution NAIP data with additional data from DTM, DSM, and
imagery for each of the two patch sizes. TWI. We investigated several classification algorithms and supplemental
For the derivative data, extraction followed a similar methodology data sources. The training and testing of the classifiers were based on 28
based on the center of each INHS data patch, after correcting for dif­ × 28 m2 and 56 × 56 m2 image patches. During training, each patch
ferences in county-level resolution in the Clearinghouse Illinois DTM carried a binary label of wetland or non-wetland.
and DSM data. DTM and DSM data were first geometrically aligned and We compare the performance of four state-of-the-art ML algorithms.
either upscaled or downscaled to 1-m resolution to match that of the Two ML algorithms were based on GPs and two were based on deep
NAIP data. Then the TWI data were calculated using the previously neural network architectures.
mentioned public ArcPy script. Performing these calculations after the Using GPs required the extraction of a set of hand-crafted features
DTM was resized to a common resolution instead of performing them on from each patch. Our first GP-based approach used the 47 features
the DTM data at its native resolution was essential for keeping the TWI proposed in a study using NAIP (Basu et al., 2015). We incorporated an
patches consistent across all counties. The TWI was calculated from a additional four features calculated from the mean of each of the TWI,
subset of the DTM data limited to a 250-m radius around each patch DSM, DTM, and DSM-DTM patches. We incorporated these additional
center. A 250-m radius was sufficiently large to reveal prominent features because we found out that using NAIP data alone would lead to

4
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

incorrect classifications in some instances, for example, for patches with decrease by at least 10-3 during five epochs, then the learning rate was
wetlands hidden in non-wetland wooded areas. We call this model Basu- divided by 10. The training ends when the learning rate reaches 10-6. To
GP. The second GP model, that we named VGG16-GP, use the features avoid overfitting, we used weight decay with a value of 10-4 and Dropout
extracted by the last convolutional layer of VGG16 (Simonyan and Zis­ (23) with 0.1 probability before each fully connected layer. We used the
serman, 2014) for the RGB, near infrared, TWI, and DSM-DTM data, Binary Cross Entropy (BCE) loss function, that is,
resulting in a total of 2,048 features (512 for each type of data). For both
BCE(y, p) = − (ylog(p) + (1− y)log(1 − p) ),
GP models, we used the Scalable Variational Gaussian Process for
Classification proposed in (Hensman et al., 2015) with a Radial Bias where y is a binary label with values 0 or 1 and p is the predicted
Function kernel with Automatic Relevance Determination, 100 inducing probability. In the last step, we augmented the training patches by
points, and a minibatch size of 1,000 samples. The Adam optimizer performing random vertical and horizontal flipping and 90◦ rotations.
(Kingma and Ba, 2014) was used to estimate the parameters of the Finally, we investigated an existing pretrained CNN model, VGG16
model, setting the learning rate to 0.005 and the number of iterations to (Simonyan and Zisserman, 2014), to serve as the baseline. We retrained
10,000. its last fully connected layers utilizing our data, the BCE loss function,
The third ML model we designed and implemented was a CNN, and the same optimization parameters as the WetSegNet architecture.
which requires large amounts of training data. CNNs perform feature We call this model VGG-Finetuned.
extraction and therefore produce a classification from the raw pixels of
each patch. Fig. 3 depicts the architecture of the CNN we developed, a 2.4. Generate segmentation maps
residual CNN based on ResNet (He et al., 2015) and referred to as
WetSegNet. One of the building blocks of the proposed architecture is Given a trained classifier, during testing, the methodology for
the residual block structure proposed in (Zagoruyko and Komodakis, obtaining a segmentation map of an image indicating the location of the
2016), which was found to provide better results for classification than wetlands included the following steps:
the original ResNet. In addition, the model was significantly downscaled
from the original 54 layers to only 28 layers to reduce its capacity and to (1) Using a sliding window, patches of either 28 × 28 m2 or 56 × 56
ensure effective training, i.e., avoid overfitting. It was evaluated that for m2 were extracted.
the problem at hand the reduced network capacity was adequate to (2) For each patch, the trained classifier calculated the probability of
provide very good performance. The use of the residual design signifi­ the center of the patch being a wetland. Doing this for all patches
cantly eased the training of the network, especially for the first layers produces a probability map with each pixel having assigned a
(see (He et al., 2015)). probability of being part of a wetland.
With the WetSegNet model, we experimented with patches of both (3) Because the patch dimensions are even numbers, with no center
sizes (28 × 28 m2 and 56 × 56 m2) and investigated the effect on the pixel, a probability map of a grid shifted by 0.5 pixels horizon­
performance of the architecture when fusing the DSM, DTM, and TWI tally and vertically was created. The probability map was then
data. Because CNNs learn to perform feature extraction from the data, registered to the original grid through bilinear interpolation.
adapting them to patches of different sizes is straightforward. In per­ (4) Finally, a threshold was applied to the probability map to classify
forming fusion of the DSM, DTM, and TWI data, we chose early fusion each pixel as wetland or non-wetland, providing a segmentation
and concatenation of the color and near infrared channels of the NAIP of the wetlands in the image. The threshold was selected exper­
image, as all of the data were image-type. In contrast, GP and similar imentally, and we found that a threshold value equal to 0.8
traditional ML classifiers (like RF and Support Vector Machines) are not produced the best results.
able to perform automatic feature extraction, and therefore the design
and extraction of new hand-crafted features was required each time we This methodology can be used with any ML classification algorithm,
wanted to fuse more data sources. As already mentioned, the incorpo­ which allowed us to the use of the four ML algorithm previously pre­
ration of elevation data allowed the algorithm to detect wetlands in sented. Of note, the sliding window mechanism for producing segmen­
those cases where the surface of the terrain was not visible (e.g., tation maps using patch-based classifiers has been described previously
woodlands). (Glumov et al., 1995; Liang et al., 2012).
WetSegNet was trained using the Adam optimizer (Kingma and Ba,
2014), setting the initial learning rate to 10-3. If the training error did not

Fig. 3. Architecture of the proposed WetSegNet CNN. All convolutional layers use a 3x3 kernel. The symbol “/2” indicates a stride 2 convolution, i.e., a convolution
that results in dividing the spatial resolution of the input by 2. If the input is a 28x28 m2 patch, the last average polling (AVG Pool) layer reduces each spatial
dimension by a factor of 3 instead of 7. WRB, wide residual block.

5
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

3. Results the features of VGG16 with the GP, VGG16-GP, the AUC increased to
93.2%. WetSegNet without the fusion of DSM, DTM, and TWI data
We evaluated and compared the performance of the four ML models performed better than the GP methods, with an AUC of 96.1% for the 28
using our data set, based on Area Under the Receiver Operating Char­ × 28 m2 patch size and 96.7% for the 56 × 56 m2 patch size (WetSegNet-
acteristic Curve (AUC), accuracy, precision, and recall. Table 2 de­ RGBI-28 m and WetSegNet-RGBI-56 m, respectively).
lineates the results by ML model, patch size, and inclusion of DSM, DTM, In Fig. 5, we compare results with each model in segmenting wetland
and TWI data. Figs. 4 through 7 illustrate how results obtained with the in a farm field that abuts a road. The GP methods (blue outlines) missed
various models in Table 2 compared to the ground truth data from INHS. a good portion of the wetland area and exhibited some false positive
The Supporting Information contains NLCD results for the images shown results. Although VGG16-GP also uses features extracted using a CNN,
in Figs. 4–7. training VGG16 with natural images (29) reduced the effectiveness of
the features the network learned from aerial images. On the other hand,
WetSegNet-RGBI-TWI-28 m (Fig. 5c) segmented all wetlands as
3.1. Ablation study of WetSegNet: Influence of patch size and fusion of
confirmed in the ground truth data (red outline) and segmented addi­
data sources
tional area as wetland. Performance of this model was assisted by the
inclusion of TWI data (Fig. 5d) where thicker white lines indicate greater
The base model, WetSegNet-RGBI-28 m, did not include elevation
wetness. The algorithm was better able to segment wetlands with the
data and employed the 28 × 28 m2 patch size. It produced results with
TWI information incorporated. The additional area the algorithm
an AUC of 96.1%. We hypothesized that introducing more information
segmented as wetland was visually indistinguishable from the area
for each pixel with the larger patch size would reduce uncertainty. Thus,
identified as wetland in the ground truth data. We hypothesize that false
we tested the model WetSegNet-RGBI-56 m, which uses the larger patch
positive results arise because the ground truth wetland is defined based
size of 56 × 56 m2. With this change, the AUC increased to 96.7%.
on U.S. Army Corps of Engineers criteria, which include the presence of
Introducing the TWI data as a feature improved the AUC further to
hydric soils based on field measurements, the presence of hydrophytic
97.4% when the patch size was 28 × 28 m2 and 98.2% when the patch
vegetation, and evidence of wetland hydrology. Our algorithm is not
size was 56 × 56 m2 (WetSegNet-RGBI-TWI-28 m and WetSegNet-RGBI-
able to evaluate these three characteristics, thus resulting in a false
TWI-56 m, respectively). This analysis demonstrates the importance of
positive classification of an area. The NLCD (U.S. Geological Survey,
terrain information in segmenting wetlands, especially in cases when the
2016) indicates no wetlands in the patch shown in Fig. 5, indicating that
ground is not clearly visible in the image data. We also investigated the
our ML-based approaches for locating wetlands offer improvements over
case in which all TWI, DSM, and DTM were used as input to the network.
the NLCD.
Although TWI is derived from DSM and DTM this is a non-invertible
Finally, we tested the use of the features extracted by our best per­
operation and therefore some information is lost in this derivation.
forming model, WetSegNet-RGBI-TWI-56 m, to train a GP classifier in a
However, including TWI, DSM, and DTM data yielded no observable
similar fashion to VGG16-GP. This approach did not yield better results
performance increase (AUC = 97.3% and AUC = 98.1% for patch sizes of
than the CNN alone, which is consistent with the findings of previous
28 × 28 m2 [WetSegNEt-RGBI-TWI-DEM-28 m] and 56 × 56 m2 [Wet­
research (Wilson et al., 2016). By training end-to-end, CNNs project the
SegNet-RGBI-TWI-DEM-56 m], respectively).
data into a space where the classes are linearly separable. Thus, the use
Fig. 4 shows a comparison of WetSegNet-RGBI-28 m, WetSegNet-
of non-linear classifiers like GPs does not significantly improve the
RGBI-56 m, WetSegNet-RGBI-28 m-TWI, and WetSegNet-RGBI-56 m-
results.
TWI. Each new modification to the input data improved the segmenta­
We created examples of segmentation maps calculated by our best
tion map obtained by the model, with the exception of a slight increase
performing model, WetSegNet-RGBI-TWI-56 m, in different scenarios
in the size of the segmented areas caused by the use of bigger patches.
and consider the underlying explanations for model performance. In
The results in Fig. 4b and d, when compared to those in Fig. 4a and c,
Figs. 6 and 7, all segmentations in the test cases are close to the ground
reveal that a larger patch size improved performance as the algorithm
truth. Sometimes, information provided to the algorithm from NAIP
was provided with more information. In this instance, the network had
imagery and the TWI data conflict. An example of this occurrence is
no false positives with 56 × 56 m2 patches, but did when the patch size
shown in Fig. 6a-b. In that case, the TWI indicates wetness in the middle
was 28 × 28 m2.
of the highway that the imagery clearly shows is present and dry. The
algorithm correctly indicates the presence of the highway, relying more
3.2. Comparison of results from different ML models on the information from the imagery. It did, however, somewhat un­
derestimate the wetland area. For the test case in Fig. 6a and 6b, NLCD
Basu-GP resulted in the lowest accuracy of all methods assessed data indicated that the wetland is only present below the highway. The
(AUC = 92.4%). Fine-tuning VGG16 (Simonyan and Zisserman, 2014), NLCD (see Supplementary Information) indicated that the area depicted
VGG-Finetuned, achieved similar results (AUC = 92.4%). When we used in Fig. 6c and 6d was entirely cultivated crops with no wetland present.
For these test cases, the ML-based model outperformed the NLCD.
Table 2 In several instances that we considered, false positive wetlands in
Comparison of the proposed models for wetland segmentation on our testing wooded areas were likely due to occlusion in areas where the model
data set. We compared all models in terms of Area Under the ROC Curve (AUC), can’t “see” the ground. This effect is evident in the image “Forest” in
accuracy, precision, and recall. Bold indicates the highest value in each measure. Fig. 7a. In this case, the method wrongly classified the tree-covered area
See text for details on the models used. All results reported as percents. surrounding the wetlands as wetlands, potentially because this area had
Model AUC Accuracy Precision Recall a strong response in the TWI (Fig. 7d). The NLCD (Supplementary In­
Basu-GP 92.35 81.59 80.94 82.27 formation) did report the presence of forested and emergent wetlands in
VGG16-GP 93.23 86.43 83.76 85.31 this example. With 56 m2 patch sizes in Fig. 7c and d, WetSegNet no
VGG16-Finetuned 92.43 85.25 81.01 83.74 longer identified two false positive wetlands as was the case with 28 m2
WetSegNet-RGBI-28 m 96.14 89.15 89.16 88.89 patch sizes for the same test case in Fig. 5c and d.
WetSegNet-RGBI-56 m 96.69 90.48 91.08 89.43
WetSegNet-RGBI-TWI-28 m 97.39 91.87 90.51 93.32
WetSegNet-RGBI-TWI-56 m 98.21 93.43 92.41 94.40 4. Discussion
WetSegNet-GP-RGBI-TWI-56 m 98.20 93.44 92.38 94.35
WetSegNet-RGBI-TWI-DEM-28 m 97.33 91.83 90.40 93.24 In this study, we performed a comprehensive comparison of several
WetSegNet-RGBI-TWI-DEM-56 m 98.09 92.89 91.93 94.12
ML approaches for generating wetland segmentation maps. We also

6
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

Fig. 4. Visual comparison of the segmentation maps obtained by (a) WetSegNet-RGBI-28 m, (b) WetSegNet-RGBI-56 m, (c) WetSegNet-RGBI-28 m-TWI, and (d)
WetSegNet-RGBI-56 m-TWI of the test image “Urban.” Blue color indicates the area segmented by each model and red indicates the ground truth (INHS). (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

developed a new architecture, WetSegNet, which outperformed all other could supplement existing wetland inventories.
methods we investigated. The approach we used in developing Wet­ For our analysis, we trained the models with a new data set that
SegNet is unique because we applied ML directly to raw pixels to contains NAIP patches of 28 × 28 m2 and 56 × 56 m2 along with TWI,
generate segmentation maps at high resolution (1 m). DTM, and DEM data for the same patches. The data set also contains
WetSegNet achieved high accuracy and performed well in most test ground-truth training data from the INHS. This unique data set is now
cases, especially with larger patch sizes and the inclusion of TWI data. publicly available (Katsaggelos et al., 2021) and may be helpful for re­
Compared to the other models studied, WetSegNet significantly out­ searchers using ML to interpret aerial imagery.
performed GP models using both hand-crafted features and features There are several challenges and sources of error with the data set
extracted from a widely used, pretrained network (VGG16). WetSegNet and models that we developed. First, the shape and location of some of
also outperformed the VGG16 network after fine tuning with our data. the wetlands in the INHS dataset were recorded years before the NAIP
Furthermore, WetSegNet detected wetlands that were undetected in the data was taken. Thus, some of the wetlands may have changed or dis­
NLCD. Information about wetland location and size from WetSegNet appeared. We worked to mitigate this issue by manually reviewing the

7
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

Fig. 5. Visual comparison of the segmentation maps obtained by (a) Basu-GP, (b) VGG16-GP and (c) WetSegNet-RGBI-TWI-28 m of the test image “Rural” and (d) by
the WetSegNet-RGBI-TWI-28 m of the corresponding TWI image. Blue color indicates the area segmented by each model and red indicates the ground truth (INHS).
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

training data set and excluding questionable data and label pairs. If some much smaller version of the proposed network, using only half the
patches remained that were not representative of wetlands, some noise number of filters in all convolutional layers. We showed an increase in
could have been introduced to the data. Although these errors should AUC from 97.1% using images from all years to 98.0 using only images
not be significant for the evaluation of the models for segmenting wet­ from the most recent year (2017). Owing to our use of imagery from
lands, they can reduce the quality of the segmentation maps produced, multiple years (including wet years, dry years, and years with varying
especially in boundary areas, which are already very difficult to delin­ image quality), the algorithm is “future proof”. That is, it can segment
eate. A second issue was that the quality of the NAIP data changed over wetlands for any year in which NAIP data are available.
time, which can have a significant impact on the quality of the results. Although the specific results of this paper should not be extrapolated
Imagery from recent years had a higher quality (Fig. 2) and thus pro­ beyond the range of the training data set within the State of Illinois, the
vided the model with more precise information. We do not have enough methodology is applicable anywhere with appropriate imagery and
images from the most recent years in our series to fully train our model ground-truthed data. The characteristics of the INHS dataset, which
on higher quality images, but we did perform an experiment using a contains wetlands that vary greatly in vegetation cover type, from

8
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

Fig. 6. Segmentation maps obtained by WetSegNet-RGBI-56 m-TWI of the test images (a) “Marsh” and (b) its corresponding TWI image, and (c) “Farm” and (d) its
corresponding TWI image. Blue color indicates the area segmented by the model and red indicates the ground truth (INHS data). (For interpretation of the references
to color in this figure legend, the reader is referred to the web version of this article.)

ponds, to emergent marshes and wet meadows, shrub-dominated wet­ from ground truth data, offers an advance over state-of-the-art methods
lands and forested wetlands support this expectation. for wetland detection, which are typically at 30-m resolution. In
contrast, WetSegNet can be used to monitor wetlands down to the 1-m
5. Conclusion scale. Fine resolution is necessary for identifying small wetlands,
which wetland inventories often fail to detect (Christensen et al., 2016;
Improving methods to develop wetland segmentation maps that are Matthews et al., 2016; Van Meter and Basu, 2015). The presence of small
automated and make use of the growing amounts of high-resolution wetlands in a landscape is important for waterfowl that use multiple
aerial imagery will enable scientists, policy makers, and other stake­ wetlands and amphibians that rely on small wetlands for breeding
holders to better track wetlands. For example, as the United States (Naugle et al., 2000; Semlitsch and Bodie, 1998; Snodgrass et al., 2000).
considers the role of its land base in providing energy, especially to grow Additionally, fine spatial resolution is essential for detecting incremen­
biofuel feedstocks, the ability to monitor wetlands with confidence can tal change at wetland boundaries due to encroaching land use and
help alleviate concerns over wetland loss while pursuing policy objec­ altered precipitation due to climate change. Given the importance of
tives that can offer other environmental benefits such as low-carbon mapping wetlands at fine spatial resolution, as additional ground truth
fuels. WetSegNet, paired with training and testing data sets derived data allow, we will test and refine our method in other regions and apply

9
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

Fig. 7. Segmentation maps obtained by WetSegNet-RGBI-56 m-TWI of test images (a) “Forest” and (b) its corresponding TWI image, and (c) test image “Rural” and
(d) its corresponding TWI image. Blue color indicates the area segmented by the model and red indicates the ground truth (INHS data). (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of this article.)

our methods to quantify changes in wetland area. Finally, we plan to curation, Formal analysis, Writing – original draft. Jeffrey Matthews:
apply a similar technique to detection of natural grasslands, another Resources, Data curation, Writing – original draft, Writing – review &
type of natural land that is difficult to characterize based on existing editing. Bradley Zercher: Resources, Data curation, Writing – original
data sets with frequent errors (Dunn et al., 2017). draft, Writing – review & editing. Liliana Sydorenko: Investigation,
Writing – original draft. Neelanshi Varia: Investigation, Formal anal­
CRediT authorship contribution statement ysis, Writing – original draft. Yuanzhe Jin: Formal analysis, Writing –
original draft. Minzi Wang: Investigation, Data curation. Jennifer B.
Santiago López-Tapia: Methodology, Formal analysis, Writing – Dunn: Conceptualization, Investigation, Writing – original draft,
original draft, Visualization. Pablo Ruiz: Conceptualization, Method­ Writing – review & editing, Supervision, Funding acquisition. Aggelos
ology, Formal analysis, Writing – original draft. Mitchell Smith: Data K. Katsaggelos: Conceptualization, Writing – original draft, Writing –

10
S. López-Tapia et al. International Journal of Applied Earth Observation and Geoinformation 105 (2021) 102581

review & editing, Supervision, Funding acquisition. Ketchum, D., Jencso, K., Maneta, M.P., Melton, F., Jones, M.O., Huntington, J., 2020.
IrrMapper: A machine learning approach for high resolution mapping of irrigated
agriculture across the Western U.S. Remote Sens. 12. Doi: 10.3390/rs12142328.
Declaration of Competing Interest Kingma, D.P., Ba, J., 2014. Adam: A Method for Stochastic Optimization. CoRR. http://
arxiv.org/abs/1412.6980.
Kloiber, S.M., Macleod, R.D., Smith, A.J., Knight, J.F., Huberty, B.J., 2015. A Semi-
The authors declare that they have no known competing financial Automated, Multi-Source Data Fusion Update of a Wetland Inventory for East-
interests or personal relationships that could have appeared to influence Central Minnesota, USA. Wetlands 35, 335–348. https://doi.org/10.1007/s13157-
014-0621-3.
the work reported in this paper. Lark, T.J., Meghan Salmon, J., Gibbs, H.K., 2015. Cropland expansion outpaces
agricultural and biofuel policies in the United States. Environ. Res. Lett. 10 https://
Acknowledgements doi.org/10.1088/1748-9326/10/4/044003.
Liang, F., Wang, D., Liu, Y., Jiang, Y., Tang, S., 2012. Fast Pedestrian Detection Based on
Sliding Window Filtering. In: Lin, W., Xu, D., Ho, A., Wu, J., He, Y., Cai, J.,
This research was supported by U.S. Department of Agriculture Na­ Kankanhalli, M., Sun, M.-.-T. (Eds.), Advances in Multimedia Information Processing
tional Institute of Food and Agriculture award 2018-10008-28530. – PCM 2012. Springer, Berlin Heidelberg, Berlin, Heidelberg, pp. 811–822.
Ma, L., Liu, Y., Zhang, X., Ye, Y., Yin, G., Johnson, B.A., 2019. Deep learning in remote
Mitchell Smith was also supported by Northwestern’s Murphey Fel­ sensing applications: A meta-analysis and review. ISPRS J. Photogramm. Remote
lows Program. We acknowledge Yikuan Li and Zhili Wang of North­ Sens. 152, 166–177. https://doi.org/10.1016/j.isprsjprs.2019.04.015.
western University, Michelle Wander and Nico Martin of the University Matthews, J.W., Skultety, D., Zercher, B., Ward, M.P., Benson, T.J., 2016. Field
Verification of Original and Updated National Wetlands Inventory Maps in three
of Illinois at Urbana-Champaign, Steffen Mueller of the University of Metropolitan Areas in Illinois, USA. Wetlands 36, 1155–1165. https://doi.org/
Illinois at Chicago, and Josh Pritsolas of GeoSpatial Mapping, Applica­ 10.1007/s13157-016-0836-6.
tions, and Research Center at Southern Illinois University Edwardsville Maxwell, A.E., Strager, M.P., Warner, T.A., Ramezan, C.A., Morgan, A.N., Pauley, C.E.,
2019. Large-Area, High Spatial Resolution Land Cover Mapping Using Random
for helpful discussions.
Forests, GEOBIA, and NAIP Orthophotography: Findings and Recommendations.
Remote Sens. 11, 1409. https://doi.org/10.3390/rs11121409.
Appendix A. Supplementary material Mohlenbrock, R.H., 2014. Vascular Flora of Illinois: A Field Guide, 4th edition. Southern
Illinois University Press, Carbondale.
Naugle, D.E., Johnson, R.R., Estey, M.E., Higgins, K.F., 2000. A landscape approach to
Supplementary data to this article can be found online at https://doi. conserving wetland bird habitat in the Prairie Pothole Region of eastern South
org/10.1016/j.jag.2021.102581. Dakota. Wetlands 20, 588–604. https://doi.org/10.1672/0277-5212(2000)020
[0588:ALATCW]2.0.CO;2.
Pekel, J.-F., Cottam, A., Gorelick, N., Belward, A.S., 2016. High-resolution mapping of
References global surface water and its long-term changes. Nature 540, 418–422. https://doi.
org/10.1038/nature20584.
Alemu, W.G., Henebry, G.M., Melesse, A.M., 2020. Land Cover and Land Use Change in Rapinel, S., Hubert-Moy, L., Clément, B., 2015. Combined use of LiDAR data and
the US Prairie Pothole Region Using the USDA Cropland Data Layer. Land 9, 166. multispectral earth observation imagery for wetland habitat mapping. Int. J. Appl.
https://doi.org/10.3390/land9050166. Earth Obs. Geoinf. 37, 56–64. https://doi.org/10.1016/j.jag.2014.09.002.
Ballerine, C., 2017. Topographic Wetness Index Urban Flooding Awareness Act Action Semlitsch, R.D., Bodie, J.R., 1998. Are Small, Isolated Wetlands Expendable? Conserv.
Support Will and DuPage Counties, Illinois Topographic Wetness Index Urban Biol. 12, 1129–1133.
Flooding Awareness Act Action Support Will & DuPage Counties, Illinois. Simonyan, K., Zisserman, A., 2014. Very Deep Convolutional Networks for Large-Scale
Basu, S., Ganguly, S., Mukhopadhyay, S., DiBiano, R., Karki, M., Nemani, R., 2015. Image Recognition. arXiv preprint arXiv:1409.1556.
DeepSat. In: Proceedings of the 23rd SIGSPATIAL International Conference on Snodgrass, J.W., Komoroski, M.J., Bryan, A.L., Burger, J., 2000. Relationships among
Advances in Geographic Information Systems - GIS ’15. ACM Press, New York, New Isolated Wetland Size, Hydroperiod, and Amphibian Species Richness: Implications
York, USA, pp. 1–10. Doi: 10.1145/2820783.2820816. for Wetland Regulations. Conserv. Biol. 14, 414–419.
Christensen, J., Nash, M., Chaloud, D., Pitchford, A., 2016. Spatial distributions of small Suloway, L., Hubbell, M., 1994. Wetland resources of Illinois: an analysis and atlas,
water body types in modified landscapes: lessons from Indiana, USA: Small Water Illinois Natural History Survey Special Publication Special Report 15. Champaign, IL,
Body Type Distributions. Ecohydrol. 9, 122–137. https://doi.org/10.1002/eco.1618. USA.
Copenhaver, K., Hamada, Y., Mueller, S., Dunn, J.B., 2021. Examining the Characteristics Stolt, M.H., Baker, J.C., 1995. Evaluation of national wetland inventory maps to
of the Cropland Data Layer in the Context of Estimating Land Cover Change. IJGI 10, inventory wetlands in the southern Blue Ridge of Virginia. Wetlands 15, 346–353.
281. https://doi.org/10.3390/ijgi10050281. https://doi.org/10.1007/BF03160889.
Cowardin, L.M., Carter, V., Golet, F.C., LaRoe, E.T., 1979. Classification of wetlands and United States Environmental Protection Agency, 2018. Biofuels and the Environment
deepwater habitats of the United States. FWS/OBS-79/31, Reprinted 1992, U.S. Fish Second Triennial Report to Congress.
and Wildlife Service, Washington, DC. United States Army Corps of Engineers, 2010. Regional supplement to the Corps of
Dunn, J.B., Merz, D., Copenhaver, K.L., Mueller, S., 2017. Measured extent of Engineers wetland delineation manual: Midwest Region (Version 2.0) (No. ERDC/EL
agricultural expansion depends on analysis technique: Ag Expansion Extent TR-10-16). United States Army Engineer Research and Development Center,
Dependence on Analysis Technique. Biofuels, Bioprod. Biorefin. 11, 247–257. Vicksburg, MS, USA.
https://doi.org/10.1002/bbb.1750. United States Fish and Wildlife Service, 2020. National Wetlands Inventory.
Environmental Laboratory, 1987. Corps of Engineers Wetlands Delineation Manual (No. https://www.fws.gov/wetlands/.
Technical Report Y-87-1). U.S. Army Engineer Waterways Experiment Station, United States Geological Survey, 2016. National Land Cover Database.
Vicksburg, MS, USA. Van Meter, K.J., Basu, N.B., 2015. Signatures of human impact: size distributions and
Glumov, N.I., Kolomiyetz, E.I., Sergeyev, V.V., 1995. Detection of objects on the image spatial organization of wetlands in the Prairie Pothole landscape. Ecol. Appl. 25,
using a sliding window mode. Opt. Laser Technol. 27, 241–249. https://doi.org/ 451–465. https://doi.org/10.1890/14-0662.1.
10.1016/0030-3992(95)93752-D. Wilson, A.G., Hu, Z., Salakhutdinov, R., Xing, E.P., 2016. Stochastic Variational Deep
He, K., Zhang, X., Ren, S., Sun, J., 2015. Deep Residual Learning for Image Recognition. Kernel Learning. arXiv preprint arXiv:1409.1556.
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), p. 770- Wolf, J., Fricker, G.A., 2013. africker/Topographic-Wetness-Index: A revised version of
778. https://ieeexplore.ieee.org/document/7780459. the topographic wetness index written for ArcPython. https://github.com/africker
Hensman, J., De, A.G., Matthews, G., Ghahramani, Z., 2015. Scalable Variational /Topographic-Wetness-Index (accessed 9.1.20).
Gaussian Process Classification. Proc. Mach. Learn. Res. 38, 351–360. In: http Wright, C.K., Larson, B., Lark, T.J., Gibbs, H.K., 2017. Recent grassland losses are
://proceedings.mlr.press/v38/hensman15.html. concentrated around U.S. ethanol refineries. Environ. Res. Lett. 12 https://doi.org/
Huang, H., Chen, Y., Clinton, N., Wang, J., Wang, X., Liu, C., Gong, P., Yang, J., Bai, Y., 10.1088/1748-9326/aa6446.
Zheng, Y., Zhu, Z., 2017. Mapping major land cover dynamics in Beijing using all Wu, Q., Lane, C.R., Li, X., Zhao, K., Zhou, Y., Clinton, N., DeVries, B., Golden, H.E.,
Landsat images in Google Earth Engine. Remote Sens. Environ. 202, 166–176. Lang, M.W., 2019. Integrating LiDAR data and multi-temporal aerial imagery to map
https://doi.org/10.1016/j.rse.2017.02.021. wetland inundation dynamics using Google Earth Engine. Remote Sens. Environ.
Illinois State Climatologist’s Office, Illinois State Water Survey, University of Illinois at 228, 1–13. https://doi.org/10.1016/j.rse.2019.04.015.
Urbana-Champaign. https://stateclimatologist.web.illinois.edu/data/ (accessed Xie, Y., Zhang, A., Welsh, W., 2015. Mapping Wetlands and Phragmites Using Publically
9.25.21). Available Remotely Sensed Images. Photogramm. Eng. Remote Sens. 81, 69–78.
Illinois State Geological Survey Prairie Research Institute, Illinois Height Modernization https://doi.org/10.14358/PERS.81.1.69.
(ILHMP): LiDAR Data. https://clearinghouse.isgs.illinois.edu/data/elevation/illinoi Zagoruyko, S., Komodakis, N., 2016. Wide Residual Networks. Proceedings of the British
s-height-modernization-ilhmp-lidar-data (accessed 9.1.20). Machine Vision Conference (BMVC) p. 87.1-87.12. BMVA Press. https://dx.doi.org/
Johnston, C.A., 2013. Wetland losses due to row crop expansion in the dakota prairie 10.5244/C.30.87.
pothole region. Wetlands 33, 175–182. https://doi.org/10.1007/s13157-012-0365- Zedler, J.B., 2003. Wetlands at your service: reducing impacts of agriculture at the
x. watershed scale. Front. Ecol. Environ. 1, 65–72. https://doi.org/10.1890/1540-9295
Katsaggelos, A.K., Smith, M., Ruiz, P., Wang, M., Dunn, J.B., 2021. WetSegNet data set. (2003)001[0065:WAYSRI]2.0.CO;2.

11

You might also like