Professional Documents
Culture Documents
EPA Public Access: Comparing Pixel-And Object-Based Approaches in Effectively Classifying Wetland-Dominated Landscapes
EPA Public Access: Comparing Pixel-And Object-Based Approaches in Effectively Classifying Wetland-Dominated Landscapes
Author manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
About author manuscripts | Submit a manuscript
Published in final edited form as:
Remote Sens (Basel). 2018 ; 10(1): 46–. doi:10.3390/rs10010046.
EPA Author Manuscript
Abstract
Wetland ecosystems straddle both terrestrial and aquatic habitats, performing many ecological
functions directly and indirectly benefitting humans. However, global wetland losses are
substantial. Satellite remote sensing and classification informs wise wetland management and
monitoring. Both pixel- and object-based classification approaches using parametric and non-
parametric algorithms may be effectively used in describing wetland structure and habitat, but
which approach should one select? We conducted both pixel- and object-based image analyses
EPA Author Manuscript
(OBIA) using parametric (Iterative Self-Organizing Data Analysis Technique, ISODATA, and
maximum likelihood, ML) and non-parametric (random forest, RF) approaches in the Barguzin
This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license
(http://creativecommons.org/licenses/by/4.0/).
*
Correspondence: [email protected]; Tel.: +1-513-569-7854.
Tedros M. Berhane https://orcid.org/0000-0003-0671-9131
Charles R. Lane https://orcid.org/0000-0003-0066-8919
Qiusheng Wu https://orcid.org/0000-0001-5437-4073
Victor V. Chepinoga https://orcid.org/0000-0003-3809-7453
Author Contributions: All authors made significant contributions to the manuscript. Charles Lane developed and supervised the
study, collected field data, analyzed the results, and co-wrote the manuscript. Tedros Berhane processed the spatial data, developed
remote sensing methodologies, analyzed results, and co-wrote the manuscript. Qiusheng Wu provided methodological development,
analyses, results interpretation, and manuscript revisions. Oleg Anenkhonov and Victor Chepinoga co-developed the study’s field
component, provided botanical expertise in collecting and interpreting field data, and contributed to results interpretation and
manuscript writing and revisions. Brad Autrey co-developed the study, collected, processed, and analyzed field data, contributed to
results interpretation and manuscript editing and revisions. Hongxing Liu conducted the initial spatial analyses and contributed to
methodological development and results interpretation, as well as manuscript editing and revisions.
Conflicts of Interest: The authors declare no conflict of interest.
Berhane et al. Page 2
Valley, a large wetland (~500 km2) in the Lake Baikal, Russia, drainage basin. Four Quickbird
multispectral bands plus various spatial and spectral metrics (e.g., texture, Non-Differentiated
Vegetation Index, slope, aspect, etc.) were analyzed using field-based regions of interest sampled
to characterize an initial 18 ISODATA-based classes. Parsimoniously using a three-layer stack
(Quickbird band 3, water ratio index (WRI), and mean texture) in the analyses resulted in the
EPA Author Manuscript
highest accuracy, 87.9% with pixel-based RF, followed by OBIA RF (segmentation scale 5, 84.6%
overall accuracy), followed by pixel-based ML (83.9% overall accuracy). Increasing the predictors
from three to five by adding Quickbird bands 2 and 4 decreased the pixel-based overall accuracy
while increasing the OBIA RF accuracy to 90.4%. However, McNemar’s chi-square test confirmed
no statistically significant difference in overall accuracy among the classifiers (pixel-based ML,
RF, or object-based RF) for either the three- or five-layer analyses. Although potentially useful in
some circumstances, the OBIA approach requires substantial resources and user input (such as
segmentation scale selection—which was found to substantially affect overall accuracy). Hence,
we conclude that pixel-based RF approaches are likely satisfactory for classifying wetland-
dominated landscapes.
Keywords
Lake Baikal; maximum likelihood; near-infrared; Quickbird; random forest; segmentation
EPA Author Manuscript
1. Introduction
Wetlands are amongst the most productive and biodiverse ecosystems on Earth [1].
However, they have been lost at prodigious rates across the globe [2], and those that remain
are imperiled. Junk et al. [3] estimated that 30–90% of global wetlands have been lost, and
that climate change and concomitant temperature and sea level rise, along with precipitation
pattern changes, will continue to stress the remaining wetlands. Davidson [2] reviewed 189
reports of wetland area changes and determined that 64–71% of wetlands have been globally
lost since approximately 1900 AD. With wetland areal loss comes loss in various ecological
and environmental functions at both local and landscape scales. For instance, wetlands are
known areas of high biogeochemical cycling (e.g., [4,5]), groundwater recharge and
stormflow attenuation (e.g., [6,7]), and habitat for many biological species (e.g., [8–10]).
EPA Author Manuscript
Wetland processes that underlie these functions vary by habitat or vegetation structure. For
instance, emergent (or non-woody) wetlands perform denitrification at different rates than
forested wetlands [11,12]. Similarly, water storage in depressional wetlands—which
decouple storm event flows—differs by wetland habitat [13]. Understanding wetland
abundance and typology is therefore important to properly managing the existing wetland
resources and their concomitant watershed functions.
Satellite remote sensing provides a useful mechanism to delineate, assess, and monitor
wetland habitats [14]. Sub-meter to coarse-resolution image data have been analyzed to
identify wetlands and demarcate wetland-upland boundaries as well as differentiate habitats
within extensive wetland systems (see expansive reviews by [14–16]). Critical decisions on
imagery acquisition include resolution, spectral bands, revisit period, and cost. The benefits
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 3
and detractions of various platforms have been—and will continue to be—assessed and
debated as an increasingly large number of satellite systems are launched (e.g., [17]).
classification [18]; random forest [19]; and artificial neural networks [20]). Other
approaches include the pixel-based unsupervised Iterative Self-Organizing Data Analysis
Technique (ISODATA) [21–24] and the supervised maximum likelihood (ML [24,25]) image
classification techniques of change-detection and pattern recognition [24,25].
In contrast with the parametric ML classifier, more recently developed approaches such as
random forest (RF) include non-parametric classification algorithms with no assumption of
Gaussian distribution of the input/predictor variables. As a powerful remote sensing image
classification tool, RF is applicable for both pixel-based or object-based classifications under
supervised or unsupervised settings [19,26,27]. Compared to ISODATA and ML, RF also
has an advantage in providing the relative importance of the input variables in predicting the
response variable by permuting the predictor variable value and measuring the error estimate
before and after the permutation [26,28,29].
EPA Author Manuscript
Numerous studies have used both pixel- and object-based image classification techniques
with RF for various natural resource management applications, including wetlands. For
instance, Husson et al. [32] used 5-cm spatial resolution true-color unmanned aircraft
systems data for mapping non-submerged aquatic vegetation, classifying water (vs.
vegetation), growth form, and dominant taxon using OBIA and RF classifiers, with overall
EPA Author Manuscript
accuracy results obtained for RF ranging from 62–90% for the growth form type to 52–75%
for the dominant taxon classifications. Mahdianpari et al. [33] used synthetic aperture radar
(SAR) data in a hierarchical object-based RF approach to discriminate eight herbaceous
wetland cover types in the Canadian province of Newfoundland with an overall accuracy of
94% achieved. Dronova et al. [31] used the 32-m Beijing-1 satellite data and fuzzy
supervised classification methods and an OBIA technique to detect changes of the major
wetland cover types (i.e., water, mudflat, vegetation, and sand) of Poyang Lake, the largest
freshwater lake–wetland system in China, with comparatively higher overall accuracy
achieved for vegetation and water (90% and 82%, respectively). Ariel et al. [34] applied a
set of spatial and spectral image-object metrics to classify water and four vegetation types
using RF with an overall accuracy of 92%. Tian et al. [35] found higher overall accuracy
using the OBIA coupled with the RF classifier (overall accuracy 93%) when compared to
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 4
support vector machine and artificial neural network approaches in classifying nine land
cover types, including wetlands.
addition to the direct use of the spectral bands of the chosen sensor, numerical band
combinations and band ratios may provide additional information [31]. For instance, the
well-known Normalized Difference Vegetation Index (NDVI [36]) can be used as a proxy
variable for indicating the presence and condition of vegetation (e.g., vigor, health, and
abundance). This index can vary by vegetation types and habitats, thus providing useful
information for improving land cover classification. Other vegetation indices have improved
classification of remotely sensed data; their use is frequently dependent on site-specific
conditions. For instance, to minimize the atmospheric aerosol scattering effect, the
atmospherically resistant vegetation index (ARVI) [37] has been derived. Similarly, soil
brightness can affect vegetation indexes, and this can be compensated by using the soil-
adjusted vegetation index (SAVI [38]). In another example, to simplify or reduce
computation time and computer processing power requirements, the infrared percentage
vegetation index (IPVI) has been used to replace the NDVI [39].
EPA Author Manuscript
Furthermore, auxiliary input variables such as digital elevation model (DEM) and spatial
metrics such as derivatives of the Grey Level Co-occurrence Matrix (GLCM) are also
extensively used in various studies for improving land cover classification and prediction
accuracies [28,31,40]. Topographic position, through its effect on hydrological processes
[41] such as the prediction of areas of soil saturation in low-lying areas [31], can influence
the distribution of wetland classes in a landscape. Rodriguez-Galiano et al. [40] have used
elevation, slope, and aspect variables derived from a digital terrain model along with Landsat
5 TM spectral data as input to a RF model to classify 14 land cover categories in Spain with
overall accuracy of 92% [40]. Wright and Gallant [42] have found DEM-derived
topographic variables to be relevant in improving wetland mapping by increasing their
accuracy in better identifying and differentiating upland areas from wetlands.
These varied random forest and object-based classification approaches and different metrics
are useful in assessing and classifying landscape components. However, the methods
EPA Author Manuscript
discussed above also suffer from limitations. It is difficult and resource-intensive to collect
sufficiently large amounts of field data for training an object-based (RF) model. Despite
widely reported improved performance of object-based over pixel-based image classification
approaches, Dronova [31] indicated results vary by data type, spatial scale, and research
objectives when applied in complex wetland systems (e.g., large wetlands with varied
wetland vegetation structure and open water).
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 5
non-parametric (RF) approaches and iteratively analyzed outputs with spatial and spectral
classification metrics seeking a parsimonious and effective wetland classification solution.
Thus, in addition to accurately classifying the wetland landscape to improve management
options, we aimed to provide an assessment and recommendation of methodological
approaches for consideration specific to classifying wetland landscapes.
In our literature review, we did not find wetland classification studies assessing and applying
highly numerous structural or habitat classes. For instance, an extensive review of OBIA for
wetland mapping [31] indicated few classified wetland systems into more than 10 or 11
classes (e.g., open water, emergent marsh, submergent vegetation, etc.). Our review
suggested that overall accuracy in wetland studies frequently fell below an arbitrary
benchmark of ~85% when the total number of wetland-specific classes exceeded four.
Thus, in our field-based analyses of a large wetland system in Siberian Russia with nearly 20
EPA Author Manuscript
different structural and vegetative habitat and wetland classes, we sought analyze the
differences in overall accuracy when comparing between three classification methods (pixel-
based ML and RF, and object-based RF), constraining our analyses to use the same input
field and remote-sensing datasets. The outcome of this study, therefore, assists end-users in
selecting (and parameterizing) the proper classifier to analyze the structure of the world’s
imperiled and complex wetlands.
2.8% before reaching Lake Baikal, where it provides ~9% of the total inflow to the lake
[46,47]. The extensive wetland area in this study (approximately 500 km2, Figure 1) is
termed the lower Barguzin Valley. The area experiences prolonged seasonal flooding
associated with valley narrowing as the river reaches the southern edge of the Barguzin
Mountain Range on the eastern Baikal periphery [46]. The regional climate is continental,
with long, cold, and relatively dry winters [46]. Land use in the region is mostly
undeveloped, with pastoral and subsistence farming. Both mining and forestry activities also
occur in the approximately 21,000 km2 watershed.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 6
angle of 20.8°, 17.8°, 26.4° and 17.0°, respectively. The Quickbird satellite collected
panchromatic data (0.6-m nominal pixel size) plus four multispectral bands (at 2.4-m
nominal pixel size): blue (B1, 450–520 nm), green (B2, 520–600 nm), red (B3, 630–690
nm), and near-infrared (B4, 760–900 nm). Nominal locational accuracy was 23 m at nadir.
The data-processing and workflow diagram is given in Figure 2. The images were converted
EPA Author Manuscript
We conducted an initial unsupervised classification of the study area using the mosaicked
four-band Quickbird TOA reflectance data (bands 1–4) and the aforementioned Iterative
Self-Organizing Data Analysis Technique clustering algorithm (ISODATA; [21,22,48]) in
ENVI (Harris Geospatial Solutions, Herndon, VA, USA, version 5.3). We arbitrarily began
with 40 spectrally based thematic classes. We reduced these to 18 reasonably distinct classes
for field analyses; 95% of all pairwise comparisons between unsupervised classes had
Jeffries–Matusita separability values ≥1.8, indicating effective class fidelity [12,49,50]).
initial ISODATA classes during an expedition in late August 2013. Sites were selected based
on accessibility to roads, along river courses (for access via boats), and through a
combination of approaches. Following Lane et al. [12,50], two teams of ecologists and
botanists collected vegetation and habitat data from 142 vegetation plots (Figure 3). Data
collected at each 100-m2 plot included dominant species (≥5% cover) and water depth.
Photographs were taken from the center outwards at all cardinal points, as well as straight
down. If more than 5% of the plot was determined to be bare ground, open water (without
≥5% vegetation), filamentous algae, or thatch from graminoids (e.g., Carex spp.), that
information was noted and used in the classification analyses below. We located the
approximate center of each plot by averaging 20 GPS location readings of either a Trimble
Nomad or a Trimble Yuma (Sunnyvale, CA, USA), with 2–5-m real-time accuracy. The
collected species-level vegetation data were subsequently combined to genus-level relative
abundance data for use in this study. Sixteen ground control points (GCPs) were collected
and used to validate the geometric accuracy of the images. Non-ambiguous, man-made
EPA Author Manuscript
physical structures and road intersections were used for this purpose (see Figure 3); at each
GCP, 100 GPS readings were averaged using the GPS receivers noted above, and
photographs were taken of the GCP sites. Quickbird multispectral data were pan-sharpened
using the Gram-Schmidt Pan-Sharpening method in ENVI for GCP validation. All but two
GCPs were within two Quickbird-pixels (i.e., within 1.2 m) from their true ground location.
Due to significantly high computational power and processing time requirements, the
original non-pan-sharpened multispectral bands were used for classification.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 7
Index (DVI; [51]), Soil Adjusted Vegetation Index (SAVI; [38]), the Infrared Percentage
Vegetation Index (IPVI; [39]), Normalized Difference Water Index (NDWI; [54]), Water
Ratio Index (WRI; [55]), and various Band Ratios (i.e., Blue/Green, B1/B2; NIR/Blue,
B4/B1; NIR/Green, B4/B2; NIR/Red, B4/B3; Red/Blue, B3/B1; and Red/Green, B3/B2;
[52]). Although certainly not a complete list of metrics, these vegetation indices and band
ratios were selected to assist in discriminating amongst similar wetland habitats typically
differentiated by botanists and ecologists based on vegetation structure. We additionally
included eight spatial metrics of texture (Contrast, Correlation, Dissimilarity, Entropy,
Homogeneity, Mean, Second Moment and Variance) calculated using the QuickBird NIR
band, the native DEM value, and an additional 10 topographic metrics (Aspect, Cross-
sectional Convexity, Longitudinal Convexity, Maximum Curvature, Minimum Curvature,
Plain Convexity, Profile Convexity, RMS Error and Percent Slope) calculated using DEM
data to further differentiate among vegetation and habitat classes in the study area [56].
DEM-based metrics were derived using elevation data from the Advanced Spaceborne
EPA Author Manuscript
class (i.e., five and two points per class for training and testing, respectively). The one
exception is for Class 10, which only had four points available for training. The partition
into the training and testing datasets was performed by visually inspecting the point
distribution so that the training and/or the testing datasets were distributed in space to
maximize distances between points thereby minimizing the chances of spatial
autocorrelation. We then delimited regions of interest (ROIs) for each of the 18 thematic
classes from the 125 field survey sites used in the balanced approach. Each ROI was
approximately 30 pixels in area and most, though not all, consisted of a single thematic
class. In certain cases, the ROIs—which otherwise were centered around the field-based data
point—were moved slightly to maintain thematic homogeneity while staying within the
approximate bounds of the field-based sampling area.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 8
highest predictive power was found with the following settings: two variables (mtry), 1000-
tree maximum (ntree), and 1000 bootstrap (or OOB) samples in the classification; these
values were used in the final model run. The accuracy of the pixel-based RF model was
assessed with pixels from the 30% hold-out ROIs across the study area. RF was conducted
on the combination of four data layers (Quickbird bands 1–4) as well as bands 1–4 plus 33
additional input variables described above (see Table 1). Random forest usefully provides
quantitative variable importance measures, the Mean Decrease in Gini (MDG) and the Mean
Decrease in Accuracy (MDA) values. MDG informs the accuracy of a particular class due to
utilization of a given variable while the MDA provides the difference between OOB error of
the original dataset and the OOB error from random permutations of a set of input predictor
variable values [26].
to classify the ROI training data into 18 supervised classes using the predictor variables that
yielded the highest overall accuracy with the most parsimonious pixel-based-RF model (i.e.,
Quickbird band 3, WRI, and mean texture computed from Quickbird band 4; see results
further described below). These layers were chosen because preliminary results using the
pixel-based-RF model demonstrated only slight changes in accuracy (both positively and
negatively affecting our results) with additional metrics. Furthermore, as many spectral
metrics were highly correlated (e.g., |r| values ≥ 0.89; see Table 2), we sought to develop a
parsimonious model that balanced processing requirements and accuracy and hence we
initially limited the analyses to a three-layer stack. However, recognizing that additional
bands may potentially improve the overall ML accuracy, two additional variables that
comprised the five most important variables as determined by MDG, (i.e., Quickbird bands 2
and 4, see below) were also analyzed along with the aforementioned three predictors.
Segmentation scale affects the object size and hence the properties of discrete objects in the
study area. We analyzed multiple scales in turn (e.g., 5–100; Figure 4) and determined via
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 9
accuracy assessment that segmentation at the spatial scale of 5 provided the optimum image-
object size. Color (multispectral heterogeneity) and shape (a function of smoothness and
compactness heterogeneity) are the primary image-object features that are used for image-
segmentation and for improvement in primitive image-object creation [30]. We segmented
color (i.e., spectral band information) and shape (compactness and smoothness metrics) to
EPA Author Manuscript
generate optimum pixel groups (i.e., image-objects) exhibiting intra- and inter-object
spectral homogeneity and heterogeneity [60]. Initial analyses of the study area informed our
decision to use shape and color segmentation weights of 0.1 and 0.9, and compactness and
smoothness parameters of 0.5, respectively. Upon segmentation completion, we calculated
the spatial and spectral metric average values for each of the 5,191,948 segmented objects in
the study area. We then utilized the RF approach described above to classify the study area.
Similar to the pixel-based ML approach, OBIA-RF was performed using three- and five-
stack layers to provide the opportunity to contrast amongst the methods in wetland
classification.
the remaining 30% were used for validation of the pixel-based ML and RF and OBIA-RF
approaches described above. We also conducted and report overall accuracy [62]. We
quantitatively assessed if the observed difference in the classification accuracies between the
three classifiers were statistically meaningful using McNemar’s test [63,64].
3. Results
3.1. Field Data Collection
Fifty-six different genera were identified as occurring at >5% frequency across the sampling
locations. Members of the genera Carex and Equisetum were more commonly found (46 and
25 times, respectively), followed by Calamagrostis, Myriophyllum, Agrostis, Nymphoides,
and Potentilla. Twenty-three genera were only encountered a single time. We also noted the
>5% abundance of open water (49 sites), thatch (33 sites), and bare ground (22 sites).
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 10
(Figure 5a,b and Tables 4 and 5). This highest overall accuracy was achieved using the
aforementioned three variables despite the fact that WRI and mean texture are highly
correlated (|r| = 0.99; see Table 2). The overall accuracy, however, decreased to 80.1% when
the 22 non-correlated variables were considered with the exclusion of WRI and inclusion of
mean texture; accuracy dropped to 72.7% when mean texture was excluded (see Table 4).
These results indicated it is paramount to include both mean texture and WRI for
meaningful improvement in prediction accuracy of the wetland classifications—although
mean texture without WRI improved the overall accuracy (e.g., from 72.7% to 80.1%).
A benefit of the RF approach is the derivation of variable importance factors based on the
MDG values. We assessed MDG based on 100 RF runs. For the pixel-based RF, Quickbird
bands 2, 3, and 4, WRI, and mean texture were consistently ranked as the five most
important variables, the rank of importance for the remaining variables changed with
different RF runs of the 100 iterations (Table 6). As the analytical process for ML and OBIA
EPA Author Manuscript
can be laborious, we used the MDG to inform our selection of variables to use in the ML
and OBIA approaches.
The ML approach was conducted using the same three-layer stack image as the RF
approach, and resulted in an overall accuracy of 83.9% (Table 7). ML analysis using the five
most important predictor variables from the pixel-based RF analysis (Quickbird band 2–4,
WRI, and mean texture) resulted in a decrease in overall accuracy from 83.9% to 80.2%.
Perfect producer’s (PA) or user’s accuracy was calculated for the pixel-based RF and ML
approach class confusion matrix (Tables 5 and 7) using three bands for 11 and 8 of the 18
classes, respectively. Comparatively higher classification error using pixel-based RF
approach occurred with Class 3 (66.9% error), Class 6 (34.5% error), and Class 14 (35.6%
error). Greater PA errors in the ML analyses occurred with Class 3 (30.2% error), Class 5
(33.3% error), Class 6 (52.5% error), Class 9 (61.9% error), and Class 14 (48.4% error). The
EPA Author Manuscript
errors suggested open water, emergent, and submerged vegetation were more difficult to
classify with remotely sensed data.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 11
exception to this is at segmentation scale of 50 where the overall accuracy was higher than at
segmentation scale of 30.
In addition, we saw an improvement in overall accuracy from 84.6% (with three layers and
OBIA) to 90.4% (with five layers and OBIA; Table 8). The likelihood of diminishing returns
EPA Author Manuscript
with additional data layers (and concomitantly the substantial resource requirements)
through this approach informed our decision to use no more than five layers in our OBIA.
4. Discussion
4.1. Random Forest in Wetland Classification
Targeting overall accuracy above an arbitrary benchmark acceptable value of 85% in
classifying complex wetland systems can be challenging [31], a challenge exacerbated by
the inclusion of numerous wetland classes (i.e., 18 in this study). Studies using high-
resolution satellite data may consider the utility of “high-resolution” ecological data (i.e.,
field-based community data that go beyond structural classifications) in their systems of
interest—although we hasten to add that the level of specificity is dependent on the
questions asked and management purpose for which the data are often requested. High
accuracy is also dependent on abundant field-based data, as well as cloud-free imagery and a
EPA Author Manuscript
Accuracy is furthermore a product of the analytical approach employed and the number of
predictor variables. In this study, we found the greatest overall accuracy to be achieved with
the OBIA and RF classifier (~90%) using five predictor variables (and segmentation scale
5); this decreased to <85% when three predictors were used. Pixel-based RF achieved 88%
overall accuracy using only three predictors, and, as Duro et al. [65] also reported, required
substantially less user interaction and processing time than conducting the OBIA.
Furthermore, though differences are evident when visually comparing between the
approaches (e.g., Figure 6), McNemar’s test found no significant overall difference between
the varied approaches (see Table 5). Thus, although any of the approaches may be employed,
our findings suggest a pixel-based RF approach may be best suited when considering
classifying the wetland-dominated landscape.
EPA Author Manuscript
That random forest has many advantageous features to consider in remote sensing
applications is becoming increasingly evident [29,40,57,66]. Some of the advantages of RF
include the fact that it is freely available (e.g., in the R statistical package) and relatively
user-friendly with limited user selections required (in our instance, the number of variables
for each node in a random subset of data (mtry) and the number of trees to grow in a forest
(ntree) were specified). RF is computationally efficient in handling large datasets, even those
with substantial “noise” and the presence of outliers [29,40]. The fact that RF uses the input-
predictor variables with replacement but not deletion (i.e., bagging) apparently allows more
specificity when classifying, and as a non-parametric approach there are fewer restrictions
on the distribution of the input data layers. Lastly, RF output includes the generation of
variable importance (MDG) and unbiased internal estimates of error (MDA), which assists
in the interpretation of the results [40,57]. Limitations of RF include the “black box” nature
of the model, where limited information about the relationships between the predictor
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 12
variables and the response variables can be discerned [66]. Furthermore, RF results change
when an unbalanced study design is employed [27,28]. We endeavored to sample at least
five sites, and as many sites as ultimately possible for each class, during our short sampling
window for the field-based expedition. However, following the suggestions (i.e., [27,28]),
we ultimately did not use some of the data points acquired during the sampling period,
EPA Author Manuscript
However, the benefits of RF appear to outweigh the limitations, and consequently, numerous
studies have used RF as the classifier of choice for both pixel- and object-based image
classifications for various applications using data from multiple remote sensing platforms.
For instance, Mallinis et al. [67] conducted a RF analysis using Quickbird to delineate
vegetation polygons of dominant species for national forest database creation, and Smith
[68] classified land cover classes using SPOT. The literature is increasingly replete with the
use of RF in classifying landscapes (e.g., [19,25,69–71]).
Table 6). It is not surprising that Quickbird band 4, mean texture (derived from band 4), and
WRI were the three most important variables in class discrimination. As water column
absorption of energy in the NIR spectrum (i.e., band 4) is high [72], the abundance of water,
a factor controlling wetland vegetation community development, would feature prominently
in categorizing the studied system. Similarly, photosynthetically active vegetation reflects
more energy in the near-infrared portion (e.g., [73]) while absorbing more energy in the blue
and in the red-light regions of the spectrum. The WRI assess the abundance of the green, red
and near-infrared band values to the blue band; with abundant water associated
approximately with lower (<2.5), barren land with intermediate (2.5–4.0) and vegetated
areas with higher (>4.0) WRI values. The Barguzin Valley is dominated by an extensive and
spatially connected surface water hydrological system (submergent marshes, rivers, lakes,
etc.) and the role of WRI is in differentiating the valley’s aquatic habitats from the uplands.
This is particularly true in differentiating mixed pixels belonging to the water-soil and water-
vegetation traditional zones. As wetland vegetation was the main study object in this
EPA Author Manuscript
analysis, differences in spectral reflectance (Quickbird bands 2–3) would logically follow as
a second group of important variables in group determination. It was, however, somewhat
surprising that the abundant additional metrics that we calculated in an effort to provide
greater discernment between wetland classes were not particularly relevant to the
classification, as evidenced by their lack of relative importance and the decrease in overall
accuracy when additional metrics were included (for the pixel-based approaches). Both the
RF and ML approaches decreased in overall accuracy with increased predictors beyond our
three-stack parsimonious model.
One potential reason for the lack of specificity with increased metrics is simply a limitation
of the input data set. Quickbird has limited spectral bands (only four plus the panchromatic
band), thus limiting to a degree the possible permutations of the data, at least the data based
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 13
on Quickbird’s available spectra (limited vis-à-vis more current satellite sensors). Of the 37
different predictor variables we explored, only 22 were unique enough to warrant their
inclusion when running our non-correlated model (see Table 2). Thus, although deriving the
additional variables was not particularly laborious in the pixel-based assessments, it may not
be warranted, at least in this study area, given the response across the breadth of classes.
EPA Author Manuscript
In addition, it is worth noting that model parsimony is generally desirable. For instance,
lower OOB-error estimates (i.e., better models) occur in RF classification with a lower
number of variables (mtry) used for splitting the trees at each node and with a lower number
of trees (ntree) constructed [74].
the number of predictors. The OBIA approach requires substantial user inputs and
processing; it is too time-consuming to consider the OBIA across the breadth of metrics (37)
considered in this analysis.
However, OBIA may be useful for analyses, depending on characteristics of the “objects”
within the study area. Thus, it is possible to harness the power of OBIA through iterative
approaches of using different study-area segmentation parameters and scales to better
generate image-object identification for subsequent input as predictor variables. For
instance, in our study area, we found that a segmentation scale of 5 provided the greatest
number of objects and was also the most informative; others may find a different scale
useful [31,75,76]. Similar to this study’s exploration of segmentation scale on overall
accuracy (see Table 8), Stumpf and Kerle [28] found decreasing model performance at larger
segmentation scales. However, the smaller the segmentation scale, the more computational
resources are required for data processing time and data storage for intermediate geospatial
EPA Author Manuscript
products. This factors into end-user decisions when processing high-resolution multispectral
remote sensing data, particularly when additional derived and auxiliary dataset that are
relevant for better wetland detection and delineation are integrated.
We therefore followed others (e.g., [28,29,31,32]) and first separated the image into
segments and then used the RF classification algorithm to assign wetland classes based on
the spectral data of the homogenous image-objects. We found that combining both OBIA
and RF for segmentation and classification purposes, respectively, allowed the utilization of
the strengths of each approach for better wetland delineation and mapping. Based on the
results from our study, the hybrid OBIA-RF approach can likely be applied in similar
settings where spectral information has more influence than shape in wetland class
distinction and should be further explored, especially as processing power continues to
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 14
increase and non-proprietary software packages and algorithms are becoming more
available.
5. Conclusions
EPA Author Manuscript
The use of remote sensing to classify wetland landscapes is critical to the understanding of
the structure that begets wetland functions, especially as the world’s finite wetland resources
are lost [77]. Field-based data plus robust spectral analyses from satellite platforms provide
useful information for effective management. There are many approaches that can be used to
classify the landscape with remotely sensed data, from established and relatively simplistic
(e.g., ISODATA) to the novel and increasingly complex (e.g., random forest, neural
networks, etc.). Similarly, many different band combinations creating spectral metrics are
used to further discriminate amongst landscape features (e.g., NDVI). To facilitate
subsequent remotely sensed wetland analyses, we conducted both pixel- and object-based
analyses using parametric and non-parametric approaches and progressively incorporated
spatial and spectral metrics. The pixel-based maximum likelihood and random forest
approaches performed well using a parsimonious Quickbird band 3, WRI, and mean texture
variables, with accuracies of 83.9% and 87.9%, respectively. The inclusion of an additional
two predictors to our “parsimonious” model degraded our accuracy using pixel-based
EPA Author Manuscript
maximum likelihood and random forest approaches. However, when an object-based image
analysis approach was coupled with the random forest analysis, our overall accuracy
increased from 84.6% with three predictors to 89.6% with five predictors; more predictors
allowed greater discrimination between the >5 million objects in the analysis. Segmentation
scale had an effect on these results; larger objects resulted in smaller overall accuracies.
Fewer objects did, however, decrease the processing and time requirements for the study—
these trade-offs are thus important to consider when analyzing wetland-dominated
landscapes. In addition, the object-oriented approach requires repeated iterations and
additional subjective parameterization (e.g., compactness and smoothness values).
Furthermore, OBIA was conducted using commercial software, an expense that these results
suggest might not be warranted (as we found no significant difference between the overall
accuracies in our analyses). Thus, we conclude that the random forest algorithm warrants its
increased use as an analytical approach in effectively assessing and mapping wetland
resources, be it through either pixel- or object-based approaches.
EPA Author Manuscript
Acknowledgments
This paper has been reviewed in accordance with the U.S. Environmental Protection Agency’s peer-review policies
and approved for publication. Mention of trade names or commercial products does not constitute endorsement of
recommendation for use. Statements in this publication reflect the authors’ personal views and opinions and should
not be construed to represent any determination of policy of the U.S. Environmental Protection Agency. The
research of O.A. Anenkhonov and V.V. Chepinoga were carried out using the framework of projects
0337-2016-0001 and 0347-2016-003 accordingly, supported by the Russian Federal Budget.
References
1. Mitsch WJ, Gosselink JG. The value of wetlands: Importance of scale and landscape setting. Ecol.
Econ. 2000; 35:25–33.
2. Davidson NC. How much wetland has the world lost? Long-term and recent trends in global wetland
area. Mar. Freshw. Res. 2014; 65:934.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 15
3. Junk WJ, An S, Finlayson CM, Gopal B, Květ J, Mitchell SA, Mitsch WJ, Robarts RD. Current
state of knowledge regarding the world’s wetlands and their future under global climate change: A
synthesis. Aquat. Sci. 2012; 75:151–167.
4. Ramesh Reddy, K., DeLaune, R. Biogeochemistry of Wetlands. CRC Press; Boca Raton, FL, USA:
2008.
EPA Author Manuscript
5. Marton JM, Creed IF, Lewis DB, Lane CR, Basu NB, Cohen MJ, Craft CB. Geographically isolated
wetlands are important biogeochemical reactors on the landscape. BioScience. 2015; 65:408–418.
6. Winter TC, Rosenberry DO. The interaction of ground water with prairie pothole wetlands in the
Cottonwood Lake area, east-central North Dakota, 1979–1990. Wetlands. 1995; 15:193–211.
7. Shaw DA, Vanderkamp G, Conly FM, Pietroniro A, Martz L. The fill-spill hydrology of prairie
wetland complexes during drought and deluge. Hydrol. Process. 2012; 26:3147–3156.
8. Kirkman LK, Golladay SW, Laclaire L, Sutter R. Biodiversity in southeastern, seasonally ponded,
isolated wetlands: Management and policy perspectives for research and conservation. J. N. Am.
Benthol. Soc. 1999; 18:553–562.
9. Gibbons JW, Winne CT, Scott DE, Willson JD, Glaudas X, Andrews KM, Todd BD, Fedewa LA,
Wilkinson L, Tsaliagos RN, et al. Remarkable amphibian biomass and abundance in an isolated
wetland: Implications for wetland conservation. Conserv. Biol. 2006; 20:1457–1465. [PubMed:
17002763]
10. The United States Environmental Protection Agency (U.S. EPA). Connectivity of Streams and
Wetlands to Downstream Waters: A Review and Synthesis of the Scientific Evidence; EPA/600/
R-14/475F; U.S. EPA Office of Research and Development; Washington, DC, USA: 2015.
11. Ullah S, Faulkner SP. Use of cotton gin trash to enhance denitrification in restored forested
EPA Author Manuscript
20. Costa H, Foody GM, Boyd DS. Using mixed objects in the training of object-based image
classifications. Remote Sens. Environ. 2017; 190:188–197.
21. Ball, G., Hall, J. ISODATA, A Novel Method of Data Analysis and Pattern Classification,
Technical Report. Stanford Research Institute; Menlo Park, CA, USA: 1965. p. 1-61.
22. Jain, A., Dubes, R. Algorithms for Clustering Data. Prentice Hall; Englewood Cliffs, NJ, USA:
1988.
23. Earth Resources Data Analysis System (ERDAS). ERDAS Field Guide. 5. ERDAS Worldwide
Headquarters: Buford Highway; NE, USA; Atlanta, GA, USA: 1999. p. 227-232.
24. Lillesand, T., Kiefer, R., Chipman, J. Remote Sensing and Image Interpretation. 7. John Wiley &
Sons, Inc; Hoboken, NJ, USA: 2015. p. 544-558.
25. Akar Ö, Güngör O. Classification of multispectral images using random forest algorithm. J.
Geodesy Geoinform. 2012; 1:105–112.
26. Belgiu M, Drăguţ L. Random forest in remote sensing: A review of applications and future
directions. ISPRS J. Photogramm. Remote Sens. 2016; 114:24–31.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 16
27. Millard K, Richardson M. On the importance of training data sample selection in random forest
image classification: A case study in peatland ecosystem mapping. Remote Sens. 2015; 7:8489–
8515.
28. Stumpf A, Kerle N. Object-oriented mapping of landslides using random forests. Remote Sens.
Environ. 2011; 115:2564–2577.
EPA Author Manuscript
29. Puissant A, Rougier S, Stumpf A. Object-oriented mapping of urban trees using random forest
classifiers. Int. J. Appl. Earth Obs. Geoinform. 2014; 26:235–245.
30. Benz UC, Hofmann P, Willhauck G, Lingenfelder I, Heynen M. Multi-resolution, object-oriented
fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. Remote
Sens. 2004; 58:239–258.
31. Dronova I. Object-based image analysis in wetland research: A review. Remote Sens. 2015;
7:6380–6413.
32. Husson E, Ecke F, Reese H. Comparison of manual mapping and automated object-based image
analysis of non-submerged aquatic vegetation from very-high-resolution UAS images. Remote
Sens. 2016; 8:724.
33. Mahdianpari M, Salehi B, Mohammadimanesh F, Motagh M. Random forest wetland classification
using ALOS-2 L-band, RADARSAT-2 C-band, and TerraSAR-X imagery. ISPRS J. Photogramm.
Remote Sens. 2017; 130:13–31.
34. Ariel E, Salas L, Boykin K, Valdez R. Multispectral and Texture Feature Application in Image-
Object Analysis of Summer Vegetation in Eastern Tajikistan Pamirs. Remote Sens. 2016; 8:78.
35. Tian S, Zhang X, Tian J, Sun Q. Random Forest Classification of Wetland Land covers from Multi-
Sensor Data in the Arid Region of Xinjiang, China. Remote Sens. 2016; 8:954.
EPA Author Manuscript
36. Tucker CJ. Red and photographic infrared linear combinations for monitoring vegetation. Remote
Sens. Environ. 1979; 8:127–150.
37. Kaufman YJ, Tanre D. Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE
Trans. Geosci. Remote Sens. 1992; 30:261–270.
38. Huete AR. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 1988; 25:295–309.
39. Crippen R. Calculating the vegetation index faster. Remote Sens. Environ. 1990; 34:71–73.
40. Rodriguez-Galiano VF, Ghimire B, Rogan J, Chica-Olmo M, Rigol-Sanchez JP. An assessment of
the effectiveness of a random forest classifier for land-cover classification. ISPRS J. Photogramm.
Remote Sens. 2012; 67:93–104.
41. Bwangoy J, Hansen M, Roy D, DeGrandi G, Justice C. Wetland mapping in the Congo Basin using
optical and radar remotely sensed data and derived topographical indices. Remote Sens. Environ.
2010; 114:73–86.
42. Wright C, Gallant A. Improved wetland remote sensing in Yellowstone National Park using
classification trees to combine TM imagery and ancillary environmental data. Remote Sens.
Environ. 2007; 107:582–605.
43. Brunello, A. Lake Baikal: Experience and Lessons Learned Brief. Tahoe-Baikal Institute; South
EPA Author Manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 17
52. Yang C, Everitt J, Bradford J. Comparison of QuickBird satellite imagery and airborne imagery for
mapping grain sorghum yield patterns. Precis. Agric. 2006; 7:33–44.
53. Liu G-R, Liang C-K, Kuo T-H, Lin T-H, Huang S-J. Comparison of the NDVI, ARVI and AFRI
vegetation index, along with their relations with the AOD using SPOT 4 vegetation data. Terr.
Atmos. Ocean. Sci. 2004; 15:15.
54. Gao B. NDWI-A normalized difference water index for remote sensing of vegetation liquid water
from space. Remote Sens. Environ. 1996; 58:257–266.
55. Wuest B, Zhang Y. Region based segmentation of Quickbird multispectral imagery through bands
ratios and fuzzy comparison. ISPRS J. Photogramm. Remote Sens. 2009; 64:55–64.
56. Haralick RM, Shanmugam K, Dinstein IH. Textural features for image classification. IEEE Trans.
Syst. Man Cybern. 1973; SMC-3:610–621.
57. Liaw A, Wiener M. Classification and regression by randomForest. R News. 2002; 2:18–22.
58. Foody GM, Campbell NA, Trodd NM, Wood TF. Derivation and applications of probabilistic
measures of class membership from the maximum likelihood classification. Photogramm. Eng.
Remote Sens. 1992; 58:1335–1341.
59. Foody GM, Palubinskas G, Lucas RM, Curran PJ, Honzak M. Identifying terrestrial carbon sinks:
Classification of successional stages in regenerating tropical forest from Landsat TM data. Remote
EPA Author Manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 18
73. Beeri O, Phillips R, Hendrickson J, Frank AB, Kronberg S. Estimating forage quantity and quality
using aerial hyperspectral imagery for northern mixed-grass prairie. Remote Sens. Environ. 2007;
110:216–225.
74. Pelletier C, Valero S, Inglada J, Champion N, Dedieu G. Assessing the robustness of Random
Forests to map land cover with high resolution satellite image time series over large areas. Remote
Sens. Environ. 2016; 187:156–168.
EPA Author Manuscript
75. Ouyang Z, Zhang M, Xie X, Shen Q, Guo H, Zhao B. A comparison of pixel-based and object-
oriented approaches to VHR imagery for mapping saltmarsh plants. Ecol. Inform. 2011; 6:136–
146.
76. Vo Q, Oppelt N, Kuenzer C. Remote sensing in mapping mangrove ecosystem-An object-based
approach. Remote Sens. 2013; 5:183–201.
77. Creed IC, Lane CR, Serran JN, Alexander LC, Basu NB, Calhoun AJK, Christensen JR, Cohen
MJ, Craft C, D’Amico E, et al. Enhancing protection for vulnerable waters. Nat. Geosci. 2017
EPA Author Manuscript
EPA Author Manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 19
EPA Author Manuscript
EPA Author Manuscript
EPA Author Manuscript
Figure 1.
Location of Lake Baikal and the lower Barguzin Valley study area and Quickbird imagery
boundary.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 20
EPA Author Manuscript
EPA Author Manuscript
Figure 2.
Field and remote sensing data-processing workflow.
EPA Author Manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 21
EPA Author Manuscript
EPA Author Manuscript
EPA Author Manuscript
Figure 3.
Field survey (n = 142) and ground-control sites (n = 16) within the lower Barguzin Valley
study area overlain on an image composite (Quickbird bands 2, 3, and 4).
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 22
EPA Author Manuscript
EPA Author Manuscript
EPA Author Manuscript
Figure 4.
Example of image-objects created at different segmentation scales using parameters of shape
= 0.1 and compactness = 0.5. The number of objects created across the study area decreased
with increasing segmentation scales: 5 (5,191,948 objects), 10 (1,474,823 objects), 15
(711,351 objects), 30 (204,026 objects), 50 (81,026 objects), and 100 (23,091 objects).
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 23
EPA Author Manuscript
EPA Author Manuscript
EPA Author Manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 24
EPA Author Manuscript
EPA Author Manuscript
EPA Author Manuscript
Figure 5.
(a) Genus-level wetland and aquatic habitats classification map (pixel-based RF approach
using three-layer stack). Four focal areas (5A–5D) are shown in finer detail in (b). Percent
values given in parentheses represent the approximate abundance of each genus or habitat
found in the field-based analyses for each class. (b) Finer-detailed genus-level wetland and
aquatic habitat classification thematic maps developed using the pixel-based RF approach
for the areas of interest shown by white-colored squares in (a). Classes correspond to the
legend in (a).
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 25
EPA Author Manuscript
EPA Author Manuscript
EPA Author Manuscript
Figure 6.
Contrasting the results between the varied methods using a three-layer predictor dataset for
inset-5D in Figure 5a: (A) Quickbird imagery color composite of bands 2, 3, and 4; (B)
pixel-based maximum likelihood classification; (C) object-based random forest
classification; and (D) pixel-based random forest classification. Classes correspond to the
legend in Figure 5a.
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 26
Table 1
Description of the input predictor variables used in this study (B1–B4 are Quickbird multispectral bands,
EPA Author Manuscript
while B5–B37 are identifications assigned to the derived spatial and spectral matrices).
Ratio Transformation Ratio of reflectance spectra [52]; B14–B19 ρB/ρG; ρR/ρB; ρR/ρG; ρNIR/ρB; ρNIR/ρG; ρNIR/ρR
Texture Metrics Texture variables (contrast (B20), correlation (B21), Source: Harris Geospatial, Texture Metrics Background.
dissimilarity (B22), entropy (B23), homogeneity Available online: www.harrisgeospatial.com/docs/
(B24), mean (B25), 2nd moment (B26), and variance backgroundtexturemetrics.html (accessed on 26 December
(B27)) computed as a measure of Gray Level Co- 2017).
occurrence Matrix (GLCM) using Band4 (Harris
Geospatial Solutions, Herndon, VA, USA, version
5.3).
Topography Metrics Advanced Spaceborne Thermal Emission and Source: Harris Geospatial, Topographic Modeling
Reflection Radiometer (ASTERGDEM)-Global Background. Available online: www.harrisgeospatial.com/
Digital Elevation Model (GDEM). GDEM-derived docs/backgroundtopographicmodeling.html(accessed on 26
variables (aspect (B28), cross-sectional convexity December 2017).
(B29), DEM (B30), longitudinal convexity (B31),
maximum curvature (B32), minimum curvature
(B33), plan convexity (B34), profile convexity
(B35), RMS error (B36), and slope (%; B37) (Harris
Geospatial Solutions, Herndon, VA, USA, version
5.3)).
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
EPA Author Manuscript EPA Author Manuscript EPA Author Manuscript
Table 2
Correlation matrix of predictor variables. Linear correlations (|r| ≥ 0.89) are shaded grey. Quickbird bands 1–4 are designed B1–4; see text for additional abbreviations.
Predictor B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12 B13 B14 B15 B16 B17 B18 B19 B20 B21 B22 B23 B24 B25 B26 B27 B28 B29 B30 B31 B32 B33 B34 B35 B36 B37
B1 1.00
Berhane et al.
B2 0.96 1.00
B11 −0.50 0.31 −0.49 −0.93 0 −1.00 0.03 −1.00 −0.96 −0.96 1.00
B12 0.28 0.4 0.26 0.85 0.02 0.94 0 0.96 1 1 −0.96 1.00
B13 0.62 0.75 0.63 0.99 −0.02 0.95 −0.05 0.94 0.85 0.85 −0.94 0.85 1.00
B14 −0.68 −0.86 −0.74 −0.82 0.05 −0.77 0.11 −0.71 −059 −0.59 0.71 −059 −0.85 1.00
B15 0.93 0.94 0.99 0.65 −0.07 0.62 −0.12 0.56 0.32 0.32 −0.56 0.32 0.68 −0.77 1.00
B16 0.9 0.84 0.95 0.51 −0.06 0.51 −0.10 0.46 0.2 0.2 −0.46 0.2 0.53 −0.59 0.96 1.00
B17 0.46 0.6 0.46 0.98 0 0.95 −0.02 0.95 0.91 0.91 −0.95 0.91 0.98 −0.75 0.51 0.37 1.00
B18 0.41 0.53 0.4 0.95 0.01 0.95 −0.01 0.96 0.93 0.93 −0.96 0.93 0.95 −0.68 0.46 0.33 0.99 1.00
B19 0.17 0.31 0.14 0.85 0.03 0.85 0.02 0.88 0.93 0.93 −0.88 0.93 0.84 −0.54 0.2 0.06 0.94 0.96 1.00
B20 0.16 0.2 0.17 0.28 0.01 0.25 0.03 0.25 0.24 0.24 −0.25 0.24 0.27 −0.23 0.17 0.12 0.27 0.25 0.23 1.00
B21 −0.17 −0.16 −0.17 −0.18 0 −0.28 −0.01 −0.28 −0.26 −0.26 0.28 −0.26 −0.19 0.14 −0.19 −0.20 −0.17 −0.19 −0.14 −0.16 1.00
B22 0.21 0.26 0.22 0.38 0.01 0.42 0.03 0.42 0.41 0.41 −0.42 0.41 0.38 −0.31 0.24 0.19 0.38 0.38 0.35 0.87 −0.38 1.00
B23 0.33 0.36 0.35 0.45 0 0.6 −0.04 0.59 0.56 0.56 −0.59 0.56 0.47 −0.38 0.39 0.38 0.45 0.47 0.39 0.36 −0.54 0.69 1.00
B24 −0.24 −0.27 −0.25 −0.40 −0.01 −0.51 −0.02 −0.52 −050 −0.50 0.52 −050 −0.41 0.32 −0.29 −0.26 −0.41 −0.42 −0.37 −0.53 0.55 −0.87 −0.89 1.00
B25 0.63 0.74 0.62 1 −0.02 0.94 −0.05 0.93 0.85 0.85 −0.93 0.85 0.99 −0.82 0.65 0.51 0.98 0.95 0.85 0.28 −0.18 0.38 0.46 −0.40 1.00
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
B26 −0.34 −0.36 −0.35 −0.42 0.01 −0.57 0.04 −0.56 −052 −0.52 0.56 −052 −0.44 0.36 −0.40 −0.39 −0.41 −0.44 −0.34 −0.30 0.6 −0.62 −0.98 0.85 −0.43 1.00
B27 0.15 0.2 0.15 0.28 0.01 0.27 0.02 0.27 0.26 0.26 −0.27 0.26 0.28 −0.24 0.16 0.11 0.28 0.26 0.24 0.79 −0.09 0.75 0.37 −0.50 0.28 −0.30 1.00
B28 −0.01 0 −0.01 0 −0.01 0.02 −0.01 0.02 0.02 0.02 −0.02 0.02 0.01 −0.01 0 0 0.01 0.01 0.01 0.01 −0.07 0.04 0.04 −0.05 0 −0.05 0 1.00
B29 −0.01 −0.01 −0.01 0.01 0 0.01 0 0.02 0.02 0.02 −0.02 0.02 0.01 0 −0.01 −0.01 0.02 0.02 0.03 0.02 −0.02 0.02 0.02 −0.02 0.01 −0.02 0.02 −0.01 1.00
B30 0.18 0.19 0.19 0.15 −0.02 0.14 −0.07 0.13 0.08 0.08 −0.13 0.08 0.16 −0.16 0.19 0.17 0.13 0.12 0.08 0.13 −0.09 0.15 0.14 −0.13 0.15 −0.14 0.09 0.03 0.02 1.00
B31 0 0.01 0.01 −0.01 0.01 0 0 0 0 0 0 0 0 −0.01 0.01 0.01 −0.01 −0.01 −0.01 0.02 0 0.01 0 0 0 −0.01 0.02 0.02 0.08 0.09 1.00
B32 −0.01 −0.01 0 0.01 0.02 0.01 0.02 0.01 0.01 0.01 −0.01 0.01 0.01 0 0 0 0.01 0.01 0.02 0 −0.03 0 0.01 −0.01 0.01 −0.02 0 0.12 0.19 0.08 0.73 1.00
B33 0.02 0.01 0.02 −0.01 0 0 −0.02 −0.01 −0.01 −0.01 0.01 −0.01 −0.01 −0.01 0.01 0.01 −0.02 −0.02 −0.02 0.03 0.02 0.02 0 0 −0.01 0 0.03 −0.09 0.19 0.05 0.72 0.09 1.00
B34 −0.01 0 0 −0.01 0 −0.01 0 −0.01 −0.01 −0.01 0.01 −0.01 −0.01 −0.01 0 0 −0.01 −0.01 −0.01 −0.02 0.01 −0.02 −0.02 0.01 −0.01 0.01 −0.02 0.01 −0.83 −0.02 −0.07 −0.15 −0.17 1.00
B35 0 0 0.01 −0.01 0.02 0 0 0 −0.01 −0.01 0 −0.01 0 −0.01 0.01 0.01 −0.01 −0.01 −0.01 0.01 0.03 0 0 0.01 −0.01 0.01 0.01 0 0.06 0.06 0.77 0.57 0.54 −0.07 1.00
Page 27
EPA Author Manuscript EPA Author Manuscript EPA Author Manuscript
Predictor B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12 B13 B14 B15 B16 B17 B18 B19 B20 B21 B22 B23 B24 B25 B26 B27 B28 B29 B30 B31 B32 B33 B34 B35 B36 B37
B36 0.01 0.01 0.01 0.04 0.01 0.04 −0.02 0.04 0.04 0.04 −0.04 0.04 0.03 −0.02 0.01 0.01 0.04 0.04 0.04 0.01 0 0.02 0.03 −0.03 0.04 −0.02 0.02 −0.01 0.06 −0.01 −0.02 0.19 −0.20 0.06 0.01 1.00
B37 −0.02 −0.02 −0.02 0.02 0.01 0.01 0.03 0.02 0.02 0.02 −0.02 0.02 0.01 0.01 −0.01 −0.02 0.02 0.03 0.03 −0.03 −0.03 −0.01 0.01 −0.01 0.02 −0.02 −0.02 0.15 0 0.02 0.02 0.67 −0.66 0.02 0.03 0.27 1.00
B1 = Quickbird band 1, B2 = Quickbird band 2, B3 = Quickbird band 3, B4 = Quickbird band 4, B5 = ARVI, B6 = BNDVI, B7 = DVI, B8 = GNDVI, B9 = IPVI, B10 = NDVI, B11 = NDWI, B12 = SAVI, B13 = WRI, B14 = B1/B2, B15 = B3/B1, B16 = B3/B2, B17 = B4/B1,
B18 = B4/B2, B19 = B4/B3, B20 = Texture (Contrast), B21 = Texture (Correlation), B22 = Texture (Dissimilarity), B23 = Texture (Entropy), B24 = Texture (Homogeneity), B25 = Texture (Mean), B26 = Texture (2nd moment), B27 = Texture (Variance), B28 = Aspect, B29 =
Berhane et al.
Cross-sectional convexity, B30 = DEM, B31 = Longitudinal convexity, B32 = Maximum curvature, B33 = Minimum curvature, B34 = Plain convexity, B35 = Profile convexity, B36 = RMS error, B37 = Slope (%).
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Page 28
Berhane et al. Page 29
Table 3
McNemar’s chi-squared test summary of the classification accuracy differences observed by the three
EPA Author Manuscript
classifiers. No significant differences were found between the three classification methods we employed when
using either a three- or five-layer stack.
Chi-Squared p-Value
Classifier
Pixel-Based ML Object-Based RF Pixel-Based ML Object-Based RF
Three-Layer Stack
Five-Layer Stack
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Berhane et al. Page 30
Table 4
Pixel-based random forest classification accuracy on training and testing datasets with various input variable
EPA Author Manuscript
combinations (including 95% confidence interval, CI). The data are sorted based on overall accuracy of the
testing data.
Testing Data
Training Data
Predictor Variables Overall Accuracy
OOB Error (%)
% 95% CI
All non-correlated variables (with WRI, not mean texture; 22 variables) 1.3 72.7 72.6 72.9
All non-correlated variables (with mean texture, not WRI; 22 variables) 0.3 80.1 80.0 80.2
All non-correlated variables including both WRI and mean texture (23 variables) 0.4 80.1 80.1 80.2
All (37) variables 0.8 84.6 84.5 84.7
Ten most important variables 1.0 84.6 84.5 84.6
Five most important variables 0.9 84.9 84.8 85.0
Fifteen most important variables 1.2 85.6 85.5 85.8
Most parsimonious model (three variables: B3, WRI, and mean texture) 1.4 87.9 87.8 88.0
EPA Author Manuscript
EPA Author Manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
EPA Author Manuscript EPA Author Manuscript EPA Author Manuscript
Table 5
Pixel-based random forest classification class confusion matrix (pixel-counts) for genus-level wetland classes and aquatic habitats (three-layer stack). PA,
UA, and OA are producer’s, user’s, and overall accuracy, respectively. See the legend in Figure 5a for additional information regarding the wetland class
community composition.
Berhane et al.
Wetland Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
1.00 55 1.00 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 0 58 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 5 0 21 0 0 0 9 0 0 0 0 0 0 0 0 0 0 0
4 0 0 30 68 0 0 0 0 0 0 0 0 0 0 0 0 0 0
5 0 0 12 0 63 19 0 0 0 0 0 0 0 0 0 0 0 0
6 0 0 0 0 0 40 0 0 0 0 0 0 0 0 0 0 0 0
7 1.00 0 0 0 0 0 51 0 9 6 0 0 0 0 0 0 0 0
8 0 0 0 0 0 2 0 58 0 5 0 0 0 0 0 0 0 0
9 0 0 0 0 0 0 0 0 53 0 0 0 0 15 0 0 0 0
10 0 0 0 0 0 0 0 2 0 51 0 0 0 0 0 0 0 0
11 0 0 0 0 0 0 0 0 1.00 0 62 0 0 0 0 0 0 0
12 0 0 0 0 0 0 0 0 0 0 0 61 0 0 0 0 0 0
13 0 0 0 0 0 0 0 0 0 0 0 0 84 0 0 0 0 0
14 0 0 0 0 0 0 0 0 0 0 2 0 0 40 0 0 0 0
15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 65 0 2 0
16 0 0 0 0 0 0 0 0 0 0 0 3 0 7 0 44 3 0
17 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 53 0
18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 56
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
PA (%) 90.2 98.1 33.1 100 100 65.5 84.3 97.0 84.2 82.9 96.9 95.3 100 64.4 100 100 90.0 100
UA (%) 97.2 100 59.7 69.3 67 100 76.8 89.3 78.0 96.7 98.4 100 100 95.2 97.0 77.1 100 98.3
OA (%) 87.9
Page 31
EPA Author Manuscript EPA Author Manuscript EPA Author Manuscript
Table 6
The importance of each predictor variable (based on MDG) for both pixel-based RF classification approaches (cumulative count of importance out of 100
iterations; only the top 20 important variables are included). Abbreviations are found within the text and Table 1.
Predictor
1st 2nd 3rd 4th 5th 6th 7th 8th 9th 10th 11th 12th 13th 14th 15th
Variable
Quickbird B1 - - - - - - - 1.00 5 12 30 45 5 2 -
Quickbird B2 - - - - 100 - - - - - - - - - -
Quickbird B3 - - 1.00 99 - - - - - - - - - - -
Quickbird B4 - 100 - - - - - - - - - - - - -
ARVI - - - - - - - - - - - - - - -
B1/B2 - - - - - - - - - - - - - - -
BNDVI - - - - - 51 49 - - - - - - - -
DVI - - - - - - - 57 18 19 3 3 - - -
GNDVI - - - - - - - 12 24 25 25 14 - - -
IPVI - - - - - - - - - - - 1.00 10 18 20
NDVI - - - - - - - - - - - - 5 20 22
B4/B1 - - - - - 49 51 - - - - - - - -
B4/B2 - - - - - - - 14 25 28 15 17 1.00 - -
B4/B3 - - - - - - - - - - - - 6 24 19
B3/B1 - - - - - - - - - - - 8 68 19 4
B3/B2 - - - - - - - - - - - - - - -
SAVI - - - - - - - - - - - - 4 17 35
WRI - - 99 1.00 - - - - - - - - - - -
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
Mean texture 100 - - - - - - - - - - - - - -
NDWI - - - - - 1 - 16 28 16 27 12 1 - -
Page 32
EPA Author Manuscript EPA Author Manuscript EPA Author Manuscript
Table 7
Pixel-based ML classification class confusion matrix (pixel-counts) for genus-level wetland classes and aquatic habitats (three-layer stack). PA, UA, and
OA are producer’s, user’s, and overall accuracy, respectively. See the legend in Figure 5a for additional information regarding the wetland class
community composition.
Berhane et al.
Wetland Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
1.00 51 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 0 55 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 9 4 44 0 20 10 0 0 0 0 0 0 0 0 0 0 0 0
4 0 0 17 68 0 0 0 0 0 0 0 0 0 0 0 0 0 0
5 0 0 2 0 42 0 0 0 0 0 0 0 0 0 0 0 0 0
6 0 0 0 0 1.00 29 0 0 0 0 0 0 0 0 0 0 0 0
7 1.00 0 0 0 0 21 61 0 20 1.00 0 0 0 0 0 0 0 0
8 0 0 0 0 0 1.00 0 54 0 0 0 0 0 0 0 0 0 0
9 0 0 0 0 0 0 0 0 24 0 1.00 0 0 27 0 0 0 0
10 0 0 0 0 0 0 0 6 0 61 0 0 0 0 0 0 0 0
11 0 0 0 0 0 0 0 0 19 0 62 0 0 0 0 0 0 0
12 0 0 0 0 0 0 0 0 0 0 0 58 2 0 0 0 0 0
13 0 0 0 0 0 0 0 0 0 0 0 6 82 0 0 0 3 0
14 0 0 0 0 0 0 0 0 0 0 1.00 0 0 32 4 0 0 0
15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 61 0 0 0
16 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 44 0 0
17 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 55 0
18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 56
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
PA (%) 83.6 93.2 69.8 100.0 66.7 47.5 100.0 90.0 38.1 98.4 96.9 90.6 97.6 51.6 93.8 100.0 93.2 100.0
UA (%) 100.0 100.0 50.6 80.0 95.5 96.7 58.7 98.2 46.2 91.0 76.5 96.7 90.1 86.5 100.0 93.6 100.0 98.2
OA (%) 83.9
Page 33
Berhane et al. Page 34
Table 8
Object-based random forest classification accuracy on training and testing datasets with various input variable
EPA Author Manuscript
combinations. Quickbird bands 2, 3, and 4 are represented as B2, B3, and B4, respectively; also included are
WRI (Water Ratio Index) and mean texture.
Testing Data
Training Data
Predictor Variables Overall Accuracy
OOB Error (%)
% 95% CI
Scale 5
0.3 90.4 90.3 90.4
B2 + B3 + B4 + WRI + mean texture
EPA Author Manuscript
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
EPA Author Manuscript EPA Author Manuscript EPA Author Manuscript
Table 9
Object-based random forest classification (segmentation scale of 5) confusion matrix (pixel-counts) for genus-level wetland classes and aquatic habitats
(three-layer stack). PA, UA, and OA are producer’s, user’s, and overall accuracy, respectively. See the legend in Figure 5a for additional information
regarding the wetland class community composition.
Berhane et al.
Wetland Class 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
1.00 60 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
2 0 59 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 1.00 0 37 0 0 0 33 0 0 0 0 0 0 0 0 0 0 0
4 0 0 25 67 0 0 0 0 0 0 0 0 0 0 0 0 0 0
5 0 0 0 1.00 63 0 0 0 0 0 0 0 0 0 0 0 0 0
6 0 0 1.00 0 0 46 0 5 0 0 0 0 0 0 0 0 0 0
7 0 0 0 0 0 0 28 0 0 0 0 0 0 0 0 0 0 0
8 0 0 0 0 0 15 0 55 0 36 0 0 0 0 0 0 0 0
9 0 0 0 0 0 0 0 0 62 0 0 0 0 21 0 0 0 0
10 0 0 0 0 0 0 0 0 0 26 3 0 0 0 0 0 0 0
11 0 0 0 0 0 0 0 0 0 0 59 0 0 0 0 0 0 0
12 0 0 0 0 0 0 0 0 0 0 0 63 0 0 0 0 0 0
13 0 0 0 0 0 0 0 0 0 0 0 0 84 0 0 4 0 0
14 0 0 0 0 0 0 0 0 1.00 0 2 0 0 41 0 0 0 0
15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 65 0 13 0
16 0 0 0 0 0 0 0 0 0 0 0 1.00 0 0 0 40 0 0
17 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 36 0
18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 11 56
Remote Sens (Basel). Author manuscript; available in PMC 2018 April 26.
PA (%) 98.4 100.0 58.4 98.5 100.0 75.4 45.9 91.7 98.4 41.9 92.2 98.4 100.0 66.1 100.0 90.9 60.4 100.0
UA (%) 100.0 100.0 52.0 72.7 98.4 88.5 100.0 52.0 74.7 89.7 99.7 100.0 95.5 93.2 83.8 97.6 100.0 83.9
OA (%) 84.6
Page 35