IEEE Geoscience and Remote Sensing Magazine - March 2022 - 290

OPPORTUNITIES
Ecological models typically use a combination of species occurrence
data and relevant environmental information to make
predictions about suitable habitat for a species [1]. Further,
these models can use information about habitat suitability
to predict distributional shifts in response to environmental
change [2]. The central assumption of ecological models
is that the observed distribution of a species provides useful
information about the environmental requirements of
that species. Most ecological models are restricted to building
correlative relationships between the species and data
within a theoretical environmental feature space, and the
representations are not always fully representative of actual
geographic space [3]. The reliability and interpretation of
model results depend heavily on the methods used to account
for the presence/absences data, the accuracy of the
environmental variables employed as predictors, and the
methods harnessed to parameterize and evaluate the representations
[4]. Satellite remote sensing can be used to
gather environmental data across entire landscapes, with
spatial and temporal coverage that is generally unattainable
with traditional field methods. The ability to observe
biophysical variables continuously and without the bias of
interpolation has transformed the way researchers examine
habitats and model ecological processes [5], [6].
Remote sensing can be used to measure biophysical
characteristics of a habitat because of the physical relationship
between the features and the radiance detected by instruments
in different spectral regions [7]. Many of these
biophysical variables can be directly measured by remote
sensing systems without the need for ancillary data, such as
soil moisture (SM) content, using microwave remote sensing,
and the leaf area index, with optical remote sensing
[8]. In many cases, remote sensing provides the only means
of estimating biophysical characteristics of habitats across
broad areas [9], a perspective that is critical for detecting
environmental changes that result from human and natural
processes [5]. The effective integration of species observations
and remote sensing data relies heavily on the scale
and quality of the information being input into a model
[10]. Predictions are optimized when remotely sensed environmental
data used to build and train a model accurately
represent suitable habitat conditions for a species [11],
[12]. While there are clear advantages for assimilating satellite
remote sensing and ecological niche modeling, there
has been limited collaborative research between the fields
to ensure that both communities understand the respective
limitations and important fundamental assumptions
about the data and models [10]. There is an opportunity for
more coordinated and process-based research that explicitly
examines the challenges of combining remote sensing
data and ecological models for habitat prediction [5].
FRAMEWORK AND OBJECTIVES
Acknowledging this research gap, we sought to conduct
the first comprehensive review of the integration of remote
290
sensing and machine learning to predict the habitat of the
world's most destructive migratory pest, the desert locust
(Schistocerca gregaria) [13]. The habitat for the desert locust
can be ephemeral [14] and extends across numerous countries
in Northern Africa, the Arabian Peninsula, and Southwest
Asia. The combined effect of optimal weather events,
soil conditions, and vegetation dynamics can promote desert
locust outbreaks characterized by the rapid reproduction
and destructive aggregation of otherwise harmless,
solitary insects [15]. The expansive range of the desert locust
has made satellite remote sensing the primary tool for
monitoring potential habitat [16]. In recent years, several
studies have combined remote sensing data and machine
learning models to make predictions about potential desert
locust habitat. The focus of that research has been to use
machine learning to identify the critical value range for remotely
sensed environmental indicators to show a correlative
relationship with recorded locust observations. Various
combinations of algorithms, absence generation methods,
and remote sensing explanatory variables have been used
to make predictions with sometimes inconsistent results.
Central to our research objectives was to explore whether
the strengths of deep learning offered a new opportunity for
predicting suitable habitat for the desert locust. Deep learning
models are designed to enhance the efficacy of feed-forward
artificial NNs, which have been shown to better capture nonlinear
relationships when modeling environmental phenomena
than traditional models [17], [18]. Additionally, in
contrast to traditional machine learning approaches, deep
learning models have the capacity to self-learn important
latent relationships within a data set [19]. Since all previous
habitat classification research has focused on using traditional
machine learning methods to identify correlative
relationships between locust observations and remotely
sensed habitat variables, our research is distinct in that it
attempts to use deep learning to extract additional hidden
features within habitat variables for improved classification
at a regional scale.
To ensure a robust analysis of the available remote sensing
data and ecological models, we evaluated the predictive
performance of six machine learning models in two separate
study regions, using desert locust observations and remotely
sensed habitat variables during a period from 2009
to 2019 (Figure 1). Our models included the traditional
kNN, decision tree (DT), support vector machine (SVM),
random forest (RF), and maximum entropy (MaxEnt). Additionally,
to examine potential advantages of deep learning
in comparison, we employed multiple sequential DNN
models of varying architectures. Since numerous methods
have been used in past research to account for the absence
data class in machine learning models, we generated three
types of pseudoabsences by using methods common for
ecological modeling to observe how each impacted classification
performance. Also, we selected one model (MaxEnt)
that required only presence data to predict suitable habitat
to account for any potential bias in our generation methods.
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2022
IEEE Geoscience and Remote Sensing Magazine - March 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2022
