IEEE Geoscience and Remote Sensing Magazine - March 2022 - 308

DISCUSSION
DIFFERENCES IN MODEL PERFORMANCE
The performance of a machine learning algorithm is strongly
influenced by the data and parameters used to conduct the
analysis [87]. Therefore, assertions about the applicability
of an algorithm for studying the potential distribution
of a specific species should be cautiously done. Moreover,
performance is a function of the metric used to define it.
We were able to achieve similar accuracy results in the two
regions, with the same best- (kNN and DNN) and lowest(DT)
performing models. However, for classifying suitable
habitat, the aim is to find a model that can accurately distinguish
the Present (1) class from absence or background
data, and the best-performing models were different for
these metrics. Across the metrics, the DNN was the best
performing for the Niger study site, while the kNN was the
best in Sudan. While each had clear advantages per study
area, the models still represented the top two for numerous
metrics in both locations, suggesting that they may offer
the highest potential for distinguishing between the two
classes across multiple regions. The kNN and DNN models
are arguably the simplest and most complicated models
that were tested and have very different predictive approaches
to solving the classification problem.
For each sample, the kNN can build local approximations
for defining the optimal classification based on what
it learns from memorizing the data set. This means the
model can decipher complex relationships within data
by solving multiple local calculations of low complexity
[101]. While the sample sizes were comparable between
study areas, the Sudan site had fewer observations during
the project period. The reduced number of samples may
have contributed a slight advantage to the model, as the
nature of kNNs can benefit from smaller data sets [86].
The optimal performance in both study regions seems to
suggest that the remote sensing variables that were used
demonstrated an environmental gradient at the present
locations and that outliners were not a significant factor
for the models. When evaluating the temporal relationship
between coarse SM and desert locust hoppers, Gomez et al.
[32] had similar results, with weighted kNNs and RFs outperforming
other models at a continental scale. This may
suggest that kNNs could be successfully applied in other
regions and across varying scales to predict suitable habitat
for the desert locust.
Deep learning models have demonstrated that they can
outperform shallow and more traditional machine learning
models when classifying complex, nonlinear, and redundant
data sets [102], [103]. Increasingly, deep learning
models have been used to predict habitat for numerous
species [104]-[106], as their capacity to understand intricate
and expansive dynamics within an environmental feature
space offers advantages over other machine learning
models [93]. However, the relationship between the sample
size and the depth of a network can vary across application
308
areas and scales of measurement [107]. In our work, the
optimal architectures for the DNN varied across the data
sets and study areas, limiting the assumptions that can be
made regarding deeper architectures being better equipped
to predict suitable habitat for the desert locust. When DNNs
outperformed the others, the majority were models with
three or more hidden layers; however, there were a few bestperforming
DNNs that needed only one or two hidden layers
to produce superior results. Interestingly, for DNNs, the
model GridSearch results optimized the logistic regression
cost function for 14 of the 18 evaluated data sets. Logistic
regression is effectively a linear model for classification
that is also referred to as the logit model or MaxEnt classifier.
DNNs use the function as a weight optimization solver to
determine the probability of occurrence for a class by building
a logit variable that contains the natural log of the odds
of the class occurring or not. A maximum-likelihood estimation
algorithm is then applied to estimate the probabilities
[101]. Piou et al. [26] and Chen et al. [35] obtained high
scores when using a logistic regression model independent
of a DNN in recent research to predict suitable habitat for
the desert locust. While both the DNN and kNN performed
well, a major disadvantage for the DNN was its extensive
computation time compared to the kNN. The fitting of multiple
architectures and parameters for each DNN required
4+ h to run on a personal workstation, whereas the kNN
could be processed in a few minutes, as it had significantly
fewer parameters to assess.
For the AUC scores, MaxEnt had a clear advantage over
the other models when using the random pseudoabsence
generation method. A key distinction of the MaxEnt model
is that unlike the random pseudoabsence generation
method that creates random samples equal to the number
of presence samples, the MaxEnt model oversamples the
background with 10,000 random points. Therefore, during
model training, the MaxEnt model had more opportunities
to learn what is not a presence observation based on the
values in the background data. The imbalance between the
classes may have accounted for this increased performance
[30]. For the MaxEnt application, raster grid files for the explanatory
variables can be directly input so that relationships
between locust observations and environmental indicators
can be mapped throughout an entire study region,
and the relative probability of occurrence can be depicted
(Figures 7 and 8). The maps show the average predicted
probability for suitable habitat in each region for the winter
and summer data sets, with warmer colors defining areas
where the probabilities are highest. In both study regions,
there were marked differences in the location and prevalence
of suitable habitat between the seasons. Using annual
data sets to generate the models may not be able to differentiate
the contributions of each season, so future research
should consider primarily using seasonal or monthly data
sets to evaluate correlative relationships.
Unlike the other models that used PCA to address collinearity
between the environmental variables, MaxEnt is
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
