IEEE Geoscience and Remote Sensing Magazine - March 2022 - 314

processing to complete. Despite these restrictions, continued
improvements and new capabilities will lead to the
increased use and relevance of cloud computing in remote
sensing. There may also be opportunities to combine the
strengths of GEE with the high predictive performance of
open-source platforms, such as H2O Automated Machine
Learning [33]. This web-based platform automates many
important machine learning steps, including algorithm selection,
feature generation, hyperparameter tuning, iterative
modeling, and model assessment. Given a training data set,
it will apply multiple models with different combinations
of parameters and build two ensemble models from all the
models and the highest-performing models. The ensembles
may be better equipped to distinguish classes and ensure
more accurate habitat classifications across an extended
range for the desert locust.
CONCLUSION
We have delineated several important considerations when
integrating satellite remote sensing data and environmental
models within the context of data science best practices.
Through our evaluation of six models in two study areas,
we demonstrated the strengths and weakness of the models
for predicting desert locust habitat at a regional scale. The
best-preforming models across numerous evaluation metrics
were the kNN and the DNN. In the Niger study area,
the kNN had an average accuracy score of 88%; precision,
recall, and F-1 scores of 89%; and an AUC of 88% across
nine data sets. The DNN had an average accuracy score of
88%, a precision score of 88%, a recall score of 90%, an F-1
score of 89%, and an AUC of 88%. In the Sudan study area,
the kNN had an average accuracy score of 88%, a precision
score of 87%, a recall score of 90%, an F-1 score of 88%, and
an AUC of 88% across the data sets. The DNN had an average
accuracy of 88%, a precision score of 86%, a recall score
of 90%, an F-1 score of 89%, and an AUC of 88%. MaxEnt
had the highest average AUC score, 92%, but seemed to
gain this advantage based on the less optimal performance
of the other models using the generated random absences.
It is worth noting that the DT and RF models may have
been disadvantaged by our application of PCA. The bootstrapping
nature of these models is generally robust to multicollinearity
in environmental variables, so the transformation
was not required. We applied PCA to keep the data
sets consistent across models, but performance may have
changed if the tree models had the original five variables for
the classifications. For the three temporal ranges, the summer
and winter data sets performed better than the annual
data sets, although overall none of the data sets performed
poorly. The environmental normalization and geographic
exclusion pseudoabsences resulted in higher precision and
recall results for the classifications when compared to the
generated random absences. For the top-performing models,
the absence generation method did not appear to significantly
influence classification capacity. However, for
several of the data sets, we observed that the SVM showed
314
improved performance with the geographic and random
absences compared to the environmental absences.
The good results for the kNN and DNN at both study sites
suggest that the models may be options for predicting suitable
habitat in other regions impacted by the desert locust.
However, if the goal is to build one model that can be trained
and transferred to other areas, the advantage resides with the
DNN. Once we have a trained DNN, the model can be used
to make predictions with new data in other locations. However,
the kNN simply memorizes a data set and then classifies
based on similarities between samples. This does not
facilitate the development of a transferable model because
each model is trained only to the data set it memorizes. One
potentially significant challenge for using DNNs for habitat
classification is that because each node within a model
acts as a feature extractor creating new variables to learn the
data, we can no longer identify the contribution of specific
habitat variables to the overall classification. This ambiguity
can limit the interpretation of results and potentially obscure
useful information for locust management. Whether we can
accept a model's good results without knowledge of the exact
structure of the function used to obtain them is probably
situationally specific. However, this constraint is an important
consideration for any future research that may incorporate
variables beyond ecological indicators, such as land
use change and socioeconomic data, into models to predict
locust habitat. This may require DNN models with featurebased
multilevel architectures to help elucidate which types
of variables most influence model results within data fusion.
Another important consideration is the limitation of a
regional-scale analysis for mapping species habitat. When we
are attempting to decipher habitat patterns with coarser remote
sensing data, we effectively create a spatial mismatch
as we increase the grain size of a species' relationship to its
environment in the model. We achieved good classification
results for all the models, so there must have been enough
patterns within the regional-scale data for the models to distinguish
between the two classes. However, it is important
to remain cognizant of the fact that coarse, remotely sensed
variables homogenize landscapes within a pixel to a degree
that may not fully represent the actual environment where
the in situ locust data were captured. This oversimplification
of the landscape may in some ways make it easier for models
to draw relationships between samples and pixel values, particularly
with a species like desert locust hoppers that tends to
have similar distributions from year to year within recession
areas. One advantage that our project had for combatting
this was an extended study period (2009-2019), which may
have enabled more natural variation within the landscapes
to be expressed in the annual and seasonal data sets. Testing
the performance of the trained models with the same locust
data but finer resolution remote sensing variables may offer
insight into the influence that scale had on the model results.
There are many questions remaining about the influence
of rising global temperatures on future invertebrate pest populations
and the impact that increased cyclone activity in the
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
