IEEE Geoscience and Remote Sensing Magazine - March 2022 - 306

notable difference between the annual and seasonal data
sets. For the Sudan site, the DNN and kNN had the highest
recall rates for the Present (1) class, while the SVM had the
fewest false negatives for the Absent (0) class. Similar to
the precision results, there were inconsistencies between
model performance for predicting each class in the study
areas. In Niger, the DNN had the highest average for both
classes, while the kNN and DT had comparable results in
Sudan. Again, the best-performing models tended to have
the highest average recall rates regardless of the absence
generation method, and the environmental normalization
absence data sets outperformed the others.
The F-1 score calculates the precision and recall for a
model and applies a penalty if there is an imbalance between
the two values. An ideal model would have high
positive predictive performance and avoid erroneously labeling
false negatives, so this metric can offer great insight
into total classification performance. The DNN achieved
the highest F-1 scores for the Niger site, while the kNN was
the most effective in the Sudan study area (Table 8). The
models were successful using multiple combinations of
absences and temporal data sets. Interestingly, the model
discrepancies for the prediction of each class that were
noted in the individual precision and recall results were
not observed in the F-1 score. These models were the best
at predicting both classes in their respective study areas.
This may suggest that they offer the best overall classification
performance for correctly identifying suitable habitat.
Also, while the successes were regionally specific, the
differences between the models in each area were not significant.
Although it did not have the top performance,
the SVM tended to best separate the classes with the geographic
exclusive and random absences. This may suggest
that the ±2 standard deviations used to generate the environmental
normalization absences did not provide a wide
enough margin for the SVM model to divide the two classes.
The tree-based models were not effective at defining
either class in the data sets, compared to the other models.
The AUC metric differs from the model accuracy score in
that instead of evaluating a model's ability to predict a particular
class, we compare a model's score above a defined
threshold (0.5). Scores below the threshold are considered the
negative class/background data, and scores above the threshold
represent the probability that a sample is the positive class/
locust habitat. In both locations, MaxEnt had a clear advantage
over the other models when using the random absence
generation method (Tables 9 and 10). For the remaining data
sets, the DNN had the highest AUC scores at the Niger study
site, while the kNN preformed the best in Sudan. The model
with the lowest values in both areas was the DT.
TABLE 7. THE MODEL RECALL RESULTS AND AVERAGE PSEUDOABSENCE PERFORMANCE.
STUDY AREA: NIGER (1 = PRESENT, 0 = ABSENT)
Annual EnvNrm_1
Annual EnvNrm_0
Annual GeoExcl_1
Annual GeoExcl_0
Annual_Random_1
Annual_Random_0
Winter_EnvNrm_1
Winter_EnvNrm_0
Winter_GeoExcl_1
Winter_GeoExcl_0
Winter_Random_1
Winter_Random_0
Summer_EnvNrm_1
Summer_EnvNrm_0
Summer_GeoExcl_1
Summer_GeoExcl_0
Summer_Random_1
Summer_Random_0
Mean_EnvNrm_1
Mean_EnvNrm_0
Mean GeoExcl_1
Mean GeoExcl_0
Mean Random_1
Mean Random_0
kNN
0.9
0.93
0.84
0.84
0.82
0.74
0.97
0.92
0.97
0.92
0.9
0.84
0.9
0.97
0.9
0.98
0.81
0.73
0.92
0.94
0.9
0.91
0.84
0.77
DT
0.89
0.91
0.79
0.78
0.74
0.7
0.95
0.94
0.95
0.9
0.84
0.82
0.91
0.92
0.9
0.93
0.73
0.69
0.92
0.92
0.88
0.87
0.77
0.74
SVM
0.93
0.82
0.91
0.8
0.91
0.62
0.94
0.92
0.89
0.87
0.67
0.84
0.91
0.97
0.89
0.87
0.84
0.78
0.93
0.9
0.9
0.85
0.81
0.75
RF
0.91
0.93
0.78
0.78
0.78
0.73
0.97
0.95
0.93
0.95
0.85
0.87
0.91
0.95
0.91
0.93
0.79
0.76
0.93
0.94
0.87
0.89
0.81
0.79
DNN
0.92
0.94
0.84
0.85
0.88
0.67
0.9
Annual_EnvNrm_1
Annual_EnvNrm_0
Annual_GeoExcl_1
Annual_GeoExcl_0
Annual_Random_1
Annual Random 0
0.98 Winter_EnvNrm_1
Winter_EnvNrm_0
0.93 Winter_GeoExcl_1
Winter_GeoExcl_0
0.9
0.89 Winter Random_1
0.82 Winter Random_0
Summer EnvNrm_1
Summer EnvNrm_0
Summer GeoExcl_1
Summer GeoExcl_0
Summer Random_1
Summer Random_0
0.91
1
0.9
0.98
0.84
0.76
0.94 Mean EnvNrm_1
0.95 Mean EnvNrm_0
Mean GeoExcl_1
Mean GeoExcl_0
0.89
0.91
0.87 Mean Random_1
0.75 Mean Random_0
STUDY AREA: SUDAN (1 = PRESENT, 0 = ABSENT)
kNN
0.91
0.92
0.84
0.8
DT
0.86
0.75
0.99
0.94
0.97
0.97
0.86
0.85
0.97
1
0.85
0.74
0.82
0.71
0.96
0.95
0.89
0.84
0.85
0.77
0.87
0.86
0.75
0.82
0.76
0.8
0.97
0.92
0.75
0.82
0.76
0.8
0.95
0.98
0.97
0.97
0.86
0.85
0.93
0.92
0.82
0.87
0.79
0.82
SVM
0.79
0.96
0.94
0.73
0.84
0.8
0.96
0.94
0.96
0.94
0.64
0.94
0.97
1
0.87
0.62
0.79
0.62
0.91
0.97
0.92
0.76
0.76
0.79
RF
0.9
0.92
0.75
0.82
0.84
0.85
0.99
0.93
0.98
0.97
0.88
0.89
0.95
0.98
0.87
0.69
0.74
0.69
0.95
0.94
0.87
0.83
0.82
0.81
DNN
0.9
0.91
0.9
0.79
0.81
0.73
0.98
0.92
0.98
1
0.92
0.87
0.97
1
0.9
0.74
0.74
0.62
0.95
0.94
0.93
0.84
0.82
0.74
306
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
