IEEE Geoscience and Remote Sensing Magazine - March 2023 - 110

We perform ASOS according to [21] to create highresolution
sensitivity maps of our test samples to assess
whether specific regions are sensitive toward wilderness
or nonwilderness. The authors provide neural network
architecture, a specific training procedure, and the approach
for sensitivity analysis. We proceed in the same
way and refer to their work for a detailed description of
the methodology.
The neural network consists of a modified form
of the U-Net of [33] and a simple classifier network,
shown in Figure 7. The activation map at the interface
of these two networks has three channels and is the
same size as the input image. The model has about
2 million parameters.
CLASSIFICATION
We run the same training procedure as in [21] and use a
maximum learning rate of 0.01, a weight decay of 0.0001,
and a batch size of 32. We train the model for 50 epochs on
an Nvidia Quadro RTX 4000 (8 GB GDDR6) for about 6 h.
The achieved overall classification accuracies are 92%
for the training set, 82% for the validation set, and 74% for
the test set. The confusion matrix of the test set is shown
in Table 3.
SENSITIVITY ANALYSIS
After the model has been trained, we run the ASOS analysis
on all correctly classified training samples. For this, specific
pixels are occluded in the activation maps by setting them
to zero. This changes the classification
scores, and the deviations are a
measure of the sensitivities of the occluded
parts. The occlusions are determined
from the activation space
in which each of the three axes corresponds
to one of the three channels
of the activation maps. Pixels that are
close to each other in the activation
space are occluded simultaneously.
The resulting sensitivities are visualized
in Figure 8. Low-density areas
are not included in this mapping.
Knowing the sensitivities, we can
predict sensitivity maps for any input
image. A selection of the test samples
is shown in Figure 9.
Wilderness Proxy
Wilderness Area Background
FIGURE 6. Baseline results for the test images with the IDs 351819, 39516, 19442, and
900000068. Row-wise, from left to right: input image, wilderness proxy, and model's output.
Input images are the single temporal subsets of the Sentinel-2 images for the corresponding IDs.
110
DISCUSSION
With the semantic segmentation experiment,
we study the concept of wilderness
on a pixel scale. The model used
in this experiment outputs densely predicted
segmentation maps for a given
test image with a shape of 1,920 × 1,920,
as shown in Figure 6. The first sample
shows parts of the Rio Novo National
Park, where the park is seen as divided
in two by a river. We see that
the model predicts the whole patch
(except for the deforestation areas) as
wilderness, while the WDPA proxy
contains annotation only on the right
side of the river. The second sample
shows the Phu Pha Man District in
Thailand. Here, the proxy loosely
aligns with the models' output. The
output segmentation map here filters
out the areas with a human influence
(i.e., a suburban area surrounded by
cultivated lands). The third sample
is located at Santa Bárbara Ecological
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE MARCH 2023

IEEE Geoscience and Remote Sensing Magazine - March 2023

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2023