IEEE Geoscience and Remote Sensing Magazine - September 2022 - 267

◗ Fewer than 100 pixels would hardly be statistically
representative.
◗ To assess how fast AL methods improve classification
performance and to evaluate their capacity to reduce annotation
cost, the lower the number of pixels selected at
each step, the better. For instance, it would be very interesting,
from an operational context, to know that only
one step querying 100 pixels is enough to reach high
performance instead of one step querying 200 pixels,
Impact of Preprocessing Methods on AL Methods
To assess the influence of preprocessing methods on Core Set and Hierarchical
performances, we carried out ten independent runs on Indian Pines
◗ without preprocessing
◗ with subsampling (25% of the unlabeled pool)
◗ with our clustering method (using 500 clusters).
◗ For a given budget of pixels to label, having few pixels at
each step allows benefiting from an updated measure of
the acquisition function.
We recall that at each step t, a classifier is trained from
scratch on the current training data set L ,t and that new
pixels are added to the next training data set L . For each
t 1+
benchmark data set, we performed a fivefold cross validation
with different splits of the ground truth to assess AL
algorithms' sensitivity to the initial training data set. We ran
The results in Figure S1 show that both preprocessing techniques had a low
influence on classification performance. Subsampling and clustering slightly
increased the performance of Core Set. We think that in this case, it may have
improved the robustness of Core Set to outliers. As a matter of fact, by taking
only a subsample of the pool, there are fewer outliers. Similarly, the preclustering
0.525
0.475
0.425
0.375
0.35
0.4
0.45
0.5
13 5
Steps
(a)
0.525
0.5
0.475
0.45
0.425
0.375
0.4
13 5
Steps
(c)
hierarchical_subsample
hierarchical_cluster
hierarchical_all
79
0.32
0.34
0.36
0.38
0.4
0.42
0.44
0.46
0.48
13 5
Steps
(d)
FIGURE S1. Accuracy metrics (mean and standard deviation over 10 runs) over the first 10 steps of the Core Set and Hierarchical methods on
the Indian Pines data set. coreset_all/hierarchical_all corresponds to Core Set/Hierarchical without preprocessing. coreset_subsample/hierarchical_subsample
corresponds to Core Set/Hierarchical applied on a random subsample of the pool (25% of the total number of pixels in
the pool). coreset_cluster/hierarchical_cluster corresponds to Core Set/Hierarchical applied after the preprocessing routine with 500 clusters.
(a) OA-Core Set. (b) mIoU-Core Set. (c) OA-Hierarchical. (d) mIoU-Hierarchical.
(Continued )
SEPTEMBER 2022 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
267
core set_subsample
core set_cluster
core set_all
79
0.475
0.45
0.425
0.4
0.375
0.35
0.325
0.3
13 5
Steps
(b)
core set_subsample
core set_cluster
core set_all
79
hierarchical_subsample
hierarchical_cluster
hierarchical_all
79
OA
OA
mloU
mloU

IEEE Geoscience and Remote Sensing Magazine - September 2022

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