OSDA-ST: Advancing Open Set Domain Adaptation Through Selective Thresholding in Remote Sensing

Segmenting and classifying satellite and aerial images is a crucial task for understanding various environmental, urban, and agricultural phenomena in remote sensing. Traditional segmentation methods often rely on large, labeled datasets, which can be challenging to obtain for every possible domain or geographic region. Unsupervised Domain adaptation (UDA) attempts to solve this problem by transferring pixel-wise semantic knowledge from a labeled source domain to an unlabeled target domain. However, current UDA methods for remote sensing assume a corresponding label space between domains, which is often unrealistic in real-world scenarios where the target domain may contain novel or unknown classes that are absent in the source domain. This paper addresses the limitations associated with Open-Set Domain Adaptation for Semantic Segmentation (OSDA-SS) by proposing OSDA-ST, a framework performing max-logit score (MLS) thresholding, source-only pretraining, and a dynamic source selection score for unknown class estimation. We test these modifications on the remote sensing dataset of Potsdam and Vaihingen for the first time. Compared to state of the art, our methods achieved 39.9 H-score, a significant improvement of 5.82% and 7% on average. Results indicate that these modifications are successful in reducing misclassification of pseudo-labels thereby providing an adaptive approach to thresholding as compared to previous state of the art.