SemiSiROC: Semisupervised Change Detection With Optical Imagery and an Unsupervised Teacher Model
Lukas Kondmann, Sudipan Saha, Xiao Xiang Zhu
Abstract
Change detection is an important yet challenging task in remote sensing. In this paper, we underline that the combination of unsupervised and supervised methods in a semi-supervised framework improves change detection performance. We rely on Half-Sibling Regression for Optical Change Detection (SiROC) as an unsupervised teacher model to generate pseudo labels and select only the most confident pseudo labels for pretraining different student models. Our results are robust to three different competitive student models, two semi-supervised pseudo label baselines, two benchmark datasets and a variety of loss functions. While the performance gains are highest with a limited number of labels, a notable effect of pseudo label pretraining persists when more labeled data is used. Further, we outline that the confidence selection of SiROC is indeed effective and that the performance gains generalize to scenes that were not used for pseudo label training. Through the pseudo label pretraining, SemiSiROC allows student models to learn more refined shapes of changes and makes them less sensitive to differences in acquisition conditions.