Stochastic and Confidence-Aware Network (SCAN)-based Semi-Supervised Domain Adaptation for Satellite Imagery Segmentation

Satellite imagery segmentation serves a crucial role in environmental monitoring and sustainable development. However, the domain gap and the high annotation cost are two major challenges that limit system performance in a practical environment. To address these issues, this paper proposes a confidence-aware semisupervised domain adaptation method for satellite imagery segmentation. Unlike conventional approaches that directly incorporate labeled target data into the segmentation loss, our method leverages the labeled target data to refine the confidence of pseudolabels generated by a teacher model. This strategy alleviates the bottleneck of semisupervised domain adaptation, where the model is overfit with a small set of labeled data. Furthermore, we introduce a sparsity constraint to select compact features that generalize robustly across both domains. The proposed method is evaluated on the LoveDA, SyntheWorld, and Open Earth Map datasets in two scenarios. The experimental results demonstrate that using only 5% of labeled target data enables our approach to achieve 98.04% and 93.78% of fully supervised performance on the LoveDA and Open Earth Map datasets, respectively. Moreover, qualitative experiments show that our method does not overfit the training data and provides reasonable results. The code for this paper is available at https://github.com/Vo-Linh/SF-CAN.git.