Abstract
Unsupervised Domain Adaptation (UDA) for semantic segmentation of Remote Sensing Images (RSIs) aims to adapt a model trained on labeled source-domain data to a target domain without requiring target-domain annotations. Existing self-training (ST) approaches adapt the model by generating pseudo labels for target-domain samples. However, due to the inherent domain gap, these pseudo labels often contain substantial noise, which in turn limits the effectiveness of such methods. We observe that noisy pseudo labels typically exhibit higher entropy. Motivated by this observation, we propose entropy-guided segmentation (EntSeg) framework, a novel entropy-guided framework for cross-domain RSIs semantic segmentation, that comprising two key models. The entropy-weighted pseudo-label denoising (EPLD) module dynamically adjusts the learning weights of noisy pixel locations (i.e., pixels with high entropy) to mitigate the adverse influence of noisy pseudo labels on domain adaptation. Nevertheless, the dynamic weighting mechanism of EPLD also results in insufficient contextual learning for target-domain data, especially at locations assigned low weights. To address this issue, we introduce entropy-guided masked image consistency (EMIC), which employs an entropy-guided strategy to mask high-entropy regions and learns spatial contextual relationships in the target domain through masked image consistency, thereby enhancing the model’s robustness in visual recognition. Experimental results demonstrate that the proposed method achieves highly competitive performance. The code will be released at https://github.com/woldier/EntSeg.
Funding Information
  • National Natural Science Foundation of China (U25A20402)
  • National Key Research and Development Program of China (2023YFF0615800)
  • National Natural Science Foundation of China (2024ZHCG0191)
  • National Natural Science Foundation of China (2026YFHZ0220)

This publication has 57 references indexed in Scilit: