TPL-DA: A Novel Threshold-Free Pseudolabel Learning Framework for Domain Adaptive Semantic Segmentation of High-Resolution Remote Sensing Images
| Autor: | Yan Ren, Jie Long, Xiaowen Gao, Ming Zhang, Guoqing Liu, Nan Su |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2025 |
| Předmět: | |
| Zdroj: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol 18, Pp 1926-1945 (2025) |
| Druh dokumentu: | article |
| ISSN: | 1939-1404 2151-1535 |
| DOI: | 10.1109/JSTARS.2024.3502075 |
| Popis: | Semantic segmentation techniques for remote sensing scene understanding have significantly advanced, enhancing the refined Earth observation. However, most methods highly depend on extensive annotated data, leading to performance deterioration in complex high-resolution remote sensing cross-domain scenes, where variations in image conditions and environments are prevalent. Domain adaptive semantic segmentation (DASS) has been proposed to mitigate the reliance on dense and costly annotations, typically using stagewise training. This article addresses three key challenges in existing DASS methods: 1) insufficient warmup training, limiting potential performance gains; 2) rigid pseudolabel threshold settings in self-training (ST) result in performance bottlenecks; 3) entropy-based prediction bias alone fails to effectively identify high-confidence noise early in ST. To address these issues, we propose a novel threshold-free pseudolabel learning framework, TPL-DA. During the warmup stage, we introduce a multiview bidirectional consistency learning mechanism within a teacher–student architecture. This mechanism employs a bias-free data augmentation strategy, fostering consistent bidirectional predictions in teacher–student networks, thereby enhancing domain generalization and feature robustness. Our multiscale context-enhanced prediction module further amplifies this. In the ST stage, we propose a dynamic threshold-free pseudolabel learning strategy that utilizes well-aligned feature prototypes in the feature space to guide pseudolabel generation in the probability space, eliminating the threshold constraints. In addition, we model uncertainty using relative entropy and incorporate it into the optimization objective to manage high-confidence noise. Extensive experiments on the LoveDA, Potsdam, and Vaihingen datasets demonstrate that TPL-DA consistently outperforms existing methods and popular benchmarks, significantly enhancing DASS performance across diverse cross-domain scenes. |
| Databáze: | Directory of Open Access Journals |
| Externí odkaz: |
|