Advancements in Semisupervised Remote Sensing Segmentation Using Adaptive Patch Augmentation and Class Ranking Weight

Traditional deep semantic segmentation methods rely heavily on large-scale, densely labeled data, which is costly and time-consuming to obtain. Semi-supervised frameworks have emerged as a promising alternative, reducing the dependency on pixel-level annotations while enhancing segmentation performance. However, these methods still face challenges in data augmentation and class imbalance. For instance, traditional augmentation techniques such as CutMix are limited by their reliance on single random local contexts, which restricts perturbation strength and hinders model generalization. Moreover, the long-tailed class distribution in remote sensing (RS) images is often overlooked, leading to pseudo-labels that are biased toward majority classes, further exacerbating class imbalance. To address these challenges,we propose a novel semi-supervised framework for RS semantic segmentation, named APRW, which incorporates adaptive patch augmentation and class rank weighting. The adaptive patch augmentation module dynamically generates adaptive patch masks by comparing the predictions of weakly and strongly augmented inputs. These masks focus on high-confidence regions while preserving diverse boundary structures, effectively expanding the perturbation space and enhancing the model’s adaptability to varied inputs. The class rank weighting module maintains memory banks for both labeled and unlabeled data, dynamically calculates class weights based on the relative ranking of pixel confidences, and adaptively fuses these weights. This design mitigates biases caused by long-tailed distributions, improves the segmentation of rare classes, and enhances overall segmentation accuracy. Extensive experiments demonstrate that our method consistently outperforms existing approaches across multiple RS datasets, including DFC22, iSAID, MER, MSL, Vaihingen, and GID-15, showcasing superior generalization and higher segmentation precision. Implementation is released at https://github.com/135az/APRW.