SS-KAN: Self-supervised Kolmogorov-Arnold networks for limited data remote sensing semantic segmentation

Self-supervised learning has emerged as a powerful approach for remote sensing image segmentation. However, its effectiveness significantly diminishes in remote sensing scenarios with extreme label scarcity (no more than 5 % of samples annotated). This limitation arises from two primary challenges: (i) insufficient exploitation of hierarchical representations in unlabeled data and (ii) irreversible information loss that occurs when adapting features from pretraining to downstream tasks. To tackle these challenges, this study proposes SS-KAN, an enhanced self-supervised learning framework based on the Kolmogorov-Arnold Network (KAN). The framework includes two key innovations: First, it features a depthwise KAN module, which combines depthwise separable convolutions with dilation rates and learnable activation functions within a multi-scale branch structure. This design enables the creation of context-aware feature representations from unlabeled images. Second, the framework develops a dual-branch adaptation strategy during the fine-tuning phase. This strategy utilizes a bifurcated structure that maintains spatial semantics through standard convolution while incorporating KAN-driven decomposable nonlinearity into residual identity mappings. As a result, it effectively enhances the hierarchical feature representation and reduces feature degradation during domain transfer, particularly in scenarios with limited data labels. Extensive experiments conducted on three benchmarks demonstrate that SS-KAN outperforms other state-of-the-art methods, even when trained with only 1 % labeled data. Ablation studies further confirm the importance of both the depthwise KAN and dual-branch adaptation modules. Our findings suggest that integrating KAN adaptive nonlinearity through learnable activation functions with depthwise convolutional operations and identity mappings opens new possibilities for data-efficient remote sensing image segmentation. Our code will be available from https://github.com/zhangjy2008327/SSKAN.