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

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Self-supervised learning (SSL) is effective for remote sensing image segmentation but struggles with extreme label scarcity.
• Key challenges include poor exploitation of hierarchical representations and information loss during feature adaptation.

Purpose of the Study:

• To propose SS-KAN, an enhanced SSL framework using Kolmogorov-Arnold Networks (KAN) to address label scarcity in remote sensing image segmentation.
• To improve the exploitation of hierarchical representations and minimize information loss in low-data regimes.

Main Methods:

• Developed a depthwise KAN module combining depthwise separable convolutions with learnable activation functions for context-aware features.
• Introduced a dual-branch adaptation strategy for fine-tuning, preserving spatial semantics and enhancing hierarchical features via KAN nonlinearity.
• Utilized multi-scale branches and residual identity mappings to reduce feature degradation during domain transfer.

Main Results:

• SS-KAN significantly outperforms state-of-the-art methods in remote sensing image segmentation, even with as little as 1% labeled data.
• Experiments on three benchmarks validate the framework's effectiveness in data-scarce conditions.
• Ablation studies confirm the critical contributions of the depthwise KAN and dual-branch adaptation modules.

Conclusions:

• The proposed SS-KAN framework demonstrates superior performance in data-efficient remote sensing image segmentation.
• Integrating KAN's adaptive nonlinearity with depthwise convolutions and identity mappings offers a promising direction for future research.
• The findings highlight the potential of SS-KAN for practical applications where labeled remote sensing data is limited.