Leveraging Kolmogorov-Arnold Networks (KANs) for Mixed-Domain Satellite Imagery Segmentation

Semantic segmentation of satellite imagery requires models that capture global context while preserving sharp object boundaries. Convolutional Neural Networks (CNNs) excel at local feature extraction, but often struggle with long-range dependencies. Transformers provide global context but may blur edges and rely on opaque classifier heads. This study aims to develop an interpretable hybrid segmentation model that improves boundary accuracy and generalization across mixed-domain satellite imagery. This study presents SwinKANet, a hybrid segmentation model that combines a transformer encoder with boundary-aware decoding and an interpretable prediction head. SwinKANet employs a Swin Transformer (SwinV2-Tiny) encoder to extract multi-scale features, while a Convolutional Block Attention Module (CBAM) at the bottleneck refines channel and spatial responses. Skip connections equipped with SharpBlock units enhance edge detail, and an FPN-like lateral fusion module aligns and merges decoder features. The conventional multilayer perceptron head is replaced with a Kolmogorov–Arnold Network (KAN) head, enabling flexible function approximation and class-wise interpretability. We evaluate SwinKANet on a mixed-domain LoveDA dataset (urban + rural) for diverse spatial learning and on the urban-only ISPRS Vaihingen dataset for city-scale benchmarking. SwinKANet achieves 0.5269 mIoU on LoveDA and 0.7645 mIoU on Vaihingen, delivering sharper boundaries and more consistent class regions than CNN, Mamba, and transformer baselines. The KAN head further enhances explainability by revealing feature contributions for each class, supporting interpretable remote sensing applications.