HSS-KAMNet: A Hybrid Spectral–Spatial Kolmogorov–Arnold Mamba Network for Residential Land Cover Identification on RS Imagery

초록

Accurate identification of residential areas from remote sensing (RS) imagery plays a pivotal role in urban planning, infrastructure development, and environmental monitoring. Existing convolutional neural network (CNN) and transformer-based models have demonstrated strong performance on the EuroSAT benchmark; however, they often struggle to effectively capture both long-range dependencies and fine-grained spectral–spatial features, particularly in complex residential environments. To address this gap, we propose a novel hybrid spectral–spatial Kolmogorov–Arnold Mamba network (HSS-KAMNet) that integrates three core cutting-edge paradigms—Kolmogorov–Arnold networks (KANs), state-space models (Mamba), and hybrid multiscale attention mechanisms—into a unified architecture. HSS-KAMNet leverages a spectral–spatial KAN encoder with learnable B-spline activations to extract adaptive interpretable features from all 13 Sentinel-2 spectral bands. A dual-branch Mamba processor employs spatial and spectral Mamba streams to efficiently model long-range dependencies, with a cross-modal fusion mechanism enhancing joint feature representation. A novel hybrid attention mechanism fuses CNN-local and transformer-global features, enabling robust context awareness. We further introduce a spectral–spatial boundary loss to enforce spectral consistency, sharpen spatial boundaries, and regulate activation smoothness. Evaluated on different benchmark datasets for residential land cover classification, HSS-KAMNet achieves a new state-of-the-art accuracy of 99.42%, significantly improving boundary delineation and generalization to unseen regions. Our results demonstrate that HSS-KAMNet not only advances classification accuracy but also enhances model interpretability, making it a practical and explainable tool for residential area mapping from RS imagery. © 2008-2012 IEEE.

키워드