HA-ViTNet: Dual-Domain Collaborative Learning for Semantic Segmentation of High-Resolution Remote Sensing Images

Le résumé fourni par la source

Semantic segmentation of high-resolution remote sensing images is critical for land use monitoring and urban planning. Although deep learning has improved segmentation performance, existing methods still struggle with fine-grained boundaries and spatial fragmentation. We propose HA-ViTNet, a dual-domain collaborative framework. It integrates an Adaptive Feature Transformer Block (AFTB) to model long-range dependencies in the spatial domain, enhancing intraclass coherence while suppressing noise. To further capture structural context, we design a Spatial RWKV Attention Block (SRAB), which simulates temporal dependencies within spatial sequences. This pseudo-temporal modeling feeds global context back into the spatial stream, achieving a fine balance between local detail preservation and global semantic understanding. Additionally, we introduce a new UAV-based remote sensing dataset with diverse landforms from the Sichuan Basin, China. Experiments on two public benchmarks and our dataset show that HA-ViTNet outperforms state-of-the-art segmentation models in accuracy and efficiency, while effectively addressing hybrid architectural heterogeneity.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.