RS3-RWKV: Leveraging RWKV for Efficient Remote Sensing Semantic Segmentation

Semantic segmentation of high-resolution remote sensing images is of great significance for accurately extracting geographic information and serving land resource management, urban planning, and natural disaster monitoring. However, mainstream Convolutional Neural Networks and Transformer models currently used to process remote sensing images generally face a trade-off between computational complexity and accuracy. The recently emerging Receptance Weighted Key Value (RWKV) model, with its efficient linear attention mechanism, has demonstrated outstanding performance in natural language processing, but its potential in the field of remote sensing images has not yet been fully explored. To address the unique challenges of remote sensing images, such as multi-scale targets, diverse spatial distributions, and complex long-range dependencies, this paper innovatively proposes the RS$^{3}$-RWKV framework for semantic segmentation of remote sensing images based on the RWKV architecture. To effectively capture spatial continuity and cross-scale features in remote sensing images, we design the Proximity-Sensitive WKV attention mechanism (PS-WKV) and Scale-Adaptive shift mechanism (SA-Shift), which significantly enhance the model's ability to perceive global spatial information and adapt to multi-scale targets while reducing computational complexity. The PS-WKV mechanism improves long-range spatial dependency modeling by emphasizing the proximity of neighboring tokens, and the SA-Shift mechanism dynamically adapts to varying target scales, allowing the model to better handle multi-scale targets in remote sensing images. Experimental results on the LoveDA and ISPRS Potsdam datasets demonstrate that this method achieved a better balance between segmentation accuracy and computational efficiency compared to existing CNN, Transformer, and Mamba methods, showcasing significant performance advantages while ensuring model efficiency, and holds promising application prospects.