Spatial Reduction Attention in Multiscale Vision Transform for Surface Water-Land Interface Zone Segmentation

Water segmentation is important for applications in flood prevention, water resource management, and urban planning. The accurate identification of water-land interface zones and the delineation of edges between water and land in remote sensing satellite imagery, however, present significant challenges for traditional segmentation methods. This research aims to enhance the precision of segmentation, particularly in identifying water and land interface zones, while also reducing computational demands to enable real-time analysis on edge devices. This article introduces a novel spatial reduction attention (SRA) mechanism within the multiscale vision transform framework, which is proficient at capturing both local and global features. The proposed multiscale multihead attention mechanism, enhanced with multiscale projection and SRA, aids in learning features from various receptive fields, thereby increasing computational efficiency. The integration of dual-branch channels for multispectral imagery and color attributes significantly improves the model's recognition capabilities. In the evaluation of water segmentation, the proposed method significantly outperforms advanced models, achieving a 10.1% improvement in mean intersection over union and a 6.7% increase in mean F 1-score. This performance underscores the model's efficacy in accurately identifying water-land interface zones and highlights its potential in improving both the accuracy and efficiency of water segmentation in satellite imagery.