A study on remote sensing image semantic segmentation techniques based on multi-scale spatial context enhancement and feature fusions

Abstract

To address core issues in the semantic segmentation of high-resolution remote sensing images, including insufficient semantic representation, inflexible feature fusion, and spatial detail loss, this paper proposes a CNN-Transformer hybrid architecture. A Spatial Context Enhancement Module (SCEM) and a Hybrid Refined Feature Fusion Module (HR-FFM) are designed to enhance the detail-preserving capability for multi-scale objects. Experimental results show that the proposed method effectively improves boundary integrity and small-object recognition while retaining strong semantic extraction capabilities.