Wavelet-Enhanced CNN-Transformer Hybrid Network for Semantic Segmentation of Multi-Scale Geospatial Objects

Semantic segmentation of remote sensing imagery has achieved pixel-level precision in land cover classification through deep learning and computer vision technologies, providing automated and high-accuracy interpretation solutions for environmental monitoring, urban planning, and disaster assessment. However, the complex backgrounds, significant scale and orientation variations, and high intra-class variance inherent in remote sensing images often limit the performance of conventional segmentation methods. To address these challenges, this work proposes a novel Semantic-Enhanced Network (SEnetformer) for high-resolution remote sensing image segmentation, which integrates wavelet-transform convolution and transformer-based attention mechanisms to tackle the critical issues of spectral-spatial ambiguity and multi-scale object variability in geospatial analysis. The proposed architecture introduces three key innovations: a wavelet-transform guided decoder that hierarchically enhances low-level texture representation while suppressing high-frequency noise, a Multiscale Weighted Transformer Block (MWTB) that establishes long-range dependencies through local-global class-aware feature alignment, and a Hierarchical Multiscale Attention (HMA) mechanism for stochastic fusion of cross-scale encoder-decoder features to dynamically adapt to object size variations. Extensive experiments on the LoveDA, ISPRS Vaihingen, and UV-Beijing datasets demonstrate that these components collectively enable precise boundary delineation and robust semantic reasoning in complex scenarios.