Adaptive Wavelet-Enhanced Diffusion Model with Cross-Domain Conditional Fusion for Remote Sensing Semantic Segmentation

Semantic segmentation is a crucial task in complex remote sensing image processing with broad application prospects. Current mainstream approaches rely on deep learning methods based on convolutional neural network (CNN) and vision transformer (ViT). However, enormous scale disparities between categories and significant grayscale variations remain persistent obstacles to comprehensive feature representations of complex remote sensing images. The CNN-based methods are limited by the receptive field of convolution, struggling to capture global features, while the ViT-based approaches often partition images to reduce computation cost, leading to severe loss of local details. In contrast, generative diffusion models iteratively model the underlying data distribution, achieving stable features and demonstrating superior performance on complex tasks. Inspired by this, we propose an adaptive wavelet-enhanced diffusion model with cross-domain conditional fusion for accurate remote sensing semantic segmentation. Specifically, based on a wavelet-driven adaptive enhancement module, our framework first introduces an adaptive wavelet-enhanced conditional diffusion model. This model captures stable features while suppressing high-frequency noise interference during segmentation, and refines the prediction through multi-step denoising, dramatically simplifying segmentation for complex remote sensing scenarios. To overcome the difficulty of capturing comprehensive feature representation, we further propose the cross-domain conditional fusion module to dynamically fuse spatial and frequency domain features through multiple spatial and frequency branches. Experiments on the Vaihingen, LoveDA, and AIR-PolSAR-Seg datasets demonstrate that our method outperforms state-of-the-art approaches, validating its robustness and effectiveness.