Multisource Noise-Reduced Diffusion Feature Fusion for Hyperspectral and SAR Image Classification

Advances in deep learning have significantly enhanced the potential of remote sensing image classification, particularly for hyperspectral images (HSIs) and synthetic aperture radar (SAR) data. However, existing methods often overlook modality-specific noise (e.g., speckle noise in SAR), leading to suboptimal deep feature fusion performance. To address this challenge, this paper introduces the noise-reduced diffusion feature fusion method (NCDFF) for HSI and SAR image classification. The proposed method leverages a denoising diffusion model to extract noise-reduced multisource diffusion features, which captures the joint latent structure of HSI and SAR data by simulating forward and backward diffusion processes. These diffusion features, together with the raw image data, are processed using a feature fusion and classification network, which combines convolutional neural networks and a Transformer to capture local and global dependencies. In addition, a multisource feature fusion module is designed to enhance deeper feature interactions and promote complementary information exchange. Experiments conducted on the Augsburg dataset demonstrate that NCDFF outperforms state-of-the-art methods in hyperspectral and SAR image classification.