Multi-scale Diffusion Features Fusion Network For Hyperspectral Image Classification

Hyperspectral images (HSIs) contain rich spectral and spatial information, which provides valuable data foundation for the classification of ground objects. However, most of the existing classification methods are devoted to constructing discriminative models for feature extraction, and lack comprehensive interpretation of hyperspectral data, especially in the case of limited samples. Diffusion model, as a novel paradigm in generative model, has been widely concerned in various fields due to its powerful data modeling capabilities. In order to facilitate the application of diffusion features to hyperspectral image classification (HSIC), we propose a multi-scale diffusion features fusion network (dubbed ‘MDFFNet’). The proposed MDFFNet adopts a hierarchical diffusion feature preprocessing and multi-scale feature fusion to efficiently learn the latent spectral-spatial feature representations from diffusion models, and then the fused diffusion features are used for subsequent classification. Experimental evaluations are carried out on two benchmark HSI datasets, and the results show impressive classification performance even with a limited number of training samples.