FedDA-HSI: Federated Class-Aware Framework for Hyperspectral Image Classification With Diffusion Augmentation

Abstract

Hyperspectral images (HSIs) have demonstrated remarkable potential in remote sensing applications due to their rich spectral and spatial characteristics. However, the scarcity of labeled HSI data from individual sources significantly hinders the deployment of deep learning models in this domain. Leveraging multisource HSI can mitigate single-source data limitations, but direct data sharing poses serious privacy challenges. Federated learning (FL) provides a compelling framework for privacy-preserving collaborative learning, yet practical challenges remain, particularly data scarcity and the nonindependent and identically distributed (non-IID) nature of client data. In this work, we introduce a diffusion-based data augmentation framework, namely FedDA-HSI, tailored for federated HSI (FedHSI) classification. FedDA-HSI addresses two core issues in real-world FedHSI deployments. First, we propose a federated diffusion model capable of generating high-fidelity synthetic hyperspectral data locally, enabling data augmentation without compromising privacy, while allowing the aggregated global model to capture shared knowledge across clients. Second, we incorporate a class-aware augmentation strategy to alleviate data imbalance by injecting synthetic samples of missing or underrepresented classes, effectively improving interclient data diversity and mitigating non-IID effects. Extensive experiments on benchmark HSI datasets validate that our approach significantly improves classification performance under heterogeneous conditions.