S3RNet: Sparse Spatial-Spectral Representation With Hybrid Knowledge Distillation for Efficient Multispectral and Hyperspectral Image Fusion

Multispectral image (MSI) and hyperspectral image (HSI) fusion is a key enabler for real-time and near real-time remote sensing analysis, where low-resolution HSI (LR-HSI) must be combined with high-resolution MSI (HR-MSI) under tight computational and bandwidth constraints. Although recent deep learning based fusion methods have achieved impressive performance, many of them rely on deep and heavy architectures that are difficult to deploy on resource-limited platforms and are highly sensitive to noise corruption. To address these issues, we propose the Sparse Spatial-Spectral Representation Network (S3RNet). The proposed framework introduces three main components as follows. 1) A multibranch fusion network that captures specialized and complementary spatial-spectral features across parallel branches. 2) A dense feature aggregation block that exploits dense connectivity for efficient feature extraction and propagation. 3) A spatial-spectral adaptive weight block that learns content-aware sparse representations to suppress noise and enhance robustness. In addition, we develop a hybrid online knowledge distillation strategy that transfers knowledge from a high-capacity teacher network to a lightweight student network, striking a favorable balance between reconstruction fidelity and model complexity. Extensive experiments and analyses demonstrate that S^{3}$RNet delivers state-of-the-art fusion accuracy, strong robustness to various noise levels, and competitive runtime, making it well suited for practical remote sensing applications that require efficient MS-HS image fusion.