Representative Spectral Correlation Network for Multi-source Remote Sensing Image Classification

Fusing hyperspectral imagery (HSI) with SAR or LiDAR data promises richer land-cover classification, but spectral redundancy and cross-source heterogeneity have stymied progress. In a new paper accepted to IEEE TGRS 2026, researchers from Ocean University of China and Mississippi State University introduce the Representative Spectral Correlation Network (RSCNet) to tackle these bottlenecks head-on. Their key innovation is a Key Band Selection Module (KBSM) that adaptively picks only the most task-relevant spectral bands under cross-source guidance, sidestepping the information loss of conventional PCA-based reduction. This yields compact, discriminative representations that align with semantic cues. Paired with a Cross-source Adaptive Fusion Module (CAFM) that applies attention-weighted local-global refinement, RSCNet achieves superior interaction between HSI, SAR, and LiDAR features.

Tested on three public benchmark datasets (including Houston and Trento), RSCNet consistently outperforms existing state-of-the-art approaches—GAN-based, CNN-based, and transformer-based fusion models—while requiring substantially lower computational overhead. The authors note that the band subsets learned by KBSM are highly interpretable and often correspond to known vegetation or mineral absorption features. By slashing redundancy and enhancing cross-modal alignment, RSCNet offers a practical path toward more accurate and efficient multi-source remote sensing classification. The full code is publicly available on GitHub, enabling reproduction and adaptation by the community.