ISSP-Net: An Interactive Spatial-Spectral Perception Network for Multimodal Classification

Coordinated and complementary spatial-spectral information is represented by the panchromatic (PAN) and multispectral (MS) images. The optimal utilization of the advantages of these images has become a subject of intense research interest. This article introduces the interactive spatial-spectral perception network (ISSP-Net) for multimodal remote sensing image classification, addressing the challenge of optimal utilization of complementary information from PAN and MS images. First, the pixel-guided spatial enhancement module (PGSE-Module) improves spatial location interaction using the spatial location enhancement learning strategy (SLEL-Strategy) and the cross-spatial aggregation learning strategy (CSAL-Strategy), integrating multiscale contextual information and emphasizing pixel-level features. Second, the time-frequency collaborative spectral enhancement module (TFCSE-Module) distinguishes useful frequency domain features through channel separation, lightweight convolutions, and adaptive Fourier transform learning. This approach enables comprehensive utilization of both primary and auxiliary information from multimodal data. Finally, experiments on four datasets demonstrate the ISSP-Net’s state-of-the-art performance in classifying MS and PAN images, with good generalization to hyperspectral (HS) and LiDAR data. The code is provided at: https://github.com/sun740936222/ISSP-Net.