Cross-Band Fusion for Multispectral Scene Classification

With the daily acquisition of vast amounts of remote sensing images by satellites, effectively utilizing these images for Earth observation has become a significant research focus. Multispectral image scene classification plays a fundamental role in advancing the understanding of remote sensing imagery. Existing methods for multispectral image scene classification do not sufficiently explore the interrelationships between different bands, which limits the performance of models in this task. Unlike these, we propose a cross-band fusion method for multispectral scene classification. Specifically, we employ two appropriate visual backbones to extract band-level visual features from different bands, followed by a cross-band fusion module to fuse the features from various bands adaptively. This approach fully explores the complementary information between different bands, achieving enhanced accuracy in multispectral scene classification. Extensive experiments on a commonly used dataset BigEarthNet validate that our proposed cross-band fusion method outperforms state-of-the-art (SOTA) methods.