Multisource Remote Sensing Data Classification Based on Quadruple Agent Attention Assisted Modal Feature Perception

The combined application of hyperspectral images (HSIs) and light detection and ranging (LiDAR) data in land cover classification has always faced the challenge of balancing the heterogeneity and complementarity among heterogeneous data. Therefore, this has also become a hot research topic for the application of the fusion of various remote sensing data. This letter proposes a novel quadruple agent attention fusion and modal feature perception network (QATMPNet) framework to address this. First, the proposed method employs a 2-D inverted bottleneck convolution to extract features from two kinds of heterogeneous data. Subsequently, feature mapping is applied to construct a four-channel input interaction to enhance feature representation. Finally, a quadruple agent attention-guided fusion module (QAT) is designed to integrate features, and a multimodal attention mechanism is utilized to capture multimodal data interactions. The effectiveness of the proposed method was verified through experiments on three public remote sensing datasets.