Hierarchical Gated Network for Multimodal Remote Sensing Imagery Classification with Limited Data

The effective fusion of multimodal data can significantly advance Earth observation-related tasks. However, the collection and annotation of spatial-aligned remote sensing (RS) data across different modalities are resource-intensive and laborious. This hinders the exploitation of complementary information between modalities, leading to unsatisfactory performance in data-limited scenarios. To solve this problem, we propose a hierarchical gated network (HGN) for multimodal RS imagery classification. Specifically, HGN incrementally integrates multimodal information through gating mechanisms at both the feature and decision levels. For the input data of each modality, the corresponding feature representations are extracted in parallel. During this process, feature gates between different convolution blocks are utilized to control the fusion flow across single-modal branches. Finally, logits derived from single-modal and fused representations are selectively combined through the decision gate. This hierarchical fusion framework enables adaptive cross-layer interactions between different modalities, thereby facilitating the exploitation of complementary information. Thorough experiments on the Houston2013 and Augsburg multimodal datasets show that HGN achieves state-of-the-art performance.