Existing methods for remote sensing imagery often ignore spatial layout and higher-order label dependencies, limiting performance. We propose the Spatially Modulated Dependency Hypergraph Network (SMD-HGNet), which integrates Spatially Modulated Dependency Loss (SMDL) and a Multi-view Dynamic Hypergraph Network (MDHN). SMDL uses class activation maps to estimate spatial proximity between label pairs and adaptively reweight dependencies, strengthening related labels and suppressing irrelevant co-occurrences. MDHN models superpixels as hypergraph nodes from appearance and location views and applies self-attention-guided hypergraph convolution to propagate semantics. Experiments on the UCMerced Land-Use (UCM) and Aerial Image Dataset (AID) benchmarks show that SMD-HGNet significantly outperforms comparative multi-label image classification methods in learning spatially-aware label depend-encies.