Vision Foundation Model-Driven Dynamic Heterogeneous Graph Neural Network for Remote Sensing Scene Classification
Abstract
Remote sensing scene classification (RSSC) requires capturing local semantic features, global contextual dependencies, and semantic associations among visual entities across different scales. Recent hierarchical graph methods have demonstrated the effectiveness of multiscale graph representations, but their hierarchies are usually derived from segmentation-based construction, and their graph connectivity is typically fixed during feature updating. These designs may remain constrained by low-level visual cues and lack flexibility when node features evolve. To address these issues, this letter proposes a dynamic heterogeneous graph neural network (DHGNN) for RSSC. DHGNN extracts high-level semantic representations from a frozen vision foundation model (VFM) and organizes three-level visual nodes into a unified heterogeneous graph. A relation-aware dynamic heterogeneous graph propagation (RDHGP) mechanism is introduced to dynamically refine intralevel connectivity. Experiments on the AID and NWPU-RESISC45 (NWPU) datasets demonstrate the effectiveness and generalizability of DHGNN. Compared with the strongest baselines, DHGNN improves overall accuracy (OA) by 0.65 percentage points on AID under the 20% training ratio and improves mF1 by 1.63 percentage points on NWPU under the 20% training ratio. The source code is available at: https://github.com/Ozeankeeper-DY/DHGNN