DAPGNet: Dynamic Adaptive Physics-Guided Graph Diffusion Network for Hyperspectral Image Classification

Hyperspectral image (HSI) classification often struggles with mixed pixels, spectral variability, and irregular boundaries. While graph neural networks (GNNs) have been used to model pixel relations, existing methods typically rely on data-driven topology construction that ignores the inherent physical properties of contiguous spectral bands. This paper asks whether explicitly injecting a structure-constrained physical prior into the graph learning process can improve classification accuracy and robustness.

The authors propose DAPGNet, a dynamic adaptive physics-guided graph diffusion network. The framework consists of three main components:

1. Multiscale Spectral Physical Prior Encoder: This module processes contiguous spectral responses using parallel 1D convolutional branches (with varying kernel widths) to capture fine-grained spectral variations and local smoothness. These are fused into node-wise physical-prior representations.
2. Physical-Prior-Aware Graph Construction: Instead of relying solely on spatial proximity or feature affinity, the graph topology is constructed by combining spectral-spatial similarity, physical-prior consistency, and spatial distance. This ensures that nodes are connected based on both their observed features and their underlying physical characteristics.
3. Physics-Gated Graph Diffusion: During message propagation, the model uses a physical gate to interpolate between graph-aggregated features and projected physical-prior features. Learned edge weights are transformed into additive attention biases, ensuring that the physical prior continues to guide the network throughout the diffusion process.

Experiments conducted on four benchmark datasets (Indian Pines, WHU-Hi-LongKou, Houston2013, and Houston2018) demonstrate that DAPGNet consistently achieves superior performance in Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficients compared to representative baselines, including CNNs, Transformers, and Mamba-based architectures. Specifically, the method improves AA by 3.64 to 7.31 percentage points over the strongest competing methods. Ablation studies confirm that each component—the physical-prior extraction, prior-aware topology, and physics-gated propagation—contributes significantly to the model's success.