Multi-Scale Graph Wavelet Convolution for Hyperspectral-LiDAR Urban Scene Classification

The Houston 2013 benchmark conducted in Houston describes a methodology that constructs a sparse pixel graph exclusively over labeled pixels. This approach employs learnable multi-scale spectral filtering using Chebyshev-approximated graph wavelets and incorporates a lightweight Multi-Layer Perceptron (MLP) head for classification purposes. The training procedure integrates feature MixUp at labeled nodes, a composite loss function combining focal loss and label smoothing, exponential moving average (EMA) weight tracking, an AdamW optimizer with warm-up and cosine scheduling, and mild graph augmentation techniques such as random edge drop and addition with re-normalization. Implementation executed precisely in accordance with the original methodology, with five independent runs producing an overall accuracy (OA) of 90.99% ± 0.41%, average accuracy (AA) of 92.11% ± 0.35%, and a Kappa coefficient (κ) of 0.9022 ± 0.0044. These results demonstrate not only high accuracy but also minimal variability, providing reassurance of the robustness of the methodology.