Hyperspectral images capture hundreds of narrow color bands per pixel, which is great for telling materials apart but hard to classify cleanly when pixels blend or sit on a boundary. Graph-based classifiers connect pixels and pass information between neighbors, but they usually build those links from spatial closeness or learned similarity, and mostly ignore the physics carried in the smooth run of adjacent spectral bands.

DAPGNet tries to put that physics back in. It encodes the continuous spectral response into a physical prior, then builds a sparse graph that blends spectral-spatial similarity, agreement with that prior, and spatial distance. During message passing, a physical gate mixes graph features with the prior features. On four standard datasets it reports the best overall and average accuracy, improving average accuracy over the strongest competitor by roughly 3.6 to 7.3 points.

Those numbers are from a cut-off abstract, so read the paper for the ablations and full comparison.