Mask-Guided Diffusion-Consistency Network for Multimodal Hyperspectral Point-Cloud Mineral Mapping

Multimodal hyperspectral point-cloud mineral mapping is complicated by spectral ambiguity and uneven class distributions. Graph encoders integrate complementary observations, but feature aggregation does not necessarily yield spatially consistent class probabilities. We introduce a Mask-Guided Diffusion-Consistency Network combining multimodal graph encoding, conditional semantic-mask denoising and k-nearest-neighbour probability refinement. Noise-prediction and mask-reconstruction supervision regularize semantic estimates, while neighbourhood propagation encourages spatial continuity. The components are jointly trained with base and refined classification losses and perturbation consistency. Across five independent training runs on Tinto, the model achieves 80.93±0.31% overall accuracy, compared with 74.99±0.54% for a matched-budget backbone (mean ± sample standard deviation). The difference between means is 5.94 percentage points; the 10.46-point difference from the original backbone also includes continued-training gains. Five-run class-wise results reveal lower mean recall for Mafic B and Felsic B despite improved aggregate accuracy. Existing single-run ablations favour the complete model, but the 0.02-point validation fusion margin does not establish a meaningful direct diffusion effect. VNIR and LWIR experiments support within-dataset applicability. These results support combined semantic regularization and spatial refinement, while paired repeated ablations, quantitative boundary assessment and cross-site validation remain necessary.