Innovative convolutional neural networks and high-quality hyperspectral datasets for non-destructive detection of early mold in wood

Wood readily develops mold under warm and humid conditions, yet early colonization is often visually unapparent and its spectral cues can be confounded by wood species and surface texture. SSAM-Net is a two-stage visible-to-near-infrared hyperspectral framework that first identifies wood species at the pixel level and then grades mold severity using a species-specific multi-scale 1D convolutional network with channel–spectral joint attention. A pixel-wise dataset (400–1000 nm) was constructed from Cunninghamia lanceolata , Pinus massoniana , and Ulmus pumila inoculated with Trichoderma harzianum , Penicillium citrinum , Cladosporium cladosporioides , and Fusarium oxysporum , and labeled as No growth, Mild growth, or Severe growth using visually and microscopically verified criteria derived from the VTT mold growth index. Species identification achieved a pixel-level hold-out accuracy of 91.48%. With Savitzky–Golay smoothing, severity grading achieved pixel-level hold-out accuracies of 88.35%, 93.10%, and 89.63% across the three species. The complete SSAM-Net achieved an end-to-end pixel-level hold-out accuracy of 88.24%. SHAP identified model-important spectral cues mainly in the visible and red-to-NIR transition regions, while board-scale maps from same-batch samples qualitatively illustrated the spatial back-projection of predicted severity patterns. These results indicate that the three predefined mold-severity classes can be differentiated at the pixel level under controlled experimental conditions, providing a basis for the further development of non-destructive screening and monitoring methods for wood-surface mold.