A multimodal imaging dataset for quality grading of Canadian wild rice kernels using RGB and VNIR hyperspectral data

This article presents a multimodal dataset of Canadian wild rice (Zizania palustris) kernels, combining high-resolution RGB imagery and visible-near-infrared (VNIR) hyperspectral reflectance data for post-harvest quality characterization. The dataset comprises paired RGB images and VNIR hyperspectral scans of individual wild rice kernels representing eight quality categories: Healthy-Large, Healthy-Medium, Discoloured-Low, Discoloured-High, Broken-Low, Broken-High, Insect-damaged, and Unhulled. RGB images were acquired under controlled laboratory conditions and processed into standardized 512 × 512 pixel single-kernel patches, with additional augmented variants generated using controlled geometric and photometric transformations. VNIR hyperspectral data were acquired in the 400-1000 nm wavelength range using a hyperspectral imaging system, followed by radiometric correction, automated kernel segmentation, and per-kernel spectral extraction. The dataset includes raw hyperspectral files, dark and white reference frames, reflectance-corrected hypercubes, segmentation masks, per-kernel mean spectral profiles, paired RGB images, and associated metadata. This resource is intended to support research in automated grain grading, defect detection, and multimodal machine-learning applications in spectroscopy, computer vision, and post-harvest quality assessment.