Toward accurate prediction of apple firmness and brix across countries, seasons and cultivars with hyperspectral imaging

Traditional apple maturity assessment methods are destructive and time- and labour-intensive, yielding only population-level approximations. Hyperspectral imaging provides a non-destructive alternative to assess individual fruit, but progress has been constrained by the lack of large, diverse datasets that support robust model generalisation. This study presents a multi-cultivar, multi-season, multi-country hyperspectral apple dataset to enable generalisable prediction of soluble solids content (Brix) and firmness. Using this dataset, we adopt an iterative modelling framework to evaluate deep learning architectures, image resolutions, cultivar encoding, seasonal effects, and feature-specific models. Wavelength and spatial region importance were also analysed. The best predictive performance was achieved using Vision Transformer (ViT) models trained on edge-cropped 40 40 pixel images with explicit cultivar encoding, with Brix and firmness modelled independently. Although seasonal specificity was observed, models trained across all three seasons achieved the strongest overall performance. A 50% reduction in spectral wavebands did not compromise prediction accuracy. Key wavelength ranges contributing to Brix and firmness prediction were identified across the visible–near-infrared spectrum. Spatial regions were unimportant for Brix prediction but showed relevance for firmness. The optimised ViT model achieved firmness prediction performance comparable to previous studies (RMSE = 0.76 kgf, R = 0.63), while Brix prediction accuracy was lower (RMSE = 0.91 Brix, R = 0.75), likely reflecting increased biological and environmental variability captured in the dataset. Overall, this work demonstrates that hyperspectral imaging combined with deep learning and large, diverse datasets enables robust, non-destructive prediction of apple quality attributes across production conditions.