MTARM-Net: A region-aware joint classification–regression network for hyperspectral moisture estimation and visualisation in air-cured cigar leaves

Moisture content is a critical physiological indicator in the cigar leaf air-curing process, with hyperspectral imaging (HSI) technology providing an effective solution for its non-destructive estimation. However, owing to inherent differences in the physiological structure of leaf mesophyll and veins, moisture distribution and spectral curves of leaves exhibit spatial heterogeneity, limiting the model’s prediction accuracy for moisture content. Driven by this limitation, this study collected spectral data and corresponding moisture content for leaf mesophyll and vein regions, using manually annotated vein structures. Furthermore, a Multi-Task Adaptive Region Modulation Network (MTARM-Net) was proposed to simultaneously achieve the spectral classification of distinct regions and predict their corresponding moisture contents. This network integrates the advantages of Convolutional Neural Networks (CNNs) and transformer architectures for feature extraction. Furthermore, it incorporates an adaptive region information modulation (ARIM) module that performs affine transformations on regression features based on the classification output. The experimental results demonstrate that MTARM-Net achieves high prediction accuracy for both region classification and moisture content prediction. Specifically, the accuracy (ACC t ) and coefficient of determination (R 2 t ) on the test set reached 99.2% and 0.977, respectively. Compared with the single-task models, the ACC t and R 2 t improved by 1.22% and 1.45%, respectively. In addition, the visualisation results of the moisture distribution in cigar leaves across different air-curing days further demonstrated the superiority of MTARM-Net. These findings provide a pathway for a novel, non-destructive method for detecting moisture content and visualising moisture distribution during the air-curing process of cigar leaves.