AI-Enhanced Hyperspectral Imaging Approach for Precise and Non-destructive Phase-Specific Detection of Drought Stress in Sweet Basil

Accurate and non-destructive detection of drought stress is essential for precision irrigation and sustainable crop management. In sweet basil ( Ocimum basilicum L .), mild drought and drought recovery can share visual symptoms but require different irrigation decisions. This study aimed to develop an AI-enhanced hyperspectral imaging framework to classify four water status phases: well-watered, mild drought, drought recovery, and severe drought. A paired visible and near-infrared (VNIR, 400–1000 nm) and shortwave infrared (SWIR, 900–1700 nm) dataset was acquired across all four phases, with each plant imaged and phase labels verified by leaf water content. The pipeline used NDVI–REP kernel density estimation to characterize spectral distributions and pixel-level K-means clustering to delineate regions of interest. To address class imbalance, autoencoder (AE)- and generative adversarial network (GAN)-based augmentation were applied to the training set only. Four machine-learning models and a one-dimensional convolutional neural network (1D-CNN) were then trained and evaluated under a plant-level data split. SWIR-based models consistently outperformed VNIR-based models. The SWIR 1D-CNN with GAN augmentation achieved the best whole-leaf accuracy of 0.800, balanced across all four phases including the management-critical MD and DR classes. Cluster-level analysis identified the leaf margin as the most informative region in the SWIR domain (accuracy 0.892). By resolving physiologically distinct phases that conventional binary stress detection cannot separate, the framework provides phase-specific information that enables precision irrigation in sweet basil.