Hyperspectral imaging combined with feature selection and deep learning for early detection of huanglongbing in citrus

• A hyperspectral framework integrating preprocessing, feature selection, and deep learning was developed for citrus Huanglongbing detection. • A dual-path feature selection strategy (SPA and CARS) effectively reduced spectral redundancy while preserving discriminative information. • Optimized spectral bands improve deep learning classification performance. • SPA–MSC–1DResNet achieves 98.06% accuracy in HLB identification, establishing a robust hyperspectral–deep learning framework for citrus disease detection. Early and reliable detection of citrus huanglongbing (HLB) is critical for mitigating its severe impact on the global citrus industry. In this study, a hyperspectral data–driven diagnostic framework is proposed for HLB detection. Hyperspectral reflectance data of citrus leaves, including both HLB-infected and healthy samples, were acquired using a hyperspectral imaging system. Three spectral preprocessing methods (MSC/STD/SNV) were applied to the raw hyperspectral reflectance data respectively. Meanwhile, competitive adaptive reweighted sampling (CARS) and the successive projections algorithm (SPA) were employed to extract key informative spectral bands. The results indicate that the selected feature bands account for 20.45%–35.79% and 3.41%–9.65% of the full spectral bands for CARS and SPA, respectively. Subsequently, the optimized spectral subsets were fed into five deep learning models (1D-CNN/LSTM/1DResNet/Transformer/TCN) for classification. The experimental results demonstrate that the SPA–MSC–1DResNet combination achieved the best performance, with an accuracy of 98.06%, and both recall and F1-score reaching 98.11%. This study provides an effective solution for addressing the high-dimensional redundancy problem in hyperspectral data and demonstrates the potential of hyperspectral techniques for early diagnosis of citrus HLB.