Identification of tea confusing diseases based on NIRS and deep learning

Abstract

To address the challenges in manual identification and the low efficiency of conventional detection for three morphologically similar tea diseases (anthracnose, leaf scorch, and nutrient deficiency), near-infrared spectroscopy (NIRS) integrated with chemometrics and deep learning was employed to achieve rapid and accurate disease identification. Nine spectral preprocessing approaches, including first derivative, multiplicative scatter correction (MSC), and wavelet transform, were compared. Characteristic wavelengths were extracted using the successive projections algorithm (SPA) and adaptive successive feature selection (ASFS), respectively. Combined with partial least squares (PLS), random forest (RF), and one-dimensional convolutional neural network (1D-CNN), a hybrid recognition model for tea diseases was developed. The results indicated significant spectral discrepancies among the three diseases at 1400 nm, 1200 nm, 900 nm, and 1700 nm. MSC, standard normal variate (SNV), and mean centring (MC) achieved the optimal preprocessing performance. ASFS outperformed SPA in comprehensively capturing subtle compositional variations. Among all hybrid models, the ASFS+1D-CNN model attained 100% recognition accuracy, enabling complete discrimination of anthracnose, leaf scorch, and nutrient deficiency. This study established a non-destructive, rapid, and high-precision NIRS-based identification protocol for similar tea diseases, providing technical support for the early diagnosis and precise prevention and control of tea diseases.