Enhancing early detection of mango anthracnose using hyperspectral imaging and 3D-CNNs

• A novel 3D-CNN achieves 95.24% accuracy in classifying mango anthracnose stages. • Grad-CAM reveals model focuses on disease-specific wavelengths and spatial regions. • Offers an interpretable AI solution for early plant disease management. Anthracnose, a severe postharvest disease in mangoes caused by Colletotrichum gloeosporioides , leads to substantial economic losses due to its latent infection. Conventional spectral-based detection relies typically on averaged one-dimensional (1D) spectral data, neglecting spatial information related to disease infection. This study introduces an early detection method combining hyperspectral imaging with a three-dimensional convolutional neural network (3D-CNN) to simultaneously extract spatial and spectral features. The developed 3D-CNN model achieved 95.24% accuracy in distinguishing healthy, asymptomatic, and symptomatic mango samples, outperforming both 1D-CNN (92.52%) and traditional machine learning models (65.31%–90.48%). Gradient-weighted class activation mapping (Grad-CAM) interpretation revealed that the 3D-CNN focuses on distinct spectral wavelengths for different infection stages: 1067–1161 nm for healthy, 1217–1298 nm for asymptomatic, and 1204–1373 nm for symptomatic, corresponding to biochemical changes during infection. In contrast, the 1D-CNN utilized the same wavelengths across all stages (1000–1117 nm). Spatially, the 3D-CNN also exhibited selective focus on different fruit regions consistent with infection status, such as the center of healthy fruit and infected areas of symptomatic samples. This interpretable approach offers a powerful tool for early anthracnose detection and holds significant potential for improving mango disease management.