Apple anthracnose detection by deep learning-based hyperspectral image

The postharvest quality and safety of apples are critical to the global consumption of fruits. Early and accurate disease detection is crucial for preventing its spread and optimizing plant protection strategies. Hence, the work proposes an early detection method of apple anthracnose based on hyperspectral imaging (HSI) multi-channel feature extraction combined with deep learning (DL). Data were collected from healthy apple samples, and the inoculated apples were divided into asymptomatic and symptomatic groups based on infection symptoms. To analyze the subtle spectral changes, the extracted region of interest (ROI) was decomposed into multi-band images as global spectral features. Contrast-limited adaptive histogram equalization (CLAHE) was used to enhance band image information, while the Otsu algorithm was used to segment and extract high-gray-level and low-gray-level regions as local spectral features. An early disease detection model was constructed by integrating global and local features and combining CNN, CNN-LSTM, and CNN + SE algorithms. The fused full-spectral CNN + SE model achieved an accuracy of 97.78%. Successive projections algorithm, competitive adaptive reweighted sampling (CARS) and uninformative variable elimination algorithm were used to select features from full-spectral features. The CARS algorithm combined with CNN + SE model achieved an accuracy of 97.04%. The results demonstrated that HSI combined with DL can effectively realize the early detection of apple anthracnose, providing an effective solution for non-destructive and accurate detection of early diseases in fruits and vegetables.