Multi-Disease Identification in Single Citrus Leaf Image Using Sigmoid Activation and Optimised Threshold

Disease classification in plant leaf images for early detection and prevention using deep learning techniques has been largely focused upon recently. Nevertheless, identifying multiple diseases in a single leaf image is still unexplored. Bounding box predictions using object detection techniques have been employed for this purpose. However, such information is not always readily available. This study focuses on identification of multiple diseases in a single leaf image of citrus fruits using state-of-the-art deep learning techniques. The considered leaves have been infected with none, one, two or all three severe diseases – Anthracnose, Melanose, and Bacterial Brown Spot. The probabilities of occurrence of these diseases are predicted for each leaf image using a sigmoid activation layer and a threshold value has been used to decide on whether the disease is present or absent in the leaf. The threshold value has been optimised during the training time using the meta-heuristic ant lion optimisation algorithm. The outcomes highlight the superiority of the DenseNet121 architecture in terms of classification accuracy, as it accurately identifies overlapping disease classes present in citrus leaves, achieving 98.6% accuracy with a threshold probability value of 0.284. Following closely, the MobileNetV2 architecture showcase comparable accuracy values, however, requiring significantly less number of training parameters.