DEEP LEARNING-BASED CLASSIFICATION OF ALTERNARIA ALTERNATA BLACK SPOT DISEASE IN PERSIMMON FRUIT USING FINE-TUNED CONVOLUTIONAL NEURAL NETWORKS

Abstract

Early and accurate detection of fungal diseases in fruit crops is critical for reducing postharvest losses and ensuring food security in resource-limited agricultural regions. This study presents a deep learning-based approach for classifying Alternaria alternata black spot disease in persimmon (Diospyros kaki) fruit using fine-tuned convolutional neural networks (CNNs). A novel field-collected dataset comprising 3,784 high-resolution images (2,145 diseased and 1,639 healthy) was acquired from persimmon orchards in the Swat region of Pakistan using smartphone cameras under varying illumination and field conditions. Three pretrained CNN architectures—InceptionV3, VGG16, and ResNet50—were systematically fine-tuned via transfer learning and evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis. InceptionV3 achieved the highest test accuracy of 96.91% (precision: 0.9764, recall: 0.9530, F1-score: 0.9642), outperforming VGG16 (90.56% accuracy, 0.9891 recall) and ResNet50 (60.77% accuracy, 0.5670 recall). The superior performance of InceptionV3 is attributed to its multi-scale inception modules that capture both fine-grained lesion textures and holistic fruit morphology simultaneously. VGG16 demonstrated high recall (0.9891), making it suitable for screening applications where missing diseased fruit carries higher cost. These findings establish a practical benchmark for automated persimmon disease diagnosis and provide a foundation for field-deployable smartphone-based detection systems in precision agriculture.