Development of an Optimized Deep Learning Technique for Tomato Leaf Diseases Recognision

Abstract

Tomatoes are among the most widely cultivated and consumed vegetables globally, valued for their rich nutritional content and versatility in culinary applications. However, tomatoes are bedevilled with diseases and pests that wipe out approximately half of farmers’ harvests every year. Recent advancements in deep learning provide promising solutions for automating disease recognition; however, challenges still persist in achieving high accuracy and hyperparameter tuning. Hence, this research optimized Convolutional Neural Network (CNN) with Hippopotamus Optimizer (HO) for Tomato Leaf Diseases Recognition. Four thousand five hundred and thirty-six (4536) images of tomato leaf were downloaded from kaggle.com. The acquired images were grouped into four (4) classes: bacterial spot, early blight, leaf mould, and healthy leaves. The images were preprocessed by cropping to remove unwanted elements, converting to gray-scale for colour complexity reduction, normalizing and filtering to reduce noise. An optimized Convolutional Neural Network (CNN) using Hippopotamus Optimizer (HO), (HO-CNN) was developed. The HO-CNN was employed to select optimal values of number of neurons and dropout rates for CNN hyperparameters; The HO-CNN was implemented using MATLAB R(2023a). The evaluation metrics used were False Positive Rate (FPR), Specificity (Spec), Sensitivity (Sen), Precision (Pre), Accuracy (Acc), and Recognition Time (RT), and the HO-CNN was compared with the traditional CNN. The FPR, Spec, Sen, Pre, Acc, and RT for HO-CNN were 2.59%, 97.41%, 95.11%, 95.67%, 96.55% and 38.16 s, respectively. The corresponding values for CNN were 5.29%, 94.71%, 90.61%, 91.14%, 93.17% and 81.66 s, respectively. A Hippopotamus optimized-Convolutional Neural Network (HO-CNN) improves tomato leaf disease recognition accuracy by about 3.6%, reduces false detections by approximately 51%, and decreases recognition time by nearly 53% compared to the traditional CNN. The HO-CNN developed can be applied for tomato leaf disease recognition in real-world agricultural development.