Smart Tomato Disease Diagnosis: A Hybrid Approach Using Zebra Optimization Segmentation and Compact Attentional Residual Convolutional Network

Tomato crop diseases caused by pathogens such as fungi, bacteria, and viruses can severely impact the health and yield of tomato crops. Early detection of these diseases is crucial to prevent significant crop losses and ensure optimal growth conditions. While existing tomato leaf disease detection systems demonstrate high accuracy, they face challenges due to high computational requirements, limiting their efficiency and practical deployment in real-time applications especially in resource-constrained environments such as mobile devices or field systems. To overcome these challenges, this work proposes a hybrid model called Smart Tomato Disease Diagnosis using Zebra Optimization Segmentation and Compact Attentional Residual Convolutional Network (ZOS-CRCN). The process begins with the input of tomato leaf images, which undergo preprocessing, including resizing to pixels and normalization using the Tanh-estimator method. This step standardizes the images and mitigates the influence of outliers for stable model training. Next, the Zebra Optimization Algorithm (ZOA) is applied for precise segmentation which isolates regions of interest (ROI) that contain diseased areas. ZOA simulates natural foraging and defense behaviors for effectively optimizing the segmentation process. For feature extraction and classification, the Compact Attentional Residual Convolutional Network (CRCN) is utilized. This lightweight network focuses on important image areas using attention mechanisms and reduces computational complexity through residual connections. The CRCN classifies the diseases into ten categories such as Healthy Tomato Leaf, Early Blight, Late Blight, Septoria Leaf Spot, Yellow Leaf Curl Virus Disease, Bacterial Spot, Target Spot, Spider Mite (Two-Spotted), Leaf Mold, and Mosaic Virus. The proposed ZOS-CRCN model achieves 99.99% accuracy, an execution time of 89ms, an AUC of 0.995, and a high F1-score of 99.99%. By this, it demonstrates its effectiveness for real-time disease detection in resource-limited environments.