Comparative Analysis of Customized Deep Learning Models for Multi-Disease Detection in Tomato Leaves for Precision Agriculture

Abstract

Tomato leaf diseases are a major threat to crop production and food security particularly in the developing world where early detection of the disease is a problem. The main aim of this research is to detects and classifies tomato leaf diseases using different advanced Convolutional Neural Network architectures in real time using a custom dataset manually collected from real tomato plants. In this work I used EfficientNetB3, MobileNetV3, ResNet50, DenseNet121 and InceptionResNetV2 with Squeeze Excitation (SE) blocks, Spatial attention mechanisms, advanced augmentation technique, and transfer learning to improve model robustness. The dataset consists of nine tomato leaf image classes as a part of multi-disease detection, including diseased and healthy leaves with both front and back-side images of leaves to cover all symptoms of the disease. In previous studies on this related task only the front side of the leaf was used for training but disease symptoms are present on both the front and back sides. According to those previous studies, if the user takes a backside image, the model cannot properly detect the disease. I trained these models on many fronts and back 7,200 labeled leaf images to ensure strong performance and real-time disease detection. This study allows users to detect tomato leaf diseases in real time using images of both the front and back sides. After evaluating different five architectures as EfficientNetB3, ResNet50 and InceptionResNetV2 perform almost equal in terms of accuracy, precision, recall and F1-score. EfficientNetB3 is only 44.65 MB which is very smaller then ResNet50 and InceptionResNetV2 and MobileNetV3 is very less then another model. Although successful, its weaknesses are that it excludes certain diseases such as Mealybug Infestation and does not identify the severity of the disease stage by stage. The ultimate goal is to create an Android application to serve rural farmers. In this study provides an accurate and scalable solution for early disease detection in agriculture sector which helping reduce crop losses, improve food sustainability and assure achieving farmer profit.