Tomato Blight Classification Using Transfer Learning and Fine Tuning

Plant diseases are a huge hazard to the world’s food supply. Farms are frequently exposed to pest and disease outbreaks. Tomatoes are the second-largest vegetable crop grown worldwide. Early identification of plant blight helps farmers to take proper care of tomato plants hence they can secure their agricultural income. Farmers need a quick, consistent, and reliable method to identify plant diseases early. So, we have prepared a React Native Mobile application through which farmers can take a picture of their tomato crop leaf, and our CNN model predicts the type of blight leaf has, if any, or shows it is healthy. CNN deep learning models are widely used for image classification. Lightweight Pre-trained models are readily available and are trained on ImageNet data. We have prepared different models with pre-trained models using transfer learning and fine-tuning techniques. We have selected MobileNetV2, NASNetMobile, and EfficientNetV2B0 pre-trained models for the study as these are lightweight models and which was the basic requirement as we have built the mobile application. We have compared the performance of the models for different learning rates, batch size, and dropout with different metrics such as accuracy, loss, precision, recall, f1-score, and AUC for the 100 epochs. The best classification performance is shown by the model using EfficientNetV2B0 with a training accuracy of 99.95%, validation accuracy of 98.12%, and testing accuracy of 97.58% and the size of the model is 30.6 MB which is easily deployable for the mobile application.