Pixels to Pathogens: A Deep Learning Approach to Plant Pathology Detection

It is known that accurately identifying, early and timely treatment and elimination of the plant diseases is essential for crop protection and healthy crop growth. In traditional or conventional methods, identification and classification were done by testing in laboratories or through visual inspection by farmers. Now going through the testing in labs is very time consuming, while the visual inspection requires enough experience and knowledge. To solve this problem, our study proposes a robust plant pathogen detection method based on a Deep Learning approach on a large dataset containing about 38 categories of different species like Maize, Potatoes, Tomatoes, Bell Pepper, Peach, Strawberry etc. and diseases like rust, molds, blight (late and early). This crop disease detection model leverages the power of the EfficientNetB3 architecture, a state-of-art convolutional neural network(CNN). The main backbone is served by EfficientNetB3and then it is fine-tuned using different hyperparameters and other regularization techniques like weight decay, dropout method and optimizers like RAdam,to enhance the overall accuracy coupled with dynamic learning rate adjustment. In the testing set of the dataset, the proposed model shows encouraging accuracy of about 99.25%, high precision of about 97.35%. A thorough evaluation of the model’s functionality is given by the help of training and validation line chart and loss chart that gives the in-depth information on the prediction. And then we implemented the detection model in our mobile application whose interface screen shots are given below. In the application the image can be taken by camera or fed from folders and it will detect the type of disease.