Tomato Leaf Disease Diagnosis Using Bayesian Convolutional Neural Networks

Plant disease diagnosis is a very important task for maintaining agricultural productivity. Convolutional neural networks (CNNs) are popular tools in the field of image recognition and have been widely used in plant disease diagnosis. However, they are prone to small data set overfitting and their inability to quantify uncertainty in their predictions can lead to overconfidence and errors. To address these issues which hinder the construction of practical plant disease diagnostic models we propose the use of Bayesian convolutional neural networks (BCNNs). BCNNs are robust against overfitting regardless of the size of the data set and the probability distribution of the models' weight parameters can be estimated to obtain predictive distributions whose standard deviation expresses their predictions' confidence. In this work, we develop a BCNN for plant disease diagnosis and perform comparative experiments between conventional CNN and our BCNN for plant disease diagnosis. Our results show that our BCNN model is more robust to overfitting than the CNN model on small data sets and that the confidence in its predictions can be when diagnosing plant diseases.