Transfer Learning for Plants’ Disease Classification with Siamese Networks in low data regime

Timely disease detection in plants remains a challenging task for farmers. They do not have many options other than consulting fellow farmers. Expertise in plant diseases is necessary for an individual to be able to identify the diseased leaves. For this, Deep Convolutional Neuronal Networks based approaches are readily available to find solutions for various problems related to plant disease detection. Actually advanced deep CNN-based models successfully performed good accuracy. However, due to a smaller number of image samples available in the datasets, there exist problems of over-fitting obstructing the performance of deep learning approaches. In this work, we used a Siamese convolutional neural network (SCNN) model with different Transfer Learning (TL) models to classify plants diseases. In our approach, we extend the insufficient and various volume data by species using data augmentation techniques. Experiments are performed on a publicly available dataset open access series of imaging studies (Plant Village), by using the proposed approach, an excellent test accuracy of 96.77% is achieved for the classification of plants disease using variant training sample size especially those on low data regime. We proceed to compare Transfer Learning with Siamese Network with their state-of-the-art most CNN architectures and discovered that the proposed model using Siamese Network outperformed the state-of-the-art models in terms of performance, efficiency, and accuracy.