Disease and Crop Classification Based on Deep Learning Approaches

Abstract

Plant disease causes annual crop losses that economically affect both growers and countries, and the resulting increase in agrochemical use in recent decades has been disconcerting. This study aims to develop and employ deep learning models to help individual growers correctly identify specific crops and diseases with five classification methods conducted using two approaches. More specifically, a convolutional neural network (CNN) was trained from scratch using three architectures, namely a classical CNN and both inception and residual blocks. Alongside, the transfer learning approach was examined using state-of-the-art VGG16 and MobileNet models. The novel raw data used was provided by an industrial party and consists of infected crop images extracted from a system used by growers and agronomic organisations. All images were taken in cultivation conditions and represent different parts of various infected plants. The data preparation progressed through different stages to clean out any irrelevant images, and augmentation techniques were applied for various purposes. Due to hardware limitations, the training phase involved three subsets of the dataset, while the testing phase involved two sets utilised to evaluate the performance of the resulting models. The first test-set was a holdout set from the trained data, and the second test-set having been extracted with a later date than the first. The highest accuracy achieved by the classical CNN was 80% in the test-set of first subset and 63% in the other test data. Intensive comparisons of all three subsets across all trained models to explore the performance of the different deep learning models have been conducted. This study suggests that, further research on images of healthy plants could also be included to improve the applicability of the models. Further enhancement, extending the model to be able to identify more than single disease type by adding annotated infected plant images would be a distinctive feature in the future.