Classification and severity prediction of maize leaf diseases using Deep Learning CNN approaches

Abstract

Maize (zea mays) is the staple food of Southern Africa and most of the African regions. This staple food has been threatened by a lot of diseases in terms of its yield and existence. Within this domain, it is important for researchers to develop technologies that will ensure its average yield by classifying or predicting such diseases at an early stage. The prediction, and to some degree classifying, of such diseases, with much reference to Southern Africa staple food (Maize), will result in a reduction of hunger and increased affordability among families. Reference is made to the three diseases which are Common Rust (CR), Grey Leaf Spot (GLS) and Northern Corn Leaf Blight (NCLB) (this study will mainly focus on these). With increasing drought conditions prevailing across Southern Africa and by extension across Africa, it is very vital that necessary mitigation measures are put in place to prevent additional loss of crop yield through diseases. This study introduces the development of Deep Learning (DL) Convolutional Neural Networks (CNNs) (note that in this thesis deep learning or convolution neural network or the combination of both will be used interchangeably to mean one thing) in order to classify the disease types and predict the severity of such diseases. The study focuses primarily on the CNNs, which are one of the tools that can be used for classifying images of various maize leaf diseases and in the severity prediction of Common Rust (CR) and Northern Corn Leaf Blight (NCLB). In essence the objectives of this study are: i. To create and test a CNN model that can classify various types of maize leaf diseases. ii. To set up and test a CNN model that can predict the severities of a maize leaf disease known as the maize CR. The model is to be a hybrid model because fuzzy logic rules are intended to be used with a CNN model. iii. To build and test a CNN model that can predict the severities of a maize leaf disease known as the NCLB by analysing lesion colour and sporulation patterns. This study follows a quantitative study of designing and developing CNN algorithms that will classify and predict the severities of maize leaf diseases. For instance, in Chapter 3 of this study, the CNN model for classifying various types of maize leaf diseases was set up on a Java Neuroph GUI (general user interface) framework. The CNN in this chapter achieved an average validation accuracy of 92.85% and accuracies of 87% to 99. 9% on separate class tests. In Chapter 4, the CNN model for the prediction of CR severities was based on fuzzy rules and thresholding methods. It achieved a validation accuracy of 95.63% and an accuracy 89% when tested on separate images of CR to make severity predictions among 4 classes of CR with various stages of the disease’ severities. Finally, in Chapter 5, the CNN that was set up to predict the severities of NCLB achieved 100% of validation accuracy in classification of the two NCLB severity stages. The model also passed the robustness test that was set up to test its ability of classifying the two NCLB stages as both stages were trained on images that had a cigar-shaped like lesions. The three objectives of this study are met in three separate chapters based on published journal papers. Finally, the research objectives were evaluated against the results obtained in these three separate chapters to summarize key research contributions made in this work.