Application of Real-Time Deep Learning in integrated Surveillance of Maize and Tomato Pests and Bacterial Diseases

Abstract

Limited access to agricultural expertise and reliable crop disease diagnostic technologies by small-scale farmers in Kenya greatly hinders  food production and security in the country. This study was aimed at investigating the potential of the use of machine learning (ML) for  real-time diagnosis of common tomato and maize diseases and pests using crop images captured by mobile phone cameras. Images  were acquired from farmers' fields in two counties in Kenya and used for training and testing two Convolution Neural network (CNN)  models for the classification of six classes of tomato crop disease and pest infections and a binary classifier for the identification of fall  armyworms in maize fields. Classification accuracies of 97.08% for the tomato model and 100% for the maize Fall Army Worm models  were recorded. The image dataset and code used for training and evaluating the models have been published in publicly accessible repositories. The recorded results strongly suggest the high potential of using ML tools to complement or supplement human extension  services to small-scale farmers.