Model For Detecting Common Bean Fungal Leaf Disease Using Deep Convolutional Neural Network

Abstract

Agriculture forms the basis of food security and economic growth in most countries. Pest and diseases remain to be a significant challenge and a big hindrance to the success of small-scale farming. Pest and diseases are responsible for heavy losses through death of crops and reduced productivity. In Kenya, common bean is the most important pulse and is the third most important food crop. Fungal based angular leaf spot and rust are two major diseases of common beans in the tropics and sub-tropics. Therefore, there is a need to provide a reliable and accessible technical solution for farmers to detect early detection of common bean leaf fungal diseases in Kenya. The main objective of the current study is to develop a deep convolution neural networks model for detection of common bean fungal leaf diseases in Kenya. The data for training was extracted from the GitHub data (Al. Lab. Makerere, 2020). Testing was done using SoftMax activation function in the output layer to provide a range of probabilities to the various output options. The initial TensorFlow model was built using the CRISP-DM methodology. The ResNet-50 model was adopted and custom layers were built using transfer learning. The TensorFlow Lite framework was used to convert and optimize the model. Float16 quantization was used to optimize the model. Performance metrics, including accuracy, precision, and recall, were used to evaluate the model.