Development of a deep learning model for the classification and detection of maize leaf diseases.
Abstract
Maize leaf blight and maize streak virus are global diseases that severely impact
maize crops, resulting in significant yield losses. Early detection of these diseases
is crucial for preventing their spread and minimizing economic losses. This project
aims to develop a deep learning model for automated detection using leaf images.
The proposed methodology encompasses image acquisition, preprocessing, feature
extraction, model training, and disease classification. The dataset was generated
by scientists from the Makerere Artificial Intelligence Lab and the National Crops
Resources Research Institute (NaCRRI) in Uganda, affiliated with the National
Agricultural Research Organisation (NARO).
Feature extraction employs convolutional neural networks (CNNs), specifically
MobileNet and a custom-built model. MobileNet leverages pre-training on
large-scale image datasets like ImageNet, while the custom model is developed from
scratch. Transfer learning fine-tunes these models using disease-specific features.
Performance evaluation utilizes a comprehensive dataset of annotated maize leaf
images infected with the diseases, measuring accuracy, precision, recall, and
F1-score. Generalization ability is assessed using an independent test dataset.
The MobileNet achieved 98.75% accuracy on the test dataset, and the Custom Model
achieved 99.69%. Additionally, a YOLOv8 model was trained for object detection.
The results demonstrate the high accuracy and promising potential of the
custom-built model for automated detection of maize leaf blight and maize streak
virus. Early identification facilitates timely interventions, preventing disease spread
and enhancing maize crop management. This technology contributes to increased
agricultural productivity.