Identification of maize leaf diseases using red, green, blue-based images with convolutional neural network (CNN) and residual network (ResNet50) models

Abstract

Maize (Zea mays) is a crucial global staple crop that serves as a primary source of food and income, especially for smallholder farmers. However, it is susceptible to diseases that drastically reduce yields if not controlled. Traditional methods of disease detection of visual inspections are often inaccurate and uncertain. Recent advances in computer vision and deep learning techniques have shown promise in improving image recognition for crop disease detection. This study aims to develop models for detecting maize leaf diseases at the subfieldlevel using red, green, and blue (RGB)-based images using convolutional neural network (CNN) and residual network (ResNet50) models. A dataset of 1500 maize leaf images representing seven categories of maize disease symptoms was collected from the maize fields in Mopani District, Limpopo, South Africa. The data were processed to train and compare two deep learning models, CNNs and ResNet50. Both models demonstrated good classification accuracy with ResNet50 outperforming CNN, achieving an accuracy of 78.76% compared to 71.01% for CNN. The findings underscore ResNet50 enhanced capability to classify maize leaf diseases more accurately than CNN, attributed to its deeper architecture. This study illustrates the potential for deploying deep learning model in detecting maize leaf diseases. This study supports the transformative potential of deep learning in advancing agricultural practices, serving as a vital tool for early disease detection and contributing to food security in maize-producing regions, particularly smallholder farming systems. Therefore, this study trains the models that can be included in the mobile applications to be used to detected diseases in a sub-field level of the smallholder farms.

HIGHLIGHTS • Maize disease detection with CNN and ResNet50 models using RGB images. • Dataset included both nutrient deficiencies and disease symptoms. • Image preprocessing and data augmentation to increase training data and reduce overfitting. • Demonstrated the potential of deep learning for multi-disease detection.