Abstract:
Maize is one of the most widely grown cereal crops in Ethiopia and the world. Maize common
rust, which is a common problem in the study area, is caused by the fungus Puccinia sorghi,
and it can severely affect the yield and quality of maize. It typically causes for the losses of 23-
65%. The traditional severity quantification of rust involves manual diagnosis and
identification, which can be time-consuming, expensive and sometimes leading to errors. In
this research, we proposed to develop a CNN-based model for quantifying maize common rust
severity and recommend suitable fungicide doses.
The proposed system includes data pre-processing, image segmentation, image augmentation,
model training and feature extraction, severity classification, performance evaluation and
pesticide dose recommendation. A dataset of 8000 maize leaf images, which were collected
from Haramaya University Rare Research Center, were used for training the CNN algorithm
and develop models. After exhaustive training of algorithm in different ratios, we found
optimum results at 70%, 15%, and 15% for training, testing and validation datasets,
respectively.
In this study, we evaluated the performance of five deep learning models (CNN types) -
Resnet50, VGG19, VGG16, CNN, and AlexNet - for classifying the severity of maize common
rust and recommending pesticide doses. Severity refers to the degree of damage caused by a
disease or pest to a particular crop. We conducted two experiments, one using dropout, batch
normalization, and early stopping techniques and the other without them, to test the impact of
regularization on model performance. Our results show that the models trained with
regularization achieved better accuracy and convergence speed than the ones without it. Among
the models, Resnet50 exhibited the best overall performance, with an average F1-score of 97%
and a mean absolute error of 0.096, followed by VGG-19 (F1-score 95%), AlexNet (F1-score
94%), VGG-16 (93%), and CNN (F1-score 92%). Using a CNN model with a Gradio interface
accurately recommended the appropriate dose of fungicide based on the severity of rust in the
maize leaves. Our findings suggest that deep learning models can be effective tools for maize
disease diagnosis and management, and that regularization techniques can enhance their
robustness and generalization performance
