Maize is a critical contributor to global food security but has consistent threats from many plant diseases that affect productivity. The ability to rapidly and accurately detect diseases in maize has great importance for understanding crop loss and promoting sustainable agricultural solutions. This paper provided a comprehensive comparative study of recent deep learning architectures for classifying four different states of maize leaves: three diseased states and a healthy state. A total of eleven models from the ResNet, DenseNet and EfficientNetV2 family with a specific set of parameters were trained and tested in a repeatable way. While all tested architectures produced high levels of accuracy and were all considered reasonable deep learning architectures for predicting maize leaf state, the most accurate was the EfficientNetV2-L architecture with an accuracy of 98.84% and an F1-score of 98.34%. The study also attempted to draw attention to tradeoff between predictive performance and computational cost. Specifically, results showed positive correlations between predictive performance and computational costs and demonstrated that all models improved predictive performance with increasing costs. Models such as DenseNet-169 and ResNet-50, also demonstrated reasonably low resource costs and strong predictive performance are interesting options. The results of this study provide an evidence-based approach for a researcher to select a deep learning model to automate the detection of diseases in maize, and all of the results offered interesting results that could be used for potential practical applications to guide the deployment of smart agricultural technologies.
Derinsu et al. (Sat,) studied this question.