Development and Evaluation of a Fine-Tuned EfficientNet-B0 Model for Maize Disease Detection

Abstract Maize is the staple crop for millions of people in Sub-Saharan Africa, particularly for Zambia. Unfortunately, maize crops are exposed to several serious threats due to their susceptibility to foliar diseases like Maize Rust, Leaf Blight, Leaf Spot, Maize Streak Virus, and Maize Lethal Necrosis that may lead to great yield losses. Conventional methods of crop disease identification consist of field surveys that are not only subjective but also difficult to conduct for smallholder farmers. This paper presents the design and evaluation of a highly optimized version of the EfficientNet-B0 Convolutional Neural Network for the automatic detection of maize leaf diseases using maize leaf images obtained from real-world scenarios. The proposed model utilized the concept of transfer learning with ImageNet pre-trained weights and was trained on the Mendeley Maize Crop Disease (Leaf) Dataset which consists of 30,120 images in nine maize disease classes. The developed fine-tuned EfficientNet-B0 yielded 97.57% classification accuracy, macro precision of 97.61%, macro recall of 97.64%, and macro F1-score of 97.61%. From these results, it is evident that transfer learning and fine-tuning greatly boost maize disease classification accuracy while ensuring high computational efficiency. This study makes a significant contribution to precision agriculture as it offers an accurate and computationally efficient AI-based maize disease classification model, which could help smallholder farmers in early maize disease classification.