A Deep Learning and YOLO-Based Hybrid Framework for Accurate Classification and Region-of-Interest Detection of Maize Leaf Diseases
TL;DR
Results provide validation that combining classification and object detection has benefits on the diagnostic reliability and decision-making capability which can be used for actionable decisions in agriculture.
Abstract
Maize leaf diseases are also major causes of loss in the productivity of the crop, and they are also visually similar to each other, which frequently causes a misclassification under manual inspection. The paper fills this gap of the necessity of a practical and automated solution by coming up with a hybrid of deep learning frameworks with the ability to not only classify diseases with high accuracy but also localize region of interest (ROI) with high precision. The aim is to strengthen the early detection with an integrated system that is comprised of both the advanced models of transfer learning and the YOLO-based detection. The dataset is processed systematically to enhance normalization, preprocessing, and augmentation of the dataset to enhance its strength in a variety of imaging conditions. Several deep learning models VGG16, ResNet50, InceptionV3 and MobileNetV2 and lightweight and ensemble versions were tested. The results indicate that Ensemble Model 1 (VGG16 + MobileNetV2) performs the best, with the accuracy of 96% and the precision, recall, and F1-score of 95%, 95%, and 95%, respectively, being higher than individual classifiers and Ensemble Model 2 do. The ability to mix complementary feature extraction tasks of deep and lightweight networks is seen to be powerful in this superior performance. YOLO component is highly effective in detecting diseased regions on the leaf with high confidence identified and produces clear and understandable bound-box results. Results provide validation that combining classification and object detection has benefits on the diagnostic reliability and decision-making capability which can be used for actionable decisions in agriculture. The area of their activity is also the real-time deployment, applicability to cross crops, and integration with mobile or IoT-based smart farming systems, which allows extending the disease management solutions to scale and efficiency.