Deep Feature-Driven Learning Framework for Visual Diagnosis of Sunflower Crop Diseases Detection

Abstract

Agricultural production is exposed to losses when plants became infected with diseases from the environment, which in turn threatens food security; therefore, having the ability to detect plant diseases early on gives farmers the opportunity to minimize the loss they incur. Sunflower (Helianthus annuus) is Ranking among other oilseed crops worldwide for total agricultural production, and it has a high risk of being affected by diseases that cause reduced crop yield and quality, e.g., Downy mildew, grey mould, and leaf scars. Disease detection using conventional methods usually involves visual inspections by trained personnel; this process, however, is very time consuming, subjective, and could produce numerous errors. To counteract these issues, this paper proposes a hybrid learning framework for plant disease detection that combines deep feature-driven image analysis with a combination of two classifiers used in an ensemble mode to create a single final classification. The use of MobileNetV2 will act as the feature extraction process for the hybrid learning framework, while the classifier portion of the hybrid learning framework will consist of two K-Nearest Neighbour classifiers. Classification using SVM and RF classifiers will be completed using a combination of ensemble voting. Grad-CAM (gradient-weighted class activation maps) will be used to improve the interpretability of the disease classification results by identifying affected areas on images. Test results for the hybrid learning framework reveal an overall accuracy of 98.8%, compared to the accuracy of the two classifiers used separately: 95.8% for CNN, 93% for SVM, and 96.0% for RF. The confusion matrix for the two classifiers shows a significant number of accurate classifications with minimal misclassifications. The proposed hybrid learning framework for plant disease detection provides an effective method for detecting plant diseases in real time, interpretable results, and potential for scalability.