Powdery Mildew Disease Classification in Laboratory and Real-Field Images using Convolutional Neural Networks for Precision Agriculture

Precision agriculture encounters challenges in detecting plant diseases. Powdery mildew is a fungal disease that impacts various crops. Accurately identifying this disease in agricultural field images is crucial for effective crop management. This research presents a systematic approach to assess the generalization capabilities of convolutional neural network (CNN) models, specifically DenseNet121, EfficientNetV2B0, InceptionResNetV2, InceptionV3, MobileNet, and ResNet50V2, in classifying powdery mildew disease when transitioning from laboratory images to real-field images. The experiment is organized into three steps, utilizing distinct subsets derived from the PlantVillage and Kaggle plant disease recognition datasets. PlantVillage supplies laboratory images, while the plant disease recognition dataset provides real-field images. The data is divided into five subsets, allowing for a comprehensive analysis across six training and evaluation scenarios. The results indicate that while all models achieve perfect accuracy in controlled laboratory conditions, their performance significantly declines when assessed on real-field data, underscoring the challenges of generalizing to complex, real-world environments. However, models trained on real-field data, or a combination of both datasets, demonstrate strong performance in real-field applications, with some CNN models achieving perfect accuracy. InceptionResNetV2 and InceptionV3 consistently excelled across various scenarios, positioning them as promising candidates for smart farming applications aimed at accurately classifying powdery mildew disease. This approach holds substantial potential for the agriculture industry, establishing a solid foundation for future studies.