On-Device Plant Disease Detection with CNN and Random Forest Ensembles

The study’s key contribution is an end-to-end, fully offline plant-disease detection system that combines features from six CNN architectures with a Random Forest classifier. Trained on a hybrid 81,686-image, five-crop dataset—including 10,000 field-collected images—the ensemble achieved 98.4% test accuracy and was deployed in an Android app for practical diagnosis in low-connectivity farming environments.