2Department of Computer Engineering, Dokuz Eylül University, İzmir, 35390, Türkiye; Information Technologies Research and Application Center (DEBTAM), Dokuz Eylül University, İzmir, 35390, Türkiye
Abstract
Automatic and early diagnosis of plant diseases is crucial for protecting crops and enhancing agricultural productivity. This paper introduces an efficient and lightweight deep-learning model to detect plant diseases from high-resolution images captured by digital cameras. The proposed model uses a convolutional neural network architecture that obtains high accuracy (96.55%) with significantly fewer parameters, decreasing computational complexity while maintaining strong feature extraction capabilities. To optimize performance, the network is downsampled using a fixed 2×2 MaxPooling2D operation, ensuring minimal loss of critical information. A real-world dataset to classify apple leaf images into diseased or healthy, is used to train a classifier model for specifically detecting Black Rot (Botryosphaeria obtusa), Cedar Rust (Gymnosporangium juniperi-virginianae), and Scab (Venturia inaequalis). Additionally, a mobile application is presented that provides farmers and agricultural workers with real-time disease identification and decision-support tools. The results of the experiments demonstrate that our approach outperformed state-of-the-art models (87.23% accuracy) while maintaining a capability for mobile deployment. These contributions position our model as a practical, accurate, and computationally efficient solution for plant disease detection in agricultural fields.