Leaf Disease Detection and Pesticide Recommendation Using Pretrained CNN Models in Keras on the Augmented New Plant Diseases Dataset

Timely and precise identification of diseases in plant leaves is crucial for protecting crops and promoting sustainable agricultural practices. In this work, a custom Convolutional Neural Network (CNN) model built with Keras is presented to identify 38 different classes of leaf diseases from the Augmented New Plant Diseases Dataset. The model architecture consists of several convolutional layers with increasing filter sizes (from 32 to 512), ReLU activation functions, and max pooling for spatial downsampling. It concludes with fully connected layers and includes dropout to reduce the risk of overfitting. Training is performed using the Adam optimizer with a learning rate of 0.0001, and the loss is calculated using the sparse categorical cross-entropy function. To improve the model's robustness and adaptability to diverse lighting and environmental conditions, data augmentation methods are applied. The results confirm the model's high accuracy, with a training accuracy of 98.37% and a test accuracy of 96.29%, supporting its application for real-time detection of plant diseases at scale. Furthermore, a rule-based system is incorporated to recommend suitable pesticides based on the detected disease category. This approach emphasizes the value of deep learning in advancing smart agriculture and providing automated support for disease identification and treatment decisions.