Plant and disease classification with explainable AI in web-based application

Abstract

The application of deep learning for plant and disease recognition has become increasingly popular; however, most studies address these tasks separately due to limitations in dataset availability for model training and testing. This study aims to overcome such constraints by developing a multi-output classification framework that simultaneously predicts both plant species and associated diseases. Three state-of-the-art architectures were employed: NASNetMobile, a CNN-based model; a hybrid CNN-LSTM; and a CNN integrated with an Attention mechanism. A combined dataset of over 15.678 images was compiled from multiple public sources, covering 10 plant species and 27 disease classes with diverse real-world conditions. The training process was done using two different approaches: with and without data augmentation. Model performance was evaluated using accuracy, precision, recall, and F1-score. The results show that NASNetMobile without data augmentation achieved the highest performance, with an F1-score of 99.79% for plant classification and 98.54% for disease classification, outperforming CNN-LSTM (98.86% and 95.2%) and CNN-Attention (98.65% and 93.3%). These findings demonstrate that lightweight yet robust architectures such as NASNetMobile can effectively bridge the gap between laboratory-trained models and field-ready agricultural applications, supporting the advancement of precision agriculture. To enhance interpretability, Local Interpretable Model-Agnostic Explanations (LIME) and Eigen-CAM were applied, providing intuitive visualizations that help users understand model predictions. The best-performing model was deployed in a web-based proof-of-concept application, developed using Streamlit. This work provides one of the first multi-output explainable frameworks for plant and disease classification deployable in a web-based system.