Ensemble Learning Framework for Image-Based Crop Disease Detection Using CNN Models

Abstract

Crop diseases pose a significant threat to global food security, causing substantial yield losses estimated at 10-40% annually. Traditional methods of disease identification, reliant on visual inspection by farmers or experts, are often subjective, time-consuming, and limited by the availability of specialists. This study proposes an ensemble learning framework for robust image-based crop disease detection, specifically designed to address the challenges of heterogeneous, non-Independent and Identically Distributed (non-IID) agricultural datasets in decentralized environments. Utilizing the Plant Village dataset, we implement a stacking ensemble model integrating diverse Convolutional Neural Networks (CNNs) such as VGG (Visual Geometry Group), ResNet, and Inception as base learners, with a meta-learner to optimize prediction fusion. The system employs comprehensive data preprocessing, including resizing, normalization, noise removal, segmentation, and augmentation, to enhance robustness against real-world variability. Transfer learning with ResNet50 was adopted as a baseline model. The baseline ResNet50 achieved 59% test accuracy across seven grape and potato disease classes. The ensemble model improved performance, attaining 63% accuracy with average precision, recall, and F1-scores of 56%, 52%, and 52% respectively. Class imbalance remained a limiting factor for certain categories. The ensemble learning approach outperformed individual models, demonstrating improved generalization across diverse datasets. Although computational demands and imbalance challenges persist, the system provides a promising AI-driven pipeline for accurate crop disease diagnosis, supporting sustainable agricultural practices.

Code and data availability

The paper's crop disease detection models were trained on the public PlantVillage dataset, which the authors explicitly state is available on Kaggle. This is the image input used directly for the paper's analysis. No author code, trained models, or other paper-specific assets are disclosed.