Paper Details

Name: Daisy Wadhwa, Arvind Kumar & Kamal Malik
Country: India

Year: 2025
Volume: Volume-12, Issue-2 (July-December)
Page Number: 153-175
DOI: https://doi.org/10.5281/zenodo.17346239

ABSTRACT
The prevalence of crop diseases presents a major challenge to global food security and
agricultural sustainability, causing significant yield losses and economic damage.
Conventional disease detection methods, which rely on manual inspection, are often
inaccurate, time-consuming, and impractical for large-scale implementation. While deep
learning, especially Convolutional Neural Networks, has shown promise in automating
disease classification, the practical deployment of these models is hindered by high
computational demands. This study proposes an efficient MobileNetV2-based deep learning
model tailored for classifying 36 healthy and unhealthy categories across 16 plant species.
The dataset used in this research combines real-field and lab-controlled images from multiple
public sources, enhancing the model’s generalizability. The model was trained with six
different optimizers, and Nadam was identified as the most effective, yielding 93.51% test
accuracy. To further enhance performance, Optuna-based hyperparameter tuning was
employed. The fine-tuned model attained 98.82% test accuracy, with precision, recall, and
F1-score of 0.9882 and ROC AUC of 0.9999, reflecting a 5.68% improvement over the
baseline model. The findings emphasize the feasibility of deploying a lightweight, highperformance model for real-time crop disease detection. By offering a scalable and
computationally efficient approach, this study advances sustainable agriculture, enabling
timely disease identification and improved crop management.
KEYWORDS:
Deep learning, convolutional neural networks, lightweight model, transfer learning,
sustainable agriculture, plant disease