Abstract:
The accurate and early detection of plant diseases is critical for ensuring
agricultural productivity and sustainability, particularly for medicinal plants
like Soursop (Annona muricata), which holds significant therapeutic value.
Traditional methods for disease identification are often labor-intensive,
subjective, and inaccessible to farmers in remote areas. To address these
challenges, this study proposes an edge-deployable lightweight ensemble model
combining multiple Convolutional Neural Network (CNN) architectures for realtime Soursop leaf disease detection. A comprehensive methodology was adopted,
involving data collection, preprocessing, model training, ensemble fusion, and
TensorFlow Lite (TFLite) optimization for mobile deployment. Four state-of-theart CNN models—VGG19, ResNet101, InceptionV3, and DenseNet201—were
individually trained and evaluated. The experimental results demonstrate that
VGG19, InceptionV3, and DenseNet201 achieved outstanding validation
accuracies of 99.84%–100%, with high AUC scores of 1.00, indicating excellent
model performance. In contrast, ResNet101 underperformed, achieving only
32.90% validation accuracy despite a high AUC score, highlighting issues of
overfitting or dataset mismatch. The ensemble model, developed through a softvoting strategy, achieved a perfect 100% validation and testing accuracy, with
an AUC score of 1.00, outperforming all individual models. This lightweight
ensemble approach ensures high classification performance while maintaining
low computational complexity, making it suitable for real-time applications in
smart farming environments. The final system is optimized for mobile devices,
enabling farmers to perform offline disease detection efficiently without
dependency on cloud services. Overall, the study demonstrates a robust, scalable,
and sustainable solution for advancing precision agriculture, empowering
farmers with accessible AI-driven tools to improve crop health monitoring and
preserve valuable medicinal plant resources.
