1. Introduction
Smallholder farmers often lack timely access to plant pathologists, and by the time visible symptoms are widely recognised, foliar diseases such as late blight can have already spread across a field, making on-device automated diagnosis attractive.
2. Methodology
A dataset of 14,200 tomato leaf images spanning ten classes, nine disease categories plus healthy leaves, was assembled from field collection across three districts and augmented with rotation, brightness and occlusion transformations. A MobileNetV3-Large backbone pretrained on ImageNet was fine-tuned with a two-phase schedule, first freezing the backbone then unfreezing the final 40 layers.
3. Results
The fine-tuned model achieved 96.3 percent accuracy on a held-out test set, with per-class F1-scores above 0.93 for all ten categories except leaf mould, which showed some confusion with early blight due to visually similar early-stage lesions. On-device inference on a mid-range Android phone averaged 78 milliseconds per image.
4. Conclusion
Transfer learning enables accurate, lightweight disease classifiers deployable directly on farmer smartphones without connectivity requirements. Future work will expand the dataset to cover additional regional cultivars and co-occurring nutrient deficiencies.
References
[1] Mohanty S. P. et al., Using deep learning for image-based plant disease detection, Frontiers in Plant Science, 2016. [2] Howard A. et al., Searching for MobileNetV3, ICCV, 2019. [3] Ferentinos K. P., Deep learning models for plant disease detection and diagnosis, Computers and Electronics in Agriculture, 2018.