Sugarcane (Saccharum officinarum) diseases including smut, yellow leaf disease, pokkah boeng, mosaic, brown spot, grassy shoot, sett rot, brown rust, and banded chlorosis collectively cause yield losses of 20-50% per annum across India, disproportionately affecting the 5.5 million smallholder farmers who cultivate over 5.5 million hectares annually. This paper describes a mobile AI application designed to enable real-time, non-expert identification of all eleven sugarcane leaf categories (nine diseases, healthy, and dried) directly from smartphone images. The application uses an EfficientNetB4 convolutional neural network pre-trained on ImageNet-21k and fine-tuned on a purpose-built dataset of 6748 field-collected images from Maharashtra, India. The model achieves a macro-average F1-score of 96.2% and an overall accuracy of 97.5% on a held-out test set, with per-class accuracy ranging from 95.4% (Banded Chlorosis) to 99.3% (Healthy). The application operates fully offline through INT8-quantised TensorFlow Lite inference, achieving average latency of 2.1–3.8 seconds on mid-range and low-end Android smartphones, with a disk footprint of only 18.4 MB. A pilot field trial with 18 farmers yielded 95.8% real-world classification accuracy and a task completion rate of 94.4% among first-time users, with an average time-to-diagnosis of 8.4 seconds compared with 3.2 days under the prior extension-officer model. Upon disease confirmation, the system retrieves localised remedial recommendations from an integrated agronomy database in Marathi and English. A user feedback loop enables weekly model retraining for continual improvement. The system is designed for Maharashtra’s cooperative agricultural system. Current constraints includes agricultural sector with single climatic zone.
Thite et al. (2026) studied this question.