An Offline Mobile Application for Plant-Disease Detection Using MobileNetV3 and TensorFlow Lite

Abstract

Crop losses to plant disease fall hardest on smallholder farmers who lack both expert diagnosis and reliable connectivity. This paper presents an offline mobile application that diagnoses leaf disease on-device using a MobileNetV3-Large classifier deployed through TensorFlow Lite, and pairs each diagnosis with a treatment recommendation validated with agricultural scientists at TNAU, in English and Tamil. Trained by transfer learning on a 19,606-image, six-class dataset, the model attains a dataset-weighted test accuracy of 94.4% (95% CI ± 0.6) and a validation accuracy of 95.7%, with per-class accuracy from 97.2% (healthy leaf) to 91.0% (bacterial spot). Inference runs in 38 ms on a Redmi Note 10 and 29 ms on a Samsung Galaxy A54 from a 5.5 MB model, entirely offline. Addressing review, we add 95% confidence intervals, a confusion matrix that localises the dominant disease confusions, a condition-robustness analysis, on-device CPU/RAM/battery profiling, and a comparison against lightweight baselines. The result is a reproducible, field-oriented precision-agriculture tool rather than a prototype.