Smartphone-based Leaf Disease Detection via TinyML on-Farm Cameras

Thesis Abstract

The rapid adoption of smartphones among smallholder farmers provides an opportunity to leverage TinyML-enabled on-device inference for real-time leaf disease detection, addressing the persistent constraints of expert diagnosis, delayed treatment, and limited access to centralized crop health services in rural farming communities. This study investigates the efficacy of on-device machine learning models integrated into on-farm cameras for accurate leaf disease classification, coupled with actionable feedback for farmers. The aim is to develop and validate a low-latency, energy-efficient smartphone solution that sustains high diagnostic accuracy under field conditions, enabling timely agronomic decisions and reducing yield losses. The specific objectives are (1) to design a robust data collection protocol that captures diverse leaf images across prevalent crops (e.g., tomato, pepper, maize) and disease spectra (e.g., bacterial speck, septoria leaf blotch, powdery mildew) under varying illumination and backgrounds; (2) to curate a representative, balanced dataset of 50,000 labeled images and implement data augmentation and domain adaptation techniques to enhance generalization; (3) to develop TinyML models (e.g., quantized MobileNetV3 and EfficientNet-Lite variants) optimized for mobile hardware with target latency under 120 milliseconds per inference and model size below 5 MB; (4) to evaluate model performance against cloud-based baselines using metrics such as accuracy, F1-score, precision-recall, and area under the ROC curve, across cross-location validation; (5) to assess user acceptance and practical utility through a field trial with 150 farmer participants, examining decision impact and economic outcomes; and (6) to formulate deployment guidelines and ethical considerations for on-device plant health diagnostics. The methodology adopts a mixed-methods research design anchored in a pragmatic paradigm. The population comprises smallholder farms in two agroecological zones with diverse crop portfolios. A stratified sampling approach yields a sample of 30 villages, with 5–6 farms per village, totaling approximately 180 farms. A two-phase data collection process is employed an image acquisition phase to build the dataset and a field trial phase to test the deployed application. Data collection instruments include smartphone cameras (48–64 MP primary sensors with consistent autofocus), standardized camera settings, a field protocol app for tagging metadata (crop type, disease label, growth stage, weather conditions), and a survey instrument to capture farmer usability and decision outcomes. Validity and reliability are established through inter-rater labeling checks (Kappa > 0.80), cross-validation with expert agronomists, and test-retest reliability of the app under varying conditions. Data analysis proceeds in two streams quantitative analysis utilizes cross-validated, on-device inference metrics, paired t-tests comparing on-device versus cloud accuracy, and multivariate regression to identify determinants of diagnostic performance (lighting, leaf angle, background complexity). Qualitative insights from farmer interviews are analyzed via thematic analysis guided by the Technology Acceptance Model (TAM) and the Diffusion of Innovations framework to contextualize adoption drivers. The study incorporates a theoretical basis with two relevant theories Activity Theory to frame tool use within on-farm practices and the Theory of Planned Behavior to interpret intention to use the technology. Model development uses transfer learning with pre-trained lightweight architectures, followed by quantization-aware training to meet device constraints. The model is evaluated using held-out test sets across crops, with additional cross-site validation to assess robustness. Feature importance analyses identify spectral and texture cues most influential for disease discrimination. An ablation study quantifies contributions of data augmentation, domain adaptation, and model size constraints. For the field trial, the on-device detector guides farmers to recommended actions (e.g., targeted fungicide application, crop sanitation), and outcomes are tracked for 6–8 weeks to estimate agronomic and economic impact. Expected findings include high in-field diagnostic accuracy (target >92% overall across main diseases), low inference latency (<120 ms per image) and small model footprint (<5 MB), with performance gains from domain adaptation and augmentation techniques. The study anticipates that on-device inference will reduce the time to diagnosis by an average of 60–120 minutes per day and decrease unnecessary pesticide usage by up to 25%, contributing to cost savings and environmental benefits. The contribution to knowledge encompasses (i) a validated TinyML deployment framework for leaf disease detection on smartphones in resource-constrained settings, (ii) empirical evidence on the effectiveness of on-device diagnostics in improving agronomic decisions, and (iii) methodological guidance for integrating user-centered design and theoretical models into ICT-driven crop-health interventions. The main conclusion posits that smartphone-based TinyML on-farm cameras can provide accurate, rapid, and scalable leaf disease detection, with positive implications for yield protection and sustainable crop management. Recommendations include expanding disease spectra, extending the approach to additional crop-pathogen systems, refining user interfaces for broader literacy levels, and establishing data governance protocols to protect privacy and ensure equitable access for smallholder farmers.