Smartphone-Based Plant Disease Diagnostics via AI-Driven Image Analytics

Thesis Abstract

This study addresses the escalating need for rapid, scalable, and accessible plant health diagnostics to mitigate yield losses and reduce reliance on laboratory assays. It proposes an AI-driven, smartphone-based framework for diagnosing leaf and fruit diseases from visual symptoms, leveraging on-device inference to enable real-time decision support for farmers in resource-limited settings. The aims are to (i) develop a mobile application capable of classifying plant diseases with high accuracy, (ii) quantify the diagnostic performance across crops common in smallholder systems, and (iii) evaluate user acceptance and practical impact on management practices. Specific objectives include assembling a annotated image dataset of 25,000 labeled plant-leaf samples spanning five crops (banana, tomato, pepper, maize, and cassava) collected under diverse agronomic conditions; training and validating a convolutional neural network (CNN) model optimized for mobile deployment using transfer learning and lightweight architectures (e.g., MobileNetV3, EfficientNet-Lite) to achieve a minimum top-1 accuracy of 92% on held-out test sets; implementing on-device inference with a cloud-assisted fallback for low-connectivity contexts; conducting a field trial with 200 farms to assess diagnostic accuracy, time-to-decision, and changes in disease management practices; and exploring user experience and adoption drivers through mixed-methods analyses. Methodologically, the research adopts a predominantly diagnostic-analytic design, with a cross-sectional validation phase and a longitudinal follow-up for impact assessment. The population comprises farmers and extension agents in tropical agro-ecologies, with a stratified sample of 1,200 image captures per crop across 300 farms for the validation dataset, and 200 farms enrolled in the field trial. Data collection instruments include standardized smartphone image capture protocols, disease labels verified by plant pathologists, expert-annotated ground truth datasets, and structured surveys supplemented by semi-structured interviews to capture usability and behavioral changes. Validity and reliability are ensured through inter-rater agreement (Cohen’s kappa > 0.80) on label quality, k-fold cross-validation (k=5) during model development, and test-retest reliability assessments of user-reported outcomes. Data analysis employs a suite of techniques model performance metrics (accuracy, precision, recall, F1-score, ROC-AUC) evaluated on stratified test sets; comparison against conventional diagnostic methods using paired t-tests and non-parametric equivalents; regression analyses to identify predictors of diagnostic adoption and subsequent management actions; ANOVA to examine performance variations across crops and environmental conditions; and thematic analysis of interview transcripts to elucidate user experience factors, guided by the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT). The study anticipates key findings CNN-based classifiers achieving ?92% top-1 accuracy across crops, with minimal drop-off (<3%) in field conditions when optimized for on-device inference; improved time-to-decision from minutes to seconds; statistically significant alignment between image-based diagnoses and pathologist labels (Cohen’s kappa ? 0.85); meaningful improvements in disease management decisions, evidenced by increased timely interventions and reduced unnecessary pesticide applications; and strong user acceptance, moderated by perceived usefulness and ease of use. The contribution to knowledge lies in delivering a validated, scalable, ICT-enabled diagnostic tool that integrates image analytics with mobile sensing to democratize plant health information, provide crop-specific disease profiles, and bridge the gap between research and practice in smallholder farming systems. The conclusion emphasizes that smartphone-based AI diagnostics can deliver rapid, accurate, field-ready disease identification with measurable impacts on management decisions, input efficiency, and yield resilience, particularly when embedded within farmer support ecosystems and extension services. The study recommends incorporating continuous learning pipelines to update models with new disease phenotypes, expanding crop coverage to legumes and tubers, integrating environmental metadata (temperature, humidity) to enhance predictive capabilities, and fostering multi-stakeholder platforms that couple diagnostics with advisory services, thereby strengthening crop protection strategies and food security in developing regions.