Smartphone-based Plant Disease Diagnosis Using Deep Learning and Leaf Imaging

Thesis Abstract

The rapid spread of plant diseases and limited access to expert phytopathology services in diverse agricultural settings necessitate scalable, field-portable diagnostic tools that can operate with minimal specialized equipment. This study addresses the need for an accessible, ICT-driven approach to early disease detection by leveraging smartphone-based leaf imaging and deep learning classifiers to diagnose common crop diseases with high accuracy, facilitating timely management decisions and reducing yield losses. The aim is to develop and validate a mobile diagnostic framework that integrates standardized leaf image capture, automated disease recognition, and user-guided decision support. Specific objectives include (1) constructing a diverse, labeled image dataset of 25 crop species across 15 disease classes under varying light and background conditions; (2) developing a convolutional neural network (CNN) architecture optimized for mobile inference, incorporating transfer learning from ImageNet and domain-adaptive fine-tuning to counteract environmental variability; (3) evaluating model performance against expert diagnoses using metrics such as accuracy, precision, recall, F1-score, and area under the ROC curve; (4) implementing on-device inference algorithms with model compression (quantization and pruning) to ensure real-time diagnosis on mid-range smartphones; (5) assessing user experience and diagnostic impact in field trials with agricultural extension agents and smallholder farmers; and (6) conducting a cost-benefit analysis to estimate potential losses avoided through early interventions. The methodology adopts a mixed-methods research design comprising four phases. Phase I involves dataset construction from field and greenhouse experiments, where approximately 15,000 leaf images are collected from 25 crops, with expert-annotated labels validated through consensus diagnoses by plant pathologists using microscopic and PCR-based confirmation for a subsample of 10%. Phase II develops a mobile-ready CNN model, employing a ResNet-50 base with domain adversarial training to improve robustness to lighting and background variation; model selection uses k-fold cross-validation (k=5) and statistically compares architectures with McNemar’s test. Phase III implements on-device deployment on representative Android devices (Pixel 6, Samsung Galaxy A54) and app-level usability testing with 60 farmers over 8 weeks, gathering qualitative feedback via semi-structured interviews and quantitative usability scales. Phase IV analyzes diagnostic performance and user impact using inferential statistics multivariate ANOVA to assess factors affecting accuracy across devices and crops, logistic regression to determine predictors of correct diagnosis in field settings, and a thematic analysis of interview transcripts to extract barriers and enablers of adoption. The study anticipates achieving an overall test accuracy exceeding 85% and per-class F1-scores above 0.80 for the majority of high-incidence diseases, with model robustness demonstrated across diverse environmental conditions. Expected findings include that domain-adaptive transfer learning substantially improves cross-crop generalization and that on-device inference maintains latency under 2 seconds per image in typical field scenarios. The research contributes to knowledge by integrating a scalable, user-centered, ICT-driven plant disease diagnostic pipeline with empirical validation across multiple crops and environments, and by delineating a practical framework for deploying deep learning models to resource-constrained settings. Theoretical grounding draws on the Technology Acceptance Model (TAM) to interpret user adoption patterns and the Unified Theory of Acceptance and Use of Technology (UTAUT) to examine performance expectancy and effort expectancy influences, supplemented by a pragmatic interpretive lens linking diagnostic accuracy to agricultural decision-making under uncertainty. The main conclusion is that smartphone-based diagnostics, when paired with domain-adaptive CNNs and streamlined user interfaces, offer a cost-effective, timely, and accurate approach to plant disease management for both large-scale farms and smallholder contexts. Recommendations include expanding the disease and crop catalog, integrating audio-visual disease cue cues for multi-modal diagnosis, linking the app to extension services and fertilizer/pesticide recommendation modules, and pursuing longitudinal field trials to assess long-term agronomic and economic benefits.