REAL TIME PEST AND DISEASE RECOGNITION MODEL FOR SELECTED VEGETABLE CROPS IN SRI LANKA : A CASE STUDY 

In Sri Lanka, agriculture is the primary source of livelihood and economic support, with vegetable farming playing a vital role in food security and significantly contributing to export earnings. Despite challenges such as outdated practices, limited technology, weak extension services, poor input supply, labor shortages, soil degradation, and poor information dissemination, pest and disease (P&D) infestations remain a major issue. Uncontrolled P&D causes significant yield losses and economic damage. The major barrier to addressing P&D challenges is the lack of context-specific, actionable information for timely, high-quality decisions, while accurate identification and effective management are  essential to minimize yield and economic losses. This study aims to design and develop a  framework for real-time detection and prevention of P&D attacks on vegetable crops in Sri Lanka using the Design Science Research (DSR) methodology. The framework integrates  severity-based matrices, ontology-based knowledge modeling, a custom Machine Learning (ML) model, and a mobile application prototype for real-time P&D management. Based on literature review, field visits, and formal and informal discussions, five vegetable crops and  their P&D were selected: potato (late and early blight), salad cucumber (whitefly and downy  mildew), brinjal (shoot and fruit borer and little leaf disease), chili (thrips and collar rot), and tomato (whitefly and tomato spotted wilt virus).The study initially focused on potato but  shifted to low-country vegetables due to COVID-19 challenges. Severity-based impact  matrices were developed by mapping crop growth stages to disease stages and pest  infestation ratios, assigning severity levels based on affected disease stages or pest ratios, with both reflecting potential yield impact if left uncontrolled. The matrices were developed, reviewed and validated by domain experts. A domain-specific, context-aware crop ontology was designed to recommend remedies based on severity by integrating matrices and relevant data on crops, pests, diseases, and remedies. The ontology was implemented in Protégé and  validated through HermiT reasoner, OOPS! pitfall detection, evaluated against 540 competency questions using SPARQL and DL queries, and externally validated by 15 agricultural instructors. A ML model was developed to identify disease stages, initially evaluating pre-trained architectures (MobileNetV2, DenseNet121, EfficientNetB0, ResNet50, VGG16). A customized lightweight CNN was adopted for edge-based deployment, achieving 93% test accuracy on a self-collected, expert-verified dataset of 3,000 augmented images of downy mildew on salad cucumber covering two consecutive stages.  An integrated framework combining the ontology and customized ML model was developed and refined with user inputs to improve accuracy and interactivity. A low-fidelity mobile prototype with Expert, Discover, and Diagnose modes was developed to deliver the framework to end users and validated by domain experts. Despite the limited stage-wise data availability, the framework demonstrates high scalability and generalizability. This study presents a standardized severity framework, a validated crop ontology, a high-performance custom ML model, and a user-informed hybrid system, advancing precise, context-aware P&D management in vegetable crops in Sri Lankan context. Future focus of study will include the expansion of the ontology and ML model, deploying the prototype via “Govi Nena” platform, conducting field level validations, exploring IoT integrations, and pursuing national collaborations to enhance impact on Sri Lanka’s agriculture. 

Keywords: Disease Stage Classification, Hybrid Framework, Machine Learning, Ontology, 
Precision Agriculture, User Centric Mobile App