An Integrated Smart Framework for Post-Harvest Optimization and Market Intelligence

Abstract

Post-harvest losses in Sri Lanka's fruit and vegetable supply chain remain a critical challenge, attributed to the absence of integrated quality assessment tools, real-time market intelligence, and accessible decision support systems tailored to local conditions. Existing approaches address these problems in isolation, leaving smallholder farmers without a unified platform for quality grading, price forecasting, post-harvest advisory, and cultivation planning. This paper presents CropShield, a novel four-component AI framework designed to address these gaps for Sri Lankan agricultural stakeholders. The first component employs YOLOv8 for fruit detection followed by MobileNetV2 fine-tuned on a papaya dataset for defect classification across six categories and maturity classification across three stages, with Grad-CAM explainability and a Random Forest recommendation engine integrated with Department of Agriculture knowledge. The second component delivers price forecasting across eleven crop varieties using a hybrid ensemble of ARIMAX, XGBoost, and LightGBM trained on HARTI market data from 2008 to 2025. The third component provides a bilingual post-harvest risk advisory assistant supporting Sinhala and English, integrating real-time weather data with an NLP-driven prediction engine, with SHAP-based explainability for transparent advisory outputs. The fourth component implements a Random Forest-based crop suitability and yield estimation model using district-level agronomic data with SHAP explainability for interpretable crop recommendations. The defect detection model achieved 95.65% accuracy and F1-score under clean conditions and 93.48% under robust augmentation, while the price forecasting model achieved R2=0.986 and MAPE=3.16%. CropShield delivers a scalable, explainable, and farmer-accessible platform for evidence-based agricultural decision-making across Sri Lanka