An AI- and IoT-Enabled Smart Food Donation Management System with Explainable Food-Risk Prediction, Safe-Time Forecasting and Automated NGO Assignment

Abstract

Surplus cooked food from restaurants, hotels and cafés is often discarded because donors cannot verify its safety or reach an NGO in time. This paper presents a Smart Food Donation Management System that combines low-cost IoT sensing, explainable machine learning and automated NGO allocation in one web platform. An ESP32 node with DHT22 and MQ-135 sensors streams storage conditions, and high gas readings flag or block spoiled food. Six algorithms were compared with 5-fold cross-validation; the selected Gradient Boosting model classifies each donation as low, medium or high risk from temperature, humidity, food age and food type. A safe-time forecast estimates when food will become high risk, and exact Shapley values explain each prediction. An allocation module offers each donation to one NGO at a time, filtering by remaining capacity and by whether the NGO can arrive before the food becomes unsafe. On 600 heldout records the model achieved 97.0% accuracy, a macro F1-score of 0.957 and 98.7% high-risk recall, with no high-risk record classified as low risk. In a 50-day simulation the allocation method delivered 87.2% of meals safely, against 83.3% for nearest-NGO allocation, and all 45 functional tests passed.