EcoScanner AI: An Explainable Computer Vision Framework for Real-Time Waste Classification and Carbon Mitigation Auditing in Circular Economies

Municipal solid waste streams remain undercharacterised at point-of-disposal, impeding material recovery and climate-positive interventions. We present EcoScanner AI, an explainable computer vision (XAI) system that performs real-time waste detection and quantifies carbon mitigation potential for circular economy decision support. The core architecture fine-tunes a YOLOv8s model on a Roboflow-hosted YOLO-format export of the TACO litter dataset, achieving a [email protected] of 37.66%, a precision of 93.82%, and an end-to-end inference latency of 12.47 ms (11.2 ms pure model forward pass) on a T4-class GPU. Explainability operates at two levels: Level 1 (Spatial) via bounding-box localisation and annotated overlays, and Level 2 (Contextual) by mapping detected classes to material categories with class-specific CO2 mitigation coefficients. The system implements a formal carbon mitigation accounting model that aggregates per-detection CO2-avoidance contributions, exposed through a Streamlit web application serving as an Environmental Decision Support System (EDSS), with SQLite persistence, bcrypt-backed identity verification, and a global leaderboard for behaviour reinforcement. A 7-day observational feasibility pilot with 18 participants recruited as a convenience sample from household and campus networks recorded 142 scan events and 3.42 kg of cumulative CO2-equivalent mitigation. Engineering for high availability includes model-resource caching and a cron-based keep-alive strategy. We report latency, throughput, and usability outcomes from this pilot and conduct a systematic ablation study justifying each architectural design choice.