Smart Waste Sorting using AI-Assisted Image Recognition and IoT Sensors

Thesis Abstract

This study addresses the escalating challenge of municipal solid waste misclassification and contamination that undermines recycling efficiency, resource recovery, and urban sustainability in rapidly urbanizing environments. It investigates how an integrated system of AI-assisted image recognition and Internet of Things (IoT) sensors can enhance waste sorting accuracy, reduce cross-contamination, and improve material recovery rates. The aim is to develop and validate a scalable, real-time waste sorting framework that combines computer vision, sensor data fusion, and decision-support algorithms to automate material separation at point of waste input and during material handling processes. Specific objectives are (1) to design a multimodal dataset comprising household and commercial waste images annotated by material class (e.g., paper, plastic, metal, glass, organics) and synchronized with IoT sensor readings (weight, moisture, and location); (2) to develop a convolutional neural network (CNN) architecture augmented with attention mechanisms for robust classification across varying lighting and occlusion, and to implement edge computing capabilities for on-site inference; (3) to implement a data fusion model that integrates visual features with sensor metadata using a probabilistic framework to enhance decision accuracy; (4) to evaluate the system in a real-world pilot across 12 municipal bins and a compact transfer station, with performance benchmarks including classification accuracy, contamination rate reduction, and throughput; and (5) to assess socio-technical impacts, including user acceptance, operational cost implications, and energy efficiency. The methodology adopts a mixed-methods research design combining empirical performance evaluation with qualitative stakeholder insights. The population includes municipal waste streams from a mid-sized city with diverse waste composition. A stratified random sample of 12 collection bins and 2 transfer stations is selected, with continuous data collection over a 6-month pilot period. Data collection instruments comprise (i) a labeled image dataset captured by overhead cameras and smartphone uploads, (ii) IoT sensor arrays measuring weight, humidity, temperature, and proximity, (iii) an on-site edge-computing device hosting the AI inference engine, and (iv) semi-structured interviews and focus groups with facility operators and municipal staff. Validity and reliability are addressed through cross-validation on a held-out test set, k-fold cross-validation for model tuning, inter-rater reliability for annotation, and calibration of sensor readings. The analysis employs a hierarchical CNN with transfer learning using ImageNet-pretrained weights, fused with a Bayesian network to integrate sensor features, and an ensemble voting mechanism to produce final waste class decisions. Model performance is assessed using accuracy, precision, recall, F1-score, and area under the ROC curve, while contamination rate and sorting throughput are analyzed via time-to-sort metrics and process capability indices. Statistical validation includes repeated-measures ANOVA to compare performance across bin types and environmental conditions, with post hoc tests for significant differences. The study also employs regression analysis to explore relationships between system accuracy, energy use, and throughput, supplemented by thematic analysis of interview data to elucidate adoption challenges and facilitators. Expected findings include a significant increase in bag-level sorting accuracy (target >92%), substantial reductions in cross-contamination between recyclable and residual streams, and improved overall material recovery yield by 8–12% relative to baseline manual sorting. The integration of image-based classification with sensor data is anticipated to outperform image-only and sensor-only baselines, particularly under variable lighting and clutter. The pilot is expected to reveal cost-neutral or short payback periods within 3–5 years when considering labor savings, reduced contamination penalties, and energy efficiency gains from optimized routing. Contribution to knowledge encompasses (i) a transferable ICT-driven framework for smart waste sorting that combines state-of-the-art computer vision with IoT-enabled environmental sensing, (ii) empirical evidence on the effectiveness and economic viability of edge-accelerated AI inference in municipal waste facilities, and (iii) theoretical advancement through an integrative model that links technological capability with organizational implementation, drawing on diffusion of innovations and socio-technical systems theory. The main conclusion anticipates that AI-assisted image recognition, when fused with IoT sensor data and deployed at source-level and transfer stations, can substantially improve sorting accuracy and resource recovery while delivering measurable efficiency gains. Recommendations include scaling to multi-city pilots, enhancements to adaptive learning with drift detection to accommodate evolving waste streams, investment in energy-efficient hardware, and the development of standardized data protocols to facilitate cross-case benchmarking and policy integration.