Leveraging AI for Material Identification in Unauthorized Dumps for Circular Economy Applications

Our research presents an advanced method for semantic segmentation of aerial images to detect specific waste types in illegally dumped construction waste. This method focuses on supporting circular economy efforts. Utilizing Meta's Segment Anything Model (SAM), we developed accurate masks from drone-captured imagery, producing a dataset of over 46,000 manually labeled masks as the base database. We then fine-tuned the ResNet-50 classification model, integrating its predictions with these masks to create a comprehensive waste stream map. This map provides critical insights into waste composition, aiding environmental and economic decision-making, such as cleanup operations and resource recovery. Our approach achieved 86% overall detection accuracy, with higher performance on common waste types. These findings offer new opportunities for waste stream analysis, contributing to more efficient waste management and understanding untapped resources on the ground. They also support both municipal and national policies for sustainable waste and resource management.