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

Aerial imagery has emerged as a powerful tool for environmental analysis and decision-making, enabling us to gain valuable insights. We present a comprehensive approach for performing semantic segmentation on aerial images of illegally dumped construction waste. We focus on the detection and analysis of the waste content to utilize it for circular economy. Leveraging the Segment Anything Model (SAM) developed by Meta, we produced highly accurate masks from aerial drone images. We created a dataset of over 46,000 manually labeled masks, which serve as ground truth for training and evaluation. Then we fine-tuned the ResNet-50 classification model together with the deep learning model. Our methodology combines the prediction of the classification model with these detailed masks to produce the final waste stream map. The map offers a comprehensive understanding of the open area allowing for further potential stocks analysis and economic evaluation. Overall, we achieved 86% detection accuracy on our full dataset, where for common classes the accuracy is higher. The waste identification can be used for economic and environmental decisions-making necessity of cleanup operations. The results also allow better planning of potential untapped stocks and treatment of different waste streams, aiding in local circular economy and waste management strategies. Our model development can serve the waste management and recycling sectors as well as municipal and national policy makers.