Exploring street-level landscape patterns around illegal dumping monitoring locations using unsupervised computer vision

This study presents a computer vision-based approach that characterises landscapes associated with frequent illegal dumping sites using street-level images. We collected street-level images from frequent illegal dumping monitoring locations in Seoul, South Korea, and classified them through feature extraction, dimensionality reduction, and unsupervised clustering. Gradient-weighted Class Activation Mapping (Grad-CAM) was then applied as a post-hoc interpretability tool to examine visual regions associated with each cluster. To provide an external assessment of the clustering results, we conducted a manual annotation-based validation experiment using 200 randomly sampled images. The results showed a moderate but meaningful correspondence between the unsupervised typologies and human annotations, particularly when noise images were excluded. We identified three recurrent landscape patterns among the clustered images: low-rise mixed-use landscapes with commercial signs, dense aged low-rise residential landscapes, and vegetation-dominated, poorly managed landscapes. These findings suggest that the proposed approach can serve as a scalable screening tool that complements field-based audits by helping prioritise locations for further inspection.