Dynamics of street views and socio-economic conditions in profiling illegal dumping ‘black spots’: An LLM-enabled study in Hong Kong

Illegal dumping remains a persistent urban problem. Previous research has established that a neighborhood’s socioeconomic status and certain urban features, observed from a bird’s-eye view, influence dumping behavior. However, environmental criminologists contend that granular, eye-level street views offer more immediate and relevant environmental cues for potential offenders. This study aims to develop an explanatory model to profile illegal dumping ’black spots’ in urban areas by employing street view analytics. The innovative aspect of this approach lies in leveraging emerging large language models (LLMs) to extract street-level cues, which are then combined with census-based socioeconomic indicators using a spatially adaptive Geographic Random Forest. The model achieved a predictive accuracy of R2 = 0.7574 and an RMSE of 0.9368 on the held-out test set. Local feature analysis revealed that compact hotspot clusters with visible waste or dense vegetation significantly increase illegal dumping risk. Compared to traditional computer vision methods, LLMs proved more efficient in extracting meaningful features without manual annotation or specialized training. These findings demonstrate that integrating scalable, LLM-derived environmental cues with spatial machine learning enables more targeted and effective interventions for urban waste management.