Performance comparison of ML models.

Illegal dumping is a major challenge of municipal solid waste management. A significant portion of waste is dumped illegally in cities worldwide. This improper waste disposal creates serious ecological degradation, environmental hazards, public health risks, and urban planning challenges. This study presents a Positive-Unlabeled (PU) learning framework for detecting illegal dumping sites using GeoAI. Seven machine learning algorithms were trained on 70% of 341 confirmed illegal dumping sites across 10 m × 10 m resolution grids covering Khulna City Corporation, Bangladesh. The remaining 30% of the data was reserved for independent testing. The best model was selected using a weighted composite score combining the Area Under the Curve (AUC) and F1 score from spatial cross-validation, and the selected best model was subsequently evaluated on an independent hold-out test set. Random Forest achieved the highest AUC (0.883) and F1 score (0.820), and consistently outperformed other models across all weighting schemes. The three most influential predictors of illegal dumping were proximity to roads, drains, and buildings. The predicted risk map shows that the very high and high-risk zones are concentrated along roads and urban centers. This study introduces the first GeoAI framework for illegal dumping site detection in Khulna, Bangladesh. This framework can also be applied in other cities to detect illegal dumping sites at the community level. This methodology can help municipal authorities to develop a waste management plan that addresses both illegal dumping challenges and long-term infrastructure planning.