Improving UAV-Based Detection of Low-Emission Smoke with an Advanced Dataset Generation Pipeline

This study improves the UAV-based detection of low-emission smoke by enhancing the dataset generation pipeline, addressing a key limitation of previous approaches, namely the lack of high-quality, automatically annotated datasets for small-scale chimney emissions. Traditional methods for smoke detection, including deep learning models trained on manually labeled images, suffer from inconsistency due to the transient nature of smoke and annotation subjectivity. To overcome this, we refine an existing motion-based annotation framework by integrating an LightGBM-based classifier, which filters out false positives caused by background motion. The proposed method was evaluated using a dataset of 634 stabilized UAV video sequences (10 s each), covering diverse environmental conditions, including winter scenes and oblique views. The improved pipeline generated 23,548 high-confidence training annotations, reducing false positives by 37% compared to the baseline motion-based approach. The trained YOLOv11 model achieved 0.98 precision and 0.91 recall on the refined dataset, significantly outperforming the previous version. Unlike prior UAV-based smoke detection studies, which primarily focus on large-scale wildfire smoke, this work targets low-intensity emissions from residential heating, ensuring the precise localization of individual pollution sources. By improving annotation quality and reducing manual labeling efforts, this study enables the more robust and scalable detection of urban air pollution sources, with potential applications in regulatory enforcement and environmental monitoring.