Contrail Detection and Classification using Computer Vision with Ground-Based Cameras

The impact of aviation-induced contrails on climate change is significant, with contrail cirrus clouds substantially contributing to positive radiative forcing. Traditional contrail detection methods using satellite imagery face limitations due to low spatial and temporal resolution and atmospheric conditions that can obscure contrails. Groundbased cameras complement satellite data, enhancing the understanding of contrail formation and evolution. Improved detection and classification can facilitate more precise climate impact assessments and inform effective strategies to mitigate aviation-induced cloudiness. This paper presents an enhanced approach for contrail detection and classification using high-resolution ground-based cameras and a tailored Mask R-CNN model. Camera units were deployed at multiple locations to collect a diverse dataset of images capturing contrails in various developmental stages. Contrails were annotated and categorized into four classes: short, thin, wide, and cirrus. This reflects the lifecycle of a climate-impacting contrail cirrus. Our model incorporates modifications such as Focal Loss to address class imbalance and an improved mask head with increased feature map resolution and additional architectural enhancements to improve segmentation performance. The model was trained with about five thousand labelled images, cross-validated and evaluated using both quantitative metrics and qualitative assessments. The quantitative results indicate that the enhanced Mask RCNN outperforms the original Mask R-CNN, achieves precision comparable to YOLOv11-seg, and surpasses it in recall, thereby detecting more contrails overall. Qualitative evaluation highlighted its effectiveness in detecting contrails during their early stages, but also revealed challenges in classifying more evolved contrails with complex cloud structures. Subjective factors in defining class boundaries, particularly between thin, wide, and cirrus contrails, contributed to misclassifications, indicating that further refinement of the labelling process and model training could enhance performance in these areas. Herewith we demonstrate that our model outperforms baseline implementations and YOLOv11-seg models, particularly in detecting and classifying early-stage contrails. The approach not only identifies individual contrails but also classifies them into developmental stages. This new possibility is offering a foundation for next studies for example for the validation of numerical weather predictions, or as decision support about climate criticality when mitigating contrails.