High-Resolution Wintertime Satellite Detection and Segmentation of Coal Power Plant Plumes in Ulaanbaatar, Mongolia

Abstract

Ulaanbaatar, Mongolia, experiences some of the world's most severe wintertime air pollution. Residential-district fine particulate matter (PM2.5; particles ≤2.5 μm in diameter) concentrations can reach up to 620 μg/m3, which is up to 80 times World Health Organization guidelines. Coal combustion accounts for 27-92% of fine particulate matter depending on location. However, no existing study has successfully separated power plant emissions from residential coal burning, leaving the specific contribution of coal-fired power plants unquantified. This study investigates whether high-resolution satellite imagery combined with deep learning can reliably detect and delineate coal power plant plumes in Ulaanbaatar, and how well those models transfer to Bishkek Combined Heat and Power Plant (CHP), Kyrgyzstan, without location-specific fine-tuning. Adapting the methodology of Scott et al. (2023), we developed a two-stage pipeline using PlanetScope imagery (3 m/pixel): a ConvNeXt Large classifier for binary plume detection and a DeepLabV3+ segmentation model with Resnet-101 encoder for pixel-level plume delineation. The models were trained on 110 manually annotated wintertime images from Ulaanbaatar and evaluated on an independent test set from Bishkek CHP. The classification model achieved an F1 score of 0.933 with perfect recall (1.000) on the Ulaanbaatar validation set, demonstrating reliable plume detection. The segmentation model attained a pixel accuracy of 96.7% and a validation Intersection over Union (IoU) of 0.448, indicating moderate boundary delineation accuracy. Cross-location evaluation on Bishkek CHP revealed substantial performance degradation (IoU dropping to 0.319), confirming that domain shift remains a significant barrier to geographic transferability without fine-tuning. These results demonstrate the feasibility of satellite-based coal plume monitoring in Central Asia. However, they also emphasize the need for location-specific adaptation and larger, more diverse training datasets to ensure effective cross-regional generalization.