Abstract. This paper considers the problem of automated assessment of atmospheric transparency in urban conditions based on images obtained from an unmanned aerial vehicle, using computer vision and deep learning methods. The study explores an approach focused on analyzing visual signs of smoke near the horizon, where the concentration of aerosol pollution is usually most pronounced. For the experimental study, a specialized dataset was formed, including aerial photographs of the urban atmosphere of Almaty taken in January-February 2024, followed by spatial division of the images into nine sectors and manual visual assessment of the transparency level on a discrete scale. This method of marking allowed us to record the spatial heterogeneity of pollution within a single frame and take into account the differences between the sky background, the horizon line, and urban development. Based on the pre-trained MobileNetV2 architecture, two model variants were implemented — classification and regression — which made it possible to compare discrete and continuous approaches to the interpretation of visual information. A comparative analysis showed that the classifier provides higher accuracy of strict class matching (83.9%), while the regression model, when rounding predictions to whole values, demonstrates higher accuracy within a tolerance of ±1 class (97.2%) and a lower level of systematic errors. The results confirm the promise of using UAVs in combination with computer vision methods for local monitoring of atmospheric transparency and highlight the potential of this approach as a supplement to traditional ground-based environmental monitoring systems in urban environments, which is particularly relevant given the limited density of stationary stations.

Keywords: air quality monitoring, UAV data; computer vision; atmospheric transparency, smog; deep learning.