Manuscript received May 6, 2026; revised June 5, 2026; accepted July 9, 2026; published November 14, 2026.

Abstract—Air pollution is one of the major global challenges affecting environmental quality and human health. The conventional monitoring stations are limited in number and expensive to deploy in highly populated cities. Although fixed sensors give accurate measurements, their limited coverage restricts fine-grained spatial analysis. As a result, visual air quality estimation has gained attention by using image-based cues such as haze density, contrast reduction, and color distortion. However, most existing solutions formulate this problem as image-level classification or regression. In this paper, a lightweight air-pollution detection framework based on the official Ultralytics You Only Look Once (YOLO)26n architecture is introduced. Unlike classical whole-image classification, the problem is formulated as a single-region sky-localization detection task with the model jointly localizing the dominant sky region and identifying the sky pollution level. A systematic data processing pipeline was developed, and it includes duplicate elimination using Message-Digest Algorithm (MD5) hashing, sky area annotation, and six-class pollution level labeling. The proposed model is compared with recent object detectors like YOLOv11n, YOLOv12n, Roboflow Detection Transformer (RF-DETR), Mobile Network V3 Single Shot MultiBox Detector Lite (MobileNetV3-SSDLite), and RetinaNet, under the same experimental conditions. The experimental results demonstrate that the proposed YOLO26n-based model achieves mean Average Precision (mAP)@0.5 of 0.953 and mAP@0.5:0.95 of 0.914 with 2.51 million parameters and 5.8 Giga Floating Point Operations (GFLOPs). Overall, the framework provides a practical balance between accuracy, efficiency, and interpretability to scalable levels of monitoring air pollution.

Keywords—air pollution detection, Ultralytics You Only Look Once (YOLO)26n, sky-region localization, sky image analysis