Air Pollucion Level Estimacion In Public Areas Using Image Segmentation And Validation Against Air Quality Sensor Data

Abstract

This study develops an air pollution level estimation system based on digital image processing, integrated with validation using a CO₂ sensor, and implemented as a Python-based desktop application. The proposed method employs image segmentation in the HSV color space using a multi-zone thresholding approach to detect smoke, haze, and other visual pollutants. Feature extraction is then performed by calculating the pollution area ratio, texture entropy, mean intensity, standard deviation, and average saturation. These features are combined into a composite score that serves as the basis for classifying air pollution levels into three categories: Low, Moderate, and High. In parallel, the system analyzes images of the CO2 sensor display by detecting the color ratio of green, yellow, and red indicators to estimate atmospheric CO₂ concentration levels. Cross-validation between the two methods was conducted using 100 paired images to evaluate the consistency of classification results. Evaluation results based on a confusion matrix indicate an overall system accuracy of 70.00%, with a Macro Precision of 37.90%, Macro Recall of 45.93%, and Macro F1-Score of 36.70%. The High category demonstrated the best performance, achieving an F1-Score of 81.53%, while the Moderate category obtained an F1-Score of 28.57%. The Low category could not be effectively detected due to data distribution imbalance. Cross-validation results showed a 37% agreement rate between image segmentation and CO2 sensor measurements, indicating that further improvements are required. This result also reflects the fundamental difference between visual pollution indicators and direct gas concentration measurements. The developed system is equipped with an interactive graphical user interface, confusion matrix visualization, and a multi-sheet Excel report export feature that includes detailed results, performance evaluation, sensor analysis, as well as automatically generated conclusions and recommendations.