Aplicación de modelos de aprendizaje de máquina e imágenes digitales para el pronóstico de la calidad del aire.

Abstract

Air pollution represents serious impacts on the health of the population; particularly exposure to high concentrations of tropospheric ozone can cause respiratory and cardiovascular problems, which is why an efficient air quality monitoring system is of great benefit to human health and air pollution control. In this research, air quality forecasting is addressed through the application of machine learning models and digital images as a function of the air quality index (AQI) for the pollutant O3, making use of the convolutional neural network technique, under a VGG16 modeling architecture. To evaluate the proposed method, a dataset was created containing a total of 366 images, recorded at 07:00 and 12:00 hours of the day, in the direction of four localities of the city of Bogotá, during a period of time established from March to October 2021.The set of images was classified against the data collected from the monitoring stations of: Las Ferias, Puente Aranda, Centro de Alto Rendimiento and Fontibón. The training and validation of the model was run at 50 epochs and a batch size of 64 samples. The model identified three categories of air quality index (AQI); good, acceptable and harmful to the health of sensitive groups. The results of the model performance metrics for these categories reflect in terms of accuracy (70%), (50%) and (46%) respectively, however, it is established that the performance can be improved by applying strategies of oversampling and algorithm refinement. The proposed model is valid in contrast to what has been identified in the literature and can be a promising tool for digital image classification.