Pronóstico de las concentraciones de SO2 y NO2 en Ecuador a partir de imágenes satelitales Sentinel 5P, mediante técnicas de Machine Learning

Abstract

Air pollution has become one of the main environmental problems worldwide due to its effects on both the environment and health in general. Both national and international governments have implemented efforts to measure and control air pollutant emissions from anthropogenic sources by installing atmospheric monitoring networks. However, not all cities and countries have these monitoring tools. For this reason, the use of satellite images has been gaining strength in recent years as it allows us to obtain satellite information from areas that do not have terrestrial monitoring and to be able to use this data for control, prevention and research purposes. Through this information we can perform analysis and modeling of emissions and behavior of atmospheric pollutants. Due to the need to be able to prevent society and take preventive measures regarding the emissions of atmospheric pollutants, the scientific community in recent years has proposed different mathematical models and unsupervised learning models that allow predicting the emissions of atmospheric pollutants. For them it is necessary to take into account the external variables that affect the behavior of pollutants depending on the study area, since the geographical location, topography, and meteorological conditions directly or indirectly influence this behavior, for this reason researchers generally design models for specific regions. There is no method to establish which meteorological variables should be used in the prediction of pollutants, the background to be used are the previous studies carried out, observing the results obtained to know the influences of these variables on the behavior of pollutants. The present work proposes two prediction models for the concentration of NO2 and SO2 for the three most important cities of Ecuador, based on information from Sentinel-5P, Giovanni NASA and ERA 5 satellite images. The first proposed model uses Recurrent Neural Networks using the number of lags or dummy variables created that are used to find relationships between concentration and meteorological variables, which provide information to the neural network to make the prediction. It was proposed to predict air pollution up to 5 days ahead with the use of different structures looking for the best one for the forecast. The second proposed model uses the Random Forest method taking into account two important characteristics, the maximum depth of each tree and the minimum number of samples to be considered Leaf Nodes. These two features give us two Bryam Montesdeoca Jara Dayana Ortiz Morocho iv perspectives about random forests looking for the best prediction model. It can be said that the prediction through the Random Forest Regression algorithm was the one that showed the best performance R2=0.98 and the error metrics MAPE, RMSE and PBIAS were lower in this method with values of 7, 3.67, 0.68, respectively. , emphasizing the different data sets, the prediction for the city of Cuenca was the best, followed by the city of Guayaquil, which slightly exceeds the predictions for Quito. This shows that the prediction of air quality is effective, showing satisfactory results and opening doors to new research in order to be able to anticipate the measurements of concentrations of polluting gases in the air and thus be able to make preventive decisions for both health and the environment.