Forecasting Air Quality in Kiev During 2022 Military Conflict Using Sentinel 5P and Optimized Machine Learning
Mohammad Mehrabi, Marco Scaioni, M. Previtali
Abstract
Recent studies have demonstrated that the Ukraine-Russia war has incurred evident changes to anthropogenic activities in the Kiev metropolis. Hence, this work employs Sentinel 5P imagery and a novel artificial intelligence model for predicting air pollution in Kiev. A well-known machine learning model, namely multi-layer perceptron neural network (MLPNN) is coupled with electromagnetic field optimization (EFO) algorithm to predict the daily concentration of particulate matter 2.5 (PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> ). Initially, a dataset is prepared by collecting eleven meteorological, atmospheric, and temporal factors from remote sensing and ground measurements. Then, principal component analysis (PCA) is used to determine the most contributive factors and create a reduced dataset. Four scenarios are defined by considering the reduced/original dataset, along with, predicting the current day/one-day-ahead PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> . A sensitivity analysis revealed that the most accurate results were achieved for predicting one-day-ahead PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> using the reduced dataset. After adjusting the EFO-MLPNN hybrid model, its performance is compared to classical MLPNN and adaptive neuro-fuzzy inference system (ANFIS). According to the results, the EFO-MLPNN with root mean square error (RMSE) 6.68 μg m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-3</sup> and Pearson correlation coefficient (R <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</sub> ) 0.82 outperformed both MLPNN and ANFIS outcomes. These findings infer that by optimizing the MLPNN by EFO, it’s prediction accuracy can be improved. The proposed hybrid model is therefore recommended for more practical air quality estimation and decision-making in the studied site. Lastly, a monolithic neural-based formula is extracted from the EFO-MLPNN hybrid for the explicit prediction of PM <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2.5</sub> .