چكيده لاتين
Abstract
Aerosols play an important role in the Earthʹs climate system and human health. AOD, which represents the amount of aerosols in a column of the atmosphere, has special importance in air quality studies and in assessing the effects of air pollution on human health. Satellite data, especially the MAIAC product, are a valuable source for AOD analysis. However, these data face the challenge of gaps caused by remote sensing limitations, surface conditions, cloud cover, and algorithmic constraints. The aim of this study is to provide a framework and recommendations for filling the AOD gaps in the MAIAC algorithm considering the type of gap. The gaps were classified into three types based on the quality flags (QA) of the MAIAC product: type I gaps (pixels without data), type II gaps (pixels with high uncertainty), and type III gaps, corresponding to spurious edges caused by differences in regional models. In this research, simulations were conducted to evaluate AOD values for all three gap types.
To estimate the gap values, statistical methods such as Kriging and GWR and machine learning methods such as XGBoost and SVR were used. Geostatistical approaches captured the spatial distribution of AOD, while machine learning models such as XGBoost and SVR did not have a direct understanding of the spatial dimension and considered the neighborhoods as input features.
The results showed that XGBoost provided stable performance and was less affected by the size of the neighborhood. SVR was accurate in homogeneous data conditions and small gaps, but its performance declined when dealing with large gaps or sparse data. The GWR method was highly sensitive to neighborhood size; in scenarios with fewer neighboring data points, its stability and accuracy decreased. The Kriging method provided stable and robust performance and, even in large gaps and heterogeneous AOD conditions, maintained a controlled error and a high percentage of predictions within the expected error range.
In type II gaps, the results showed that only the XGBoost method was able to fill these gaps. Simulations and evaluations indicated that the estimation of these gaps was also influenced by spatial conditions and the level of regional pollution.
The findings indicate that given the nature of the data and spatial variability, filling AOD gaps with high accuracy is possible, and the choice of the appropriate method should be based on the type of gap, data distribution, and spatial conditions. Finally, for small gaps, all methods performed well, but estimating gaps larger than 10 km² is not recommended.
Keywords: AOD, Gap, XGBoost, SVR, GWR, Kriging
