تخمين محصول عمق نوري هواويز MODIS و بهبود قدرت تفكيك زماني آن‌‌، بر اساس داده‌هاي ماهواره زمين ثابت METEOSAT-9 با استفاده از يادگيري ماشين

AOD (Aerosol Optical Depth) is a key parameter for measuring particulate matter in the air. There are two main methods for estimating AOD: ground-based AERONET stations and satellite observations. Ground stations provide high accuracy but have limited spatial coverage, making it difficult to estimate AOD over large areas. Satellite products, on the other hand, offer broader coverage at lower costs but rely on physical methods and require a lot of physical data for AOD estimation. Despite advances in satellite technology, developing a method that can provide real-time AOD estimates with high temporal resolution remains essential. The goal of this research is to use machine learning and the reflectance data from the geostationary MSG-2 satellite’s SEVIRI sensor to improve the temporal resolution of the Terra sensorʹs MAIAC-AOD product in two different regions. Initially, to train and eva‎luate the algorithms using AERONET stations, data from Italy were utilized. Four machine learning methods—Support Vector Machines (SVR), Random Forest, XGBoost, and Multilayer Perceptron—were trained on the country’s stations. The results compared with AERONET data showed that XGBoost performed better than the other methods. To improve model accuracy, meteorological data were used alongside the SEVIRI sensorʹs reflectance data, leading to significant improvements. XGBoost achieved an RMSE of 0.04 and an R² of 0.70, while SVR showed the weakest performance with an RMSE of 0.08 and an R² of 0.48. To eva‎luate temporal generalization, models trained on data from 9:30 UTC were tested at 10:30 UTC. The XGBoost model had an R² of 0.8 and a correlation coefficient of 0.87 at 9:30 (training time), and an R² of 0.68 and a correlation coefficient of 0.75 at 10:30 UTC (eva‎luation time for temporal generalization). These results demonstrate the superior performance of XGBoost compared to the other methods. The models were also trained for the Isfahan region, where XGBoost again outperformed the other methods. In this region, the models were trained on data from 07:30 UTC (Terra satellite overpass time) and were then eva‎luated for temporal generalization using data from 10:30 UTC (Aqua satellite overpass time). The results showed an R² of 0.51 and a correlation coefficient of 0.75.

5