Aerosol fine fraction (FMF) is a key parameter for distinguishing between anthropogenic and natural aerosols, but it has significant uncertainty in satellite retrieval, especially on land. This dataset integrates physics and deep learning (Phy DL) methods, based on MODIS data, to retrieve aerosol micromodule fractions at the global land scale, and generates 20-year (2001-2020) Phy DL aerosol micromodule fractions (500 nm) at daily time resolution and 1 ° spatial resolution
(1) In this study, MODIS C6.1 L1B MOD02SSH data (i.e. top of atmosphere (TOA) reflectance of bands 1 to 7), MODIS C6.1 L3 MOD09CMG data (surface reflectance of bands 1 to 7), and MODIS C6.1 L3 MOD08 daily data were obtained from 2001 to 2020 to retrieve FMF.
(2) Due to insufficient level 2.0 data as training data for modeling, we used the 1.5-level SDA FMF dataset generated from data from 1170 AERONET sites worldwide as the foundational data for further modeling and validation.
(3) Due to the influence of meteorological factors on FMF, the five meteorological variables (i.e. 2-meter temperature, PBLH, surface pressure, 10 meter wind component, and 2-meter dew point temperature) all come from the fifth generation product of the European Centre for Medium Range Weather Forecasts (ERA5), which provides hourly data since 1950 with a spatial resolution of 0.25 °. Then calculate the relative humidity based on the 2-meter dew point temperature and air temperature (Tetens, 1930). Due to the high temporal and spatial resolution of MODIS data, only meteorological data collected between 10:00 and 11:00 local time was used and resampled to 1 °× 1 ° to obtain daily averages.
In this study, we combined a physical model and a deep learning model using a cascade mode, where the output of the physical model is used as the input of the deep learning model. The physical model used is LUT-SDA (Yan et al., 2017). LUT-SDA is designed for satellite FMF retrieval when there are only two wavelengths of AOD (such as DT AOD products). Firstly, a minimum AOD of three wavelengths is required to obtain the AE derivative (α '). Then, the AE (αf) and FMF of the fine mode AOD can be calculated.
Based on the analysis of 361089 data samples from 1170 AERONET sites worldwide, the Phy DL FMF dataset is comparable to the measurement results of the Aerosol Robot Network (AERONET). Overall, the root mean square error (RMSE) of Phy DL FMF is 0.136, the correlation coefficient is 0.68, and the proportion of results within the expected error (EE) range of ± 20% is 79.15%. In addition, off-site validation of the Surface Radiation Budget (SURFRED) observation results showed that the root mean square error of Phy DL FMF was 0.144 (72.50% of the results were within ± 20% of the expected error (EE) range). The performance of Phy DL FMF is superior to other deep learning or physical methods (such as the spectral deconvolution algorithm proposed in our previous research), especially in forest, grassland, cultivated land, as well as urban and barren land types. As a long-term dataset, Phy DL FMF is able to show a significant downward trend in the overall terrestrial regions worldwide (with a significance level of 95%)
| # | number | name | type |
| 1 | 42030606 | National Natural Science Foundation of China | |
| 2 | 91837204 | National Natural Science Foundation of China |
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
CSTR:11738.11.NCDC.ZENODO.DB6527.2024
10.12072/ncdc.zenodo.db6527.2024