A global land aerosol fine-mode fraction dataset (2001–2020) retrieved from MODIS using hybrid physical and deep learning approaches

Abstract

Abstract. The aerosol fine-mode fraction (FMF) is valuable for discriminating natural aerosols from anthropogenic ones. However, most current satellite-based FMF products are highly unreliable over land. Here, we developed a new satellite-based global land daily FMF dataset (Phy-DL FMF) by synergizing the advantages of physical and deep learning methods at a 1∘ spatial resolution covering the period from 2001 to 2020. The Phy-DL FMF dataset is comparable to Aerosol Robotic Network (AERONET) measurements, based on the analysis of 361 089 data samples from 1170 AERONET sites around the world. Overall, Phy-DL FMF showed a root-mean-square error (RMSE) of 0.136 and correlation coefficient of 0.68, and the proportion of results that fell within the ±20 % expected error (EE) envelopes was 79.15 %. Moreover, the out-of-site validation from the Surface Radiation Budget (SURFRAD) observations revealed that the RMSE of Phy-DL FMF is 0.144 (72.50 % of the results fell within the ±20 % EE). Phy-DL FMF showed superior performance over alternative deep learning or physical approaches (such as the spectral deconvolution algorithm presented in our previous studies), particularly for forests, grasslands, croplands, and urban and barren land types. As a long-term dataset, Phy-DL FMF is able to show an overall significant decreasing trend (at a 95 % significance level) over global land areas. Based on the trend analysis of Phy-DL FMF for different countries, the upward trend in the FMFs was particularly strong over India and the western USA. Overall, this study provides a new FMF dataset for global land areas that can help improve our understanding of spatiotemporal fine-mode and coarse-mode aerosol changes. The datasets can be downloaded from https://doi.org/10.5281/zenodo.5105617 (Yan, 2021).

Deep learning in airborne particulate matter sensing: a review

“…Since pollution forecasting involves a high-dimensional parameter space, with partial amounts of data, the ability of deep learning to integrate data and physics models could provide suitable architecture for highly accurate and efficient forecasting models. Indeed, preliminary work in this emerging field has been carried out using the outputs of the physical model as the inputs for the deep-learning model, for the purpose of predicting fine-mode fractions (FMF) of aerosols over land, and achieving an accuracy higher than current methods [124]. The authors found that, compared with more traditional physical-based and deep-learning-based FMF results, the their PINN was more accurate at predicting FMF values over five land types (e.g., barren land, croplands, forests, grasslands and urban area).…”

Comment on essd-2021-426

“…Environmental covariates selected in this study contain 12 covariates in three categories (meteorology, surface information, and topography). Covariates are selected based on two criteria: first, each variable is considered important to AOD and has a vital influence on AOD formation, accumulation and migration process, referring to existing research and expert experience (Zhao et al, 2019;Chen et al, 2020;Yan et al, 2022); the second, the data is released to the public for free, which means that the data set is freely available on the national or global scale (Li et al, 2020). The detailed information is listed in Tab.1.…”

Full-coverage 250 m monthly aerosol optical depth dataset (2000–2019) amended with environmental covariates by an ensemble machine learning model over arid and semi-arid areas, NW China

“…The AERONET AOD data have a low uncertainty (0.01-0.02) and are considered the highest accuracy AOD data available; these data are widely used as a reference in RS AOD product validations (Almazroui, 2019). In this study, data from a total of 12 AERONET sites in northwest China were selected, most of which were from the third version of the level 2.0 AERONET AOD, except for the Mt_WLG station data (Level 1.5) (Yan et al, 2022;Giles et al, 2019). Related information about these AERONET sites is available in Table S2 and Fig.…”

“…The environmental covariates selected in this study comprised 12 covariates in three categories (meteorological parameters, surface properties, and terrain factors). The covariates were selected based on two criteria: first, each variable had to be considered important to the AOD and to have a vital influence on the AOD formation, accumulation, and migration processes, referring to existing research and expert experience Chen et al, 2020;Yan et al, 2022); and second, the data must be freely released to the public, meaning the datasets must be freely available on the national or global scale (Li et al, 2020). Detailed information on these covariates is listed in Table 1.…”

“…For dynamic covariates, the averaging method was adopted to obtain monthly average data. The static variables, similar to the baseline conditions, played an initial constraint role in the downscaling of the monthly AODs, while the dynamic variables played a more dynamic evolution role (Yan et al, 2022). Notably, the relevant operations are not limited to ArcGIS, and relevant open-source software such as QGIS could also be implemented.…”