Addressing observational gaps in aerosol parameters using machine learning: Implications to aerosol radiative forcing

Abstract

Listen

Translate article

The role of atmospheric aerosols has remained hugely uncertain due to lack of long-term systematic observations. This is particularly the case in the global emission hotspot the Indo-Gangetic Plain. In this paper, we apply a machine learning (ML) approach to fill gaps in aerosol observations by combining meteorological reanalysis, and satellite-observations. Aerosol Optical Depth (AOD), Single Scattering Albedo (SSA), and Asymmetry Parameter (AP) are critical parameters for understanding the atmospheric aerosol properties, radiative forcing and climate patterns. We considered AOD, SSA, and AP over an urban site, Kanpur, in the Indo-Gangetic Plain, which is a global hot-spot of high aerosol loading. We used ML technique to fill the observational gaps in AOD, SSA, and AP at 440 nm. Accurate ground-based daily observations of AOD, SSA, and AP from AERONET (Aerosol Robotic Network) over a period of ∼2 decades (2001–2022) are used in the analysis, which have ∼ 37%, ∼62%, and ∼58% data gap, respectively. To reduce observational gaps, an ML model was trained with reanalysis (meteorological parameters) and satellite (angstrom exponent and absorbing aerosol index) datasets by optimally tuning the hyperparameters. The model tested on unseen data showed that it could simulate the AOD (coefficient of determination, R2 = 0.51 and refined index of agreement, RIA = 0.66), SSA (R2 = 0.60 and RIA = 0.70), and AP (R2 = 0.64 and RIA = 0.71), successfully with moderate accuracies: root mean square error of 0.25, 0.018, and 0.018, respectively. This magnitude of error in AOD is comparable to the errors, 0.26 and 0.25, in satellite (MODIS-Moderate Resolution Imaging Spectroradiometer) derived and reanalysis (MERRA-2- Modern-Era Retrospective analysis for Research and Applications, Version 2) AODs, respectively for the study region. When MODIS AOD was included as a predictor in the ML model, it could reproduce AOD with much smaller errors of 0.14, reproducing 84% of total variability. Thus, based on available input data, we reduced observational gaps of daily AOD, SSA, and AP by ∼10% (∼2 years), ∼23% (∼5 years), and ∼21% (∼4 years), respectively, at the study region. Aerosol Radiative Forcing (ARF) at the top-of-atmosphere was estimated using AOD, SSA, and AP both with and without gap-filling. We found that data gaps can significantly impact the estimations of aerosol climate forcing. A considerable change in ARF, ∼ ± 3 Wm−2 indicating that the observational gaps can mislead our understanding of aerosols' impact on climate. This points toward the need for more comprehensive measurements with minimal observational gaps. Our study highlights the possibility of compensating observational gaps by an optimised ML model and that such innovative approach can potentially lead to better understanding of impact of aerosols on climate. Further, the present study could be of immense interest for the community as our validated approach can be applied for other pollutants and for other data-scarce regions in the South and Southeast Asia.