Model-enforced post-process correction of satellite aerosol retrievals

Abstract

Abstract. Satellite-based aerosol retrievals provide a timely view of atmospheric aerosol properties, having a crucial role in the subsequent estimation of air quality indicators, atmospherically corrected satellite data products, and climate applications. However, current aerosol data products based on satellite data often have relatively large biases compared to accurate ground-based measurements and distinct uncertainty levels associated with them. These biases and uncertainties are often caused by oversimplified assumptions and approximations used in the retrieval algorithms due to unknown surface reflectance or fixed aerosol models. Moreover, the retrieval algorithms do not usually take advantage of all the possible observational data collected by the satellite instruments and may, for example, leave some spectral bands unused. The improvement and the re-processing of the past and current operational satellite data retrieval algorithms would become tedious and computationally expensive. To overcome this burden, we have developed a model-enforced post-process correction approach to correct the existing operational satellite aerosol data products. Our approach combines the existing satellite aerosol retrievals and a post-processing step carried out with a machine-learning-based correction model for the approximation error in the retrieval. The developed approach allows for the utilization of auxiliary data sources, such as meteorological information, or additional observations such as spectral bands unused by the original retrieval algorithm. The post-process correction model can learn to correct for the biases and uncertainties in the original retrieval algorithms. As the correction is carried out as a post-processing step, it allows for computationally efficient re-processing of existing satellite aerosol datasets without fully re-processing the much larger original radiance data. We demonstrate with over-land aerosol optical depth (AOD) and Ångström exponent (AE) data from the Moderate Imaging Spectroradiometer (MODIS) of the Aqua satellite that our approach can significantly improve the accuracy of the satellite aerosol data products and reduce the associated uncertainties. For instance, in our evaluation, the number of AOD samples within the MODIS Dark Target expected error envelope increased from 63 % to 85 % when the post-process correction was applied. In addition to method description and accuracy results, we also give recommendations for validating machine-learning-based satellite data products.

Deep-learning-based post-process correction of the aerosol parameters in the high-resolution Sentinel-3 Level-2 Synergy product

“…The core idea in the model-enforced post-process correction model is to improve the accuracy of the approximate retrieval (Eq. 2) by machine learning techniques (Lipponen et al, 2021). With Eqs.…”

“…In Lipponen et al (2021), we proposed a model-enforced machine learning model for post-process correction of satellite aerosol retrievals. The key idea in the model-enforced approach is to exploit also the model and information of the conventional retrieval algorithm and train a machine learning algorithm for correction of the approximation error in the result of the conventional satellite retrieval algorithm.…”

“…The advantages of the model-enforced post-process correction approach are improved accuracy over the existing data products and the possibility to correct the existing products by a simple postprocessing step without need for any reprocessing of the existing retrieval algorithms, which are usually managed and operated by the algorithm development teams. In Lipponen et al (2021), the model-enforced approach was combined with a random forest regression algorithm for post-process correction of MODIS AOD and AE products using collocated MODIS and AERONET aerosol data for training the correction model for the approximation error in AOD and AE in the MODIS DT over land product. The post-process correction was found to yield significantly improved accuracy over the MODIS AOD and AE retrievals, and the correction approach resulted in better accuracy retrievals than the fully learned machine learning approach.…”

Bayesian uncertainty quantification in aerosol optical depth retrieval applied to TROPOMI measurements

“…The assessment of uncertainty estimates on the variety of satellite aerosol retrievals, also in the pixel level, are discussed by Sayer et al (2020). On the other hand, there are studies to correct inaccuracies in the results of satellite aerosol product by model enforced post-processing using machine learning techniques as presented by Lipponen et al (2021). The ongoing research with uncertainty analysis and improved uncertainty estimates benefits the data users as well as the retrieval algorithm developers.…”

High‐Resolution Post‐Process Corrected Satellite AOD

“…Recent advances in methods to combine conventional retrieval algorithms and machine learning have significantly improved satellite AOD estimate accuracy (Lipponen et al, 2021(Lipponen et al, , 2022. For example, the post-process correction approach for satellite AOD retrieval uses a machine learning-based model to predict the retrieval error in the satellite AOD and uses that prediction to correct the retrieval.…”