Accurate monitoring of atmospheric carbon dioxide(CO_(2))is crucial for understanding the global carbon cycle and informing climate policy.Satellite-based remote sensing provides a promising means to obtain global mea...Accurate monitoring of atmospheric carbon dioxide(CO_(2))is crucial for understanding the global carbon cycle and informing climate policy.Satellite-based remote sensing provides a promising means to obtain global measurements of the column-averaged CO_(2) dry air mole fraction(XCO_(2)).However,traditional retrieval algorithms are computationally intensive due to their reliance on iterative radiative transfer simulations.In this study,we introduce the Spectrum Transformer(SpT),a novel neural network model that employs a Transformer-based architecture to enable fast and accurate XCO_(2) retrievals directly from satellite-measured spectra.Unlike previous machine learning approaches,the SpT model effectively handles data drift caused by increasing atmospheric CO_(2) levels without requiring synthetic future data or additional assumptions.Trained exclusively on historical OCO-2 spectra and retrievals from 2017 to 2019,the SpT model demonstrates unbiased generalization to data from 2020 to 2022,achieving high accuracy(root mean square error[RMSE]∼1.5 parts per million[ppm])in“future”retrievals.Through periodic fine-tuning with minimal new data(<10%of all available data),the model maintains even higher accuracy(RMSE∼1.2 ppm),demonstrating its applicability for ongoing missions up to the most recent measurements(2024 April 1).The SpT model reduces computational time from minutes to milliseconds per retrieval,offering an important advancement over traditional methods.Validation against TCCON ground-based measurements confirms the model’s ability to capture seasonal and regional variations in XCO_(2),highlighting its potential for real-time global CO_(2) monitoring.展开更多
基金supported by the National Natural Science Foundation of China(grants nos.52276077 and 52120105009).
文摘Accurate monitoring of atmospheric carbon dioxide(CO_(2))is crucial for understanding the global carbon cycle and informing climate policy.Satellite-based remote sensing provides a promising means to obtain global measurements of the column-averaged CO_(2) dry air mole fraction(XCO_(2)).However,traditional retrieval algorithms are computationally intensive due to their reliance on iterative radiative transfer simulations.In this study,we introduce the Spectrum Transformer(SpT),a novel neural network model that employs a Transformer-based architecture to enable fast and accurate XCO_(2) retrievals directly from satellite-measured spectra.Unlike previous machine learning approaches,the SpT model effectively handles data drift caused by increasing atmospheric CO_(2) levels without requiring synthetic future data or additional assumptions.Trained exclusively on historical OCO-2 spectra and retrievals from 2017 to 2019,the SpT model demonstrates unbiased generalization to data from 2020 to 2022,achieving high accuracy(root mean square error[RMSE]∼1.5 parts per million[ppm])in“future”retrievals.Through periodic fine-tuning with minimal new data(<10%of all available data),the model maintains even higher accuracy(RMSE∼1.2 ppm),demonstrating its applicability for ongoing missions up to the most recent measurements(2024 April 1).The SpT model reduces computational time from minutes to milliseconds per retrieval,offering an important advancement over traditional methods.Validation against TCCON ground-based measurements confirms the model’s ability to capture seasonal and regional variations in XCO_(2),highlighting its potential for real-time global CO_(2) monitoring.
