Five years of GOSAT-2 retrievals with RemoTeC: XCO ₂ and XCH ₄ data products with quality filtering by machine learning

Accurately measuring greenhouse gas concentrations to identify regional sources and sinks is essential for effectively monitoring and mitigating their impact on the Earth's changing climate. In this article we present the scientific data products of XCO 2 and XCH 4 , retrieved with RemoTeC, from the Greenhouse Gases Observing Satellite-2 (GOSAT-2), which span a time range of 5 years. GOSAT-2 has the capability to measure total columns of CO 2 and CH 4 to the necessary requirements set by the Global Climate Observing System (GCOS), who define said requirements as accuracy<10 ppb and <0.5 ppm for XCH 4 and XCO 2 respectively, and stability of <3ppbyr-1 and <0.5ppmyr-1 for XCH 4 and XCO 2 respectively. Central to the quality of the XCO 2 and XCH 4 datasets is the post-retrieval quality flagging step. Previous versions of RemoTeC products have relied on threshold filtering, flagging data using boundary conditions from a list of retrieval parameters. We present a novel quality filtering approach utilising a machine learning technique known as Random Forest Classifier (RFC) models. This method is developed under the European Space Agency's (ESA) Climate Change Initiative+ (CCI+) program and applied to data from GOSAT-2. Data from the Total Carbon Column Observing Network (TCCON) are employed to train the RFC models, where retrievals are categorized as good or bad quality based on the bias between GOSAT-2 and TCCON measurements. TCCON is a global network of Fourier transform spectrometers that measure telluric absorption spectra at infrared wavelengths. It serves as the scientific community's standard for validating satellite-derived XCO 2 and XCH 4 data. Our results demonstrate that the machine learning-based quality filtering achieves a significant improvement, with data yield increasing by up to 85 % and RMSE improving by up to 30 %, compared to traditional threshold-based filtering. Furthermore, inter-comparison with the TROPOspheric Monitoring Instrument (TROPOMI) indicates that the quality filtering RFC models generalise well to the full dataset, as the expected behaviour is reproduced on a global scale. Low systematic biases are essential for extracting meaningful fluxes from satellite data products. Through TCCON validation we find that all data products are within the breakthrough bias requirements set, with RMSE for XCH 4 < 15 ppb and XCO 2 < 2 ppm. We derive station-to-station biases of 4.2 ppb and 0.5 ppm for XCH 4 and XCO 2 respectively, and linear drift of 0.6 ppb yr −1 and 0.2 ppm yr −1 for XCH 4 and XCO 2 respectively. For XCH 4 , GOSAT-2 and TROPOMI are highly correlated with standard deviations less than 18 ppb and globally averaged biases close to 0 ppb. The inter-satellite bias between GOSAT and GOSAT-2 is significant, with an average global bias of −15 ppb. This is comparable to that seen between GOSAT and TROPOMI, consistent with our findings that GOSAT-2 and TROPOMI are in close agreement.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.