A blended TROPOMI+GOSAT satellite data product for atmospheric methane using machine learning to correct retrieval biases

Satellite observations of dry-column methane mixing ratios (XCH4) from shortwave infrared (SWIR) solar backscatter radiation provide a powerful resource to quantify methane emissions in service of climate action. The TROPOspheric Monitoring Instrument (TROPOMI), launched in October 2017, provides global daily coverage at a 5.5ĝĝ7ĝkm2 (nadir) pixel resolution, but its methane retrievals can suffer from biases associated with SWIR surface albedo, scattering from aerosols and cirrus clouds, and across-track variability (striping). The Greenhouse gases Observing SATellite (GOSAT) instrument, launched in 2009, has better spectral characteristics and its methane retrieval is much less subject to biases, but its data density is 250 times sparser than TROPOMI. Here, we present a blended TROPOMI+GOSAT methane product obtained by training a machine learning (ML) model to predict the difference between TROPOMI and GOSAT co-located measurements, using only predictor variables included in the TROPOMI retrieval, and then applying the correction to the complete TROPOMI record from April 2018 to present. We find that the largest corrections are associated with coarse aerosol particles, high SWIR surface albedo, and across-track pixel index. Our blended product corrects a systematic difference between TROPOMI and GOSAT over water, and it features corrections exceeding 10ĝppb over arid land, persistently cloudy regions, and high northern latitudes. It reduces the TROPOMI spatially variable bias over land (referenced to GOSAT data) from 14.3 to 10.4ĝppb at a 0.25ĝ ĝĝ0.3125ĝ resolution. Validation with Total Carbon Column Observing Network (TCCON) ground-based column measurements shows reductions in variable bias compared with the original TROPOMI data from 4.7 to 4.4ĝppb and in single-retrieval precision from 14.5 to 11.9ĝppb. TCCON data are all in locations with a SWIR surface albedo below 0.4 (where TROPOMI biases tend to be relatively low), but they confirm the dependence of TROPOMI biases on SWIR surface albedo and coarse aerosol particles, as well as the reduction of these biases in the blended product. Fine-scale inspection of the Arabian Peninsula shows that a number of hotspots in the original TROPOMI data are removed as artifacts in the blended product. The blended product also corrects striping and aerosol/cloud biases in single-orbit TROPOMI data, enabling better detection and quantification of ultra-emitters. Residual coastal biases can be removed by applying additional filters. The ML method presented here can be applied more generally to validate and correct data from any new satellite instrument by reference to a more established instrument.