Reconstructing monthly 0.25° terrestrial evapotranspiration data in a remote arid region using Bayesian-driven ensemble learning method

This study presents a Bayesian-driven ensemble learning framework to minimize errors in evapotranspiration (ET) data assimilation, thereby generating a continuous-scale ET estimation with improved accuracy for the data-scarce arid Central Asia (CA). Five state-of-the-art gridded ET products were extensively tested and utilized to reconstruct new long-term ET data at grid scale, including two Global Land Data Assimilation System (GLDAS)-Land Surface Models (LSM) (Catchment and NOAH), two reanalysis products (European Centre for Medium-Range Weather Forecasts Reanalysis 5 (ERA5) and Modern-Era Retrospective Analysis for Research and Applications - ver. 2 (MERRA2)), along with the Global Land Evaporation Amsterdam (GLEAM) satellite ET product ver. 3.6b. Taylor Skill Score was employed to evaluate the input gridded ET datasets, using in situ measurements across four arid biomes - cropland, shrubland, farmland, and grassland -as benchmarks. Bayes-Markov Chain Monte Carlo (Bayes-MCMC)-derived logarithmically transformed predictors from the posterior ensemble anomalies were standardized to resolve the mismatches in the cumulative distribution functions of ET data relating to the in situ observations. We found that the constrained residual errors were more profound between 45th and 95th percentiles for the rescaled total ET anomalies across all the input ET datasets at 95 % confidence interval. Moreover, the weighted ensemble estimates derived from the Bayes-MCMC were instrumental in correcting ET anomalies, leading to the robust reconstruction of a more reliable ensemble mean ET data for Central Asia (CAET(ens)) at monthly temporal and 0.25 degrees spatial resolutions, spanning from 2003 to 2020. Overall, CAET(ens )exhibited superior performance, achieving the highest averaged Kling-Gupta Efficiency of 0.75 and registering the lowest Root Mean Square Error of 6.85 mm/m compared to the standalone gridded ET products. Comparisons of spatial distributions of fine-scale features, and evaluations using standard metrics revealed that all ET products varied significantly (p < 0.05) by biome types. Remarkably, only CAET(ens) exhibited a reasonable low positive bias and pronounced ability to resolve the complex spatiotemporal patterns associated with the input gridded ET products. The study bridges the gap in continuous ET monitoring for sustainable water resources management in CA.