An improved process-based evapotranspiration/heat fluxes remote sensing algorithm based on the Bayesian and Sobol’ uncertainty analysis framework using eddy covariance observations of Tibetan grasslands

Abstract

• Improved an AET process retrieval algorithm by incoporating radiation and soil water constraints. • Determined the most sensitive parameters of P-LSH algorithm using Sobol’ sensitivity analysis. • Derived the parameter posterior distributions to reduce P-LSH uncertainty using the DE-MC. Actual evapotranspiration (AET) is an important component of the water cycle on the Tibetan Plateau. Remote sensing retrieval of AET requires further improvement in this region. The objective of this study is to improve the Process-based Land Surface evapotranspiration/Heat fluxes algorithm (P-LSH) by incorporating the impacts of incoming shortwave radiation and soil moisture on canopy conductance and optimizing the parameters. Based on observations from three FLUXNET towers on the Tibetan Plateau, five key parameters in the improved P-LSH algorithm were determined by Sobol’ sensitivity analysis. Among them, the parameter b 3 , which defines the relationship of potential canopy conductance ( g 0 ) and Normalized Difference Vegetation Index (NDVI), was identified as the parameter to which the algorithm was most sensitive. We then utilized the Differential Evolution Markov Chain (DE-MC) method to analyze uncertainty of model parameters, and posterior distributions of parameters were used to correct the response curve of g 0 to NDVI in the improved P-LSH algorithm and simulate daily AET. The results show that uncertainty in five parameters is greatly reduced after the Markov process, and compared with the original algorithm, the simulation of the improved P-LSH algorithm is greatly upgraded, specifically reflects in higher R 2 (0.93 vs 0.90) and lower RMSE (11.39 W m −2 vs 15.11 W m −2 ). Finally, although reanalysis-based net radiation, vapor pressure deficit, and wind speed were major sources of uncertainty of reanalysis-driven AET estimates, reanalysis-driven AET estimates did not cause excessive errors compared to tower-driven AET estimates (R 2 = 0.87 vs 0.93, RMSE = 15.76 W m −2 vs 11.39 W m −2 ), which provided a reasonable basis for further application of the improved P-LSH algorithm to larger watershed scales.