Potential of remote sensing surface temperature- and evapotranspiration-based land-atmosphere coupling metrics for land surface model calibration

Abstract

Imperfect land physics introduce significant levels of uncertainty into current land surface models (LSMs) and can cause bias in their representation of land-atmosphere coupling strength (ρ). When LSMs are coupled with atmospheric prediction models, such errors will eventually degrade the accuracy of lower atmosphere forecasts. Here, we investigate the potential of two remote sensing (RS)-based ρ references for addressing LSM ρ bias. To minimize meteorological uncertainty and maximally attribute LSM ρ bias to land sources, we focus specifically on off-line LSM calibration forced using high-quality, observation-based meteorological data. Both ρ references are based on a newly proposed two-system approach for eliminating the impact of random errors in RS retrievals and quantified using the temporal correlations of soil moisture (SM) versus both evapotranspiration (ET) and the diurnal amplitude of surface temperature (dT). Experiments are conducted to calibrate an off-line LSM individually against each resulting ρ reference and using a combination of both dT- and ET-represented ρ references. The resulting calibrated LSM is further evaluated using independent ground-based ET observations and RS dT retrievals. Results show that although dT- and ET-represented ρ references are physically consistent across space, model calibration results based on them are quite different. Specifically, the calibration experiment targeting ET-represented ρ outperforms that targeting dT-represented ρ in ET and dT modeling. Diagnostic results indicate that the failure of dT-based calibration experiments is due to the confounding impacts of transpiration/evapotranspiration partitioning error and large dT uncertainties in LSM. However, results also confirm the potential of both dT- and ET-represented ρ references for jointly diagnosing and understanding LSM ρ bias. As a result, we suggest diagnosing LSM ρ bias using both ET- and dT-represented ρ references – but calibrating LSM using only ET-represented ρ reference data.