Sensitivity and uncertainty quantification for the ECOSTRESS evapotranspiration algorithm – DisALEXI

Abstract Evapotranspiration (ET) is a measure of plant water use that is utilized regionally for drought detection and monitoring, and locally for agricultural water resource management. Understanding the uncertainty associated with this measurement is vital for science predictions and analysis and for water resource management decision making. In this manuscript, the uncertainty in disaggregated Atmosphere-Land Exchange (disALEXI) is quantified; disALEXI is an ET algorithm that utilizes land surface temperature (LST) derived from the ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS), as well ancillary inputs for landcover, elevation, vegetation parameters, and meteorological inputs. Since each of these inputs has an associated, and potentially unknown, uncertainty, in this study a Monte Carlo simulation based on a spatial statistical model is used to determine the algorithm's sensitivity to each of its inputs, and to quantify the probability distribution of algorithm outputs. Analysis shows that algorithm is most sensitive to LST (the input derived from ECOSTRESS). Significantly, the output uncertainty distribution is non-Gaussian, due to the non-linear nature of the algorithm. This means that ET uncertainty cannot be prescribed by accuracy and precision alone. Here, uncertainty was represented using five quantiles of the output distribution. The distribution was consistent across five different datasets (mean offset is 0.01 mm/day, and 95% of the data is contained within 0.3 mm/day). An additional two datasets with low ET, showed higher uncertainty (95% of the data is within 1 mm/day), and a positive bias (i.e., ET was overestimated by an average of 0.12 mm/day when ET was low).