Sensitivity analysis on inputs to the DisALEXI algorithm using a Machine Learning interpretability method

The Atmosphere-Land Exchange Inverse model (ALEXI) and associated flux disaggregation technique (DisALEXI) are used to provide spatially distributed high-resolution evapotranspiration (ET) estimations. The ALEXI/DisALEXI modeling framework has proven reliable in determining field-scale ET across many surfaces, using remotely sensed surface variables such as thermal infrared (TIR) imagery and other ancillary data sources as input. ALEXI/DisALEXI is based on the Two-Source Energy Balance model, a physically-based algorithm designed to partition the TIR signal into canopy and soil components. The large number of non-linear physically-based equations dynamically interacting within DisALEXI make it challenging to quantify individual variables/parameters' effect on ET estimations. We propose to perform a sensitivity analysis on DisALEXI using the model-agnostic algorithm Shapley Additive exPlanations (SHAP) to explain how input variables/parameters in the DisALEXI algorithm affect DisALEXI ET for a given study domain. The algorithm's efficacy is demonstrated by applying it to four distinct agricultural sites in Central Valley, California. We find that land surface temperature (LST) and air temperature (Ta) are the most dominant inputs for soil latent heat flux (lEs) while leaf area index (LAI) is the major driver for canopy latent heat flux (lEc). However, the final ET estimate is driven by lEs and is most sensitive to LST and Ta. Subject to uncertainties, the effect of most inputs to lEs, lEc and ET are linear. Additionally, ET is sensitive to DisALEXI parameters related to the Priestly-Taylor equation and the clumping index. Though SHAP is generally used for black-box Machine Learning models, according to the authors' knowledge, this is the first study that uses SHAP to understand a physically-based algorithm such as DisALEXI.