Development of Groundwater Recharge Maps
Team is currently exploring the use of a variety of approaches that can integrate various types of observation datasets with different types of models to generate groundwater recharge maps. In the first approach, exploring the use of process models and integrate the model predictions with various types of earth observational data to map recharge fluxes. Second approach is to use data-driven models and machine learning tools, which are constrained by physics constraints, to map recharge fluxes.

Task-1: Use of process models for estimating recharge (UA)
In this modeling approach, use different water balance methods and will employ large-scale process models and machine learning tools to assimilate big hydrological data to predict recharge. Any process model can be conceptualized as a solution to a system of mass/energy balance equations. Team is currently attempting to use a variety of process models including simple water balance approaches. Dr. Pooja Preetha (summer faculty), will kick start modeling of regional-scale recharge processes using SWAT and she will continue to work with a graduate student work on different process models

Task-2: Development of physics-constrained deep learning models for estimating recharge In this modeling approach

Team is exploring the use of a physics-constrained LSTM network for predicting aggregate soil storage states and recharge fluxes in the vadose zone. Mauricio Gonzalez and Georgios Boumis (Graduate students), work on this problem. Both students have learnt basic LSTM fundamentals, and have developed codes for time series predictions in response to forcings with a range of scales of interactions. Next, Team will compile ancillary data and develop LSTM codes for recharge estimation. In particular, Team will train an LSTM model at the catchment level using evapotranspiration data and direct runoff data from the USGS Water Information System and allow a residual term in the mass balance to represent deep percolation into the water table. The physics constrained LSTM can track the major observed out-fluxes from a given catchment and thus can provide an estimate for the major unobserved out-flux (recharge to the aquifer). Team will benchmark this against inverted recharge estimates from available field data at selected sites where these are available for model evaluation. To aid in benchmarking, Georgios Boumis (PhD student) is generating recharge estimates at gauging stations using Water Table Fluctuation (WTF) and Episodic Master Recession (EMR) methods. Mauricio Gonzalez (PhD student) is currently exploring the use of different method to implement physics constraints within LSTM algorithms to model simulated time series data.

Task-3: Obtaining more accurate recharge estimates by improving evapotranspiration
estimates (UA)


This will involve development of a parsimonious representation of state-of-the-science plant-hydraulics-based soil-plant-atmosphere continuum (SPAC) model (Liu et al., 2017). Existing models of crop evapotranspiration (ET) incorrectly over attribute the role of soil moisture limitation on evapotranspiration (Liu et al., 2020). A parsimonious model that appropriately captures the coupled influence of vapor pressure deficit (VPD) and soil moisture deficit on plant hydraulics, without the need for explicit representation of root-to-leaf hydraulics, can capture the relative attribution sufficiently well.


Coarse ET estimates can be downscaled to field scale by exploiting the vegetation index–radiometric surface temperature relationship. We will fuse the infrequent Landsat-derived ET imagery and ET imagery derived from the figure below. ET estimates will be integrated within appropriate water balance models coupled with ML-based data analytic tools to obtain recharge. We will intercompare our daily 30-m ET estimates with existing ET products (e.g., MODIS ET, SSEBop, and Penman-Monteith-Leuning (PML) based ET at FluxNet towers)