Estimating Groundwater Dynamics from Artificial Intelligence and InSAR

Groundwater represents about a third of global water withdrawals, and approximately half of global irrigation water. In many arid and semi-arid regions, groundwater is rapidly being depleted, which is affecting agricultural productivity over the long term. The over-exploitation of groundwater due to the current drought episode in South-Western U.S. has already led farmers to fallow hundreds of thousands of acres of farmland. We leverage recent developments in artificial intelligence in order to improve deformation detection thresholds in Interferometric Synthetic Aperture Radar (InSAR). Our technology allows to deconvolve signal from noise in InSAR data, and lowers the detection threshold of surface deformation in spatially continuous InSAR time series of ground deformation by about an order of magnitude compared to the state-of-the-art. Phase I focused on separating ground deformation signals from atmospheric noise, and improving the time resolution of the associated InSAR time series. In this Phase II proposal, we focus on further improving detection capabilities by taking directly into account a second major source of noise in InSAR data, due to soil moisture. We specifically train models on time series of seasonal depletion and recharge in order to study variations in groundwater levels in InSAR data, with the goal of allowing governments and organizations to take informed assessments and policies related to groundwater supplies. Prolonged episodes of groundwater depletion can have drastic socioeconomic consequences, and impacts of the ongoing severe drought occurring in the U.S. are already becoming apparent. The South-Western U.S. is currently experiencing a major drought episode. In this context, reliance on groundwater resources has drastically increased, and the over-exploitation of groundwater resources during the current drought is causing hundreds of wells to dry up. Existing approaches to monitor groundwater suffer from either poor spatial or from poor temporal resolution, and we propose a new cost-effective and accurate methodology allowing for finer monitoring in time and space. We rely on deep learning to deconvolve ground deformation signals from atmospheric and soil moisture noise in interferometric synthetic aperture radar (InSAR), allowing us to reach greater detection capabilties at a much improved time resolution. In this Phase II proposal, our technical objectives are the following: (i) to first improve our synthetic InSAR time series, by adding synthetic noise induced by soil moisture; (ii) to further lower noise in our InSAR time series of groundwater deformation, by training new deep learning models on this improved synthetic data; (iii) to validate the associated displacement maps on our areas of interest, by comparing results with local instrumentation (monitoring wells, GPS) and seismic velocity maps; and (iv) to make sure that soil moisture `noise' was well removed and that there are no systematic biases, by comparing this estimated noise to published soil moisture data. The deliverables at the end of Phase II will be the following: new InSAR synthetic data that encompasses both physical signals linked to groundwater depletion and recharge and realistic soil moisture 'noise'; an improved deep learning model trained to deconvolve errors induced by soil moisture and the atmosphere from groundwater-induced inflation or subsidence; an assessment of the quality of the results (with comparison to well data, GPS data, and seismic tomography) in particular in New Mexico and California; and a comparison of our estimated soil moisture 'noise' with soil moisture data, to make sure that we did not introduce biases in the approach.
Benefits: NASA is about to launch a new satellite constellation for InSAR (NiSAR), with a planned launch date in 2023, and our proposed technology could be applied to data provided by this new constellation. Moreover, our algorithm could be interfaced with InSAR interferogram data from NASA Earthdata, and can be used in combination with NASA code libraries (isce, AriaTools). Last, the NASA Observational Products for End-Users from Remote Sensing Analysis (OPERA) project started in April 2021, and our proposed work could be coupled to OPERA products. Our other commercialization applications are targeting several sectors: i) farming; ii) state and local governments, with the goal of helping monitor groundwater resources; iii) the insurance sector, and in particular insurance products related to flooding and drought; and iv) the finance sector, to help anticipate changes in the price of oil and gas, as well as water.