Exploiting Earth Observation Data to Impute Groundwater Level Measurements with an Extreme Learning Machine

Abstract

Groundwater resources are expensive to develop and use; they are difficult to monitor and data collected from monitoring wells are often sporadic, often only available at irregular, infrequent, or brief intervals. Groundwater managers require an accurate understanding of historic groundwater storage trends to effectively manage groundwater resources, however, most if not all well records contain periods of missing data. To understand long-term trends, these missing data need to be imputed before trend analysis. We present a method to impute missing data at single wells, by exploiting data generated from Earth observations that are available globally. We use two soil moisture models, the Global Land Data Assimilation System (GLDAS) model and National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center (CPC) soil moisture model to impute the missing data. Our imputation method uses a machine learning technique called Extreme Learning Machine (ELM). Our implementation uses 11 input data-streams, all based on Earth observation data. We train and apply the model one well at a time. We selected ELM because it is a single hidden layer feedforward model that can be trained quickly on minimal data. We tested the ELM method using data from monitoring wells in the Cedar Valley and Beryl-Enterprise areas in southwest Utah, USA. We compute error estimates for the imputed data and show that ELM-computed estimates were more accurate than Kriging estimates. This ELM-based data imputation method can be used to impute missing data at wells. These complete time series can be used improve the accuracy of aquifer groundwater elevation maps in areas where in-situ well measurements are sparse, resulting in more accurate spatial estimates of the groundwater surface. The data we use are available globally from 1950 to the present, so this method can be used anywhere in the world.