Machine Learning Application on Closing Data Gaps in Groundwater Measurements

Most earth science research relies on collecting spatiotemporally varying data to understand system behavior of either the individual system components of or the entire Earth system. These spatial patterns of changing phenomena are governed by physical laws with dynamic driving forcing and heterogeneous physical and chemical properties. In hydrology, a common method of data collection is to establish a network of wells that monitor changes in the physical system, such as temperature, specific conductance (SpC) and water elevation. Networks like these are one of the best sources of reliable, long-term time series data. However, the majority of monitored time series datasets are full of gaps due to sensor failure or field measurement errors. Complete datasets (without gaps in the time series) are needed to better reveal spatiotemporal variability of the highly dynamic physical system and provide more accurate boundary conditions and validation for numerical models. In this study, we apply different machine learning based data-driven approaches to close the data gaps. We introduce artificial gaps into real data streams, apply alternative gap filling approaches to the remaining data, and compare the gap-filled approximations with the original data to evaluate the alternative approaches.