Modeling multiscale temporal hydroclimatic forcing of long-term historical groundwater levels using hybrid wavelet analysis–artificial neural network models for geohydrologic characterization

Abstract:

Long-term historical groundwater level measurements provide large datasets that give the opportunity to improve understanding of aquifer properties and geohydrologic processes using a data-driven approach. Wavelet analysis and artificial neural network techniques were used to define the response time scale of an aquifer system. The discrete wavelet transform (DWT) was used to decompose hydroclimatic forcing signals (daily time series of rainfall, potential evapotranspiration, and runoff) to develop hybrid wavelet analysis–artificial neural network (WA-ANN) models. These techniques were explored by examining the effects of using the DWT approximate and detail coefficients of hydroclimatic forcing signals to simulate groundwater-level response. Seventeen wells with at least 30 years of daily groundwater-level data were selected, representing seven principal aquifer types across the United States. The aquifers monitored by these wells were classified based on the dominant wavelet period, aquifer type, and depth of water table into three main groups: (1) confined aquifers with large dominant wavelet period exceeding 5 years (3 wells); (2) unconfined aquifers with large dominant wavelet period exceeding 5 years and a deep water table (4 wells); and (3) unconfined aquifers with short to medium dominant wavelet periods of less than 9 months and a shallow water table (10 wells). The third group was further classified by lithology, water table depth, or aquifer time constant with subtle differences in minimum and dominant wavelet periods. Each well recorded groundwater-level fluctuations characteristic of a particular time scale (high and low frequency variations in water level) that are related to the hydraulic properties and spatial scale of both the vadose zone and the aquifer system. The physical processes effecting groundwater-level fluctuations in a particular well could be inferred by distinguishing the properties which influence a particular aquifer system. This new method of studying the physics of aquifer systems, rather than simply forecasting, showed that the DWT was able to decompose hydroclimatic forcing signals effectively, indicating that hybrid WA-ANN models can be useful tools for groundwater-level modeling for the purposes of geohydrologic characterization of an aquifer system.