GROUNDWATER SYSTEM CHARACTERIZATION USING HYBRID WAVELET ANALYSIS–ARTIFICIAL NEURAL NETWORK MODELS: COMPARISON WITH A PHYSICS-BASED SYNTHETIC AQUIFER MODEL

Multiscale temporal groundwater-level fluctuations recorded in a long-term well hydrograph are attributable, in part, to variations in aquifer recharge and the scale and hydraulic properties of the aquifer. Thus, modeling the multiscale forcing-response behavior of an aquifer using data-driven methods may yield information about its geohydrologic characteristics relevant to sustainability. A synthetic basin-fill alluvial aquifer system model was developed using MODFLOW. Simulated groundwater levels were analyzed using a hybrid wavelet analysis–artificial neural network (WA-ANN) modeling technique, with the objective of using physical principles (theory) to guide interpretation of data-driven models. The synthetic aquifer model comprised five layers, consisting of an unconfined aquifer bisected by a river overlying two confined aquifers. Four 30-year time series of daily recharge—with decadal, annual, seasonal, and shorter time-scale periodicity representing annual average recharge of 41, 74, 224, and 455 mm—were used as the forcing functions. The response functions were MODFLOW-simulated water levels at an observation well in each aquifer, which were modeled again with a WA-ANN model for each well by using the discrete wavelet transform (DWT) to decompose the recharge time series. Simulated groundwater level hydrographs were interpreted in the context of the hydraulic characteristics—diffusivity and time constant—of the aquifers. Results demonstrated that DWT coefficients, representing multiscale temporal variability in recharge, are reflective of differences in aquifer properties and the timing of aquifer recharge, but do not reflect differences in the magnitude of recharge. The duration of optimum DWT periods—those providing improvement in WA-ANN model fit—were inversely and directly proportional to aquifer diffusivity and time constant, respectively, and varied with multiscale variability in groundwater levels; whereas, optimum DWT periods had similar durations for models with different annual recharge. Multiscale forcing-response behavior of an aquifer can be represented by WA-ANN models of long-term groundwater levels and related to temporal variability in recharge and differences in aquifer properties, providing a means of geohydrologic characterization.