**1. Introduction**

In the 1970s, groundwater contributed 20% of the total water usage (TWU) in Taiwan, whereas faster economic growth with industrial expansion and population increase have greatly expanded water demand in Taiwan in the 1990s, the groundwater contributing 31% of the TWU in Taiwan [1,2]. In the 2010s, the groundwater water percentage of TWU has reached 34%, achieving 64.9% in Southern Taiwan [3,4]. However, due to anthropogenic interference and climate change, the water supply is extremely unstable, groundwater withdrawals may locally exceed recharge, and undesirable conditions may develop in the aquifer and in hydraulically connected surface waters [5]. Overexploited groundwater has led to seawater intrusion, land subsidence, lowering of groundwater levels, and salinization of soil, and has reduced the well water withdrawal yields in Taiwan [6,7]. Artificial recharge has been defined, as the process of replenishing groundwater through an artificial recharge lake (ARL), to increase groundwater storage in subterranean zones [8–10]. However, the process of groundwater recharge is complex. The infiltration impact is complicated by many confounding factors such as rainfall (or applied water) intensity, micro-topography, vegetation, soil texture, and vertical and horizontal heterogeneity in soil properties [11]. Morbidelli et al. indicated that, although a variety of local infiltration models for vertically homogeneous soils with constant initial soil water content and over horizontal surfaces have been proposed, the estimate of infiltration at di fferent spatial scales (i.e., from the local to watershed scales) is a complex problem as further challenges are imposed by the natural spatial variability of soil hydraulic characteristics and that of rainfall [12]. Due to the intricate mechanism of groundwater, it is hard to simulate the hydrological phenomena in a hydro model because of the applied limitation, causing the use of extremely inconvenient models [13]. The simulation accuracy could be improved, by investing much time and many resources to establish a hydro model [14]. Thus, some researchers use statistical approaches to simulate groundwater variation. These approaches can reveal the stochastic dependence among the groundwater observations and their related variables, and commonly require a relatively fewer number of parameters than a physical-based model, with limited subjective assumptions [15]. Related research studies, such as those by Daliakopoulos et al. [16], Nayak et al. [17], Sahoo and Jha [18], and Liu et al. [19]. They used multiple linear regression (MLR) and artificial neural networks (ANNs) to establish the relationship using statistical models. For example, Daliakopoulos et al. used 17 years of data to establish ANN models according to time-lag rainfall, temperature, streamflow, and groundwater table in the Messara Valley basin (Greece). The dataset was divided into three parts for the purposes of training (11 years), cross-validation (3 years), and testing (3 years). The established ANNs included a feedforward neural network (FNN), a recurrent neural network (RNN), and a radial basis function network (RBFN). The coe fficient of determination (R2) was selected to evaluate model performance. The results showed that the R<sup>2</sup> in the FNN was between 0.592 and 0.993, in the RNN it was between 0.609 and 0.911, and in the RBFN it was 0.744 [16]. However, in statistical model development, it is more important to select highly related inputs to acquire accurate simulation results, and to exclude data having less impact on the output of the model, to reduce the freedom degree of the model and the processing. Many scholars use a sensitivity analysis (SA) as the determination method to obtain the relative importance of model inputs. Jha and Sahoo established multilayer perception (MLP), an RNN, and an RBFN to simulate groundwater variation in 17 sites in Jiangnan, Kochi, Japan. Groundwater table di fferences, past rainfall, stream stage, temperature, and seasonal dummy variables were selected as inputs. In that research, data were split into four years for the model training and validation, and the remaining two years of data for the model testing. The R<sup>2</sup> was selected for model performance evaluation. The results showed that, in 17 sites, the R<sup>2</sup> in MLP was between 0.781 and 0.971, in the RNN it was between 0.691 and 0.983, and in the RBFN it was between 0.691 and 0.983. After the ANN model was developed, an SA was conducted to identify the sensitivity of all inputs. The SA results showed that the H-4 well had the highest sensitivity for each input [20]. Ahlawat developed the relationship between precipitation and runo ff using an ANN model in the Betwa catchment, India. A sensitivity analysis was conducted for model importance evaluation. The results indicated that some of the rainfall stations could be removed from the model because of the low sensitivity [21].

#### **2. Materials and Methods**
