Precise and accurate groundwater level (GWL) prediction is vital in developing water resources management strategies since they provide reliable quantitative information (Thomas and Gibbons 2018; Wunsch et al., 2020). Recently, numerous studies have explored the GWL prediction using different numerical and data-driven models (Malekzade et al., 2019; Moghaddam et al., 2019; Adedeji et al., 2020; Almuhaylan et al., 2020; Chakraborty 2020; Rezaei et al., 2021). Needing an extensive and uncertain dataset, including hydrogeological, water budget, geophysical, is a drawback for phsics-based models. Such limitations have pushed engineers and researchers to apply data-driven methods in practice. Data-driven modeling utilizes real-time tolerance to model the hydrological events in an inaccurate and uncertain environment (Kisi et al., 2017; Dehghani and Torabi, 2021; Ghazi et al., 2021).
Hydrological and meteorological time series exhibit nonlinear time-dependent behavior, which are too complicated to solve with standard numerical and statistical models (Rajaee et al., 2019). Recently, artificial intelligence (AI)-based methods such as artificial neural networks (ANNs), group method of data handling (GMDH), gene expression programming (GEP), least-square support vector machine (LSSVM), fuzzy logic (FL), adaptive neuro-fuzzy inference system (ANFIS), model tree (MT), multivariate adaptive regression splines (MARS), and evolutionary polynomial regression have been widely employed to simulate GWL (Shiri et al., 2013; Suryanarayana et al., 2014; Kisi et al., 2017; Mohammadrezapour et al., 2018; Rajaee et al., 2019; Roshni et al., 2019, Adedeji et al., 2020; Afzaal et al., 2020; Roshni et al., 2020; Ghazi et al., 2021). These well-accepted models can cope with the complexity of GWL prediction and could relatively provide a better accuracy than numerical models.
Comparison of the several AI-based models in GWL prediction still is highly demanded. In a study, Moghadam et al. (2021) used combinations of parameters including GWL, groundwater withdrawal, recharge, precipitation (P), evapotranspiration (ET), and temperature (T) in 10 scenarios to predict GWL for five observation wells. Results pointed out that the GMDH had a better outcome Bayesian network, and ANN models in evaluating the influential variables in GWL prediction. Yin et al. (2021) assessed GWL in San Joaquin's aquifers using ANN, response surface regression, support vector machine (SVM) models and Bayesian model averaging machine learning ensemble models. Their results showed that ensemble model performance was better than stand-alone AI-based models. Also, the authors figure out that groundwater extraction for agricultural usage is the main driving force for aquifer storage changes. In another study, Shiri et al. (2020) used six AI-based models, ANN, BT, MARS, RF, GEP, and SVM, in a coastal aquifer to forecast GWL, and they figured out that GEP's outcomes were the superior one. Osman et al.'s (2021) study showed that the Xgboost model had the best results among other used AI-based models such as ANN and support vector regression to predict GWL. A brief detail of studies regarding applying the AI-based models for GWL prediction is presented in Table 1. The ANN model is the most common AI-based model for GWL prediction based on Table 1,
Table 1
Earlier research of AI-based methods for GWL prediction.
Reference | Models | | | | | | Input Variables | Lead Time | Time Interval |
ANN-based | Fl | ANFIS | GMDH | SVM | Other methods |
Gong et al. (2015) | ✔ | | ✔ | | ✔ | | GWL,SWL, P, T | GWLt+1, GWLt+2, GWLt+3 | Monthly |
Nourani and Mousavi (2016) | ✔ | | ✔ | | | | GWL, P, T, Q | GWLt+1 | Monthly |
Wen et al. (2017) | ✔ | | | | ✔ | Wavelet-ANN | GWL, P, E, T | GWLt+2, GWLt+3 | Monthly |
Huang et al. (2017) | ✔ | | | | | | GWL | GWL | Daily, Weekly, Monthly |
Kouziokas et al. (2018) | ✔ | | | | | | P, T, M | GWL | Daily |
Roshni et al. (2019) | ✔ | | | | | | GWL, P, T, Q | GWLt+6, GWLt+12 | Weekly |
Derbela and Nouiri (2020) | ✔ | | | | | | GWL, P, E | GWLt+2 | Monthly |
Naganna et al. (2020) | | | ✔ | ✔ | | GTB | GWL | GWLt+1 | Monthly |
Sahu et al. (2020) | ✔ | | | | | | GWL, P, T, SWL | GWLt+1, GWLt+2 | Monthly |
Moosavi et al. (2021) | | | | ✔ | | Wavelet-GMDH | GWL, P, T, SWL, D | GWLt+2 | Monthly |
Razzagh et al. (2021) | ✔ | ✔ | ✔ | | | | GWL, P, T, D | GWLt+1 | Monthly |
Current study | ✔ | ✔ | ✔ | ✔ | ✔ | | GWL, P, T, Et | GWLt+1, GWLt+2, GWLt+3 | Monthly |
Models |
FL: Fuzzy Logic; GTB: Gradient Tree Boosting; ANFIS: Adaptive Neuro Fuzzy Inference System; ANN: Artificial Neural Network; GMDH: Group Method of Data Handling; LSSVM: Least Square Support Vector Machines. |
Input Variables: GWL: Groundwater Level; Q: River Flow; SWL: Surface Water Level; P: Precipitation; T: Temperature; ET: Evapotranspiration; M: Moisture; E: Evaporation; D: Exploitation Well Discharge. |
[Please insert Table 1 here]
The ANN has been recently utilized for GWL prediction (e.g., Zare and Kooh 2018; Banadkooki et al., 2020; Dehghani and Torabi 2021). Likewise, ANFIS and SVM have been applied to predict GWL and indicated an improvement in accuracy and precision compared to ANN in GWL prediction (Emamgholizadeh et al., 2014; Kasiviswanathan et al., 2016; Khedri et al., 2020). Even though the ANNs, SVMs, and ANFIS have been commonly employed in GWL simulation whereas the usage of the GMDH model has seldomly been investigated on groundwater modeling procedures. However, this method has been successfully applied in civil engineering, water quality management, and soil science (Najafzadeh et al., 2013; Rahmati 2017; Mehri et al., 2019; Tayebi et al., 2019; Lin et al., 2020). One of the motivation of this study is to assess the viability of GMDH model in the GWL prediction.
The present study evaluates the ability of various AI-based models in GWL prediction. The main aim to conduct this research could be summarized as a) modeling behavior of an rising GWL in a monitoring well while the other parts of the aquifer demonstrate a severe declining GWL; b) predicting GWL at the aquifer scale using monthly GWL, P, T, ET dataset; and c) comparing efficiency of the FL, ANFIS, ANN, GMDH and LSSVM models in GWL prediction. The present study sheds light on the GWL modeling in aquifers with poor hydrological and hydrogeological datasets. The outcomes of this sort of AI-based models provide a reliable perspective for decision-makers to attain sustainable water resources management goals. Figure 1 shows the procedural outline of the applied AI-based models.
[Please insert Fig. 1 here]