Comparison of extreme gradient boosting, deep learning, and self-organizing map methods in predicting groundwater depth

Abstract

Excessive extraction of groundwater resources has led to a drop in the groundwater table. On the other hand, accurate data on spatial and temporal changes in groundwater depth (GWD) in the plains of Iran is not available. Therefore, using a new modeling method such as artificial intelligence to predict the groundwater depth can be efficient for accessing the desired data. This research aimed to compare the performance of extreme gradient boosting (EGB), deep learning (DL), and self-organizing map (SOM) methods in predicting the GWD on the Mazandaran plain. Data of the GWD from 250 piezometric wells was used as the output parameter. The effective factors in GWD fluctuations were used as input variables. Data were separated into two categories (a) 70% for training, and (b) 30% for the test stage. The modeling process was performed using the two methods of machine learning (EGB and DL), and a SOM with the same data. Three models were trained and tested and their results were compared. The modeling results in the training phase were appropriate for EGB (R-sqr = 0.97, NSE = 0.95), DL (R-sqr = 0.972, NSE = 0.81) and SOM (R-sqr = 0.88, NSE = 0.65) models. In the test phase, models EGB (R-sqr = 0.82, NSE = 0.8) DL(R-sqr = 0.74, NSE = 0.56), and SOM (R-sqr = 0.61, NSE = 0.31) had different performance.The results indicated that the EGB method possessed the maximum performance in the training and testing stages compared to the other two methods. Then, the tested EGB model was used to predict the spatial variations of GWD in the study plain, and the predicted results in the GIS were presented as a GWD map. The results of the analysis of the GWD map proved the high performance of the used methodology (R-sqr = 0.79). Finally, the proposed methodology can be used as a tool to predict the spatial variation of GWD in places without data or other plains.