Spatiotemporal Forecasting of the Groundwater Quality for Irrigation Purposes, Using Deep Learning Method: Long Short-Term Memory (LSTM)

Present study was conducted to predict the spatiotemporal groundwater suitability for irrigation purpose through deep learning method, Long-Short term memory (LSTM), in northwest of Iran. Sodium Adsorption Ratio (SAR) as a crucial irrigation water quality criterion for 101 sampling point for an 18-year data period from 2002 to 2019, was utilized as the input for deep learning model in order to forecast the irrigation water quality for the next year, 2020. To evaluate the model accuracy in spatiotemporal data forecasting, performance criteria such as Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and R were used which approved the model accuracy by 1.212, 0.312 and 0.89 of MAPE, RMSE and R, respectively. On the other hand, the model capability was assessed by RBIAS and generalization ability (GA), which results showed that LSTM model underestimated the targets with RBIAS equals to about 1.539, while had an acceptable GA, equals to 1.1832. Considering the carried-out map of irrigation water quality for the study area it was revealed that about 78% have the desirable to acceptable quality for irrigation and the about 22% are moderate to non-acceptable. The most non-acceptable points are juxtaposed to the residential area which shows the anthropogenic effect on groundwater quality through the fertilizers and the other pollutants, infiltered into the groundwater resources.