Prediction of Groundwater Level and its Correlation with Land Subsidence and Groundwater Quality in Cangzhou, North China Plain, Using Time-Series Long Short-Term Memory Neural Network and Hybrid Models

• Published in: Natural resources research (New York, N.Y.) Vol. 34; no. 3; pp. 1645 - 1666
• Main Authors: Nafouanti, Mouigni Baraka, Li, Junxia, Chakira, Hamada, Nyakilla, Edwin E., Fabiani, Denice Cleophace, Gondwe, Jane Ferah, Sallah, Ismaila
• Format: Journal Article
• Language: English
• Published: New York Springer Nature B.V 01.06.2025
• Subjects: China; Contamination; Correlation; Deep learning; Drinking water; Fluorides; Forecasting; Groundwater; Groundwater levels; Groundwater quality; Land subsidence; linear models; Long short-term memory; Mathematical models; Neural networks; Numerical models; prediction; Regression analysis; Subsidence; Support vector machines; Time series; time series analysis; Water quality; water table; China
• DOI: 10.1007/s11053-025-10474-1
• ISSN: 1520-7439

Groundwater is the primary source of drinking water in the world, but its contamination and reduction cause environmental problems. Traditional hydraulic and numerical models for assessing groundwater and land subsidence are time-consuming and expensive. Thus, this study used the long short-term memory (LSTM) neural network to predict groundwater level and employed linear regression analysis and the hybrid random forest linear regression to find the correlation between groundwater and land subsidence. The impact of groundwater level on groundwater quality was investigated by forecasting the fluoride in groundwater using the hybrid models of random forest and k-nearest neighbor (RF–KNN), random forest linear model (HRFLM), and gradient boosting support vector regression (GBR–SVR) for the prediction of groundwater fluoride. The LSTM model yielded an R2 of 0.96 in forecasting groundwater level, and the time series results from 2018 to 2022 showed a variation in groundwater level, with a decline in 2022. The LSTM model suggested that from 2024 to 2040, the groundwater level would recover progressively. The regression analysis showed an R2 of 0.99 and a p value of 0.01 for the correlation between groundwater level and land subsidence, and the HRFLM model yielded an R2 of 0.94. For predicting groundwater fluoride contamination, the hybrid RF–KNN had the highest R2 of 0.97 compared to HRFLM and GBR–SVR, with R2 of 0.95 and 0.93, respectively. This research demonstrated that hybrid models and deep learning are advanced techniques that can be applied in Cangzhou to evaluate groundwater level and land subsidence and they can be applied in areas facing similar challenges.