Predicting salinity levels in the Mekong delta (Viet Nam): analysis of machine learning and deep learning models

Salinity intrusion stands out as a severe yet escalating challenge facing the water resource management and agricultural production of the Mekong Delta in Vietnam as a result of climate change and upstream hydrological changes. This paper assesses the efficacy of six different machine learning (ML) and deep learning models (DL) for hourly prediction of salinity in the Mekong Delta at four stations (Cau Quan, Tra Vinh, Ben Trai, and Tran De). The six models are XGB, GB, SVR, LSTM, RNN and ANN. Using hourly hydrological data of 2015–2020 with upstream discharge and tidal water levels as major inputs we designed training and testing of models (training: Jan 2015-mid 2018; testing: mid 2018-Feb 2020). Our results prove that LSTM and XGB models have the best prediction. In particular, they showed good accuracy in predicting upstream salinity (RMSE: 0.25 to 0.30, R2 > 0.97) and downstream salinity (RMSE: 1.5 to 1.6, R2 > 0.88). This success is due to capacity of high temporal resolution as well as spatio-temporal dynamics of salinity variation. The LSTM structure has proven to be effective at capturing long-term temporal dependencies, such as seasonal discharge patterns, while XGB successfully models non-linear interactions between stations with the greatest success, particularly discharge-tidal level interactions. The ML/DL models are capable of successfully forecasting salinity which can open doors to data-driven water management in the Mekong Delta. Future studies should further add hydro-meteorological parameters, other hybrid architectures, and real-time prediction systems, which could be useful operationally and have wider applicability.