Employing machine learning to document trends and seasonality of groundwater-induced subsidence

Land subsidence is largely influenced by the declining trend and seasonal fluctuations of groundwater levels. This study proposes a comprehensive data-driven subsidence trend identification and forecasting system based on hydraulic head observations by integrating seasonality and trend decomposition with machine learning. The study firstly identifies the decomposition components of the deformation and hydraulic head time series data, and formulates a relationship between subsidence trend and hydraulic head to detect regions suffering from significant subsidence. Overall deformation was estimated using head changes firstly, and then subsidence trend was identified after removal of the non-trend component. Results revealed that the deformation varied with seasonal groundwater fluctuations that were dominated by a strong response during wet and dry seasons. Seasonality between subsidence and groundwater change is strongly correlated. The subsidence trend can be determined from the removal of the seasonality related to head change, especially near the proximal and mid-fan areas. A subsidence hotspot with a significantly steeper trend (> 4 cm/year) was observed in inland areas (mid-fan). The R² achieved a strong fit when model estimation was compared with the observed deformation trend for the testing period (2019–20). The subsidence forecasting system developed in this study could enhance subsidence and groundwater management to offer the potential of identifying areas experiencing increased irrecoverable subsidence by analyzing the estimated subsidence trend associated with groundwater changes.