Improved Groundwater Level Prediction in Data-Scarce Regions: A Spatiotemporal Framework Integrating Data Reconstruction and Graph Deep Learning

Reliable forecasting of groundwater level (GWL) fluctuations is essential for water resources management， yet it remains difficult in data-scarce regions where monitoring networks are sparse and meteorological and auxiliary data are limited. To address incomplete observations， this study reconstructs the GWL series using Seasonal-Trend decomposition using LOESS (STL) and the Discrete Wavelet Transform (DWT). This combination provides an effective strategy for reconstructing incomplete groundwater records while preserving dominant temporal dynamics and reducing high-frequency noise and local disturbances. The reconstructed multi-well records are then used for observation-only forecasting within a graph-based spatiotemporal learning framework， offering an alternative to approaches that rely on multi-source external inputs. Among the tested variants， the ResGCN-LSTM-Attention model achieved the best overall performance， with NSE values of 0.957 (training) and 0.954 (test)， along with well-calibrated uncertainty estimates. Residual graph convolution and attention mechanisms contributed complementary gains， capturing heterogeneous inter-well spatial dependencies and strengthening long-range temporal representation， respectively. Overall， the results show that accurate GWL forecasting can be achieved from incomplete multi-well observations alone， providing a practical solution for similarly data-scarce regions.