A mechanism-informed data-driven correction framework for deformation monitoring of high earth-rockfill dams

Accurate deformation monitoring is crucial for the construction and operational safety of high earth-rockfill dams. However, atmospheric refraction errors pose significant challenges due to their complex spatiotemporal variability, limiting the effectiveness of traditional meteorological correction methods, especially in mitigating random fluctuations and localized anomalies. Therefore, a mechanism-informed data-driven correction framework is proposed to improve the monitoring accuracy. Slope distance observations are first corrected using a physical meteorological model based on temperature, pressure, and humidity. A dynamic screening process is then employed using Mann–Kendall trend testing to identify stable reference conditions. To address residual nonlinear components, a feature-fused light gradient boosting (FF-LGB) model is developed, which integrates dynamic meteorological features (e.g., sliding averages of humidity, pressure gradients) and static spatial parameters (e.g., terrain roughness, sight obstruction index), using theoretical corrections as supervised labels. The FF-LGB model achieves excellent prediction performance, with a determination coefficient of 0.9961 and root mean square error of 0.5897 mm on the test set, outperforming eXtreme Gradient Boosting and multiple linear regression. Results show a mean absolute error reduction to 0.81 mm, representing improvements of 17.35 and 34.68% over traditional correction and uncorrected results, respectively. The proposed framework demonstrates strong robustness and generalizability, offering a promising solution for atmospheric refraction correction under complex environmental conditions.