Evaluation of Subsurface Damage for In-Service Concrete Dam by Time-Series Deep Learning with Three-Dimensional Shadow Modeling

This study integrates shadow effect simulation with surface temperature prediction methods for concrete dam monitoring using passive infrared thermography. UAV-LiDAR based three-dimensional point cloud analysis enabled accurate shadow effect simulation on concrete dam surfaces throughout diurnal cycles using geometric ray-tracing algorithms. Two prediction methods were systematically compared: heat balance analysis incorporating solar radiation modeling and Long Short-Term Memory (LSTM) neural networks, both evaluated with and without shadow effects. LSTM with shadow effects achieved the highest prediction accuracy (MAE: 1.07°C) compared to heat balance analysis with shadow effects (MAE: 2.94°C), demonstrating superior performance in capturing meteorological-thermal relationships. Both enhanced methodologies successfully distinguished surface temperature variations between damaged (spalling) and non-damaged areas with sufficient accuracy (MAE: 1.64°C) for reliable subsurface damage detection applications. The results demonstrate that LSTM effectively captures complex environmental-structural interactions through implicit learning of 11 meteorological parameters, while shadow effect incorporation significantly enhances prediction accuracy for both computational approaches. This integrated approach provides a robust and practical framework for structural health monitoring of aging concrete dam infrastructure through advanced passive infrared thermography techniques.