Toward Robust Semantic Segmentation in Levee Infrastructure Monitoring: Enhancing Accuracy with High-Fidelity Synthetic Data and Ensemble Learning