Line‐Based Levee Detection: A New Generation of Levee Detection Methods Using Digital Elevation Models

Abstract Levee location information is an essential input for accurate flood simulation. Several methods for detecting levees from LiDAR digital elevation models have been proposed in recent years. However, these detection methods are pixel‐based, treating continuous levees as discrete points, classifying each pixel as levee or non‐levee. Thus, they produce fragmented results and struggle to balance high recall (correctly identifying levee pixels) with high precision (avoiding misclassification of non‐levee pixels). In this study, we propose a framework for a new generation of line‐based detection methods that capture the linear continuous nature of levees and simultaneously optimize precision and recall. We first develop a line detection algorithm to identify all linear features that are potentially levees from DEM data. Each line is then classified as levee or non‐levee based on statistics of pixels along the line. We find that the line detection algorithm, combined with a newly developed active learning logistic regression model, results in a line‐based method that outperforms the latest pixel‐based method across all metrics. Specifically, the F1 score increased from 0.39 to 0.59 in the United States and from 0.23 to 0.54 in New Zealand. These increases are attributable to higher precision by the line detection algorithm and higher recall by the logistic regression model. By applying the newly proposed method to 31 countries, we detected a total of 54,200 km of levees, providing a new data set for global flood risk assessment.