Embankment has been one of the most important measures of flood control since the ancient times. Its extraction and detection serve an important role in flood prevention. However

traditional optical images cannot extract precise embankment threedimensional models using feature points because dikes are always surrounded by trees and roads. The airborne light detection and ranging（ LiD AR） system can directly obtain high-accuracy three-dimensional information on a river basin. Dike points can be classified directly from the original LiD AR point data using the spatial semantic feature

including the slope and topological relation between the river and bank without any additional data source support. The method presented in this paper is important to automatic embankment extraction of embankment models.This paper presents a method that performs data pre-processing

water edge judgment

and slope fitting of extracted dike points using only the LiD AR point data.（ 1） The water and non-water areas are separated using echo

intensity

and point density properties. The water edge is obtained by gridding the point data.（ 2） The searching direction of the water edge is judged and sought.（ 3） The embankment points are extracted using the least square method after excluding the water edge without embankment.The experiment data covers the cities of Changyang

Shashi

and Xianning in Hubei Province to test the validity of the algorithm process

the point densities for these three cities are 1 dots / m2

1 dots/m2

and 2 dots/m2

respectively. The checkpoints are collected by the GPS RTK distributed randomly along the river side to verify the precision of the method used in this study. The test results are: The precision of the embankment brink extraction is less than 0. 43 m

0. 48 m

and 0. 3 m in Changyang

Shashi

and Xianning City

respectively

which shows that the accuracy satisfies the one-half point interval. The error of embankment elevation is less than 0. 13 m

0. 3 m

and 0. 08 m

respectively.The test results show that the proposed method can successfully extract the embankment of the river using only LiD AR data based on its semantic information. Moreover

a relatively high accuracy result can be achieved when the dike structure is simple and the point data are dense.