Building a water feature extraction model by integrating aerial image and lidar point clouds

Abstract

An innovative model for extracting water regions from aerial images fused with light detection and ranging (lidar) data is proposed in this article. This model extracts water features from coarse to fine levels of accuracy by considering special spectral bands of existing airborne lidar systems and their spectral characteristics. The particular model consists of two parts, namely inexact water region recognition and precise water extraction. (1) A strategy of using a triangulated irregular network (TIN) is introduced to describe point clouds with a particular structure. A TIN coarsely divides the network into water and non-water regions through a threshold, which can be determined through an equation by inputting the minimum width and point density of water regions. The coarsely defined water region can be detected through overlay analysis between the aerial image and the raster surface generated from the TIN. (2) An improved mean-shift algorithm is used to remove most land pixels from the roughly recognized water to obtain precise water edges from coarse water. A new empirical formula to describe distance between multi-dimensional data is adopted. Using the mean-shift algorithm and empirical distance function, accurate water edge features are extracted from inexact water region(s). In addition, the classification field of lidar point clouds is used to remove land pixels from water features.

A case study based on a point cloud data set and an aerial image is conducted to evaluate the feasibility and accuracy of the proposed model. Spatial distances between checkpoints and extracted water edges, as well as the confusion matrix of mean-shift classification, are adopted as measurements of accuracy for the extracted water edges in two case regions. Evaluation results show that the proposed model achieved continuous water-edge features, and that spatial accuracy of water edges is 0.3 to 0.4 m, at approximately the 1–2-pixel level, which is more than four times better than the maximum-likelihood classification method. General accuracy of the confusion matrix shows that mean-shift classification in the proposed model is better than 95%, which indicates excellent results.

Acknowledgements

This work was supported by the National Science Foundation of China (No. 41101382), the National Basic Research Programme of China (No. 2013CB733204 and 2012CB957701), the Scientific Research Foundation of Key Laboratory for Land Environment and Disaster Monitoring of NASMG (No. LEDM2010B01), and the Key Laboratory of Surveying and Mapping Technology on Island and Reef of NASMG (No. 2010B12). The authors greatly appreciate the help of Ms Shumin Jing.