Use of active learning for earthquake damage mapping from UAV photogrammetric point clouds

ABSTRACT

This article presents an effective classification method for earthquake damage mapping from unmanned aerial vehicles (UAV) photogrammetric point clouds. The classification method consists of three main components: (a) construction of a point feature descriptor regarding to spectral, textural, and geometrical features, (b) optimization of collecting informative training samples through an active learning (AL) method, and (c) fine-tuning the point-based classification results with contextual information. Besides using existing spectral and geometrical features, we design a textural feature based on fractal theory to construct a point feature descriptor through linear combination. A batch-model AL method called Margin Sampling and Multiclass Level Uncertainty (MS-MCLU) is proposed based on classification uncertainty using a Support Vector Machine classifier. We use a multi-label Markov random fields to fine-tune the point-based classification results with a pairwise model. The proposed method was tested using three sets of point clouds generated from UAV images over Mirabello, Lushan, and Wenchuan earthquake scenarios in 2012, Italy, and in 2013 and 2008, China, respectively. The proposed classification method was compared with that of two other feature descriptors, i.e. spectral combined with textural features (Spe_Tex) and geometrical features (Geo). The results show that classification accuracies were improved by using the proposed point feature descriptor. Results also show that the proposed MS-MCLU AL method evidently saved the cost of collecting informative training samples and produced higher classification accuracies than a random sampling strategy. Moreover, contextual information contributed to the improvement on the point-based classification results and was suggested to be considered in earthquake damage mapping applications.

Acknowledgements

We would like to thank the anonymous reviewers and the editor for their comments. We would also like to thank Aibotix Italy for sharing the data set over Mirabello earthquake scenario in Italy and Jing Li and Hong Tang of Faculty of Geographical Science, Beijing Normal University, for providing the data sets over earthquake scenarios in China. We thank Mengmeng Li of Department of Earth Observation Science, Faculty ITC, University of Twente, for giving some comments and writing suggestions. This work is partially supported by the National Key R&D Program of China: [Grant Number 2017YFB0504101], the grants from National Natural Science Foundation of China: [Grant Numbers 41701534, 41701533], and the Innovation Leading Talent Project of Central South University: [Grant Number 506030101].

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. For the symmetric covariance matrix, only values on its triangular matrix are included, which are transformed into a vector (Permuter, Francos, and Jermyn 2006).

2. RS method was conducted with the integral RAND function in MATLAB and was independent from the feature descriptor of 3D points. Hence, the number of collected samples for each class was the same for different point feature descriptors Spe_Tex, Geo, and Spe_Tex_Geo.

Additional information

Funding

This work was supported by the National Natural Science Foundation of China: [Grant Numbers 41701533, 41701534];Innovation Leading Talent Project of Central South University: [Grant Number 506030101]; National Key R&D Program of China: [Grant Number 2017YFB0504101].