3D gray level co-occurrence matrix and its application to identifying collapsed buildings

With the remarkable progress in access to remote sensing imagery data, nowadays research very often utilizes more than one image. We are often able to use multitemporal, hyperspectral, and/or full polarization of microwave radar images. In addition, it has become the general consensus that texture analysis plays an important role in remote sensing. It has been found in several publications that texture analysis was applied to each layer separately; however, this procedure requires a significant amount of computation and produces a massive volume of data. One alternative, and perhaps a better procedure, is to arrange the images into a multi-layered structure and perform texture analysis within some sort of three-dimensional domain. This manuscript extends the concepts of the gray level of co-occurrence matrix (GLCM) texture analysis applied for a single image to a multi-layered set of images, referred to in this paper as 3DGLCM. We then presented an interpretation of the 3DGLCM within the context of building damage identification. A set of 3DGLCM-based features were computed and evaluated as well. As a result, it was observed that some texture features have certain similarities with other methods proposed in previous studies, whereas other features have not been used before. Furthermore, this paper evaluates the performance of the Support Vector Machine (SVM) classifier in learning and detecting collapsed buildings using 3DGLCM-based features. Thus, the empirical evaluation focuses on the identification of collapsed buildings caused by the 2011 Tohoku earthquake and tsunami, where individual polarized TerraSAR-X intensity images are used to compute the texture features, and the collapsed buildings caused by the 2016 Kumamoto earthquake, where LIDAR-based digital surface models are used to compute the texture features. Extensive datasets consisting of building damage states that have been visually inspected by local authorities and research teams are used to set up the training and testing subsets. Furthermore, the proposed texture features are compared with features commonly used to identify collapsed buildings. The study concludes that an SVM trained with 3DGLCM-based features identifies collapsed buildings with high accuracy and outperforms an SVM trained with common features used in previous studies.