Research of Multi-Source Data Classification Based on the Combination of Spatial-Spectral Feature and Elevation Feature

• 1、School?of?Geosciences?and?Info-Physics,?Central?South?University,?Chang sha, 410083
• 2、School?of?Geosciences?and?Info-Physics,?Central?South?University, Chang sha, 410083

Abstract：As an important means of remote sensing image information extraction, remote sensing image classification has always been important in the field of remote sensing research, providing basis analysis of the data for other applications of remote sensing images. Traditional remote sensing image classification methods determine the category of image target (object or pixel) based on its spectral characteristics . However, due to the presence of same object with different spectrums and different objects with same spectrum, if only using spectral characteristics to distinguish targets ,the classification accuracy is low, and spatial information can be added to improve the classification accuracy. In addition, we can extract the plane information of the detected object from hyperspectral image, and extract the height information from laser radar data. Therefore, in order to further improve the classification accuracy, we can combine the spectral data with the radar data. According to the above analysis, we propose a novel approach of multi-source data classification based on the combination of spatial-spectral feature and elevation feature. First, we extract the spatial-spectral feature by reorganizing the local image patch with the first d principal components (PCs) into a vector representation after using minimum noise fraction to reduce dimensionality of the original hyperspectral image. Then, we get the feature of elevation information through processing the lidar image such as filtering and inverse distance weighting interpolation. Secondly, we combine features obtained from different data sources. Finally, we embed the resulting sparse feature coding into the support vector machine (SVM) for hyperspectral image classification. Experiments show our approach can extract the effective feature to represent the image, and achieve higher classification accuracy of ground objects.

Keywords： hyperspectral image lidar spatial-spectral feature elevation feature land-covers classification