Urban classification by multi-feature fusion of hyperspectral image and LiDAR data

Abstract

Land Use/Land Cover (LU/LC) classification of urban areas is of great significance to urban studies and has become a highly important research direction. However, with continuous urbanization and more types of inner cities diversified, single remote sensing image has been unable to meet the requirements of high precision. Therefore, urban LU/LC classification by data fusion has emerged. In this study, hyperspectral images are widely used in urban LU/LC classification because of their abundant spectral information. However, an objective limitation is that similar spectral characters with different elevation cannot be distinguished. LiDAR data can obtain accurate elevation information. Therefore, such data will obtain better classification maps when merged with hyperspectral images. This work proposes an urban LU/LC classification method based on the multi-level fusion of hyperspectral imagery and LiDAR data by using the complementary of their advantages. First, the spectral, spatial, and elevation information extracted from two images are stacked to achieve level fusion. Then, the classification is divided into two frameworks. One framework classifies all pixels of the feature images, while the other uses LiDAR data to extract the building mask and classify the off-building area. Classification maps of this framework are obtained by combining the classification map of the latter framework and the off-building area. The classification results are then obtained by voting the classification results obtained by the two frameworks to complete the decision-level fusion. Finally, the conditional random fields are processed to smoothen the image and remove noise. The data set of 2013 IEEE GRSS data fusion contest was experimented on to verify the effect of the proposed algorithm. The OA was 93.22%, and Kappa was 0.93. The accuracy of the proposed method exceeded 90% in most categories, while the classification accuracy of synthetic grassland, soil, tennis court, and running track was 100%. Experiment results showed that the proposed algorithm greatly improved the classification of buildings, roads, and parking lots. In this study, hyperspectral imagery and LiDAR data are applied to classify LU/LC in urban areas. It also combines feature level and decision level and achieves good results. The following problems will be considered in future works: increasing the accuracy of building extraction to improve the effect of feature-level fusion, considering the increasing intensity of LiDAR point cloud data in feature-level fusion, and increasing the number and diversity of classifiers when using the multiple classifier classification.