Integration of Hyperspectral Image and Lidar Data through Tri-training for Land Cover Semi-supervised Classification

Abstract

The topology information derived from lidar sensors is an important complement to classification of the optic remote sensed imageries. In the paper, we investigate the joint use of hyperspectral and lidar data to extract information of spectra, texture and elevation for landcover discrimination. A semi-supervised fusion method using tri-training is proposed to integrate the different features. The DSM (Digital Surface Model) derived from the lidar data is first interpolated to the same spatial resolution as the hyperspectral image. The co-registration of DSM and hyperspectral image is followed by the spatial feature extraction using three different morphological profiles. Each kind of textural feature is concatenated with the spectral and altimetry features into a stacked feature vector, and thus three types of vectors are obtained. Subsequently, three supervised classifiers are built based on the three kinds of vectors, respectively. These classifiers are refined with help of the unlabeled samples in the tri-training process. Finally, an improvement in the final classification accuracy is achieved. The experimental results show that the proposed method can effectively integrate the information both from the spectral-positional-textural features and the labeled-unlabeled data.