Multispectral-lidar data fusion via multiple kernel learning for remote sensing classification

Abstract

The airborne multispectral LiDAR system can simultaneously acquire the spatial, spectral and elevation features of the surface, making it an advanced method to acquire geometry spectral and elevation data at the same space and time. In order to represent and fuse the multispectral LiDAR data effectively, we research on two kernel feature structures, both of which utilize the provided features and fuse them by different Multiple Kernel Learning (MKL) methods to form the basic kernels. Instead of directly stacking multidimensional features as a single feature to generate multiple kernel matrices by altering the kernel parameter (FS-MKL), Feature-Based MKL (FBMKL) is used to form the combined kernel. With a fixed kernel parameter, FB-MKL firstly generates basic kernels according to each feature space by Single-Kernel (SK) method, and then applies the state-of-the-art kernel learning methods to align the generated kernels to project the features for linear SVM classifier. To prove the validity of the model, we exploit Single Kernel (SK) and several MKL methods to conduct the classification experiments with a real airborne Multispectral LiDAR data set. The result shows that the aforementioned FB-MKL model suits multispectral LiDAR data features and achieve higher classification precision compared with the existing FS-MKL model.