Research On Urban Land Object Classification Fusing Visible/Hyperspectral And LiDAR Data

The high spectral resolution of hyperspectral images can help to understand the attribute features of objects more deeply,but the lower spatial resolution and lack of the elevation information limits its further application.The emergence of high spatial,high temporal resolution data,multi-platform and multi-view data plays an important role in making up for the lack of hyperspectral images.Therefore,the dissertation takes hyperspectral images as the core,which is supplemented by LiDAR data with elevation information and visible images with high spatial resolution.The fusion classification algorithm is deeply studied based on the characteristics of multi-source data.To solve the problems that the different size of objects in the image and different feature space of multi-source data,this dissertation introduces an extended morphological filter for feature extraction.This method can integrate different data structures,and extract multi-scale features to better distinguish different scales without any regularization parameters.In addition,the post-processing method based on the edge-preserving filter is studied and the spatial features of the visible image are integrated in this step,which effectively improves the phenomenon that the labels are discontinuous in the classification results.The experimental results show that the fusion of hyperspectral image,LiDAR and visible image can effectively improve the classification accuracy.In order to mine the internal information of different data more deeply and achieve a more comprehensive description of ground objects,this dissertation adopts the deep learning theory to design a dual-branch convolutional neural network(DB-CNN)consists of 2D-CNN branch with cascade blocks and 3D-CNN branch.The feature extraction and fusion for the two data sources effectively maintain the image-spectrum characteristic of the hyperspectral image,as well as complish the multi-scale feature extraction of LiDAR data.At the same time,it greatly reduces the human participation in the feature extraction.The experimental results show that DB-CNN achieves better performance in the end-to-end feature-level fusion of hyperspectral image and LiDAR data.Considering the problem of the high dimensionality in fusion processing result from the difference between the characteristics of multi-source data,the ensemble learning classification for multi-source data in this dissertation is studied.By means of the characteristics of ensemble learning in which different classifiers can be generated,the results of different classifiers in different feature spaces will be majority voted to accomplish the decision fusion for multi-source data.In addition,the feature extraction algorithm of multi-source data is studied and the high-dimensional features are transformed to reduce the dimension and the redundancy of the features.The algorithm is proved to be well-performing and reliable.