Research On Feature Classification Based On Fusion Of Hyperspectral And Airborne LiDAR

Land resources are the material basis on which people depend,and they are also the guarantee of the country's social and economic development.China's land resources are rich and diverse,and the key to the rational development and utilization of them is to obtain true and accurate information about the ground features,which is of great significance for the timely planning and management of land resources.With the continuous development of aerial remote sensing technology,remote sensing images are more and more widely used in the classification of ground features.However,land resources are constantly developing and changing,and the types of ground features are becoming more and more complex.A single remote sensing image technology cannot meet accurate ground feature information extraction.Therefore,the classification of ground features has shifted from a single remote sensing data source to the fusion of multi-source remote sensing data.Hyperspectral imagery has high spatial resolution and contains rich spatial spectral information,and has been widely used in land feature classification in recent years.Airborne LiDAR data has unique high-precision three-dimensional spatial information,which can provide high-level information not available in hyperspectral imagery for classification of ground features.Therefore,the fusion of hyperspectral and airborne LiDAR data is an effective method for feature classification.In this paper,the core observation area in the middle reaches of the Heihe River is taken as the research area,and the characteristics of the hyperspectral and airborne LiDAR data are used to extract the characteristics of multi-source data for fusion classification research.The paper first uses principal component analysis and minimum noise separation transform to reduce the dimensionality of hyperspectral images to obtain effective image bands containing most of the spectral information,and uses SVM,convolutional neural network and residual network algorithms to perform dimensionality reduction on the images,and compares and analyzes the classification results of different algorithms.Then,by calculating the normalized vegetation index and gray level co-occurrence matrix of the hyperspectral imagery,the spatial spectrum and geometric texture features are obtained,and a multi-source feature library is formed with the canopy height feature of the onboard LiDAR,and different feature combinations are compared to the ground The effect of discriminating types.Finally,based on the hyperspectral and airborne LiDAR fusion data,a hierarchical fusion residual network classification algorithm is designed to extract the sample features step by step to achieve the fusion of complementary information between network layers.In this paper,CASI aerial remote sensing images and airborne LiDAR canopy height point clouds are used to verify the effect of feature extraction and classification.The experimental results show that:1)Research on the feature classification method based on hyperspectral imagery proves that compared with the SVM classification algorithm,deep learning can automatically extract feature features,reducing the problem of incomplete features caused by artificial design and more accurate classification results;In the deep learning algorithm studied in this paper,DRN solves the burden brought by the increase in the number of CNN network layers,learns more discriminative features,and has better classification results.2)Combining CASI hyperspectral and airborne LiDAR data for feature classification,compared with classification using only CASI hyperspectral data,the accuracy of each object and the overall accuracy have been greatly improved;the spatial and texture features of hyperspectral images and the canopy height features of the airborne LiDAR can complement each other,providing richer and more accurate feature information for classification,and solving the problem of insufficient features of a single remote sensing data.3)The overall classification accuracy of the fused data based on the hierarchical fusion residual network reaches 97.89%,which is 10.13%higher than that based on CNN,and 5.68%higher than the overall accuracy based on DRN,the classification accuracy of various features has also been improved;The algorithm improves the classification accuracy of the fusion of hyperspectral and airborne LiDAR images by fusing output vectors of different network levels to extract complementary feature information,and provides a feasible direction for the research of feature classification algorithms.