Research On Remote Sensing Image Classification Of Land Covers Based On Multi-Scale Spectral-Spatial Features

The emergence and application of hyper spectral remote sensing technology has a history of nearly 30 years.It is one of the most important earth observation technologies.It can obtain images with nanometer resolution and has great advantages in describing image information.The classification of hyperspectral image is an important part of the application of hyperspectral remote sensing technology.At present,the main classification methods generally improve the classification accuracy by mining the spatial,radiation and spectral information of the image.However,due to the interference of climate and other factors,hyperspectral image processing faces some difficulties,such as high dimensionality and low spatial resolution,which hinder the improvement of classification accuracy.From the perspective of multi-scale,problems caused by the uneven distribution of the land covers in hyperspectral remote sensing images have been relieved,due to the benefits of multi-scale spectral-spatial features.The proposed two algorithms fuse multi-scale features at feature level and decision level respectively,and design corresponding optimization algorithms.A software system for land cover classification of hyperspectral remote sensing images is designed and developed on C#based on the two algorithms mentioned above.The effectiveness of the algorithms has been proven by using real hyperspectral remote sensing images.The main contents include:(1)A hyperspectral image classification algorithm based on multi-scale spatial spectral weighted filtering is proposed.Firstly,multi-scale spectral-spatial weighted filter is used to extract spectral-spatial features of hyperspectral images and reduce the influence of noise.Then,classification is carried out by using support vector machine at each scale.After that,the results of multi-scale classification are obtained.The label of each pixel is determined by using Markov Random Field.Experiments on Indian Pines,Pavia University and Kennedy Space Center datasets show that the application of multi-scale strategy is successful.The proposed algorithm can achieve satisfactory experimental results.(2)A hyperspectral image classification algorithm based on multi-scale spectral-spatial feature extraction is proposed.Firstly,spectral-spatial information of different scales of image is extracted by using multi-scale morphological profiles.Then,spectral-spatial features of different scales are concatenated into to a long vector.After that,low-rank embedding multiple features extraction is used to reduce the dimensionality of multi-scale features.Finally,support vector machine is employed for classification and the labels can be obtained.The algorithm is validated on the hyperspectral remote sensing data sets Indian Pines and Botswana.The satisfying results prove the effectiveness of multi-scale strategy.(3)Based on the above mentioned multi-scale classification algorithms,the software of hyperspectral remote sensing image classification and processing is designed and developed using C#.It consists of three modules,including file management,classification,and display of classification results.The system framework,process design and module work test are also presented...