Research On Classification Method Of Remote Sensing Images Based On MLR And Spatial-Spectral Feature

As remote sensing techniques develop rapidly,relative research of fields like atmosphere,landform and ocean information,whose primary works are image analysis and processing,is offered efficient and abundant material directly.However,the remote sensing images,especially the hyper spectral images,contain a large amount of data,including some time-based data like meteorological data,which means artificial processing is a hard work with high cost and low efficiency to conduct.Image processing methods,like classification algorithm,could be adopted in these steps to promote the efficiency.Nowadays,there are huge numbers of algorithms used for dividing the objects,like cloud,mountain,river and ocean,in a sensing image with relatively high accuracy.After the research of current classify techniques for remote sensing processing and analysis,this paper promotes a classification method of remote sensing images based on MLR and spatial-spectral feature.Compared with other method based on SVM and feature extracted from single pixel,there are three advantages as follows:1.This paper promotes a band selection method based on OIF.Before the band selecting by OIF,this method uses a primary selection parameter to reduce the images with indistinct structure and in noise bands.The number of final images is controlled by a choose threshold,so that the amount of original dataset is reduced to a lower level.And then,select the remaining bands through OIF,which is of a huge amount of calculation.Finally,the bands selected by this method is used for classification.Our method is proved with lower calculation cost than other method using OIF directly.2.This paper promotes a method to set up a feature dataset by region spatial feature.Firstly,divide the original images into pieces called super-pixel by image segmentation algorithm based on entropy rate.These pieces with compact structure are distributed evenly and the edges fit well.And then,calculate the variant of each region as the feature used for classification.Compared with other feature extraction methods,our method makes full use of the spatial structure of pixels and treat the pixels having similar quantity as a unit.In this way,the calculation cost is cut down and the spatial structure is integrated.3.This paper adopts an MLR classifier based on general composed kernel.The GCK is of adequate flexibility to combine the spatial and spectral feature and it is possible to extend the kernel structure in order to draw in more feature to improve the accuracy of classification.With respect to the efficiency,the kernel can conduct both mapping and inner product at the same time,which means the relative operations used for dealing with high dimension vectors are simplified.On the other hand,this method is of less training and classifying time than traditional method based on SVM,which proves this method is of higher classification efficiency and accuracy.