Research On Unsupervised Classification Method Of High Spatial Resolution Remote Sensing Image Based On Super Pixel

In recent years,the high spatial resolution remote sensing image has developed rapidly,and the high spatial resolution remote sensing image data in China has reached the centimeter level.With the improvement of spatial resolution,the high spatial resolution remote sensing image data presents the uncertainty of ground object types,the morphological diversity of similar ground objects and the complexity of ground object spatial distribution,which makes it still very difficult to realize the accurate classification of ground objects in remote sensing images of natural scenes.At present,in view of the phenomenon that the classification of land cover relies heavily on the ground sample data and too much subjective interference,the unsupervised classification algorithm is adopted in this study.The traditional k-means unsupervised classification method,which is usually based on pixel classification,can only utilize the single spectral feature of the image,so it cannot overcome the interference of the data noise points of the high-resolution remote sensing image.Moreover,the output results are highly dependent on the input parameters,and the algorithm efficiency is low,so the image classification effect is affected.In order to improve the unsupervised classification effect of images,this study introduces a super-pixel algorithm,which takes the super-pixel segmented objects as processing units,and further makes use of the spectral,geometric,texture and other feature information of the objects to improve the accuracy and efficiency of classification and bring good news to the image classification of high spatial resolution.Based on the unmanned aerial vehicle(UAV)high spatial resolution remote sensing image as data source,the image data after pretreatment,respectively adopted the combination of L0 smooth and super pixel K-means unsupervised classification method,and based on the multi-scale thewire cut(mutli-scale thewire cut,hereinafter referred to as Rcut)segmentation of K-means unsupervised classification,analysis and comparison the around the content classification accuracy,the specific research work and content are as follows:(1)Pre-processing based on UAV image.First,L0 smoothing is used for uav image to effectively reduce a lot of noise and redundant information and make the image clearer.Thenhistogram equalization is used to enhance the image and make the texture and shape of ground objects clearer.(2)K-means unsupervised classification based on L0 smoothing combined with SLIC superpixel.First,the UAV image after L0 smoothing is segmented by SLIC superpixel,and at the same time,it is organically combined with the color and plane spatial features in the image to effectively make all features in the image converge.Then,unsupervised classification of the segmented image K-means can effectively solve the problems of wrong classification,missing classification and difficulty in sample selection,and improve the efficiency and accuracy of image classification with high spatial resolution.(3)K-means unsupervised classification,which is the fusion of SLIC superpixel and multi-scale R-cut.Firstly,the UAV image after L0 smoothing is segmented by SLIC superpixel combining texture,shape and color of ground objects.Then,aiming at the selection of scale parameters in the object-oriented classification process,this paper studies the use of scale evaluation tool and RMNE method to realize the selection of the optimal scale of local objects.Finally,the image is segmented by R-cut based on the optimal scale,while K-means unsupervised classification is carried out.This method not only retains the advantages of compact structure and strong homogeneity of SLIC superpixel algorithm,but also can truly reflect the edge contour information of the object,effectively improving the accuracy of classification.