Research On The Segmentation Method Of High-resolution Remote Sensing Image Based On Super Pixels

With the development of remote sensing technology for small drones and the gradual improvement of observation systems,high-resolution remote sensing images have become easier to obtain,and my country has gradually entered the era of high-resolution remote sensing image applications.With the continuous improvement of spatial resolution,high-resolution remote sensing images can show more detailed information of features,making the relationship between target features more complex,and the amount of information about features of features increases.The extraction and analysis of high-resolution remote sensing image information brings new challenges.Traditional image segmentation methods use pixels as processing units,and only consider the spectral information of the pixels,while ignoring the effects of other features of the image.The accuracy and efficiency of high-resolution remote sensing image segmentation are difficult to meet the needs of practical applications.Super pixels are small areas composed of a series of adjacent pixels with similar features such as color,brightness,and texture.These small areas retain the main information of the original image.Using super pixels instead of pixels to represent a high-resolution remote sensing image is more in line with human visual perception,while reducing the redundancy of information and reducing the amount of data processing.Therefore,the segmentation method of high-resolution remote sensing images based on superpixels has become an important research direction.This paper mainly conducts research from two aspects:the generation of super pixels in high-resolution remote sensing images and the fusion of super pixels.The research work and conclusions are as follows:(1)With high-resolution remote sensing images as the data source,the traditional pixel-based image segmentation method has low processing efficiency and will produce a large amount of "salt and pepper" noise,so the super pixel segmentation method is introduced.The basic principles of SLIC algorithm and ERS algorithm are studied.By changing the superpixel number parameter k,the compactness parameter m,and the number of iterations p in the SLIC algorithm,the generation law of SLIC superpixels is studied,and an empirical method for preliminarily estimating the parameter k through typical ground objects is obtained,and the parameter m And the setting experience of parameter p.The segmentation results of high-resolution remote sensing images by the number of superpixel parameters k,balance parameters and distance parameters in the ERS algorithm are studied,and the generation rules and parameter setting experience are mastered.Through the comparison of the two methods,it is found that the SLIC algorithm can generate more regular superpixels,and the superpixels generated by the ERS algorithm are more dependent on the edges of the features.(2)Superpixels are also a result of over-segmentation,and superpixels need to be merged to obtain the final segmentation result.In this regard,a spatially constrained fuzzy clustering method based on the fusion of spectral differences of superpixels is proposed.By analyzing the spatial information of the image features,comparing the similarities of the regions,the fuzzy clustering model of spatial constraints(Fuzzy Clustering Model of Space,FCM)merge super pixels.Seven high-resolution remote sensing images of aerial drones with a resolution of 0.3m and two hyperspectral remote sensing images with a resolution of 2m are selected,and the initial pre-segmentation area is generated by the SLIC and ERS algorithms,and the fuzzy c-means clustering method and The method in this paper merges the divided regions.Analyze and compare the experimental results through a combination of qualitative and quantitative methods.Experimental results:show that this method is better than the fuzzy c-means clustering method,can segment the typical objects in the image more completely,and improve the over-segmentation phenomenon;the precision and recall of the segmentation results reach 0.9,indicating that the method has better performance.With high segmentation accuracy and better boundary fit,it is a high-precision segmentation method suitable for high-resolution remote sensing images.