Residential Areas And Edge Lines Of Area Feature Extraction Methods Based On The Optimized Mean Shift Algorithm

With the rapid development of aerial photography and computer application, UAV and satellite for earth observation bring us the vast amounts of remote sensing data and information. How to extract the object of our interests from these remote sensing images has been the focus of scholars’ research. Among all kinds of objects, the residential area is the focus of the target extraction because of its importance and variability. The accurate extraction of this information is of great significance in the fields of urban construction planning, land use investigation, disaster assessment and national defense and military affairs. The existing methods of residential areas extraction based on decision tree, prior knowledge, statistical model and object-oriented approach. These methods have their own limitations. The method based on the decision tree model relies on the brightness value of each band to distinguish the type of object but the phenomenon of "wrong points" and "missing points" will appear when the complex land features are encountered. The method based on prior knowledge model has a higher requirement for surveying and mapping personnel’s prior knowledge. Its research scope is more specific, so it is not universal. The method based on the statistical model is unable to accurately distinguish the "synonyms spectrum" or "objects with the same spectrum" in image processing. The object-oriented approach has an undesirable result for residential areas in the high-resolution remote sensing images because of the large differences and complex structures of the residential areas.Mean shift algorithm, which is based on kernel density but not parametric density estimation algorithm, has the advantages of small computational complexity, high operating efficiency and strong robustness etc. Mean shift algorithm is currently used for image segmentation. This article attempts to use it in the processing of remote sensing images. Using the method of "segmentation firstly, then classification, finally extraction" to draw the specific information (residential areas and edge lines of area feature). The research contents and main results are summarized as follows:(1) By introducing texture parameters to optimize the mean shift algorithm, and test by taking different values of the algorithm which involves three parameters (space bandwidth parameter hs, chrominance bandwidth parameter hr and texture parameter of bandwidth ht). By comparing and analyzing the experimental results, we find that different parameters’ values influence the results of image segmentation directly, this fact benefits to residential areas extraction.(2) A method of extracting residential areas from remote sensing image based on optimal Mean shift algorithm is proposed. The Mahalanobis distance, Maximum Likelihood Estimation and Neural Network included by statistical classification models are combined with the algorithm. Experiments and researches of this method are done in aerial images and satellite images, and the experimental results of this method testifies that the new method is more accurate, faster, and more applicable.(3) A method of extracting edge lines of area feature based on optimal Mean shift algorithm is proposed and two improvements are made. The extracting is combined with the cartographic generalization to make it more targeted. The extracting is combined with the shape feature and context feature which belong to the image features of object to make the accuracy higher.