Research On Town And Villiage Land Use Classification Methods Based On Google Earth Images

The land use analysis is greatly important for land survey and detection, meantime land use is also a kind of management method for reflecting management policies and testing the efficiency of management. Aim at the traditional pattern recognition classification, which exits the problems that there are much more images being misclassified and being missed recognized because some reasons which include spatial resolution of remote images, as well as different objects with the same spectrum and the same object with the different spectrum. Against the low classification accuracy of traditional methods .This thesis research the land objects'features analysis, features extraction and optimization, classification and accuracy assessment based on free Google Earth images.The main research work and conclusions are as follows:(1)The free and open Google images are adopted as research objects. The methods are promoted based on these images; According to the land use types and the acquired images'features, research object in this thesis are divided into three types: green land, residential areas and bare land; The new methods are developed for acquiring the land use information and town–village layout, which promote the land resources assessment and town planning.(2)Aimed at the different objects with the same spectrum and the same object with the different spectrum phenomenon, this thesis take gray feature as well as texture feature for classification features, and make a serious of experiments to ensure the parameters of gray level co-occurrence matrix in order to extract the completely and accurate textural features and image texture information; The Principal Component Analysis is adopted for optimizing the features to assure that a few features express the most objects'features, which reduce the redundancy of classification and improve the classification speed and accuracy.(3)BP neural network model is constructed in the thesis, according the experiences and analysis of many experiments, make sure the reasonable construct parameters and training parameters, after the BP network gets the high study and training level, put the classification samples into the BP network model; Confusion matrix is used to assess the accuracy of classified images and results show that BP network model is efficient for image classification and this model's accuracy is up to 88% and the kappa is 0.8145.Compared to the accuracy of Maximum likelihood classification, the BP network classification method is highly efficient. (4)The classification method based on K-means clustering algorithm is promoted, firstly change the color model of images from RGB model to CIE Lab model which is more like the human's visual, then make k-means clustering in the CIE Lab color space; in order to realize the auto segment, Hill-climbing algorithm is used for making the clustering types and clustering centers; Finally, confusion matrix is used to assess the accuracy of classified images and results show that this method's accuracy is 87.3%.