Research On Water Extraction Algorithm For Remote Sensing Images

The classification and recognition in remote sensing images is one of the research hotspots in the field of remote sensing imagery.The research on water extraction has important application value in the both military and civilian fields.This thesis analyzes the advantages and disadvantages of water extraction algorithms for remote sensing images.It is found that most of the current water extraction algorithms through analyse the relationship between the spectral points of pixels is extracted to extract water areas.The amount of data to be processed is huge and the efficiency is low,and the important spatial information is not considered.Therefore,this thesis proposes an object-oriented water extraction model for RGB remote sensing images.Super pixel blocks replace pixel points as basic processing units,which reduces the scale of data processing and improves the computational efficiency.The content of this thesis is as follows:1)During the process of acquiring super pixel blocks using the SLIC algorithm,the super pixel block is not well-matched to edges of water and other objects.This thesis proposes a SLIC segmentation algorithm(RLBP-SLIC)that combines LBP local texture features.Firstly,calculate the RLBP coding value of the pixel in the image,and the rotation invariance is transformed and homogenized.Then the local texture feature extracted from the RLBP coded value is integrated with the Lab color and spatial information of the pixel,and a new similarity metric rule is constructed.Finally,according to the new similarity measure criterion,the pixel points in the image are classified,and a super pixel block is formed.The new SLIC algorithm considers the local texture information of the image,so the obtained superpixel block edge has a higher degree of coincidence.2)Using the SVM classification algorithm to classify water and non-water blocks,it is necessary to obtain the characteristics of the samples in the data set.This thesis extracts the color features of the superpixel block and the texture features of the wavelet transform to form the 291-dimensional feature vector.This method solves the problem of data redundancy caused by the feature dimension being too high and causes the classification effect of the classifier to decrease.This thesis is based on the original random forest feature.Based on the selection algorithm,this thesis proposes a random forest feature selection method that improves the selection of optimal features in a circular manner.The method establishes a random forest by cycling.Each time a random forest is established,a part feature with the most important feature is selected as a parameter of the random forest,and a new feature vector with a lower dimension and a better classification effect is finally obtained through loop selection..3)This thesis needs to establish a data set for classification of water and non-water samples.After superpixel segmentation of images,the data set is constructed by manually marking water and non-water ultra-pixel blocks,and divided into training set and test set..In this thesis,SVM SVM is used to classify the superpixel block datasets.Finally,the effectiveness of the random forest algorithm to improve the cycle selection of optimal features is verified.Moreover,the closeness of the waters in the water extraction experiment results is higher,which further explains the effectiveness of the proposed RLBP-SLIC algorithm.