Multi-texture-model for Water Extraction Based on Remote Sensing ImageHua WANG, Li PAN, Hong ZHENGSchool of Remote Sensing and Information & Engineering, Wuhan University 129 Luoyu Road,Wuhan 430079,P.R.ChinaSchool of Electronic Information, Wuhan University 129 Luoyu Road, Wuhan 430079,P.R.ChinaAbstract:In this paper, a multi-texture-model for water extraction based on remote sensing imagery is proposed. The model is applied to extract inland water (including wide river, lake and reservoir)from high-resolution panchromatic images. Firstly directional variance is used to find river regions, and then grain table is adopted to avoid noise including objects that have similar directional variance characteristic as water surfaces. The experiment result shows that the proposed method provides an effective way for water extraction.1. IntroductionThe recognition of water from remote sensing image has drawn considerable attention in recent yeas. A large number of publications about water extraction appeared and various approaches for water extraction have been proposed. Zhou developed a descriptive model for automatic extraction of water based on spectral characteristics[1]. Barton applied channel 4 for NOAA/AVHRR to extract water[2]. Du proposed a approach for water extraction from SPOT-5 based on decision tree algorithm[3]. Li recognized and monitored clear water from MODIS[4]. Wu extracted water from Quick Bird image and used active contour model to obtain accurate position of river bank[5]; In order to extract water from high-spatial remote sensing images, He used wavelet technique to expend the information and cleaned main noise of the images, and then presented multi-window linearity reserve technique to conserve linear water[6].Recently, most research work on water extraction was forced on automatic recognition of water from remote sensing images based on spectral characteristics. However, there are some disadvantages of these methods: (1) The resolution of image used for water extraction is low. The minimum size of recognizable object is depended on the spatial resolution of sensor. Therefore it is difficult to obtain accurate position of water boundary. (2) Due to the characteristic of water itself and the sensor applied, in certain channels the spectral features of different objects are equilibrated. The equilibration leads to the phenomena of “different objects same image” or“different images same object”, which results in noise objects included in extraction result.In this paper, a multi-texture-model for water extraction based on remote sensing is proposed. The model is applied to extract inland water (including wide river, lake and reservoir) fromhigh-resolution panchromatic image. Firstly directional variance is applied to find river regions, and then grain table is adopted to avoid noise including objects that have similar directional variance characteristic as water surfaces. The experiment result shows that the proposed method provides an effective way for water extraction.This paper is organized as follows. In Section 2, the directional variance model adopted is introduced. Then, fusion of proposed grain table model with directional variance model is discussed in Section 3.The experimental results of the proposed multi-texture-model and comparative studies with single models are given in Section 4. We conclude this paper in Section 5.2. Directional Variance ModelThe aim of our research is to extract water larger than 100m2from panchromatic images. As shown in Figure 2(a), the research objects can be divided into three classes: wide river, lake and reservoir, which all represent as region in high-resolution imageries. The objects of background can be divided into two classes: building and cropland, which also represent as region.In panchromatic imagery, wide river has a similar gray level to building and cropland, though the mean grayof lake and reservoir is much lower than the background objects. Conventional methods for water extraction based on spectral characteristics are not effective in the situation. In the meantime, water body defines homogeneous areas whereas building and cropland correspond to heterogeneous regions. Therefore, we take into account the homogeneity of the image to separate wide river, lake and reservoir from background instead. To characterize the difference of homogeneity between water body and the other types of areas, we use a textual operator: the directional variance.2.1. The Directional Variance OperatorThe operator is derived from those defined by Guerin & Maitre and Airault & Jamet[10]. As shown in Figure1, the directional variance consists in computing, for each pixel M of the image, the variance of the gray levels of the image on several direction of a circle whose center is M and radius is R. Then, the direction with the highest variance value is kept. Its direction defines the direction for which image is the most heterogeneous, locally. Its variance value is the directional variance value of the pixel M.2.2. Extraction of water based on directional varianceAccording to the definition of the operator, the minimum acreage of recognizable water body is depended on the length of radius R. We have chosen a length of 10 pixels for 1m resolution. The directional variances of the five typical training samples (wide river, lake, reservoir, building and cropland) have been computed and the statistical comparison is summarized in Table1. The overall average of water directional variance is lower than the objects of background.Nevertheless, the directional variance of cropland is similar to wide river with overlapping potion over 90%.Inhigh-resolution panchromatic imagery, details inside wide river, such as boat, wave, etc, are represented clearly which result in the heterogeneous of water. In the meantime, the textures of parts of building (for example, roof ) and cropland are rather fine. In a small window, these potions define homogeneous areas with similar directional variance as wide river. The result is improved if we chosen a length of 100 pixels. The statistical comparison is shown in Table2. If the length of radius is large enough, directional variance of building is higher than other objects obviously with no overlapping portion; the difference between cropland and wide river is increased while the overlapping potion is decreased. However, increasing the radius leads to two problems which are outlined as follow:1) The size of recognizable water body increases;therefore water which has small acreage (for example narrow river) can not be detected.2) The position of water bank is not accurate although the spatial resolution of imagery is rather high.Hence, in this paper, a multi-texture-model is presented and two texture models are fused to extract water from panchromatic images. Firstly, we chose a radius of 10 pixels to extract water based on directional variance; and then, grain table is adopted to avoid noise including parts of building and cropland that have similar directional variance characteristic as water surface.3. Multi-texture-modelIn high-resolution imagery, cropland and building represents structural characteristic. According to this characteristic, grain analysis is adopted for further research on the original extraction based on directional variance. The grain table histogram is able to represent structural characteristic of the research object, which can be applied to recognize many kinds of different objects [12].3.1. Extraction of water fused by grain tableThe grain table histograms of the five typical training samples (wide river, lake, reservoir, building and cropland) are computed and correlation coefficients between them are summarized in Table3. Correlation coefficients between water classes are over 85%, however, correlation coefficients between water classes and background classes are lower than 65%.Hence, we compare the correlation coefficients of regions in extraction image base on directional variance with three water samples and two background samples respectively. If the region has a higher correlation coefficient with background classes, it will be marked background and wiped off[13].4.Experimental ResultsWe run the algorithm on several high-resolution panchromatic images. In Figure2.(a), we have been considering an aerial photograph(6126×4800) of a region in Wuhan, China, the resolution of which is 1m,including building, cropland, wide river( Changjiang river), lake, reservoir and cropland. The results of extraction based on directional variance with radius of 10 pixels is displayed in Figure2.(b), and clearly, water has been detected completely, whereas parts of building and cropland are included as noise objects in the result. Water extraction using directional variance with radius of 100 pixels is displayed in Figure2.(c)with correctness over 95%, however, small lakes are missed and the position of bank is not as accurate as Figure2.(b). Finally, in Figure2.(d), the result of Figure2.(b) is fused by grain table analysis, so that the correctness and completeness of extraction are both over 90%.5. ConclusionsBased on textural analysis of water in high-resolution panchromatic imagery, a multi-texture-model is presented for water extraction.The experimental results proved that the approach is efficient for inland water (including wide river, lake and reservoir) extraction. As the complexity and diversity of water, the rate of recognition of our algorithm fluctuates. Furthermore, the method is supervised which needs a lot of human interference to obtain training samples. Therefore, there are problems to be solved in future:1) Our further work should be extensible to multispectral remote sensing images.2) To decrease human interference, old vector will be applied to obtain training samples instead. 6. AcknowledgmentsThe work was supported by the National Key Technology R&D Program of China under grant No.2006BAB10B01.根据遥感图象的多纹理模型相关的水抽取Hua WANG, Li PAN, Hong ZHENGSchool of Remote Sensing and Information & Engineering, Wuhan University 129 Luoyu Road,Wuhan 430079,P.R.ChinaSchool of Electronic Information, Wuhan University 129 Luoyu Road, Wuhan 430079,P.R.China文摘:在本文中,提议了一个多纹理模型为根据遥感成像的水提取。