Surface Water Information Extraction From Remote Sensing Image Using Progressive Enhancement Model

With the rapid development of economy, the process of urbanization and industrialization produced water pollution is becoming more and more prominentenvironmental problems, moreover it is a serious threat to the social sustainable development. The water body information extracted precisely is important for the investigation of water resources, river comprehensive treatment, water project, flood and drought monitoring and disaster evaluation. Nowadays,it is an important way of multi-survey and supervision of water source that water body information can be quickly and precisely extracted from satellite remote sensing images. The water body information extraction accuracy problems may be obvious in areas where the non-water features includes low albedo surfaces such as asphalt roads in urban areas, and shadows from mountains and vegetation, building and clouds. Because water body and a part of non-water features both have similar reflectance, it would make mistake if we used a single and simple extraction method, thus decrease the accuracy of extraction result, also directly influences on the subsequent analysis work. Especially there is no algorithm for computer automatically to do a large of images that extract water body information from non-water features with high precision.Meanwhile, the selected threshold values are empirically, the results usually requires artificial threshold filtering. This process is not conducive to extract water body information of a large number and different timing remote sensing images. Therefore, all of this problems will bring some challenges to computer automation of water body information extraction, and this is also the main research content in this paper.In this paper, test areas, the west coast area of Pearl River delta, were selected on purpose so that the experiment consist of complex surface features, and we used three different timings Landsat TM/ETM+ images as data resource. Based on deep analysis about typical water body extraction method, we introduce a new water body information extraction method called Water Extraction Progressive Enhancement Model(WEPEM). At last, evaluation and precision analysis of result, we use visual judgement and image vectorization of human-computer interacting interpretation. Form data preprocessing to WEPEM building,mostly research details and achievements list as following:(1)Water information extracted indexes, the most normally water extraction methods are Normalized Difference Vegetation Index(NDVI), Normalized Difference Water Index(NDWI) and Modified Normalized Difference Water Index(MNDWI), those indexes using different image bands' spectral divergence between the water and the other features, to extract the water. But NDVI cannot effectively distinguish the signal from vegetation and water; NDWI will mix up the signal from building shade and water; MNDWI is the best among the three, but it also extracted some sort of mountain shade which lead to the result would not precise enough.(2) With the problems of this three methods of water information extracted, this paper presents a model of progressive enhancement of the surface water information extraction. This model, first of all, calculates three indexes(NDVI, NDWI, MNDWI) respectively; then, extracts water information progressively by setting threshold; finally, adds up three isolated results so that the information of water get progressively enhanced. This model can be used for extracting water with higher accuracy; faster speed; especially solves out the problems of those three indexes, makes the extraction of water information to achieve the best.(3) In order to test accuracy, the results extracted by progressive enhancement model and the three indexes were compared with the real situation in the selected part studied area. The result was shown that the new method can be used for extracting water in lake and fishpond study area that would promote at least 13% in overall accuracy, and at least 0.26 overall Kappa value get promoted; at least 1% overall accuracy and 0.02 overall Kappa value get promoted in river study area.The default threshold of WEPEM was shown to be stabilized with images of different locations and times compared to that of MNDWI, especially in extracting water information automatically by computer with high accuracy and with a default threshold. In this paper, we used the low spatial resolution of Landsat TM/ETM+ images, more mixed pixels would affect the accuracy of water body information extracted. We will consider the mixed pixels decomposition to solve the problem in later study.