面向不同环境背景的Landsat影像水体提取方法适用性研究

快速、准确地从卫星影像中提取水体信息一直是遥感应用的热点问题,在水资源管理、水环境监测和灾害应急管理等领域极具应用价值。虽然目前已有多种针对Landsat系列影像的水体提取方法,但由于地理位置、地形和水体形态等环境背景因素的影响,导致同种方法在不同的环境背景中呈现出不同的提取效果。本文针对人为影响严重、影像明暗对比强烈的城区（北京怀柔县城周边）以及地形起伏明显、水体细小的非城区（北京密云水库周边） 2种典型背景环境,选择波段设置略有差异的Landsat 5（2009年）和Landsat 8（2019年）卫星影像,对比了常用的指数法（NDWI和MNDWI）和分类法（最大似然法和支持向量机）在水体信息提取方面的优势和不足。结果表明：在城区背景中,SVM的准确性最高（总体精度>97%）;在非城区背景中,MNDWI与SVM的精度相当（总体精度>95%）,前者更适用于水体的快速提取,而后者提取的山间细碎河流更完整,且在Landsat 8中应用的效果更好。该研究为不同环境背景下水体提取方法的选择提供了参考。

Application of Water Extraction Methods from Landsat Imagery for Different Environmental Background

Rapid and accurate extraction of water information from satellite images has been a hot issue in remote sensing applications and has important application value in water resources management, water environment monitoring, and disaster emergency management. Although there are a variety of water extraction methods for Landsat series images, the same method can generate different extraction results in different environmental backgrounds due to the influence of environmental background factors such as geographic location, topography, and water body shape. In order to study the applicability of water extraction methods under different environmental conditions, this article focuses on two typical environments: urban areas around Huairou County, Beijing with severe human influence and strong contrast between light and dark images, and non-urban areas around Beijing Miyun Reservoir with obvious topography and small water bodies. Water index method and classification method are tested based on water extraction and accuracy verification using Landsat 5 (2009) and Landsat 8 (2019) satellite images which have slightly different band settings. The water index method includes Normalized Difference Water Index (NDWI) and Modified Normalized Difference Water Index (MNDWI), while the classification method includes Support Vector Machine (SVM) and Maximum Likelihood (ML). Our results show that SVM has the highest accuracy with overall accuracy > 97% in the urban areas. By selecting training samples reasonably and delicately, the extracted spatial pattern of water results is close to the real water distribution. It applies well to multiple-scale and complex water bodies. In the non-urban areas, SVM can completely identify the fine rivers which are usually difficult to be identified by other methods. It is also suitable for judging the shape and flow direction of small rivers between mountains, though the shadow of the mountain could be easily mixed together by mistake. Due to the difference in sensor band settings, SVM has a better performance in Landsat 8 data. MNDWI can effectively reduce the error extraction rate, resulting in an overall accuracy > 95%. It is convenient to determine the threshold value of MNDWI through visual interpretation, which is more suitable for the rapid extraction of water in the non-urban areas. The environmental background may show different effects on water body extraction. The water index method and classification method have different advantages in different environmental backgrounds. The most suitable method should vary according to the actual situation. In scenarios with higher requirements for efficiency, we should focus on the use of index method, and design a new index which can make full use of the band information. In application scenarios where higher extraction accuracy is required, classification methods can improve the accuracy of water extraction. Moreover, we cannot ignore the differences between interpretation methods in data sources from different sensors. This study provides a reference for the selection of water extraction methods under different environmental backgrounds.