美国高分辨率土地覆盖信息提取技术研究进展

Development of Extraction for High-Resolution Land-cover Information in the United States

土地覆盖数据是进行全球变化研究的基础。美国地质调查局组织的土地覆盖特征研究项目成功的开展了两次工作,建设了土地覆盖数据库,得到了多方面的认可,代表了美国相关领域的前沿技术。项目中涉及一系列技术,包括土地覆盖分类分区、遥感影像选取、数据预处理过程、图像转换变化分析、土地覆盖分类技术、数据产品验证、土地覆盖数据库建设等。另外还有关键技术研究,包括:土地覆盖多边形分析、纹理特征分析、树冠密度信息提取、城市不透性表面估算等。就美国地质调查局项目中土地覆盖遥感影像数据处理方法、技术流程、数据库建设,以及有关土地覆盖度等相关关键技术进行了介绍,期望能够反映其进展情况,对国内的相关工作起到启示作用。

Land-cover data was foundation about study for globe change.The USGS Multi-Resolution Land Characteristics(MRLC) Consortium was originally formed in 1993 in order to meet the needs of several federal agencies for Landsat 5 imagery,and land-cover information.One of the results of this consortium was the completion of a successful mapping of the conterminous United States into the National Land Cover Dataset(NLCD 1992).The growing need for current Landsat 7 data,land-cover and other geospatial data within the federal government culminated in reforming the MRLC Consortium in 2000(MRLC 2000),and derived a second-generation National Land Cover Database(NLCD 2000).This multi-layer,multisource,value-added database will include a suite of 30-meter resolution data that will serve as standardized ingredients for the production of land cover-both nationally and locally.There was systematic Image Processing for standard image data.At first,mapping zones as a re-classification stratification method that was stratified landscapes into sub-regions of similar biophysical and spectral characteristics.To meet the requirements of the land cover database,the scene selection strategy was based on image quality and vegetation dynamics of target land cover types over a growing season.For MRLC 2000 image Preprocessing,images are geometrically corrected using cubic convolution resampling.Images are radiometrically corrected using standard methods at the USGS EROS Data Center to eliminate band bias and gain anomalies.And then images are converted to at-satellite reflectance for the 6 reflective bands and to at satellite temperature for the thermal band.TC transformations distill original resolution TM bands into spectral-efficient transformations without losing important information.NLCD 2000 classification included clustering,expert system,neural network and decision tree classifiers.Land cover derivative data layers were initially assessed using cross validation in both regression and decision tree models.At the mean time,there are some method to enrich the land-cover dataset.Image segmentation is the process of separating an image into homogeneous polygons that depict distinct regions on the ground.Using the segmentation output,four shape measures are calculated including,convexity,compactness,fractal dimension,and form.Image texture is indicated by variance in pixel DN values across space.Per-pixel imperviousness and tree canopy estimates that is based on empirical relationships between tree canopy density and Landsat data,established using regression tree techniques.In this text,We are expecting that could reflect its progress situation,enlighten to the local related work.