基于Landsat TM的热带精细地物信息提取的模型与方法——以海南岛为例

应用遥感技术进行精细地物信息提取是研究生态系统结构、过程和功能的重要手段之一。由于热带地区生态系统复杂,为精细地物信息提取带来很大的不确定性,极易产生"同物异谱"、"同谱异物"的现象。研究以地处热带地区的海南岛精细地物遥感信息提取为例,在综合分析典型地物光谱特征、空间分布、斑块形状等基础上,构建和优化了水陆指数WLI(Water and Land differing Index)、乔灌草指数GSI(Grass and Shrub differing Index)、旱地-沙地指数SSI(Field and Sand differing Index),并结合新型通用植被指数VIUPD(Vegetation Index of the Universal Pattern Decomposition Method)及DEM(Digital Elevation Model)等多源数据,提出基于决策树的面向对象遥感信息提取方法。该方法首先确定要提取的对象,明确对象类别与对象隶属关系,然后逐层逐项的提取天然林、橡胶林、浆纸林等地物信息。结果表明,综合提取的精度达88%,相比传统的监督分类方法精度(66%)提高22个百分点,精度明显提高。

Models and methods for information extraction of complex ground objects based on LandSat TM images of Hainan Island, China

The service functions of different ecosystems vary widely and include water conservation, soil and water conservation, and the maintenance of biological diversity. Feature extraction from images obtained from remote sensing technology is an important method used to study the structures, processes and functions of ecosystems. Tropical evergreen ecosystems are very complex because different surface features can share the same spectra, or similar surface features can produce different spectra, making information extraction difficult. Current methods used for information extraction have some problems including the lack of application of spatial information and multi-scale image segmentation, the lack of an effective model to extract homogeneous plaque, and the lack of multi-source data support. To address these problems we carried out this study using Landsat TM data. Information on complex ground features on Hainan Island, China, was extracted after comprehensive analysis of spectral characteristics, spatial distribution, patch shape and other characteristics of typical ground objects. We developed an information extraction method based on the decision tree model, and a set of remote sensing information extraction models. These extraction models included the land and water index WLI (Water and Land differing Index), the bushes and grass index GSI (Grass and Shrub differing Index), the dry field and sand index SSI (Field and Sand differing Index), a new vegetation index VIUPD (Vegetation Index based on the Universal Pattern Decomposition method), and the DEM (Digital Elevation Model). According to the classes and interrelationship of target objects, first we established flow charts and processes for sophisticated object-based information extraction based on the decision tree. At the same time, we identified and extracted the information on surface features, taking prior knowledge of the spatial distribution and texture characteristics of ground features into account, and utilized multi-source data (DEM, planning maps and other auxiliary data). Depending on the node order, first we separated areas of water and land. We then separated areas with buildings without vegetation coverage from areas with no buildings, and finally, using the VIUPD method, the DEM and the GSI, nine different surface features including natural forest, rubber trees, pulp and paper forest, paddy field, dry land, orchards, and sand, were classified. We evaluated the precision of the classification by comparing results with high resolution remote sensing images obtained by Quick Bird and SPOT, and 300 GPS positioned points confirmed by local technical staff. The accuracy of our method was 88%, compared with traditional supervised and non-supervised classification methods that have a maximum accuracy of 66%. The accuracy was therefore significantly improved. Based on the results of this study we concluded that the WLI, the GSI, the SSI, and the VIUPD method we developed performed well in the information extraction of complex features on Hainan Island. The decision tree-based object-oriented information extraction model combined with the above models was applicable and effective in the classification of complex surface features in a tropical evergreen ecosystem when the information extraction model was tailored for tropical areas. After validation, the accuracy of feature extraction was greatly improved compared with traditional classification methods. The thresholds and parameters of the model can be adjusted for different ecosystems; therefore, this method is a valuable and widely applicable new tool for information extraction.