软硬结合分类方法提取特定地物信息*——以冬小麦为例

THE SPECIFIC LAND COVER EXTRACTION USING THE SOFT AND HARD CLASSIFICATION* ——AN EXAMPLE OF WINTER WHEAT

关键词:  混合像元软分类硬分类ISODATA支撑向量回归

Key words:  mixed pixel  soft classification  hard classification  ISODATA  support vector regression

摘要:

［目的］针对影像上纯净、混合像元共存的现象，文章结合硬分类方法和软分类方法各自的优势，提出了目标地物信息的软硬结合的分类方法。［方法］该方法将遥感影像划分为典型目标地物像元、非目标地物像元和混合目标地物像元3个部分。典型的目标地物像元和非目标地物像元，采用硬分类方法（ISODATA）聚类确定类型； 混合目标地物像元采用非线性支撑向量回归混合像元分解模型，从目标地物端元光谱库和非目标地物端元光谱库中多次随机选择像元，进行目标地物不同丰度值的混合像元模拟，构建样本库进行支撑向量回归，提取出混合像元的目标地物丰度。该文以冬小麦为研究对象，选用2006年4月7日的TM影像，采用软硬结合的分类方法进行冬小麦识别。［结果］较传统的硬、软分类方法，软硬结合分类方法精度高，总体精度达到了902%； 而软分类方法为866%，硬分类方法为816%。［结论］软硬结合的分类方法克服了硬分类方法对混合像元信息提取受到光谱不确定影响，也克服了软分类方法受到光谱异质性干扰的问题。该分类方法简便、易操作，适合单目标特定地物的信息提取。

Abstract:

Land use/cover mapping is one of the most widely used fields of remote sensing technology. At present, the methods of recognizing specific features from single phase remote sensing images mainly include hard classification and soft classification. The accuracy of land use/cover classification is influenced by the spatial resolution and spectral characteristics of image. Wheat is one of the most important crops in China with the area of 1 /5 of the whole crop area all over the country. Therefore, acquiring the information of crop acreage, especially wheat acreage timely and accurately, is important in making a method for regional economic development to guide the adjustment of planting structure and improving agricultural management. Land use/cover mapping is one of the most widely used fields of remote sensing technology. At present, the methods of recognizing specific features from single phase remote sensing images mainly include Hard Classification method (HC) and Soft Classification method (SC). And there are also some shortcomings about these two methods. HC methods, such as maximum likelihood method or minimum distance method, classify each pixel as a ground object type, and the information results of each pixel are independent and irrelevant. Therefore, the hard classification is immune to the spectral fluctuation, which is not suitable for identifying mixed pixels. Soft classification methods, such as linear decomposition method, neural network method, supervised fuzzy classification method, support vector method, decision tree method, etc., are applied to classify mixed pixel, and the extraction accuracy is higher than that of hard classification method. However, due to the characteristics of this method and the influence of spectral fluctuation, the accuracy of soft classification method is lower than that of hard classification method in the extraction of terrain objects in the pure area. To take advantage of conventional HC and SC, the soft and hard classification method (SHC) was proposed to express the variation of crop distribution by overcoming the status of coexistence of pure pixels and mixed pixel in remote sensing image. According to this method, the spectral information of study area was divided into three parts, namely, typical aim pixels area, non typical aim object pixels area and mixed pixels area. In typical aim pixels area and non typical aim object pixels area, ISODATA method of hard classification method (HC) was applied. While in mixed pixels area, nonlinear Support Vector Regression (SVR) of Soft Classification method (SC) was applied. In the application process of nonlinear SVR, we simulated mixed pixels in different abundances from end member spectrum database of target object and non target object by random selection and form sample database. Then abundance of mixed pixels of target object was extracted by SVR. The method was applied in a sub region of Beijing. One Landsat 5 TM image that was utilized by SHC, HC and SC. The image was acquired on April 7, 2006. One QuickBird image of the same time period and area as Landsat 5 TM was utilized for extracting real value of winter wheat distribution for accuracy assessment. The results showed that the crop derived from SHC could get a higher accuracy that that from SC and HC. The overall accuracy of SHC, SC and HC were 90.2%，86.6% and 81.6%, respectively. SHC method is applied to overcome the effects of uncertainty to mixed pixels by HC method and the effects of spectral variability by SC method. Therefore, our method is more flexible and suitable in single crop mapping.