不同端元模型下湿地植被覆盖度的提取方法——以北京市野鸭湖湿地自然保护区为例

植被覆盖度作为反映湿地植物生长状况的重要生态学参数,在评估和检测湿地生态环境方面起着关键的作用.以华北内陆典型的淡水湿地——北京市野鸭湖湿地自然保护区为研究对象,中等分辨率的Landsat TM影像为数据源,基于线性光谱混合模型(LSMM)对研究区的植被覆盖度进行了估算.针对湿地植被类型丰富、土地利用类型多样化的特点,利用归一化植被指数(NDVI)在反映植物生长状况、覆盖程度以及区分地表覆盖类型方面的优势,通过对原始Landsat TM影像增加NDVI数据维对影像进行维度扩展,克服了传统研究中通常从Landsat TM影像上提取3-4种端元的局限,经最小噪声分离变换(MNF变换)、纯像元指数(PPI)计算以及人机交互端元选取等一系列运算,构建以陆生植物、水生植物、高反射率地物、低反射率地物、裸露土壤为组分的五端元模型来反映研究区的地物组成；同时,以原始Landsat TM影像为基础,构建植物、高反射率地物、低反射率地物、裸露土壤为组分的四端元模型.针对两种端元模型,采用全约束下的LSMM算法进行混合像元分解以获取研究区的植被覆盖度,其次辅以研究区的纯水体信息对其进行优化.精度检验采用相同时期的高分辨率WorldView-2多光谱影像来进行.研究表明:虽然四端元模型与五端元模型对植被覆盖度的估算结果在空间上具有基本一致的分布趋势,但是前者的估算结果在数值上要普遍低于后者,在研究区的水体及其附近,四端元模型难以体现水生植物的植被覆盖信息；另外,五端元模型的估算结果与检验数据的相关系数R达到0.9023,均方根误差(RMSE)为0.0939,明显优于四端元模型的R=0.8671和RMSE=0.1711.这反映了通过对影像进行维度扩展的方法来改进端元提取的数量是可行的,而由此构建的五端元模型可以更充分的反映研究区地物之间的光谱差异,从而获得更好的估算精度.

Research on estimating wetland vegetation abundance based on spectral mixture analysis with different endmember model: a case study in Wild Duck Lake wetland, Beijing

Vegetation abundance is a critical indicator for measuring the status of vegetation. It is also important for revealing the rules of spatial change and evaluating the eco-environment of wetland. In this paper, vegetation abundance is estimated through a Linear Spectral Mixture Model (LSMM) using Landsat Thematic Mapper (TM) data within the Wild Duck Lake Wetland, a typical freshwater wetland in North China. Considering the biodiversity of wetland, it is necessary to select enough endmembers to characterize the spectral variability of features. However, a maximum of four endmembers is usually used in spectral mixture analysis when using Landsat TM data due to the high correlation coefficients among the three visible bands. Thus, the Normalized Difference Vegetation Index (NDVI), a widely used vegetation index, is included along with the six reflective bands of Landsat TM image, which extended the dimensionality of the TM data and allowed the use of five endmembers. After performed the Minimum Noise Fraction (MNF) transformation and Pixel Purity Index (PPI) algorithm, five endmembers are selected including Terrestrial plants, Aquatic plants, High albedo, Low albedo and Bare soil using n-Dimensional Visualizer Tool of ENVI software. At the same time, a four-endmember model which consists Plants, High albedo, Low albedo and Bare soil was established using the six reflective bands of Landsat TM data. For both four- and five-endmember model, we used a fully constrained LSMM algorithm to obtain the vegetation abundance of the study area. We modified our result using the pure water information of Wild Duck Lake Wetland, which was obtained by using NDVI and Normalized Difference Water Index (NDWI). The accuracy of vegetation abundance is validated by WorldView-2 multispectral image with a 1.8 meter spatial resolution through randomly selected 60 samples. The result showed that although the vegetation abundance extracted by four- and five-endmember model share almost the same spatial distribution, their values vary much from each other. Generally, the five-endmember model results a much higher value than the four-endmember model and it can reflect the vegetation abundance of aquatic plants better than the other. After the analysis of relationship between predicted and inspection values of both four- and five-endmember models, we found that the correlation coefficients are 0.8671 and 0.9023, respectively. Additionally, the Root Mean Square Error (RMSE) clearly showed that the five-endmember model can achieve a much better result than the other model: the former is 0.0939 and the latter is 0.1771. Our results suggest that it is feasible to improve the number of endmembers by extending the dimension of remote sensing image. By using NDVI as an additional data dimension along with the reflective bands of Landsat TM data, a five-endmember model can estimate vegetation abundance of Wild Duck Lake wetland better due to the inclusion of the fifth endmember. Moreover, four endmembers may be inadequate to spectrally characterize the heterogeneous landscapes of wetland. The use of five endmembers may further improve the representation of the spectral variability of land-cover types within the study area and thus potentially improve the accuracy for estimating vegetation abundance of wetlands.