MODIS NDVI时间序列在三江平原湿地植被信息提取中的应用

Application of MODIS NDVI Time Series to Extracting Wetland Vegetation Information in the Sanjiang Plain

以三江平原为研究区,利用多时相的中分辨率成像光谱仪(Moderate Resolution Imaging Spectroradiometer,MODIS)影像数据,采用一种基于归一化植被指数(Normalized Difference Vegetation Index,NDVI)时间序列的监督分类方法获取了研究区湿地植被的分布数据。监督分类以NDVI时间序列的波形所反映出的植被物候特征作为分类器,将离散的傅立叶变换应用于NDVI时间序列以减少高频噪声对分类的影响,并运用傅立叶变换后波形幅度和相位的相似性来确定像素的归属类别。根据研究区植被的物候特征的差异,区分出7种地表(沼泽、沼泽化草甸、滩地、水田、旱地、灌木和林地)的植被类型,得到三江平原2005年湿地植被的分布数据。该方法的总体分类精度达到79.67%,Kappa系数为0.752 5。研究表明,基于MODIS多时相NDVI数据,采用基于傅立叶组分的相似度分类方法可以客观、经济、快速的提取湿地植被分布数据。

The Sanjiang Plain has ever been the largest freshwater wetland in China. It has become an important commodity grain base in China after more than 40 years exploitation. At the same time, its wetland area has decreased sharply. Destruction of wetlands and some environmental problems are attracting more and more atten- tion. To conserve and manage wetland resources, it is important to investigate and monitor wetlands and their adjacent uplands. The principal advantage of remote sensing data compared to traditional observations in the field is the possibility they offer to gather synoptic information at regular time intervals over large areas. Repeated observations from satellite -borne sensors can be used to identify wetlands and separate them from other land cover types. A variety of remote - sensing techniques could be used to identify, investigate and monitor wetlands. On the basis of multi - temporal MODIS data, a kind of supervised classification method to extract wetlands＇ vegetation of the Sanjiang Plain was used in this paper. The MODIS NDVI time series data were got from 16 - day MODIS NDVI data after processing by Harmonic Analysis of Time Series （HANTS） . The curve shape of NDVI time series, which is diagnostic for certain vegetation phenology, was the primary classifier. A Discrete Fourier Filter was applied to NDVI time series data in order to minimize the influence of high - frequency noise on class assign-ments. A similarity measure based directly on the components of the Discrete Fourier Transform that introduced by Evans J P was used to determine a pixels class membership. Based on the difference between vegetation phe- nology, 7 kinds of vegetation had been classified from marsh, marshy meadow, bottomland, paddy field, farmland, bush and woodland. The spatial distribution of natural wetland （ marsh, marshy meadow, open water） and man -made wetland （paddy field） were extracted. Results showed that total classification accuracy is 79, 67% and Kappa coefficient is 0. 752 5. The analysis of terrain showed the correlation to distribution of wetland vegeta- tion and geography. Results indicated that this Fourier component similarity measure produced an objective, computationally inexpensive and rapid method of wetlands classification. Results also showed that the accuracies for vegetation of forest, farmland and paddy field are 89, 36%, 83.33% and 74.19% respectively. The accuracies for that of marsh and marshy meadow are 70% and 61,53% respectively. The main reason for the low classification accuracies for vegetation of marsh and marshy meadow was the high similarity of phenology characteristics between marsh and marshy meadow, Therefore future research should pay attention to the phenophases when the differences of NDVI values between marsh and marshy meadow are significant. Assistant data including water depth and terrain would be used to enhance the accuracy for vegetation of marsh and marshy meadow.