青藏高原小嵩草高寒草甸返青期遥感识别方法筛选

Identifying an optimal method for estimating green-up date of Kobresia pygmaea alpine meadow in Qinghai-Tibetan Plateau

小嵩草高寒草甸是青藏高原的主要植被类型,研究其返青期识别方法对于模拟及预测青藏高原植被物候变化具有重要意义。常用的植被返青期遥感识别方法主要是先对遥感植被指数原始时序数据进行拟合去噪声再求取返青期,各种方法对研究区域、研究经验、参数设置、函数初值设置等有很强的依赖性。为避免返青期识别方法在曲线拟合时对参数初值的依赖性和陷入局部最优解,本文引入了模拟退火算法对双高斯和双逻辑斯蒂函数进行参数优化,并分别对基于以上两种函数及多项式拟合的植被指数时序曲线进行对比,从而选出最佳拟合方法,最后采用最大斜率阈值法、动态阈值法和曲率法识别返青期。利用青藏高原小嵩草高寒草甸34个样本点的返青期地面观测数据及相应的8 km分辨率的NOAA归一化差值植被指数(NDVI)时序数据对以上各种组合的返青期遥感识别方案进行了测试,并选取了153个遥感实验点求取了近30年(1982年—2011年)青藏高原小嵩草高寒草甸的返青期,结果表明:采用双高斯函数拟合的NDVI曲线与原始NDVI时序数据最为接近,在此基础上采用最大斜率阈值法识别的小嵩草高寒草甸返青期及其变化趋势与地面物候观测结果最为一致;同时发现近30年青藏高原小嵩草高寒草甸的平均返青期主要集中在每年的第120—140天,并且呈逐年提前趋势,30年来提前了7天。

The Kobresia pygmaea alpine meadow is a main vegetation type in the Qinghai-Tibetan Plateau. An accurate detection of the green-up dates for K. pygrnaea is important to simulate and predict vegetation phenology shifts under the influence of climate change in the Qinghai-Tibetan Plateau. Green-up date estimation methods from remote sensing data generally include two processes： reconstruction of high-quality vegetation index time-series data through noise removal and calculation of green-up dates from the reconstructed vegetation index time series. The reconstruction methods for vegetation index time-series data can be divid- ed into two categories： filter fitting and curve fitting methods. The green-up date relrieval methods include the threshold, maximum slope, curvature, and moving average methods. The green-up date identification method is a combination of the reconstruction methods for vegetation index time-series data and the retrieval methods for green-up dates under different study conditions. The accuracy of the green-up date identification methods is usually affected by many factors, such as specific geographic location, prior experience, parameterization, and initial parameters. In this study, we adopted a simulated annealing algorithm to optimize the reconstruction process and thus avoid the problems of low efficiency and local optimum caused by traditional optimal methods. We first used the double-Ganssian, double-Logistic, and polynomial functions to reconstruct the Normalized Difference Vegetation Index （NDVI） time series. Atter evaluations with visual inspections and root mean square error, we identified the most feasible reconstruction method. We then used the maximum slope, threshold, curvature, and dynamic threshold methods to derive the green- up dates from the best reconslructed NDVI time series. The performance of these three methods for green-up date identification were tested using the green-up data from 34 ground observation samples and their corresponding National Oceanic and Atmospheric Administration NDVI time-series data at 8-kilometer resolution. We selected additional 153 samples, which were evenly distributed in the K. pygmaea alpine meadow in the Qinghai-Tibetan Plateau, to test the identified optimal green-up estima- tion method and to investigate the changes in green-up dates in the study area. The reconstructed NDVI time series with the doub- le-Gaussian function had the smallest deviation from the original NDVI time series, and the noises can be reduced effectively through the double-Gaussian fitting process. Therefore, the aforementioned method was the most suitable for describing the intra- annual growth cycle of the K. pygmaea alpine meadow. The reconstructed NDVI time series with the double-Gaussian function method indicated that the green-up dates identified with the maximum slope threshold method agreed with the observed ground phenology data. The correlation coefficients between the identified green-up dates and the observed dates were 0.823 （P〈0.001） and 0.646 （P 〈0.01） at the Haiyan and Gande stations, respectively. The average green-up dates for the K. pygrnaea alpine meadow in the Qinghal-Tibetan Plateau were mainly located between DOY （Day of Year） 120 （i.e., 30 April） and 140 （i.e., 20 May）. The green-up onset date advanced by an average of 7 days from 1982 to 2011 in the study region.