提出一种基于傅里叶谐波分析的改进算法, 引入异常值检测算法, 检测拟合过程中的异常值, 增加数据拟合的真实性; 迭代前动态估算出待处理序列点的峰值个数(即频数), 解决整个区域预设单一频数的不合理性; 引入拟合影响因子, 自动控制迭代终止条件, 避免传统方法中人为设置阈值导致的不确定性。利用2003年华北平原MODIS_EVI时间序列图像验证表明, 较之HANTS算法, 改进算法能够有效修正噪声污染像元值, 修正后的EVI时序曲线更能反映地物内在的物候变化规律, 并能够更好地保真原始曲线上的特征(点), 如作物EVI最大值、最小值出现的时间和大小关系。

This paper, based on Fourier-transform-based Harmonic Analysis Algorithm, proposes an improved algorithm to overcome the drawback of artificially setting key parameters and to reconstruct high-quality time-series data. Firstly, outlier detection algorithm, instead of threshold setting method, is used to find unreasonable data points before curve fitting. Then, regarding the inherent phenological regulation for each land cover, Numbers of Frequency (NOF) is calculated pixel by pixel by polynomial fitting, which is more reasonable than setting a unique global NOF for the whole scene which contains complex land cover types. Fitting-effect Index, instead of manually setting one fitting tolerance, is employed to decide automatically when to terminate the iteration. The improved method is validated by the MODIS_EVI time series of Huabei plain of 2003. The widely used Harmonic Analysis of Time Series (HANTS) is chosen as a comparison. The result shows that both of the two methods can reflect the phenological regulations of land covers, the reconstructed EVI temporal profile of single-season crop-land (e.g. cotton) takes on one peak pattern, and those of double-season cropland and inland water take on double peak pattern and low-steady curve. But the improved method performs better in tracking the change tendency of the original curve. More-over, the peak time and peak value of croplands are mostly consistent with the original curve, which will be useful for VI-based crop yield prediction.