Abstract: Normalized Difference Vegetation Index (NDVI) time series exactly derived from remote sensing data are usually contaminated by different types of noise. Only by adopting the local or global filtering algorithm can not eliminate all kinds of noise from NDVI time series and maintain their local or global characters at the same time. A hybrid filtering algorithm, called Hyb-F, is proposed based on Savitzky-Golay filter, Asymmetric Gaussian algorithm and the Grubbs test method. Then, Hyb-F, which is a four step method, is used to reconstruct the Landsat NDVI time series for eight land cover types within Xianghai National Natural Reserve. First, the obvious outliers are detected and removed by setting thresholds for Normalized Difference Snow Index (NDSI), standard deviation and boundaries of NDVI. Second, via combing Grubbs test and Savitzky-Golay filter, the local outliers are removed. Third, via combing Grubbs test and Assymmetric Gaussian algorithm, the global outliers are removed. Finally, the final NDVI time series are built by using the Savitzky-Golay filter. The shape of fitting curves, statistical indicators and regional application effect of Hyb-F are analyzed. The results of the proposed Hyb-F method are compared with that of only Savitzky-Golay filter or Asymmetric Gaussian algorithm. For the land cover types with vegetation, Hyb-F performs well to the samples of shrub land, grassland, forest and cropland. Results of the mentioned types show large correlation with the clean reference series. The correlation coefficients are from 0.8488 to 0.9215 and the root mean square errors are from 0.0429 to 0.1057. The results of the first two land cover types achieve the highest accuracy among all the methods discussed. For the types without vegetation, such as saline, construction land and water area, the Hyb-F method can smooth the curves well. Compared with other methods, Fyb-F takes all local and global characters and dynamic trends into consideration. It can effectively recognize different types of outliers. Besides, it can weaken the loss of peak values and model the annual growing pattern of land cover types with vegetation. When applied to the whole study region, the indicator of efficiency of reconstruction is 0.59, demonstrating the reliability of Fyb-F. Moreover, the work flow the proposed method is clear and it is easy for application.

Key words: remote sensing, hybrid filtering algorithm, Savitzky-Golay filter, asymmetric Gaussian, Grubbs test, NDVI time series