Improving the accuracy of spring phenology detection by optimally smoothing satellite vegetation index time series based on local cloud frequency

Abstract

Vegetation phenology can be extracted from vegetation index (VI) time series of satellite data. The maximum value composite (MVC) procedure and smoothing filters have been conventionally used as standard methods to exclude noises in the VI time series before extracting the vegetation phenology [e.g., National Aeronautics and Space Administration (NASA) VNP22Q2 and United States Geological Survey (USGS) MCD12Q2 phenology products]. However, it is unclear how to optimize the MVC and smoothing filters to produce the most accurate phenology metrics given that cloud frequency varies spatially. This study designed two simulation experiments, namely (1) using only the MVC and (2) using the MVC and smoothing filters together to smooth the enhanced vegetation index (EVI) time series for detecting spring phenology, i.e., start of season (SOS), over the northern hemisphere (north of 30°N) on a 5° × 5° grid cell basis by the inflection point and relative threshold algorithms. The results revealed that (1) the inappropriate selection of MVC periods (e.g., too short or too long) affected the accuracy of the SOS extracted by both phenology detection algorithms; (2) a filtering process with optimal parameters can reduce the effects of the MVC period on SOS extraction to a considerable extent, i.e., 65% and 61% for iterative Savitzky–Golay (SG) and penalized cubic splines (SP) filters, respectively; (3) optimal parameters for both the MVC and smoothing filters showed significant spatial heterogeneity; and (4) validation with ground PhenoCam data indicated that optimal parameters of the MVC and smoothing filters can produce more accurate results than official vegetation phenology products that use uniform parameters. Specifically, the R2 values of the NASA product and the USGS product were 0.58 and 0.67, which were increased to 0.70 and 0.81, respectively, by the optimal smoothing process. Optimal parameters of the MVC and smoothing filters provided by this study in each 5° × 5° sub-region may help future studies to improve the accuracy of phenology detection from satellite VI time series.