Abstract:

There are two important steps when monitoring vegetation phenology by remote sensing. The first is separation of ‘vegetation data’ from ‘noise’ such as cloud cover in the MODIS satellite data so as to obtain the best possible estimate or “reconstruction” of the normalized difference vegetation index （NDVI） and its changes over the time period of interest. The second is the extraction of estimates of vegetation phenology parameters from the reconstructed NDVI data. The comparative merits or demerits of different available NVDI reconstruction and vegetation phenology extraction methods and their suitability in different regions are still little-studied. It is necessary to carry out a comparative analysis of different methods and to compare results with field data to identify the optimal remote sensing extraction methodology for a particular region. Here we report a comparison of four extraction methods to extract the start of season （SOS） date for grassland in Xinjiang from 2001 to 2019. The four methods compared were： the asymmetric Gaussian function fitting method （asymmetric gaussian， A-G）， the double logistic function fitting method （double logistic， D-L）， Savitzky-Golay filter method （S-G） and the land cover dynamic product （MCD12Q2）. Through comparison of the extraction results obtained by the four methods， the optimal model for extracting SOS was found， and the regional variation and change over the 19-year study period in the SOS in Xinjiang Province were evaluated. The main conclusions were as follows： 1） The A-G extraction method provided the best estimate of the SOS for Xinjiang grassland. The correlation （R=0.879） between A-G estimates of SOS and the SOS recorded at field-measured points was higher and the root mean square error （RMSE=16.395 d） was smaller. Spatially， the standard deviation for SOS measured by the A-G method was less than 30 d for 82.19% of the area evaluated， compared with 76.53% and 67.98% for D-L and S-G methods， respectively. The MCD12Q2 method had a smaller standard error but there were large areas of terrain where no estimate of SOS could be provided. 2） Over the 19 years for which data were evaluated， the SOS in Xinjiang occurred as early as the 60th and as late as the 140th day of the year， with obvious regional differences， SOS being earlier in the north and later in the south. In the north of Junggar Basin and Ili river valley， the SOS was earlier than the 80th day， while in the Altai Mountain， central Tianshan Mountain and Kunlun Mountain the SOS was later than the 140th day. The SOS of different grassland types had obvious differences. For example， SOS of alpine meadow and alpine grassland was the latest， while SOS of temperate deserts was the earliest. 3） From 2001 to 2019， SOS in Xinjiang showed interannual variation of approximately 15 days between early and late years， with a trend over time to later SOS in lowland meadow， temperate desert and alpine desert， and earlier SOS in temperate grassland， alpine grassland and alpine meadow.

Key words: phenology, grassland, MODIS, Xinjiang, spatial-temporal variation, remote sensing