JIA-2019-11

2632 ZHANG Xi-wang et al. Journal of Integrative Agriculture 2019, 18(11): 2628–2643 considerable noise, although the MOD13Q1 and MYD13Q1 datasets have been processed by the maximum value composite method. Therefore, it is necessary to perform filtering and smoothing. Savitzky-Golay is a weighted average filter - based sliding window, as shown in eq. (1). N IC I n n j j i j ' i ∑ −= + = (1) where I is the original NDVI, and its value range is [–1,1]; I’ is the fitted value of NDVI; C j is a coefficient when filtering the j th NDVI; and N =2 n +1 is the number of data points contained in the sliding window. Based on the land use data and TM image, the NDVI time series data for each land use type is extracted from the MODIS products. Then they are processed using the Savitzky-Golay filtering method to eliminate jagged irregular fluctuations of the curve, as shown in Fig. 3. This figure illustrates the NDVI time series of the main land cover types in the study area during the winter wheat growing seasons from 2003–2010, which are clearly different from each other. In particular, winter wheat has two distinctive peaks which distinguish it from the other land cover types. There are also slight differences between years that reflect the phenological variations (Pan et al. 2012). 3.2. Abundance assessment fromMODIS time series Analysis of temporal change characteristics In order to more clearly observe the differences within a season, the MODIS NDVI curves of each land cover type are extracted and they are plotted together for the 2009–2010 growing season in Fig. 4. For winter wheat, Fig. 4 shows five distinctive features which differ from the other land cover types. (1) The rapid increase in NDVI values from point A to point B reflects the growth of winter wheat from sowing to emergence and tillering, whereas the NDVI values of the other land cover types decline. (2) After point B, winter wheat begins to winter and its NDVI value decreases, resulting in a distinct peak at point B. The others do not show such a decline. (3) During the period between points C and D, winter wheat begins to “wake up” from its winter phenology and it passes through the greening, jointing and booting stages. The NDVI value increases rapidly, whereas the others have not yet begun to grow and their NDVI values do not significantly increase. (4) Winter wheat begins to mature at point E, after which its leaves begin to yellow and the NDVI begins to decrease, thus forming a distinct peak at E, whereas the others continue to grow. (5) After point E, the NDVI of winter wheat declines rapidly, whereas the others continue to increase. The distinctive features captured in the time series can usually be used to identify crop types (Tao et al. 2017). Analogous to the principle of atmospheric windows in earth surface reflectance data (Liang 2004), the peaks at points B and E can be considered as reflection peaks, while the points A, C and F are absorption troughs (Pan et al. 2012). Like the vegetation index’s highlighting of vegetation information (Liu and Huete 1995), the slope of the time series curve indicates the acreage fraction within a pixel. For example, in the period from point C to point D, if the acreage proportion of winter wheat in a MODIS pixel is closer to 100%, then its NDVI time series curve should be more similar to pure winter wheat. AMODIS pixel with winter wheat mixed with other types will result in a lower slope, TM image Land use 25-m TM Membership for winter wheat Resampling Pure winter wheat samples MODIS NDVI products Time series curve Slope of key periods WW acreage fraction samples Winter wheat abundance Regression analysis Cultivated land Bayesian rules AM model Identified winter wheat Accuracy analysis Fig. 2 Flowchart for identifying winter wheat in this study. MODIS, moderate resolution imaging spectroradiometer; NDVI, normalized difference vegetation index; TM, Thematic Mapper; WW, winter wheat; AM, Abundance-Membership model.