JIA-2019-11

2639 ZHANG Xi-wang et al. Journal of Integrative Agriculture 2019, 18(11): 2628–2643 Factors impacting the abundance In previous research on the crop acreage proportion in a pixel using a vegetation index time series, the AB and EF intervals were usually used (Pan et al. 2012). The underlying principle is that the slopes of AB and EF change with the amount of other types of vegetation mixed into the pixel. However, the NDVI difference is large for AB and EF in this study area. This causes the pixel values to be very different when mixed at the same proportions as other types of vegetation. In other words, the slopes of AB and EF will be different for a pixel containing the same winter wheat acreage proportion because of the different mixtures of vegetation types. Therefore, it is most appropriate that winter wheat is relatively active, whereas other vegetation types are relatively stable, during the selected key period. In this study, R 2 reaches 0.8098 when the regression analysis is performed using only CD. If the combined effects of other periods are considered, such as introducing AB into the model, R 2 is 0.8146 when using a multiple regression analysis. R 2 is 0.8158 when using the combination of CD and EF. Thus, the most critical single period for estimating the pixel acreage proportion of winter wheat in this study area is CD. During this period, the growth of other vegetation types has not yet begun, whereas winter wheat grows very rapidly after greening, resulting in a steep slope in CD. This is similar to the vegetation index, which is constructed based on the steep slope between red and near-infrared bands in the spectral curve of vegetation. In addition, the time series curve will be significantly affected by the sowing date, seeding method and other cultivation measures. In other words, the slope of the time series curve may be different when using the different cultivation measures even if they result in the same winter wheat acreage proportion. Accounting for this requires a good understanding of the local agricultural production environment. If similar research is carried out at a large scale, phenology is also affected by climatic conditions. That is to say, the key feature points of the time series curve may be different in different areas. These factors need to be considered in generating slope data when using the proposed method. Factors impacting membership Accurate calculation of membership plays a crucial role in winter wheat identification using the proposed method. The most important step is the correct selection of training samples. The selection of samples is primarily based on spectral, spatial and other characteristics of the objects shown in the images (Li and Narayanan 2004; Mennis and Guo 2009). Sample selection is impacted by multiple factors, such as the interpreters’ familiarity with the study area and planting structure. It is also affected by local topography, climate and other conditions of the larger study area. Thus, it is necessary to select representative and sufficient samples for each class to accurately calculate the probability density function and ensure that the calculated result can indicate the relative proximity for each type of object. For the images used to calculate membership, the best choice is using completely cloud-free data when winter wheat is most clearly distinguished from other objects, such as the time corresponding to point D in this study. If the image acquisition time deviates from the optimal time for the proposed method, the overall identification acreage determined by the time series NDVI will not change, and the loss of accuracy will also be smaller compared to other Table 5 Confusion matrix of MLC, RFC and the method proposed in this study 1) Method 1) Identification 2) Reference UA (%) 3) Non-winter wheat Winter wheat MLC Non-winter wheat 78 38 67.24 Winter wheat 39 345 89.84 PA (%) 66.67 90.08 OA (%) 84.65 Kappa 0.569 RFC Non-winter wheat 85 25 77.27 Winter wheat 32 358 91.79 PA (%) 72.65 93.47 OA (%) 88.60 Kappa 0.675 Proposed method Non-winter wheat 92 18 83.64 Winter wheat 25 365 93.59 PA (%) 78.63 95.30 OA (%) 91.40 Kappa 0.755 1) MLC, maximum likelihood classification; RFC, Random Forest classification. 2) PA, producer’s accuracy; OA, overall accuracy. 3) UA, user’s accuracy.