JIA-2019-11

2640 ZHANG Xi-wang et al. Journal of Integrative Agriculture 2019, 18(11): 2628–2643 methods. The reason for this difference is that winter wheat is identified in each matrix with 10×10 membership pixels in the proposed method. This aspect reduces the interference of environmental conditions, such as terrain differences, climate differences, etc., and contributes to the higher accuracy of the proposed method relative to the other methods. 5.2. Comparison with traditional methods Ideal temporal windows for spectral features-based methods For the spectral feature-based methods, the ability to discriminate objects changes over time due to the temporal variability of different vegetation types (Nitze et al. 2015). The ideal image ensures a maximum spectral difference between the identified target and the surrounding objects and produces good identification results. However, there can also be slight spectral differences within the identified target. These differences within the target are usually caused by factors such as terrain and microclimate (Van Niel and McVicar 2004). For example, there are some differences in winter wheat between mountainous and plain areas at the same time. In the non- ideal periods, the intraclass differences will approach the interclass differences in the spectral domain; this leads to a decrease in accuracy. Therefore, the temporal separability was reflected in the classification accuracies, where the optimal choice of image dates outperformed the worst image date by 13% (Nitze et al. 2015). To increase the effectiveness of the classification process, it is necessary to optimize the image acquisition timing and frequency. In this study, we analyzed the vegetation growth process from Fig. 4 and found the ideal temporal window. During the period CD, winter wheat grows rapidly, while the growth of other vegetation just begins from point D where the largest spectral difference occurs between winter wheat and other vegetation. Therefore, the ideal time for identifying winter wheat using spectral information is also near point D. Comparison with MLC and RFC MLC is one of the most popular classifiers and it is widely used in the supervised classification of remote sensing (Liu et al. 2011; Schachtner et al. 2014). It usually performs better than the other known parametric classifiers because it takes into account the variance-covariance within the class distributions (Otukei and Blaschke 2010). RFC is a classifier that uses multiple decision trees to train samples. By reducing multivariate collinearity, RFC solves the problem of overfitting to guarantee the high effectiveness (Salles et al. 2018). We applied these two classifiers based on the same image under the same conditions. For MLC, Aa is 82.78% and As is 84.60%, while Aa and As of RFC are 86.57 and 88.60%, respectively. In the confusion matrix, OA of MLC and RFC are 84.65% (Kappa is 0.569) and 88.60% (Kappa is 0.675), respectively. These results are obviously less accurate than the results obtained with the proposed method. MLC and RFC can provide better results with ideal medium-resolution images (Baumann et al. 2011; Schmidt et al. 2015). However, in this study Landsat images with high quality only appeared on May 2, which is not the optimal date. This may be one of the reasons for the low accuracy. Despite being sub-optimal, this image acquisition time is far from the worst-case scenario, and the influenced winter wheat pixels are still closer to pure winter wheat compared to the surrounding pixels in the local space. Compared to MLC and RFC, the proposed method compares membership within a particular membership matrix (10×10 pixels) rather than using a uniform standard for the whole study area. One advantage of the proposed method is that it is not entirely dependent on spectral characteristics and can identify winter wheat in a small defined spatial range. Therefore, it can avoid the fuzziness that results from using only spectral characteristics and the accuracy is higher. Comparison with the methods which use time series images In the estimation of crop acreage using a low- resolution time series of remote sensing data, the temporal profile similarity is usually used (Sun et al. 2012; Gumma et al. 2015) and the typical result is abundance (Verhoeye and de Wulf 2002; Busetto et al. 2008). However, these traditional methods cannot account for the crop spatial distribution within pixels (Kasetkasem et al. 2005; Potapov et al. 2008). In this study, the increased temporal change characteristics first determine the pixel number (F) that belongs to winter wheat in each matrix, then they determine the ownership of each pixel according to the membership value within a limited space. Therefore, we obtain a finer scale identification result compared to the temporal profile similarity based methods. 6. Conclusion Undoubtedly, higher spatial resolution remote sensing images will generally result in higher crop identification accuracy (Cruz-Ramírez et al. 2012; Pan et al. 2012). However, it is difficult to acquire high-quality images covering an entire study area of interest in a specified short time period because of cloud cover and other weather conditions (Ju and Roy 2008; Roy et al. 2008; Schmidt et al. 2015). Thus, the pursuit of the maximum resolution level in practical applications is rarely successful. Intraclass differences will approach interclass differences when the remote sensing images are not obtained during the ideal period, and the traditional spectral-based methods do not perform well. To solve this problem, this study proposes a solution for winter wheat identification. The abundance