本文以冬小麦为测量目标，在分析研究区典型地物物候特征的基础上，提出一套基于不同种植结构下的多时相中高分辨率遥感时序数据冬小麦识别方法，并对简单种植结构和复杂种植结构下的冬小麦播种面积识别进行了试验，得到如下主要结论：1.简单种植结构条件下：1）SVM法、NDVI差值法和PCVA法这三种基于冬小麦物候关键期的变化检测方法提取出的冬小麦播种面积结果的面积总量和空间分布十分相似。其中，PCVA法测量结果总体精度为95.6%，Kappa系数为0.902，远高于SVM分类后直接比较方法（总体精度91.2%，Kappa系数0.794），同时，相比于SVM法，PCVA法漏分误差降低，生产者精度得到了很大的提高。NDVI差值法识别结果总体精度为81.6%，Kappa系数为0.613，远远低于前两种方法识别精度。2）NDVI差值频度直方图曲线很不稳定，且频度值在阈值取值范围内很高，这使得其变化阈值的选择必然对象元的最终提取数量有很大的影响；而在PCVA法中，变化向量强度的频度直方图出现两极化现象，这使得频度值在阈值取值范围内被压低摊平，阈值敏感性降低，这减小了阈值判断导致的误差，使结果更为客观，一定程度上解决了阈值难以设定的问题。3）NDVI差值法和SVM分类后比较法受影像空间异质性的影响较大，提取结果的稳定性小于PCVA法。4）SVM分类、同类别隶属度合并和变化向量分析这三个部分的结合增强了PCVA法对植被光谱差异的敏感性，使其能够监测不同季相上植被的长势变化，进而提高了农作物播种面积遥感监测的精度，是简单种植结构条件下较为优秀的作物播种面积识别方法。2.复杂种植结构条件下：1）PCVA具有较高的识别精度和结果稳定度，但由于无法获取物候特征较大的关键期时相组合，该方法只能以越冬作物为识别目标，而无法对冬小麦进行有效识别。2）多时相影像集方法对研究区影像的分类精度很低，分类总体精度为64.4%，Kappa系数为0.5657。以越冬作物为目标的识别实验表明，PCVA法明显优于多时相影像集法，其中，多时相影像集方法的总体精度为88.4%，远远低于PCVA法（96.8%），Kappa系数相差0.15（多时相影像集法为0.763，PCVA法为0.929）。主成分分析法的总体精度为70.1%，略高于多时相影像集方法（64.4%），Kappa系数为0.6343，也高于多时相影像集方法（0.5657），这表明将在分类前对多时相影像集中的主要特征进行提取和压缩能够提高目标作物的识别精度，但该方法对非植被地物中的裸地、河流水体和废水池的识别能力有所降低。3）作为冬小麦的主要同期作物，油菜对冬小麦的识别产生了很大的干扰。单纯利用作物光谱时间曲线形状特征的光谱相似度匹配指标进行作物类型识别的方法与主成分分析方法得到的地物识别精度相近，而综合了光谱时间曲线形状特征和数值特征的光谱相似度匹配指标对目标作物的识别精度有了很大程度的提高。其中冬小麦的漏分误差和错分误差在该方法中接近10%，精度的提高幅度十分明显。4）与基于像元的识别方法相比，对象和像元尺度相结合的多尺度识别方法能够得到更好的总量精度和位置精度，但该方法识别结果的椒盐现象仍然存在。5）使用历史耕地地块数据进行两阶段作物地块分割，能够在最大限度的保持稳定的耕地地块边界的同时，获得比直接分割更均一的对象，同时，二级分割过程中按对象局部调整分割参数的方法能够合理考虑各地块内部的实际情况，使“欠分割”和“过分割”现象基本消除，分割结果比全局参数结果更为合理。6）复杂种植结构条件下，冬小麦作物的总体识别精度较低，这是由于研究区地理位置与气候条件、遥感数据源本身的质量和光谱可辨识程度、同期作物油菜与冬小麦物候极为相似以及同期作物油菜光谱变异程度很大这些因素共同决定的，分类器和分类方法的改进对识别结果精度的提高并不十分理想。

In this paper, aiming at the winter wheat area extraction, a winter wheat area extraction system for different planting structures was proposed based on the phenological characteristics analysis of the main typical surface features. Through the experiments on the winter wheat area extraction on the simple planting structure and on the complex planting structure, we get the main conclusions as follows:1. Under the simple planting structure1) Three methods including SVM, NDVI difference and PCVA were adopted on the simple planting structure area as represent of the change detection methods based on the key temporal, and the results of these three methods are similar on the area and the spatial distribution. SVM and PCVA can both get a satisfied result, but the PCVA method has a better performance, the overall accuracy (95.6%) and the Kappa coefficient (0.902) of PCVA method is much better than the SVM method in which the overall accuracy is 91.2% and the kappa is 0.794. And the producer accuracy of PCVA method is better than the SVM method too. The NDVI difference method has a much worse performance than the former two methods, the overall accuracy of NDVI Difference method is 81.6% and Kappa is 0.613.2) The frequency curve of NDVI difference method is unstable, and the frequency is very high in the threshold for judging change or not, it leads to the result that the influence of threshold value on the change pixels number will very high. Compared with this, the threshold value selection in PCVA method is more objective. With the polarization phenomenon of the frequency histogram in this method, it decreased the partial frequency of change threshold value and led to a lower threshold sensitivity, thus the determination of threshold value was more objective.3) The NDVI difference method and the SVM method are greatly influenced by image spatial heterogeneity, so the stability of their results is worse than the PCVA method.4) The combining use of SVM model, the dichotomy model and the CVA model made PCVA method more sensitive to spectral changes, and improved the detection of crop growth change under different growing stages, the results shows that the accuracy on winter wheat planted area is improved on PCVA method, we can get the conclusion that the PCVA method is a excellent method under the simple planting structure.2. Under the complex planting structure1) The PCVA method has a high accuracy and a stable result, but because it is unable to obtain the key phenological multi-temporal combination, the PCVA method can only be used in the extraction of overwinter crops, but not the winter wheat.2) The multi-temporal images set method has a low classification accuracy in this area, the overall accuracy is only 64.4% and Kappa coefficient is only 0.565. Aiming at overwinter crops, the overall accuracy of the multi-temporal images set method is 88.4%, which is lower than the PCVA method (whose overall accuracy is 96.8%), the Kappa has a difference of 0.15 between the multi-temporal images set method (0.763) and the PCVA method (0.929), it shows that the multi-temporal images set method has a worse performance on overwinter crop extraction than the PCVA method.Compared with the results of the multi-temporal images set method and the PCA method, the result shows that the process of the extraction and the compression of the main features of multi-temporal images can improve the accuracy of the classification. The overall accuracy of the PCA method is 70.1%, and Kappa is 0.6343, which are better than the multi-temporal images set method (64.4% and 0.5657). But the accuracy of bareland, river and wasterwater pond of the PCA method are lower than that of the multi-temporal images set method.3) AS the main crop in the same period of winter wheat, the rape crop makes a great interference on winter wheat extraction. The results of the spectral matching method which based on the shape similarity of the NDVI curves are similar with the result of the PCA method, when combined the shape and the value features, the spectral matching method gets a much better result. The misclassification error and the overclassification error are both near 10%, it shows a great improvement of accuracy.4) The method that on multi scale can get a better overall accuracy and a better positional accuracy than the method on pixel scale, but the “salt” and “hole” phenomenon is still exist.5) The two-step segmentation based on the historical farmland parcel data can get a more homogeneous object with stable boundaries than traditional segmentation. In the process of the second segmentation, the adaptive selection of segmentation features can take a full consideration of difference parcels, it makes the “less-segmentation” phenomenon and the “over-segmentation” phenomenon are almost eliminated in the result, so the result is more reasonable than that on the overall segmentation features.6) Under the complex planting structure, due to the geographical position and climate condition of the study area, the quality and spectral resolution of remote sensing data, the similar phenophase of winter wheat and rape crop, and the spectrum difference of the rape crop pixels, the accuracy of winter wheat extraction is low. The improvement of classifier and classification method has litter influence to the accuracy improvement.