Abstract: Abstract: An accurate and rapid extraction can be highly required for the crop sown area and spatial distribution from the remote sensing images, particularly for the sustainable development of cultivated land and food security. However, winter wheat mapping using remote sensing depends mainly on optical images and complex classification at present. Besides, it is still unclear on the classification performance and time-transferring capability of existing classification with the small sample sets in the highly land-fragmentation areas. The fragmentation of cultivated land has always been the core of rural land regulation, where the land resources are wasted to reduce the cultivated land productivity in the soil fertility with the high production costs. The difficulty of crop mapping in finely fragmented areas is generally higher than that in large-scale farming areas. The applicability and stability are very important for the study of such areas. It is necessary to realize long-term large-scale crop mapping with a low dependence on the number of samples and high efficiency. Therefore, it is of practical significance to develop a new extraction with a low complexity suitable for small samples. Previous studies have shown that the accuracy of crop mapping using single-phase satellite imagery cannot fully meet the high requirement in recent years, especially in land fragmentation areas. In this study, the high-level fragmentation of cultivated land was selected as the study area in the Wancheng District, Nanyang City, China. Using the Google Earth Engine cloud computing and Sentinel-1 SAR and Sentinel-2 optical images, three advanced classifications were evaluated, including the time-weighted dynamic time warming (TWDTW), random forest (RF), and OTSU with distance measure (DSF), for the winter wheat mapping accuracy and time-transferring capability with the small sample sets in the study area. The results show that effective extraction was achieved in the sown area and spatial distribution of winter wheat in 2020, but there were some differences in the classification accuracies. The TWDTW presented the highest classification accuracy, with the Overall Accuracy (OA) and Kappa coefficients 0.923 and 0.843, respectively, followed by the RF (OA=0.906, Kappa=0.809) and DSF (OA=0.887, Kappa=0.767). The OTSU with the Euclidean Distance showed the lowest classification accuracy. When transferring to extract the winter wheat classification maps of 2021, the classification accuracy of each model decreased: The TWDTW and DSF showed better stability and classification accuracy than the RF. The TWDTW shared the highest accuracy with the OA and Kappa of 0.889 and 0.755, respectively. The classification accuracy of RF decreased significantly, and the OA and Kappa decreased by about 0.07 and 0.19, respectively, indicating the lower stability of the model. In general, the TWDTW presented low sensitivity to the training samples and spatial heterogeneity. As such, the high-precision continuous mapping was realized for the winter wheat in the agricultural areas with high spatial heterogeneity under the condition of limited samples. However, the RF was sensitive to the training samples and spatial heterogeneity. The condition of limited samples can cause low stability in the continuous winter wheat mapping in high spatial heterogeneity agricultural areas. This finding can provide important selection ideas and scientific support for continuous crop mapping with the small sample sets in the highly land-fragmentation areas.