Abstract: Accurately obtaining the spatial distribution of crops is a prerequisite for crop growth monitoring and yield prediction. At present, automatic acquisition of crop distribution is still a challenge because the processing of remote sensing image is time-consuming and the collection of sufficient training samples is laborious. How to obtain sufficient training samples in an efficient and economical way has become one of the key factors in crop mapping. By combining the phenological characteristics of winter crops with Sentinel-2 time series images, this paper proposes an automated sample generation strategy for winter crop mapping. Firstly, the normalized difference vegetation index (NDVI) time series curves are used to identify winter crops. Secondly, the differences between unknown samples and standard green chlorophyll vegetation index (GCVI) time series curves are calculated through the curve similarity measurement method, so as to assign the correct label to the unknown samples. Finally, the Random Forest model is trained with the obtained samples, which realizes the extraction of winter crops in the study area. In the final accuracy evaluation result, the overall accuracy (OA) is 98.46% and Kappa is 0.973, which shows the effectiveness of this method to realize the quick automatic winter crop mapping.

Key words: Sentinel-2, phenology, garlic, winter wheat, automated, time series