| Crop planting structure information is an important reference basis for crop growth monitoring and agricultural structure adjustment.Timely and accurate acquisition of regional crop spatial distribution and planting information is of great significance to the sustainable development of agriculture.In recent years,with the successful launch of the GF series and Sentinel series satellites,the temporal and spatial resolution of remote sensing data has been significantly improved,and the integration of multi-source remote sensing data used to obtain crop planting structure information has become one of the hotspots of research in this field.As the Heilonggang watershed is a seasonal fallow pilot area in Hebei Province,how to quickly and accurately obtain the spatial distribution information of crops in the basin with the help of remote sensing is a scientific problem that needs to be solved urgently.Based on this,this thesis selects Yongnian District of Heilonggang watershed as the experimental area,extracts vegetation index,backscatter coefficient and other features from sentinel-2,GF-1 and sentinel-1 remote sensing images,and constructs a variety of classification schemes,and deeply analyzes the influence of different feature combinations under machine learning classifier on the extraction accuracy of crop planting structure information in the experimental area.With the help of the Google Earth Engine（GEE）cloud platform,the best solution obtained in the experimental area is applied to the Heilonggang watershed to extract the winter wheat planting area,and the characteristics change of its time and space are analyzed.The main research work and results are as following:（1）Time series data of Normalized Difference Vegetation Index（NDVI）,Normalized Difference Red Edge Index（NDRE1）,Simple Ratio Index(SRre)and Red Edge Chlorophyll Index(CIred-edge)extracted from Sentinel-2 data and their combination are classified features.Random forest（RF）and support vector machine（SVM）classifiers are used to extract crop planting information in the experimental area.The results show that the combination of NDRE1 and CIred-edge time series data has the highest classification accuracy under the RF classifier,with an overall accuracy of80.86%.The NDVI data extracted by Sentinel-2 and GF-1 images are highly linearly correlated.After the conversion equation is converted,the data difference caused by different sensors can be effectively reduced.The high time resolution NDVI time series data constructed by the integration of the two data are used in crops.A higher classification accuracy is achieved in the classification.The overall accuracy under the RF classifier is 91.16%,which is 5.04%and 6.13%higher than the NDVI time series data of Sentinel-2 and GF-1 single sensors under the same classifier.（2）Using Sentienl-1 data to extract VH and VV polarization time series data and texture features and construct a variety of classification schemes,the results show that the multi-phase VH and VV dual polarization time series data has the highest classification accuracy under the RF classifier,with an overall accuracy of 80.86%.Combining the Sentinel-2 vegetation index time series data with the Sentinel-1 optimal classification scheme,the classification accuracy is significantly improved compared with the Sentinel-1 data alone,and the overall accuracy reaches 92.22%under both the RF and SVM classifiers.（3）Through the GEE cloud platform,apply the optimal classification scheme and classification algorithm obtained in the experimental area to the Heilonggang watershed to extract winter wheat planting area information.The results showed that the extraction accuracy of winter wheat from 2017 to 2019 reached more than 90%,and the planting area showed a downward trend year by year. |
