Crop Classification based Multi-Temporal Sentinel2 Data in the ShiYang River Basin

Abstract

<p>     Accurate and timely spatial distribution of crops is of great significance to agricultural security issues. They are not only important parameters for regional crop growth monitoring, yield prediction, model simulation, but also important basis for government decision-making to assess food security and predict comprehensive productivity of agricultural resources. The Shiyang River Basin is the inland river basin with the highest development intensity and the most prominent water resource conflicts in arid regions of Northwest China. The timely and accurate acquisition of crop distribution information in the watershed can provide data support for alleviating the water resource conflicts in the Shiyang River Basin. However, there is no publically available in-season crop distribution information with high spatial resolution in the watershed. To solve this problem and generate accurate ,in-season crop-type map with high spatial resolution, this study uses multi-temporal sentinel2 data to generate spectrum, texture, vegetation index, red edge multi-feature sets while exploiting random forest classification algorithm(RF).the study designed Experiments to evaluate what kind of feature is most effctitve and further dig which spectral information is most important for the training RF model to map crop types in Shiyang River Basin. The experimental results show that the spectral information outperforms other features in the highest classification accuracy, followed by vegetation index information and red edge index information, and texture information has no effect on crop classification in the watershed. Temporal information has a significant impact on crop classification. the classification accuracy increases with the progression of time, and the accuracy reaches a plateau (94%)around DOY 218. The highest accuracy for single-date classification is 87%, and for multi-temporal is 94.8%. The classification performance of SWIR1 band of Sentinel2 is better than the visible and near-infrared bands commonly used in crop classification at present, and crop classification based on the Sentinel2 Visible-NIR-SWIR1 three-band combination mode can achieve the effect of full-band classification. The Visible -NIR-SWIR1 data of sentinel2 before DOY218 was used to generate the 2019 crop distribution map of Shiyang River Basin, with an overall accuracy of 93.5%.</p>