基于多季相Sentinel-2影像的白洋淀湿地信息提取

白洋淀湿地是华北平原上重要的浅水湖泊湿地，对雄安新区绿色发展具有重要的生态价值。对白洋淀高度异质化的景观格局进行分类，能够为白洋淀湿地资源的遥感监测提供指导意义。针对湿地季节变化的特点，对白洋淀每个季节选取一期具有代表性的Sentinel-2影像，采用分类与回归树（CART）、支持向量机（SVM）、随机森林（RF）3种常用的机器学习分类器对15种季相组合实验方案进行分类，分析不同季相遥感影像及其组合对白洋淀湿地信息提取的优劣。结果表明：相较于使用单一季相影像分类，多季相影像的组合能够显著提高分类精度，春&夏季相组合能够得到最优的分类效果，相对单季影像总体分类精度提高了10.9%~25.5%，Kappa系数提高了0.09~0.29；SVM分类器的分类表现较为稳定，能够得到最高的平均分类精度，CART分类器在处理高维特征的能力不如随机森林和SVM；不同特征类型对湿地信息提取的贡献度从高到底依次是红边光谱特征、传统光谱特征、缨帽变换特征、主成分分析特征、纹理特征。实验成果能为湿地信息的遥感识别提供依据。

Information Extraction of Baiyangdian Wetland based on Multi-season Sentinel-2 Images

Baiyangdian is an important shallow lake wetland in the North China Plain， which has important ecological value for the green development of Xiong’an New Area. Wetland mapping of the highly heterogeneous landscape pattern of Baiyangdian can provide guidance for the remote sensing monitoring of Baiyangdian Lake wetland resources. In view of the seasonal changes of wetlands， a representative Sentinel-2 image is selected for each season of Baiyangdian in 2019. Three commonly used machine learning classifiers， including Classification and Regression Tree （CART）， Support Vector Machine （SVM） and Random Forest （RF）， were used to classify 15 classification scenario. The advantages and disadvantages of different seasonal remote sensing images and their combinations for extracting Baiyangdian wetland information were analyzed. The results showed that the combination of multi-seasonal images can significantly improve the classification accuracy. The combination of spring and summer images obtained the optimal classification accuracy. Compared with the single seasonal images， the overall accuracy was improved by 10.9%~25.5% and the kappa coefficient was improved by 0.09~0.29. The classification performance of the SVM classifier was relatively stable， and the highest classification accuracy can be obtained. The ability of CART classifier in processing high-dimensional features was not as good as that of random forest and SVM. The contribution of different features to the wetland information extraction was described as follows： red-edge spectral feature > traditional spectral feature > tasselled cap transformation feature > principal component analysis feature > texture feature. The research results can provide a basis for the remote sensing mapping of Baiyangdian wetland.