南美洲湿地面积广且类型多样,但湿地制图相关研究匮乏,通过遥感手段可为南美洲全域湿地制图提供科学技术支撑。本研究依托Google Earth Engine平台面向南美洲湿地提出一种多源多特征集成的湿地制图方法。研究选取南美洲典型湿地地区为研究区,首先利用已有土地覆盖数据集提出一种有效的湿地样本采集流程以保证样本质量,其次结合哨兵1号、哨兵2号和SRTM数据构建多源特征集合,并基于随机森林的递归特征消除算法(RF_RFE)进行特征优选,构建不同特征组合方案对比多源特征对湿地分类结果的影响,最后采用随机森林算法对研究区湿地进行分类提取。研究结果表明,样本采集方案可有效提高样本质量,多源特征集合能够提升湿地分类精度,特征优选能够减少特征冗余并提升分类精度。研究区分类总体精度为85.62%,Kappa系数为0.8333。

Objective：Wetlands play an important role in maintaining ecological balance, conserving water resources, recharging groundwater, and controlling soil erosion. They are often called the "kidneys of the earth" because they help purify water by filtering out pollutants and sediments. South America has a vast wetlands covering area, and a variety of wetlands types.While most of these wetlands were conserved in a relatively good condition until a few decades ago, pressures brought about by land use and climate change have threaten their integrity in recent years.Nevertheless, there are still no complete and uniform wetland maps providing information on the location, distribution, size, and changing status of wetlands in South America. Remote sensing has been an effective tool for characterizing, mapping, and monitoring the complexity and dynamics of large areas of wetlands. Despite the fact that fine wetland mapping may be done by combining data from many sources, the following two issues still persist. On the one hand, due to the complicated temporal dynamics and spectral heterogeneity of wetlands, large scale wetlands mapping remains a challenging task. On the other hand, supervised classification is a widely used technique for multi-category wetlands mapping, but choosing training samples is time-consuming and labor-intensive, there is no finer and more precise wetland extent as a reference. Method：In the study, we selected four typical wetlands study area in South America.First, an effective wetland sample collection process was proposed by using the existing land cover dataset to ensure the sample quality. Second, a multi-source feature set was constructed by combining Sentinel-1, Sentinel-2 and SRTM data.Then, feature selection is carried out based on the random forest recursive feature elimination method(RF_RFE), and we constructed a multi-feature combination scheme to compare the influence of multi-source features on wetlands classification. Finally,the random forest algorithm is used to classify wetlands in study area. Result：The research results show that sample collection process facilitates sample collection and improves sample quality. The combination of Sentinel-1 and Sentinel-2 data can improve the accuracy of land cover mapping, and terrain features help greatly improve the overall classification accuracy and the accuracy of various types of objects. For wetland categories, the addition of multi-source data features can improve the separability of wetland categories.The feature selection based on RF_RFE can reduce the feature redundancy and improve the classification accuracy. The feature optimization results show that SAR polarization features and derived texture features can be used as an effective supplement to optical features, but the dominant features account for less.The overall classification accuracy of the study area was 85.62%, and the Kappa coefficient was 0.8333. Conclusion：The study proposed an effective classification process and sample collection scheme to large-scale wetlands in South America.This study integrated Sentinel-1 synthetic aperture radar data, Sentinel-2 optical data, and terrain data to explore their importance to the extraction of different wetlands at large scale, and verified the feasibility of feature selection based on random forest recursive feature elimination method. The research results show that sample collection process facilitates sample collection and improves sample quality. The combination of Sentinel-1 and Sentinel-2 data can improve the accuracy of land cover mapping, and terrain features help greatly improve the overall classification accuracy and the accuracy of various types of objects. For wetland categories, the addition of multi-source data features can improve the separability of wetland categories.The feature selection based on RF_RFE can reduce the feature redundancy and improve the classification accuracy.