摘要
利用遥感技术对南瓮河保护区进行湿地信息提取,为寒区湿地保护和科学管理提供依据。本研究选取2021年多季相Sentinel-1雷达影像及多季相Sentinel-2影像,用集成学习XGBoost(Extreme Gradient Boosting,XGBoost)算法对多季相Sentinel-2特征变量重要性进行排序,剔除影响分类精度的变量。基于特征优选的XGBoost的机器学习算法,结合多季相雷达后向散射系数、极化分解特征及地形湿度指数(Topographic Wetness Index,TWI)对南瓮河湿地进行分类。结果表明:基于特征优选的组合中,加入TWI、雷达后向散射系数和极化分解特征分类精度最高,达到了96.03%,Kappa系数为0.95;与依次加入TWI和雷达后向散射系数特征分类精度相比,提高了3.80%和1.78%;对应的Kappa系数提高了0.40和0.17。多季相极化分解特征可提高湿地分类精度的有效性。
Remote sensing technology was used to extract wetland information from Nanweng River Nature Reserve,in order to provide a basis for wetland protection and scientific management in cold regions.The multi-seasonal Sentinel-1 radar data and multi-seasonal Sentinel-2 data in 2021 were used to rank the importance of multi-seasonal Sentinel-2 feature variables to eliminate the variables that affect the classification accuracy using the ensemble learning XGBoost(Extreme Gradient Boosting,XGBoost)algorithm.Based on the feature-optimized XGBoost machine learning algorithm,combined with multi-seasonal radar back-scattering coefficient,polarized decomposition characteristics and Topographic Wetness Index(TWI),the Nanweng River wetlands were classified.The results showed that the classification accuracy of terrain humidity index,radar back-scattering coefficient and polarized decomposition characteristics was the highest among the optimized combinations,reaching 96.03%and Kappa coefficient was 0.95.Compared with adding TWI and radar back-scattering coefficient feature in turn,the classification accuracy was improved by 3.80%and 1.78%respectively.The corresponding Kappa coefficient increased by 0.40 and 0.17.The multi-seasonal polarized decomposition features can be used to improve the effectiveness of wetland classification accuracy.
作者
李佳芪
那晓东
LI Jia-qi;NA Xiao-dong(Heilongjiang Province Key Laboratory of Geographical Environment Monitoring and Spatial Information Service in Cold Regions,Harbin Normal University,Harbin 150025,heilongjiang,China)
出处
《湿地科学与管理》
2022年第6期16-20,共5页
Wetland Science & Management
基金
黑龙江省自然科学基金优秀青年科学基金项目(YQ2020D005)。
