基于随机森林的绿洲典型湿地信息提取

Information Extraction of Wetlands in Typical Oasis Based on Random Forest Model

以塔里木河边缘的渭-库绿洲(渭干河-库车河绿洲)为研究区,采用在特征选择和分类提取方面具有明显优势的随机森林算法,对研究区内的湿地信息进行提取。基于多时相、光谱信息丰富的Landsat8 OLI数据生成包括光谱特征、植被和水体指数、盐分指数、纹理信息在内的4种特征变量;根据以上特征设计6种不同的提取方案,对绿洲内部的干旱区湿地信息进行提取并验证不同方案的提取精度,旨在选取最佳方案提高湿地信息提取的精度。结果表明:①多种特征变量的有效组合是提高湿地信息提取精度的关键,就不同特征对湿地信息提取的贡献度而言,光谱特征>植被和水体指数>纹理特征>盐分指数;②基于随机森林算法优选的特征变量提取速度最快,效果最佳,总体精度为90.09%,Kappa系数为0.882 5。提取方法在挖掘特征变量的同时,保证了湿地信息提取的准确性,提高了运行效率。湿地提取结果对当地绿洲制定科学的水肥管理措施及进行干旱状况评估具有一定的现实意义。

In this study,the Wetland in Werigan-Kuqia River Delta Oasis( WKRDO),as a typical wetland on the banks of the Tarim River,is considered as a study area. The method of Random Forests,which has obvious advantages in feature selection and classification,is chosen to extract wetland information from the study area. First of all,four different characteristic variables,including spectral features,vegetation indices,water indices,salinity indices and texture features,are generated based on landsat 8 OLI data with rich multi-temporal and spectral information,and then six different classification schemes are constructed based on the above characteristic information. At last,the random forest classifier is used to extract the wetland information of the WKRDO,and to verify the extraction accuracy of different results.The purpose is to select the best plan to improve the effect of wetland information extraction. The results show that:① The effective use of multiple feature variables is the key to improving the information wetland of extraction. For the contribution of different characteristics to the wetland information extraction,spectral features > vegetation indices and water indices > texture feature > salinity indices;② the preferred features based on the random forest algorithm are significant to extraction accuracy,with the overall accuracy up to 90.09%,and the Kappa coefficient is 0.882 5. The extracted types reflect the differences in soil properties. It shows that the random forest algorithm can effectively process the feature selection. The result has important practical significance for formulating scientific water and fertilizer management measures and evaluating drought situation in the local oasis.