Extraction Of Wetland Information In Typical Areas Of The Source Of The Yellow River Based On Multi-source Data Feature Optimization

As one of the three major ecological systems on earth,wetland has various and unique ecological environment functions,such as stabilizing climate,conserving water source,degrading pollution and protecting species diversity.The wetland at the source of the Yellow River is an important part of the wetlands on the Qinghai-Tibet Plateau and an important source of water for the Yellow River system.It has important strategic value for maintaining the ecological balance of the source region of the Yellow River and for the high-quality development of the Yellow River basin.In recent decades,with the dual impact of climate change and human activities,the wetland area in the source region of the Yellow River has been shrinking and expanding continuously.It is of great significance to obtain the accurate wetland area for the ecological environment protection in the source region of the Yellow River.Madoi County,the source area of the Yellow River,was chosen as the research area in our study,and the wetland information in the area was extracted based on Object-oriented combined with machine learning methods with the help of multispectral,radar data,and auxiliary topographic data.Firstly,the Sentinel-1A and Sentinel-2A data were used to analyze the applicability of multi-source data in the extraction of Yellow River source wetlands,and use the RF-OOB algorithm to optimize the feature space based on the preliminary screening of the Relief F algorithm.Secondly,three machine learning algorithms,KNN,SVM and RF,were used to classify wetlands on the basis of the optimal feature optimization set.The optimal classifier was then selected considering the accuracy evaluated by the confusion matrix.Finally,three temporal Landsat8 OLI data combined with Sentinel-1A data were used to extract the temporal changes of wetland information in the study area,and the following conclusions were drawn:(1)When multi-spectral data sources were used alone for classification,the overall accuracy and Kappa coefficient were 83.48%and 0.8027 respectively,which could easily lead to the misclassification of wetland categories.With the addition of terrain,radar and texture data,the overall classification accuracy was improved by 4.17%,3.19%and 1.32%,respectively.The overall accuracy and Kappa coefficient of multi-source data were 92.16%and 0.9058,respectively,indicating that multi-source data has certain advantages in the extraction of wetlands from the source plateau of the Yellow River compared with a single remote sensing data source.(2)Relief F and RF algorithm were used to optimize the feature space,and the overall accuracy reached the highest 93.06%,indicating that feature optimization can screen out important feature information,reduce feature dimensions,and improve classification efficiency and accuracy.In addition,RF has the highest overall classification accuracy and Kappa coefficient,which are 93.06%and 0.9167,respectively,higher than KNN by 3.12%and 0.0371,and SVM by 2.15%and 0.0257.(3)In 2013,the total wetland area in the study area was 3056.55 km~2.In 2017,the total wetland area was reduced to 2756.24 km~2.In 2020,the total wetland area will be 3176.87 km~2.Among them,the largest change is the herbaceous swamp.From 2013 to 2017,the range of change was-23.82%,and the area decreased by 147.12 km~2.From 2017 to 2020,the range of change was 27.63%,and the area increased by 129.95 km~2.It is preliminarily considered that the main reasons for the change of wetland area are natural factors,including precipitation and temperature.