Classification Of Dagu Estuary Wetland Based On Multi-Source Remote Sensing Data Feature Optimization

The Dagu River is the mother river of Qingdao,and its estuary wetland carries important economic and ecological functions,which bring benefits to people,but also suffers from destruction.Remote sensing images can be used to quickly understand the condition of wetlands and give timely policy protection.However,the optical remote sensing data commonly used are not enough to extract all kinds of wetland objects,and the fully polarized SAR data can extract the geo-type information effectively by using the different backshot characteristics of the ground objects,as a supplement to the spectral information.However,this will inevitably lead to an increase in the amount of data and thus increase the cost of time,so it is necessary to select a large number of remote sensing features before classification,in order to obtain a subset of the optimal features.The subset of optimal features can greatly improve the classification efficiency and simplify the feature extraction work before classification,which provides a reference for the research of the data source in wetland classification,and also provides an efficient feature selection algorithm.Based on different remote sensing data sources and their combination of features,this paper separately performs the classification work after feature extraction,feature selection and selection.Through the comparison between different feature selection methods and the results of different classifications,the best feature selection algorithm,the best combination of features and the best classification method are drawn.The main research contents and workss of this paper are as follows:(1)The filter feature selection algorithm based on multi-criteria parallel fusion has higher stability.In this paper,three single-criteried filtering algorithms are used: Relief F algorithm,distance correlation coefficient and Fisher Score,and compared with the parallel fusion algorithm of all three.By comparison,it is found that the weight calculated by the multicriteried filter feature selection algorithm comprehensively considers the advantages and disadvantages of the three,so that the weight value of the feature is more reasonable and the robustness is higher.(2)The serial combination of filtered and encapsulated feature selection algorithm based on multi-criteria parallel fusion has better feature selection results.In this paper,three characteristic selection algorithms are used for the characteristics of three sets of different data sources: the feature selection algorithm(MP-F)based on multi-criteria parallel fusion,the feature selection algorithm(SBFS-RF)based on heuristic search strategy,and the serial combination algorithm(MF-SBFS-RF)of the two.By comparing the scale,calculation cost and cross-verification accuracy of the subset of features,the MF-SBFS-RF algorithm can select the optimal subset of features.(3)The classification of object-based multi-source remote sensing data has the highest classification accuracy.In this paper,three classification methods are used: support vector machine(SVM),random forest(RF)and object-based(Ge OBIA)to classify the three optimal subsets of features selected by the MF-SBFS-RF algorithm.Experiments show that objectbased classification method is better than cell-based classification method,and the combination classification accuracy of feature combination between hyperspectral image and full polarized SAR image is the highest.