Study On Object Oriented High-Accuracy Information Extraction Of Caofeidian Seagrass Bed By Remote Sensing

Seagrass beds are among the most productive and biodiverse marine ecosystems.Seagrass beds are in a state of rapid degradation due to global climate change and human activities,and human activities are the main cause of their degradation.Therefore,it is imperative to understand the ecological status of seagrass beds and to establish effective conservation and management measures for them.The research on seagrass beds in China started in 1980 s,mainly focusing on the morphology,classification and phylogeny of seagrasses,but less on the distribution and dynamic monitoring of seagrass beds.At present,seagrass beds in China are mainly distributed in the South China Sea and the Yellow and Bohai Seas.The current research on seagrass beds is still lacking realistic area-wide monitoring data support.In this paper,we take Caofeidian seagrass beds in the Yellow and Bohai Seas as the research object,select the satellite remote sensing images with wide monitoring area and short monitoring period as the data base,adopt the object-oriented information extraction method,and determine a set of seagrass beds extraction method with high accuracy by optimizing the remote sensing image segmentation scale and feature parameters to realize the fine identification of seagrass beds and provide decision support for the management and protection of Caofeidian seagrass beds.The main research contents and conclusions of this paper are as follows..(1)Image segmentation scale preference.In this paper,the suitable segmentation range of image data is initially determined as 80-140 by the maximum area method,and to reduce the segmentation error,three segmentation scales are selected within the suitable segmentation range,namely,the beginning,middle and end,and the shape factor is determined as 0.1 to 0.3 and the tightness factor as 0.5 to 0.7.Parameter 2(ESP2)scale evaluation algorithm is combined with the ESP2 scale evaluation algorithm for group experiments,and the more suitable combinations of shape factor and tightness factor are selected by comparison and discrimination to effectively reduce the errors caused by the traditional trial-and-error method and improve the ease of use and execution efficiency of the method(2)Object-oriented classification extraction combination and parameter filtering based on multidimensional feature parameters.In this paper,spectral parameters,texture parameters,shape parameters and custom parameters are selected,and five sets of parameter combination schemes are designed:(1)spectral,texture,shape,custom;(2)spectral,texture,shape;(3)spectral,shape,custom;(4)spectral,texture,custom;(5)shape,texture,custom.The K nearest neighbor classification(KNN),categorical regression tree(CART),and random forest(RF)classifiers are combined respectively to obtain the importance of parameter types in different classifiers.The results show that the importance of parameters in KNN classification algorithm is in the order of spectral parameters > custom parameters > texture parameters >shape parameters;in CART classification algorithm,the importance of parameters is in the order of shape parameters > custom parameters > spectral parameters > texture parameters;in RF classification algorithm,the importance of parameters is in the order of custom parameters > spectral parameters > texture parameters > shape parameters.In order to efficiently remove the redundant features in the initial parameter space,this paper uses Relief F algorithm and CFS(correlation-based feature selection)algorithm for parameter screening,respectively.(1)After the Relief F algorithm ranks the data with weights,the correlation heat matrix is calculated to remove the parameters with correlation threshold greater than 0.8 and small weights,and then the classification accuracy is calculated by combining three classifiers in turn in steps of 2 to obtain the optimal number of parameters for each classifier.(2)The best classification parameters are obtained based on the CFS algorithm using the best-first search method.(3)Object-oriented seagrass bed extraction methods are constructed.(1)Six object-oriented extraction methods are constructed by two different parameter selection methods of CFS and Relief F combined with three machine learning algorithms of KNN,CART,and RF.By comparing the overall accuracy and Kappa coefficients of different methods,it can be seen that the accuracy of both parameter screening methods is improved,and only the accuracy of Relief F parameter screening method in KNN algorithm is smaller than the accuracy of spectral,texture,and custom combination,because KNN algorithm is less affected by the number of parameters when there is no shape parameter.The overall accuracy of CFS algorithm is higher than that of Relief F algorithm,among which the best performance is the combination of CFS in RF classification with an overall accuracy of93.75%.The overall accuracy and Kappa coefficient were the highest in RF-CFS algorithm when conducting the overall classification extraction of the study area,but this paper is mainly oriented to seagrass bed classification extraction,and the user accuracy of 70.69% for the extraction of tertiary seagrass using RF-Relief F algorithm was slightly higher than that of RF-CFS algorithm,so RF-Relief F classification algorithm was chosen to extract seagrass beds.(2)Object-oriented seagrass bed affiliation model construction.The variables obtained by two different parameter selection methods,CFS and Relief F,are used as the basis for constructing the affiliation degree fuzzy classification model and comparing with machine learning algorithm for analysis.(4)Object-oriented seagrass bed classification extraction results and analysis.Comparing the RF-Relief F method with the affiliation fuzzy classification method,the accuracy of the RF-Relief F method is 93.43% greater than that of the affiliation fuzzy classification CFS method 87.18%.The total area of seagrass bed extracted by the RF-Relief F method is 40.19 square kilometers,and the area value is between the area range of the measured concentrated distribution area and the peripheral distribution area of the Caofeidian seagrass bed in 2018,indicating the scattered distribution of tertiary seagrass in the Caofeidian seagrass bed.