Research On The Method Of River Network Automatic Generalization Based On Machine Learning

River network thematic cartographic generalization is the process of reducing large-scale map to small-scale map.Its purpose is to identify the elements of map as clearly as possible under different scales.The key step is selection and simplification.There are many factors that need to be considered in the selection of river network,and the relationship between these factors and the selection of river network is fuzzy and complex,which is difficult to express with a certain model.In recent years,the application of machine learning method in the cartographic generalization of river network thematic elements has gradually increased,but these studies are often limited to one machine learning method,and the feasibility of this method is verified under specific data sets.This paper selects three machine learning methods: BP neural network(BPNN),support vector machine(SVM)and decision tree to analyze the application effect of different methods in river network automatic generalization,and the simplification method of river network line feature is studied.The main contents are as follows:(1)Analysis of river network characteristics and determination of selection indicators.By summing up the expert knowledge in river network generalization and referring to the relevant research results,the spatial features,geometric features and quality features of rivers are summarized.The spatial features include river basin,river grade,main and tributary information of rivers,etc.;geometric features include river length,river flow direction,angle between rivers,etc.;quality features include connectivity between rivers and reservoirs,lakes,seasonality and navigability of rivers,etc.Combined with the research and exploration of this paper,river length,river grade,river spacing,seasonality,connectivity,catchment area,number of tributaries at the next grade,and total number of tributaries are taken as the indicators of river network selection.(2)The structural expression method of river network data is studied.In geospatial database,river network is stored in the form of river sections.Cartographic generalization is to operate river entities,so it is necessary to organize river sections,establish river entities.On this basis,Horton rule is used to code river,which can both reflect the hierarchical relationship of rivers and describe the depth of rivers.(3)The river network selection model based on three machine learning methods is researched and constructed.The construction methods of three models are studied,including the design of BPNN parameters such as the number of layers,neurons and training function,the selection of kernel function of SVM and the determination of relevant parameters,the generation and pruning of decision tree.According to the above method,real river network data are used to select and optimize the parameters of each model.(4)The advantages,disadvantages and applicability of the three machine learning methods are compared and analyzed.Using the vector river network data published by the National Geomatics Center of China,taking the generalization of 1:250000 scale map to 1:1000000 scale map as an example,three methods are used to carry out the experiment respectively,and the generalization results are compared with the actual 1:1000000 scale map.The performance of three machine learning methods is evaluated from the aspects of selection accuracy,data processing,difficulty of parameter adjustment,sensitivity of sample number,interpretability of the model.(5)The method of river network line feature simplification based on bending is studied.In view of the shortcomings of traditional methods in maintaining the geographical features of line feature,referring to the relevant research results in recent years,this paper studies the simplification of line feature with bending as the basic unit,and puts forward a bending recognition method based on judgment of inner angles sum of simple polygons.This method takes the theorem of inner angles sum of polygons as the criteria,calculating the sum of the internal angle of the polygon enclosed by the curve and the baseline to determine whether the division of the curve is complete;for the bending division results,a bending selection method considering bending area and bending baseline length is designed.Through the above research,the results show that the eight river network selection indicators used in this paper can reflect the features of river network comprehensively,and the structural expression of river network can meet the needs of river network generalization.In addition,the experiment and comparative analysis of three machine learning methods show that SVM model is more suitable for river network selection with small samples;in terms of data processing and parameter adjustment difficulty,BPNN and SVM model are more rigorous,and the implementation of decision tree model is relatively simple;in terms of model interpretation,the decision tree method has a unique advantage because of its visualized tree structure and the characteristics of the rules that can be derived.Finally,the bending recognition method proposed in this paper accords with people's cognition visually,and the simplification results keep the geographical characteristics of line feature well.