Font Size: a A A

Study On Ensemble Learning And Multi-parameters System For OSM Road Network Selection

Posted on:2019-10-17Degree:MasterType:Thesis
Country:ChinaCandidate:L H YuanFull Text:PDF
GTID:2392330545477708Subject:Cartography and Geographic Information System
Abstract/Summary:PDF Full Text Request
Automatic cartographic generalization is the most fundamental and difficult problem in cartography,whose core work is automatic road network selection.In the study of road network selection,the wisdom and experience of countless cartographers have been embodied.However,it is difficult to use quantitative empirical formula to form a unified and clear rule for the manual selection method,which consumes a lot of manpower and time.At the same time,with the rapid development of urban traffic electronic maps,the demand for automatic selection of road networks by multi-level scaling of information is increasing.An intelligent road network automatic selection method is desiderated.In the new era of big data booming,artificial intelligence and VGI data provide new ideas for solving this problem.Road network selection is essentially a classification problem.As a kind of machine learning classifier,ensemble learning can be applied not only to automatic road selection,but also to promote weak classifiers to strong classifiers under integrated thinking,which can achieve better classification effect.Therefore,to elimate the limitation of comprehensiveness,accuracy and objectivity in the current research of automatic road network selection,this paper proposed a method of automatic road network selection with multi-parameter system and ensemble learning.As the study area,Manhattan with representative roads was chosen.The main content included:(1)Construction of a multi-parameter knowledge system suitable for integrated learning.According to the characteristics of the road,combining the non-standard map of OSM with the standard map of USGS,the attributes,geometry,topology,and other parameters of the road "stroke" were extracted.A multi-parameter system combined with the classification tags was formed,which describes the road in terms of semantics,shape and interrelation from a multidimensional perspective in a more comprehensive way.Through feature selection and Gini Impurity assessment,the parameters were selected for knowledge acquisition by the integrated learning model,while those contributed less to road selection were removed.(2)Design and optimization of CART classifiers for road sample subsets.The individual CART classifier was trained and constructed from a random sampling set of road data.K-fold cross-validation was used to fully learn the samples,thus avoiding over-fitting and improving the score of model.With repeated pre-experiments,the number of CART was adjusted and determined,then the optimization of the model was achieved.(3)Integration and assessment of classifiers for automatic road network selection.The individual classifiers were integrated by voting method.Random forest algorithm and AdaBoost algorithm,which applied to the problem of automatic road network selection were implemented to predict result of each road.In the same road parameter system,those models were compared with the results from commonly used decision trees,GBDT,SVM,multi-layer perceptron,and naive Bayes.Based on the number of road strokes,the number and area of meshes and the density index of roads,evaluation and analysis were made.The results showed that approach in this study made use of multi-parameters and ensemble learning to achieve automatic road network selection.It could satisfy the requirements of automatic road network selection,which included the accuracy and shape preservation and founded a better way to deal with the problem of multi-scale roadmap zooming from an intelligent point of view.Based on pre-experiments,the multi-parameter system was refined by feature selection,which made the accuracy increased by 1.11 percentage points than before.After further comparison experiments on the same road multi-parameter system,result from random forest in ensemble learning,which accuracy rised by 2.83%was better than AdaBoost's.Compared with the algorithms of non-integrated thoughts,the advantages of random forest were also relatively obvious.At the same time,through the automatic selection from the OSM road network data to the map on a standard scale,the application value of OSM in the field of cartography has been enhanced,the advantage of rich VGI data information has been utilized,and the dependence on artificial cartography knowledge has been eliminated.
Keywords/Search Tags:Automatic Cartographic Generalization, Road Network Selection, Artificial Intelligence, Ensemble Learning, OSM Data
PDF Full Text Request
Related items