Research On Urban Land Use Function Classification Model Based On Multi-source Big Data

As the space required for urban construction and satisfying the functioning of cities,urban land belongs to important basic data in urban calculation.The classification of urban land use functions from a micro perspective can assist in the Site Planning and Regulatory Plan in urban planning.However,the current urban land function classification still faces many difficulties.The concept definition and research angle of urban land are different.There are few data sources suitable for the classification of urban land function.The difference between cities poses a challenge to the generalization performance of the algorithm.This paper aims to classify the basic units of the city at the architectural level,given the open source geographic data.The features are extracted from the built environment and spatial morphology,and the graph networks are established on account of geographic proximity.The land use functions are identified by a supervised model based on gradient boosting trees and a semi-supervised model based on graph attention model in case of insufficient labeled sample.First of all,this paper proposes a building-based feature engineering method by extracting construction environment and space morphological.The features are built from three perspectives: the built environment,the architecture’s own morphological and the surrounding space.Statistical processing on the POI data is performed by TF-IDF statistical method.The embeddedness and learning of morphological data in relation to the building itself is realized by the self-encoder with the Auto-Encoder structure.Cross combination method is used for feature expansion based on the characteristics of the building’s surrounding morphology.Secondly,this paper proposes a supervised classification model based on Gradient Boosting Trees.Ensemble learning method is generated by Gradient Boosting based on CART regression tree is used to classify urban land function.Through experiments,it is found that the effect of the supervised learning model proposed in this paper is better than the related models in recent years,especially in terms of the number of classifications and the classification accuracy of most categories.Finally,this paper proposes a semi-supervised classification model based on Graph Attention Network.By using the geographical proximity between the basic units of the city to imply the building-based connection relationships,graph attention networks can classify the land use of buildings.Experiments show that our model can achieve the similar classification accuracy with fewer labeled samples compared with supervised learning model.