Development of building type estimation model for micro land use analysis

Abstract

To consider the development of urban areas and future planning, it is important to analyze the micro land use transition of architectural units. The purpose of this study is to develop a machine learning model to estimate building type from building name and obtain micro land use transition data. Housing maps have a high potential for the analysis of transition, but such data does not include a building type. We use geo-coordinated phone book and Google Place API, which include both building names and their types, as training data. In data preprocessing, the building names are divided into words by MeCab, and each word is converted into a vector by Word2Vec. Using TF-IDF, each word was merged into a unified vector, weighted according to the frequency of it. Then, we used Random Forest, which is able to get accuracy and the results can be viewed as a decision tree. The target area is the city center, where mixed building types are observed, and the types were classified into five types. Increasing the number of urban areas in the training data improves the accuracy, but the rate of improvement decreased after about 10 urban areas. The final accuracy of more than 80% was achieved stably.