Using Convolution Neural Network to Extract Environmental Feature – taken Real Estate Appraisal as a precedent

Keywords: Convolution Neural Network, Real Estate Appraisal, Spatial Heterogeneity
Abstract Type: Paper Abstract

Abstract

In recent years, many statistical methods have been applied to the prediction of housing prices, including machine learning techniques such as linear regression, random forest, etc. Among those methodologies, deep learning has achieved outstanding performance. With excellent prediction accuracy, many researchers had applied the artificial neural network to the prediction of housing prices. However, most of them use tabular data only focusing on the feature of the house transaction record itself, ignoring the fact that the house price could be influenced by the surrounding environment, which eventually brings the problem of spatial heterogeneity to the table. Therefore, this study aims to combine the convolution neural network with multilayer perception, proposing a mixed input neural network structure that also learns the distribution of the surrounding variable. Hoping to improve the generalization of the model by downsizing the influence of spatial heterogeneity. This study chooses Taoyuan city, Taoyuan district as the main research area, research period ranging from 2015/01 to 2018/12. We use Actual Price Registration data released by the ministry of interior, R.O.C as the real estate feature data source. In regard to the environment feature, we use satellite images of Sentinel-2 and land use inventory data provided by the ministry of interior, R.O.C.