地球信息科学学报 ›› 2019, Vol. 21 ›› Issue (2): 168-177.doi: 10.12082/dqxxkx.2019.180508

• 地球信息科学理论与方法 • 上一篇    下一篇

卷积神经网络和随机森林的城市房价微观尺度制图方法

姚尧1,2(), 任书良1, 王君毅1, 关庆锋1,*()   

  1. 1. 中国地质大学(武汉)信息工程学院,武汉 430074
    2. 阿里巴巴集团,杭州 311121
  • 收稿日期:2018-10-09 修回日期:2018-11-27 出版日期:2019-02-20 发布日期:2019-01-30
  • 通讯作者: 关庆锋 E-mail:yaoy@cug.edu.cn;guanqf@cug.edu.cn
  • 作者简介:

    作者简介:姚 尧(1987-),男,广东梅州人,副教授,研究方向为空间大数据和智慧城市。E-mail: yaoy@cug.edu.cn

  • 基金资助:
    国家重点研发计划项目(2017YFB0503804);国家自然科学基金项目(41671408);国家自然科学基金青年基金项目(41801306);湖北省自然科学基金杰出青年项目(2017CFA041)

Mapping the Fine-Scale Housing Price Distribution by Integrating a Convolutional Neural Network and Random Forest

Yao YAO1,2(), Shuliang REN1, Junyi WANG1, Qingfeng GUAN1,*()   

  1. 1. School of Information Engineering, China University of Geosciences, Wuhan 430074, China
    2. Alibaba Group, Hangzhou 311121, China
  • Received:2018-10-09 Revised:2018-11-27 Online:2019-02-20 Published:2019-01-30
  • Contact: Qingfeng GUAN E-mail:yaoy@cug.edu.cn;guanqf@cug.edu.cn
  • Supported by:
    National Key Research and Development Program of China, No.2017YFB0503804;National Natural Science Foundation of China, No.41671408, 41801306;Natural Science Fund of Hubei Province, No.2017CFA041

摘要:

随着中国城市化进程的加快,城市人口的大规模集聚带来了住房紧张的问题,房价政策制定的时效性与正确性也时刻吸引着社会的关注,因此在微观尺度下对房价进行精细化制图变得愈发重要。由于数据可获取性和现有模型精度的限制,目前已有研究均较少涉及微观尺度。本研究通过将房价数据和遥感影像相融合,构建了一种基于卷积神经网络(CNN)和随机森林(RF)的遥感影像挖掘模型,以实现在不考虑其他数据的情况下,精确、合理地进行房价的微观尺度制图。本文以武汉市作为研究区,在仅有房价数据和遥感影像的情况下,利用本文所构建的模型成功得到武汉市中心城区5 m精度的精细房价图。此外,还利用其他数据源以及挖掘技术与本文所构模型进行了对比分析。结果显示,本文所构建的模型获得了最高的房价模拟拟合优度(R2=0.805),相比传统方法中的最高拟合优度(R2=0.653)其精度提升了23.28%,其制图结果可为政府部门规划决策及武汉市经济分布研究提供基础支撑。

关键词: 房价, 深度学习, 微观尺度, 卷积神经网络, 随机森林, 武汉

Abstract:

China's rapid urbanization has caused a large number of migrants to move to the city, which has also led to housing shortages. Rapid access to fine-scale house price distribution data plays a very important role in urban housing management, government decision-making, and urban economic model analysis. The availability of data and limitations of existing models make only a few studies involving the mapping of house price distribution at the microscale. By combining house price data with remote sensing images, this study builds a remote sensing image features mining model based on Convolutional Neural Network (CNN) and Random Forest (RF). The proposed CNN-based model in this paper can be applied for accurate and reasonable microscopic mapping of house prices without introducing auxiliary geospatial variables. Only using the house prices data and remote sensing images, we successfully carry out the house prices mapping with the precision of 5 meters in the downtown area of Wuhan city. By comparison with the results generated by the other three traditional mining techniques (including A: using spatial datasets extracted from auxiliary geographic dataset only, B: using original features extracted from high-resolution remote sensing images only, C: using original features extracted from high-resolution remote sensing images and auxiliary geographic dataset), the results show that the proposed CNN-based model has the highest house price simulation accuracy (R2=0.805), at least 23.28% higher than the fitting accuracies of the traditional methods (A: R2=0.592, B: R2=0.0.434, C: R2=0.653). Moreover, based on the fine-scale house price map, this study further analyzes the spatial heterogeneity distribution of housing prices in the downtown area of Wuhan city. By comparing the partial and overall similarity of the simulated house price distribution map calculated via the perceptual hash algorithm, the results also demonstrate that the house prices distribution of Wuhan city has remarkable fractal characteristics. The micro-scale house price data obtained in this study can provide a basis for microeconomics and fractal research in the urban economics. Meanwhile, this study also provides a brand-new research method for micro-scale economic analysis and resource optimization of large cities in China.

Key words: Housing price, deep learning, microscale, convolutional neural network, random forest, Wuhan