Download the Report

 

Key Highlights

Convolutional Neural Network (CNN) models were developed to use aerial and/or street view image data from publicly available sources such as Google or Microsoft Bing to construct visual desirability indicators that capture intangible attributes of residential houses. The fusion model incorporating a semi-log hedonic pricing (HP) model and the CNN models were estimated onto three single-year datasets covering houses located across 128 suburbs in Brisbane in three executive years from 2018 to 2020. The inclusion of the visual desirability indicator constructed by the CNN model using street-view images would reduce the root mean square of errors (RMSE) errors of prediction by 11 – 21 per cent, varying from year to year.

The inclusion of the visual desirability indicator constructed by the CNN model using aerial images would reduce the root mean square of errors (RMSE) errors of prediction by 20 – 29 per cent, varying from year to year. These empirical results imply that using visual data, either aerial or street-view images, significantly enhances the predictive accuracy of HP models which are the backbone of automatic valuation models currently deployed in the Australian real estate markets.

These results are robust in comparison to the HP baseline model which considers house attributes (e.g. bed rooms, bathroom, car spaces and land area), distances from facilities (e.g. shops, education centers, medical/hospital, and public transport stations), distance from noise sources such highways, and other attributes of the community such as crime rates, population density, and school quality. The CNN and fusion models developed can be deployed in an Artificial Intelligence (AI)-enabled automatic manner to provide better AVMs while further improvement is also possible. 

Authors

Chief Investigator: Associate Professor Viet-Ngu Hoang, Queensland University of Technology

 

Co-Investigators: Dr. Kien Thanh Nguyen and Dr. Andrea Blake

Executive Summary

The accurate assessment of property value is vital to the property industry, especially for the accurate assessment of risk, investment return projection, and asset management. Mass valuations produced by Automatic Valuation Models (AVMs) provide cost-effective solutions to differing stakeholders in the real estate markets.  

 

AVMs are analytics models using data from a variety of sources to predict property values. Conventional AVMs underpinning the many existing commercial solutions in Australia do not have capacity to incorporate visual data. Scholarly research reported in countries other than Australia shows that models that incorporate visual data deliver a significantly greater level of accuracy. 

 

This project has two primary aims. First, it develops an analytical framework of using visual big data from publicly available sources to construct a visual desirability indicator for each property. This visual desirability indicator captures differing dimensions of the intangible values of property such as attractiveness of street frontage, community “green space”, street aesthetics, and other attributes of urban design. This visual desirability indicator can be used to capture attributes of residential properties which are not considered in any commercial AVM solutions currently used in Australia but would be considered by property valuers when valuing a property asset.  Second, it examines empirically to what extent the constructed visual desirability indicator improves the accuracy of property value estimates using the pilot data of residential houses located in Brisbane. 

 

Specifically, the researchers developed a Convolutional Neural Network (CNN) model with a ResNet architecture which uses street-view and aerial image data from Microsoft Bing and Google to construct the visual desirability indicator. The researchers then examined the extent to which the inclusion of this visual desirability could improve the predictive accuracy of property valuation estimates in the framework of semi-log hedonic pricing (HP).  

 

The developed framework is applied to three single-year data sets covering houses located across 128 suburbs in Brisbane in three executive years from 2018 to 2020. The inclusion of the visual desirability indicator constructed by the CNN model using street-view images would reduce the root mean square of errors (RMSE) of prediction by 11 – 21 per cent, varying from year to year. The inclusion of the visual desirability indicator constructed by the CNN model using aerial images would reduce the RMSE of prediction by 20 – 29 per cent, varying from year to year. 

 

These results indicate that the inclusion of visual data, either aerial or street-view images, enhances the accuracy of existing Hedonic Pricing (HP) models. More importantly, these results highlight the potential to deploy Visual Artificial Intelligence (AI)-enabled fusion models to provide better AVMs. Another important finding is that the inclusion of both aerial and street-view data exhibits less improvement than the best individual, i.e. aerial, which is probably due to the complexity of the fusion schemes. This issue would be worthy of further academic research.  

 

It is also recommended that future research be undertaken to consider other types of visual data, such as images or videos related to residential houses. It is also derisible to deploy more advanced models to assess the robustness of empirical results of this project. 

Other resources

Media coverage

Article: How visual data can improve Automatic Valuation Models in Australia | The Australia & New Zealand Property Journal

 

Previous APREF Research

 

This research was funded by the Australian Property Research and Education Fund (APREF). Learn more about APREF here.