Source-linked AI summary
Take a Look Around: Using Street View and Satellite Images to Estimate House Prices
Stephen Law, Brooks Paige, Chris Russell
TL;DR
Urban visual qualities matter for house prices but are difficult and expensive to measure systematically. The paper learns visual responses from Google Street View and Bing aerial images alongside traditional housing attributes, using nonlinear and hybrid linear models for London. The image-augmented models improve prediction, while the hybrid approach yields interpretable neighborhood-visual proxies, though geographic specificity and image reliability limit scope.
Problem
Urban qualities such as neighborhood prestige and visual impression affect house prices but lack systematic, economical methods and data for measurement.
Method
The paper uses deep neural networks on street and aerial neighborhood images with traditional housing attributes, comparing nonlinear fusion with a hybrid linear model that extracts visual proxy variables.
Results
Image-augmented models outperform the housing-attribute baseline, aerial images outperform ground-level Street View augmentation, and interpretable visual proxies improve explainability with minor accuracy losses.
Takeaways & Limitations
The learned visual proxies can support interpretable econometric analysis and generate visual-desirability maps for urban environments similar to London.
Takeaways & Limitations
The model is geographically specific, comparative datasets lack a common diverse baseline, and image quality can be affected by obstruction, lighting, and weather.
Abstract
from arXiv · showhide
When an individual purchases a home, they simultaneously purchase its structural features, its accessibility to work, and the neighborhood amenities. Some amenities, such as air quality, are measurable while others, such as the prestige or the visual impression of a neighborhood, are difficult to quantify. Despite the well-known impacts intangible housing features have on house prices, limited attention has been given to systematically quantifying these difficult to measure amenities. Two issues have led to this neglect. Not only do few quantitative methods exist that can measure the urban environment, but that the collection of such data is both costly and subjective. We show that street image and satellite image data can capture these urban qualities and improve the estimation of house prices. We propose a pipeline that uses a deep neural network model to automatically extract visual features from images to estimate house prices in London, UK. We make use of traditional housing features such as age, size, and accessibility as well as visual features from Google Street View images and Bing aerial images in estimating the house price model. We find encouraging results where learning to characterize the urban quality of a neighborhood improves house price prediction, even when generalizing to previously unseen London boroughs. We explore the use of non-linear vs. linear methods to fuse these cues with conventional models of house pricing, and show how the interpretability of linear models allows us to directly extract proxy variables for visual desirability of neighborhoods that are both of interest in their own right, and could be used as inputs to other econometric methods. This is particularly valuable as once the network has been trained with the training data, it can be applied elsewhere, allowing us to generate vivid dense maps of the visual appeal of London streets.
1 INTRODUCTION
House prices reflect tangible property characteristics and difficult-to-measure neighborhood qualities, but city-scale urban-quality data are costly and scarce. The paper addresses this gap by learning visual responses from street and aerial images and combining them with housing attributes in interpretable and nonlinear pricing models.
- Motivation: Property prices incorporate tangible assets and intangible neighborhood qualities such as perceived safety, activity, and relationships with surrounding properties.These intangible effects challenge assessors who must assign realistic prices to what buyers are willing to pay.
- Research gap: Prestige and street aesthetics are difficult to quantify, leaving limited research, computational tools, and data for linking urban design attributes to property values.Existing quantitative support for which urban attributes produce better cities or higher property values is described as limited.
- Research gap: City-scale urban-quality measurement is costly and time-consuming, while computer-vision alternatives would require large collections of images annotated with expensive quality metrics.The annotation burden creates a chicken-and-egg problem for supervised visual-quality prediction.
- Approach: The proposed end-to-end approach learns visual urban features from street-level and aerial images using convolutional neural networks, avoiding intermediate hand-labeled quality measures.The learned image cues are used with traditional housing features to estimate London property prices.
- Approach: The paper compares a neural black-box model with a hybrid model that maps images to latent attributes before linear fusion, producing interpretable proxy variables for neighborhood visual appeal.The hybrid model supports extracting marginal visual effects and mapping visual appeal across Greater London.
- Contribution: Unlike prior image-based house-price work focused on property interiors or exteriors, this study uses street and aerial neighborhood images to measure visual desirability directly.The resulting proxy attributes are intended for direct use in existing econometric models.
2 RELATED WORK
Earlier research used hedonic models, machine learning, and image-derived features to estimate prices or urban perceptions, but this paper combines street and aerial neighborhood views with interpretable latent responses. It also evaluates nonlinear and hybrid linear pricing models, including spatial out-of-sample testing across London boroughs.
- Hedonic pricing: Hedonic pricing decomposes housing value into utility-bearing attributes by comparing properties while holding other characteristics constant.The implicit price of an attribute is inferred from the corresponding difference in real estate values.
- Terminology: The paper defines visual latent response as an overall semantically interpretable response of the visual system, distinct from a visual feature that drives classification.This distinction frames the learned output as a global interpretable response rather than an image patch.
- House-price estimation: Prior work found that artificial neural networks improved house-price estimation over OLS by capturing nonlinear relationships among traditional housing variables.Such nonlinear effects are especially relevant when neighborhood quality interacts with tangible property attributes.
- Image-based pricing: Related studies supplemented housing variables with property-photo features or used recurrent models to incorporate image information from wider districts.These approaches generally emphasized the property itself or spatially related property imagery rather than the local urban environment studied here.
- Urban visual analysis: Street-image research also estimated perceived safety, liveliness, and other urban characteristics that could serve as house-price covariates.Other work extracted visual features associated with socioeconomic factors or location attributes such as house price, density, trees, and graffiti.
- Paper positioning: This paper collects both street-level and aerial neighborhood images, compares augmented and housing-only models, and contrasts nonlinear with hybrid linear latent-variable approaches.It also tests multiple neural architectures and excludes Southwark from training to evaluate spatial generalizability.
3 METHOD AND MATERIALS
The study combines traditional housing attributes with street-level and aerial imagery to estimate London house prices. It compares nonlinear prediction with interpretable linear or additive models that expose visual desirability proxies.
- Data collection: The model estimates log house prices from housing attributes, street images, and aerial images collected across Greater London.The procedure includes data collection, training, and testing phases.
- Data collection: 130,557 property transactions were grouped to 40,470 streets, with structural, neighborhood, and location features assembled from housing and public datasets.Neighborhood features include nearby parks and shops or commercial uses within 800 meters.
- Data collection: One front-facing Google Street View image was collected for each London street using street-network locations and bearings.Images were centered on the road, oriented consistently, and collected with a 120-degree field of view before invalid-image cleaning.
- Data collection: One Microsoft Bing aerial image was collected per street at zoom level 18 to provide a roughly 150-meter neighborhood view.The aerial image location was based on the centroid of each street edge.
- Visual Feature Extraction: Street and aerial imagery are processed by convolutional networks whose learned visual responses serve as proxies for urban-environment desirability without expert annotations.The networks are trained end-to-end with a hedonic price model while controlling for individual housing attributes.
- Model Architectures: The study compares a fully nonlinear hedonic perceptron with linear or additive hedonic models that trade predictive flexibility for interpretability.The nonlinear model captures interactions and nonlinear responses, while the linear model uses weighted feature combinations and the additive model excludes interactions.
- Linear Hedonic Model: The interpretable model’s learned visual response can be read as a measure of how neighborhood visual desirability alters house prices and mapped across London.The models are evaluated with and without visual latent responses to examine their effect on house-price estimates.
4 EXPERIMENTAL RESULTS
The experiments show that combining housing attributes with street and aerial imagery improves house-price prediction, including when testing on an unseen London borough. Interpretable visual-desirability proxies retain much of the predictive benefit while enabling neighborhood-level interpretation.
- Experimental design: The experiments compare neural, linear, and generalized additive models with housing attributes, Street View, aerial imagery, and their combinations.Six data-source combinations and multiple model families are evaluated, including held-out Southwark generalization.
- Spatially missing-at-random results: 85% accuracy is achieved by the full model combining housing attributes, Street View, and aerial imagery, versus 83% with housing attributes and Street View and 84% with housing attributes and aerial imagery.Models combining multiple data sources outperform single-source models, while housing attributes alone outperform either image source alone.
- Generalization to held-out Southwark: Visual latent responses reduce the held-out-borough accuracy loss to around 5%, compared with 15% for models without these responses.All models lose accuracy when generalizing from other London boroughs to Southwark, but the loss is cut by two thirds when attributes and visual latent responses are used.
- Linear hedonic pricing comparison: Linear models with visual-desirability proxy variables improve substantially over standard housing-attribute models while approaching the accuracy of the nonlinear hedonic perceptron.This combination preserves the interpretability of a linear model while incorporating visual urban appearance.
- Generalized additive model: The visual-desirability variable is associated with approximately a 13% house-price change for a 0.1-unit change, while its full range can add or reduce house price by up to 1.85 times.Including the variable also significantly improves AIC in the linear model; the GAM with splines further improves AIC relative to the linear model.
- Generalized additive model: The additive model achieves comparable accuracy to other nonlinear models while retaining greater interpretability, and the learned visual contribution can be mapped across central London.Top-ranked imagery tends to show dense historic buildings, tree-lined streets, and greenery, whereas lower-ranked imagery tends to show wider streets, blank frontages, and little vegetation.
5 DISCUSSION AND CONCLUSION
The approach improves house-price prediction by incorporating visual information about local urban environments while retaining interpretable proxy measures. Its main boundaries are geographic specificity, imperfect imagery, limited visual explanation, and omitted environmental or wider-neighborhood cues.
- Contribution: Visual features from local street and aerial environments improve house-price prediction, while the property itself forms only a small part of aerial images and is typically absent from Street View images.The approach characterizes the neighborhood where a property sits rather than relying primarily on images of the property for sale.
- Limitations and future work: The study’s geographic specificity leaves open whether its visual-desirability proxy generalizes to different housing markets.The authors propose fitting the model to other cities to compare aesthetic preferences, including London and Kyoto.
- Limitations and future work: Further work is needed to explain the convolutional neural network visually, incorporate complementary and environmental cues, and capture wider neighborhood or district-level effects.The paper also identifies a need for more diverse methods and a consistent, open, multimodal, geographically diverse benchmark database.
- Limitations and future work: Google Street View and Bing aerial images can be affected by visual obstruction, poor lighting, and weather differences, although the authors expect these issues to add noise rather than systematic bias.The stated assumption is that such image problems are distributed fairly uniformly and are uncorrelated with price or housing attributes.
- Implications: The models can support mass appraisal for mortgage assessment and inform planning guidance aimed at improving the visual desirability of streets and new neighborhoods.These implications are presented for urban planning and the real estate sector.
- Contribution: End-to-end training avoids costly urban-data annotation while extracting visual proxies that can be reused through released code and pretrained models in environments similar to London.The proxies may also support economists estimating willingness to pay for different levels of visual desirability.