[Paper Review] Take a Look Around: Using Street View and Satellite Images to Estimate House Prices

Stephen Law, Brooks Paige+1 more|arXiv (Cornell University)|Jul 18, 2018

Housing Market and Economics34 references22 citations

TL;DR

This paper proposes a deep learning framework that fuses street-level (Google Street View) and aerial (Bing satellite) images with traditional housing features to estimate house prices in London. By training a convolutional neural network to extract visual urban quality features end-to-end, the model improves price prediction accuracy and generates interpretable proxy variables for neighborhood visual desirability, even generalizing to unseen boroughs.

ABSTRACT

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.

Motivation & Objective

• Address the lack of systematic quantification of intangible urban amenities—such as visual appeal, prestige, and neighborhood ambiance—that significantly influence house prices.
• Overcome the high cost and subjectivity of traditional urban quality data collection through automated visual feature extraction from unstructured street and satellite imagery.
• Develop a scalable, end-to-end deep learning pipeline that integrates visual features with conventional hedonic pricing models to improve house price prediction.
• Generate interpretable proxy variables for visual desirability using linear fusion of learned visual features, enabling use in downstream econometric analysis.
• Demonstrate generalization of visual quality proxies to previously unseen London boroughs, validating transferability of learned representations.

Proposed method

• Employ convolutional neural networks (CNNs) to extract deep visual features from Google Street View and Bing satellite images without requiring manual annotation of attributes like greenery or pavement width.
• Combine CNN-extracted visual features with tabular housing data (e.g., property size, age, accessibility) in a hybrid modeling approach using both end-to-end deep learning and linear hedonic models.
• Use a two-stage pipeline: first, train a CNN to learn latent representations of urban visual quality from image data; second, fuse these features with traditional housing attributes in a linear model for interpretability.
• Compare non-linear end-to-end models (e.g., multi-layer perceptron) with linear models that incorporate visual features to evaluate trade-offs between accuracy and interpretability.
• Apply Grad-CAM and other explainability techniques to visualize which image regions contribute most to high or low visual desirability scores.
• Train the model on London property data and evaluate generalization to unseen boroughs using out-of-sample prediction performance.

Experimental results

Research questions

• RQ1 Can visual features extracted from unstructured street-level and aerial images improve the accuracy of house price prediction models in urban real estate?
• RQ2 To what extent can deep neural networks learn meaningful representations of urban visual quality without requiring annotated training data for specific attributes?
• RQ3 How do linear models that incorporate learned visual features compare to black-box deep learning models in terms of predictive accuracy and interpretability?
• RQ4 Can the visual desirability proxy derived from image data be generalized to predict house prices in previously unseen London boroughs?
• RQ5 What are the most salient visual cues in street and satellite images that correlate with higher or lower property values, and how can they be interpreted?

Key findings

• The integration of visual features from Street View and satellite images significantly improves house price prediction accuracy compared to models using only traditional housing features.
• The model generalizes effectively to previously unseen London boroughs, indicating that learned visual representations capture transferable urban quality signals.
• Linear models that fuse CNN-extracted visual features with conventional attributes achieve strong performance while enabling direct interpretation of visual desirability as a proxy variable.
• The visual desirability proxy derived from the model correlates with neighborhood aesthetics and urban quality, producing dense, vivid maps of perceived visual appeal across Greater London.
• The approach avoids costly manual annotation by using end-to-end training, enabling scalable extraction of interpretable visual cues from raw, unlabeled images.
• Visual explanation methods such as Grad-CAM identify specific image regions—like greenery, pavement width, and building façade quality—that contribute most to the model’s prediction of visual desirability.