Estimating House Price from Multiple Data Sources

House prices are usually estimated based on basic house attributes, such as age. However, prices also vary depending on the different environments in their neighbourhood, such as greenery, noise of vehicles. These attributes are not generally available in traditional datasets as it takes a lot of effort to collect them. Recent research has shown that we can infer these socio-economic attributes from street view images and very high-resolution aerial imagery.

In this work, we collected street view images and aerial images of each transacted property in New York City, as well as conventional attributes, to estimate the log-transformed price of the property. For the aerial images, potential visual features were extracted using a model pre-trained on an aerial image dataset, while intermediate values were also extracted, presenting the density of cars and swimming pools. For the street view images, visual latent features were extracted by a standard ResNet-50 model. After detecting and classifying car makes and models in the street view images, we aggregated the information on detected car into census blocks.

Various combinations of features were used to compare the performance of the models. The results show that the model augmented with features extracted from street view images and very high-resolution (VHR) satellite images outperforms a model that extracts only basic features from traditional data sources.