Predicting real estate prices via satellite imagery and deep learning

Abstract

Real estate prices are estimated by licensed professionals based on manually collected survey data. However, obtaining survey data is labor intensive and the estimating process is limited by its subjective nature. To tackle these problems, we maximized the potential of satellite imagery as an alternative source using deep learning models that consider the latent features of real estate properties. We checked the transferability of feature extraction models for a metropolitan city, a medium-sized city, and a rural county, and interpreted the extracted features through visualization. Compared to the baseline models, using high-resolution satellite imagery and state-of-the-art models lead to superior or comparable performance for urban cities such that 92.5% of the variation could be explained. We also found that high accuracy was achievable even if features were extracted from a model that had been trained on data of a different city. These results may provide a first step towards a fully automated system that solves the difficulties of traditional feature extraction and prediction methods.