We introduce a novel way to enrich data driven value estimation using satellite imagery and land-use information to enable value predictions on properties with sparse data availability. We find that using the surrounding land-use labels generated by BigEarthNet slightly increases our models accuracy in predicting Airbnb properties price. This result hints at a potential correlation between spatial attributes and price values. However, we do not find improvements by using satellite imagery for the prediction task. This holds for different scales of satellite images. Also when using external information, such as location, in combination with the satellite images our proposed neural networks cannot significantly improve over techniques that just utilize external properties. We outline several potential reasons why the proposed methodology does not work as well as might be expected to inspire more robust future work.

Authors: Tobias Florin Oberkofler (Leiden University), Maxime Casara (Leiden University).