SATPOI-Net: Deep Learning Model to Predict the Point of Interest (POI) from Satellite Imagery

Point of Interest (POI) data provides valuable insights into unique offerings and experiences at various places. Tracking user visits to POIs offers valuable information for personalization and targeting. However, obtaining complete and consolidated POI data can be challenging, especially in certain areas. This study addresses this challenge by proposing an approach to enhance and expand POI categories, thereby enriching previously unavailable information. Unlike previous studies, this study uses satellite images and other geospatial data such as Open Street Map (OSM) derived information to predict POI categories using deep learning approaches. The study conducted tests on multiple global and Japan-specific POI datasets, revealing that employing regional-specific modeling and a substantial amount of ground truth data can yield promising results on POI grid categorization using the SATPOI Net model. The study further showcased the superiority of the multimodal approach over the single-modal approach, achieving an impressive 3 to 5% increase in accuracy.