Geospatial image understanding based on deep learning

Geospatial image refers to images whose geo-location plays a critical role, such as remote sensing images and street view images. Essentially, geospatial images are a kind of maps, providing overall information of a specific area or detailed record for a small spot. Users see the raw visual perception of a place via geospatial image and they know where the place is, like a New York City or a street view photo in their communities.

Normally, geospatial images have no abstract information (e.g., semantics). Given an image, viewers need to distinguish the objects, extrapolate the story, and finally, understand what is happening in that place. The human’s recognition ability is the protagonist in this process. However, this talented ability hinders the use of massive geospatial imagery. Fortunately, the emerging deep learning technique, especially in the computer vision domain, has become a powerful tool to exercise some recognition tasks once achieved by humans only.

Our research group has conducted several studies on geospatial image understanding, and will continue working on this field to build AI systems and knowledge for social good.

We project a street view image into a landcover map, combining the GeoScience and the deep learning technique.

Translate ยป