SBIR-STTR Award

There is an abundance of text documents, news articles, intelligence reports, etc. containing information that is particularly valuable in understanding and analyzing aerial/satellite imagery. These text documents can be annotated with the corresponding lat/long coordinates of the geospatial references using commercial tools, such as MetaCarta. However, this approach is 1) not scalable since new documents should continuously be annotated and 2) limited to only those geographical features that already exist in the MetaCarta gazetteer. Even if the document is annotated, numerous “links” to all the related documents from a target image is not an effective presentation method. In this project, we will develop the technology for automatically 1) finding text documents relevant to a given imagery without any a priori annotation of the documents (similar to a search engine), 2) ranking the documents based on their geospatial relevance to the imagery, 3) summarizing the documents, and 4) linking the documents to their corresponding location on the imagery. The resulting technology will allow an analyst to view a satellite image for any place in the world, automatically find and link the geospatially related documents, and then browse the summary of the documents to better understand the information shown in an image.

Keywords:
Natural Language Processing, Geospatial Information Systems, Satellite And Aerial Imagery, Text Documents, Text Summarization, Imagery Analysis, Geospatial Knowledgebase

The objective of this project is to develop a capability to automatically link news articles to satellite imagery and then to display summaries of the linked articles. In Phase I of this effort, we developed an initial approach to automatically link news articles to geographic locations and then cluster and summarize the linked articles. In our experimental evaluation, we showed that articles could be linked with high precision and moderate recall. We also demonstrated the feasibility of automatically clustering the articles into topics and generating both single-line summaries as well as longer summaries. In Phase II of this effort we plan to flesh out the techniques into a fully-functional system, called GeoXRay. In particular, we will develop automatic techniques to rapidly build the required geospatial knowledgebase, which allows the system to link articles to locations outside of the US. We will also improve the linking techniques to improve the recall of the document linking while maintaining the high-level of precision that we achieved in Phase I. Finally, we will improve the techniques for automatic clustering and summarization of the news articles linked to a location so that an analyst can quickly determine which articles are likely to contain relevant information.

Keywords:
Geospatial Data, Geospatial Intelligence, Geospatial Knowledgebase, Geospatial Reasoning, Natural La