Extracting geospatial information from aerial and satellite imagery
Automated street and building footprint extraction from satellite imagery is explored in a high performance computing environment. The performance of sequential learning models that directly extract polygons and polylines from the imagery is compared against the performance of indirect extraction method like the Hierarchical Supervision (HiSUP) framework which first performs a segmentation prior to vectorization of the results. The research used a scalable GPU infrastructure distributed across up to eight computing nodes, each equipped with four GPUs to accelerate processing and measure how performance degraded as a function of scale.  The models were trained on the Massachusetts roads and building data set, with ground truth derived from validated geojson format foundation data. Successful direct extraction of vector features from satellite imagery without intermediate representations was demonstrated. Key findings are that the method used outperformed state of the art models using intersection over union metrics and that scale does not inhibit performance. Conclusions provide valuable insights into current capabilities and limitations of state of the art methods in geospatial applications and suggest the potential benefits of integrating emerging technologies in distributed computing. The implications of the work are that we should expect more rapid and accurate fully automated updating of geographic information systems that support urban planning and disaster response.
IMPACT
Accomplishment: Developed new technology for automating geospatial information generation; Result: Improved building and road extraction performance with increased confidence levels of 20%. This can be applied to NGA foundation data creation worldwide for the military which is currently lightly automated and comes at the expense of hundreds of millions of dollars per year.
PRESENTER
Qiu, Siwei
siwei.qiu@nga.mil
571-558-1838
National Geospatial-Intelligence agency
CO-AUTHOR(S)
Duncan McCarthy
Duncan.S.McCarthy@nga.mil
CATEGORY
AI/ML for HPC
SECONDARY CATEGORY
GPU usage for HPC
SYSTEM(S) USED
Nautilus, Blueback and Raider