Automatic generation of high-quality terrain data for use in military training applications (especially, integrated Live-Virtual Training) depends heavily on gathering and acquiring high-quality and high-resolution data from a variety of appropriate sources.  One of the more accessible sources today comes from images collected by drones. Using photogrammetry, the images are often post-processed to produce detailed 3D representations of the environment. Starting the data generation process with clean data substantially reduces the need to remove anomalies or to perform unnecessary processing further in the terrain database generation pipeline. Unfortunately, quality data is not always provided or available. High resolution collection sources often produce outputs that include inaccuracies, errors, or suboptimal content. The increase in resolution also adds to the challenges in discerning between accurate and erroneous aspects of data. Recent collections also highlight the need for more sophisticated geometric analysis methods and tools to detect and remove anomalies that are geometrically joined with or close to good data. Therefore, the input data must be further analyzed, cleaned/corrected, and prepared to make it usable in the terrain generation process. This paper describes some of the key techniques and the analysis process used to (a) improve the collection process and ensure data collection approaches yield the highest quality data possible, (b) leverage AI/ML techniques to detect, then fix (or at least reduce) anomalies in suboptimal input/collected data before the data is further propagated through the pipeline, and (c) combine both AI/ML and non-AI/ML automation to extract the desired content needed in generating high-resolution terrain data.

Keywords
AI, AUTOMATION, CONTENT GENERATION, GEOSPATIAL DATA , MACHINE LEARNING, TERRAIN

Additional Keywords
photogrammetry, data collection/preparation