Predicting High Quality Air-Risk Maps From Historical Data and Geospatial Context

Airspace risk modeling is a vital component used in Operational Risk Assessment (ORA) processes for both strategic airspace and tactical Unmanned Aircraft System (UAS) mission planning. For accurate airspace risk modeling, this paper proposes a new data-driven method for estimating ambient airspace risk, capable of producing scenario, time, and condition-dependent risk maps. The method proposed can be used for more accurate risk estimation and mitigation as part of ORA processes, such as those recommended by the Joint Authorities for Rulemaking on Unmanned Systems (JARUS) Working Group 6 (WG-6). Specifically, we show that special-purpose vision transformer neural networks trained on large amounts of authoritative historical radar track data, in combination with the H3 geospatial indexing system, can produce high-quality conditional risk maps in near real-time. Additionally, as outlined by JARUS, robustness is crucial. We therefore apply methods from Bayesian deep learning for producing uncertainty estimates in order to provide a higher level of assurance to our estimated risk profiles. While there exist several proposed ways for estimating the ambient probability of near mid-air collision (NMAC) in the literature, we make several novel and important extensions, including a design based on fast spatial H3 indexing and Graphical Processing Unit (GPU)-accelerated parallel computation; as well as learning to combine external context and historical data into accurate risk maps. Our method can easily be extended to ingest and integrate other types of context or historical data from various sources; thus, paving the way towards more holistic risk profiles.