Improved Geolocation for Over the Horizon Radar (IGOR)

The primary objective of the Improved Geolocation for Over the Horizon Radar (IGOR) Phase II SBIR is to develop and demonstrate a prototype system that processes OTHR radar returns to produce improved geolocation estimates for targets illuminated in the radar’s surveillance region. To accomplish this objective, EXP will enhance the IGOR Deep Learning algorithms developed during Phase I and evaluate geolocation performance improvements using both high-fidelity surrogate data and as well as collected operational OTHR data. \n\n During Phase I of the IGOR SBIR, EXP developed proof-of-concept results, which used a deep learning-based model of the OTH propagation channel to improve the precision of range and Doppler target measurements. Traditionally, significant effort is invested in modelling the ionosphere explicitly. Sophisticated physics-based forward models requiring large amounts of computation must be supplied with accurate ionosphere state information (e.g. altitude, shape, electron content, etc.) to guess-and-check against known targets and iteratively arrive at a time-of-flight to ground range mapping. Even the best iterative solvers are still subject to residual biases and unknown/unseen ionospheric state deviations which results in tens of kilometers of range uncertainty. Moreover, the backpropagation step (mapping time-of-flight to ground range) has many degenerate solutions which can be highly dependent on subtle variations of the ionospheric state over a large sea/land area. \n\n The IGOR deep learning (DL) model directly learned a parameterized implicit representation of the ionospheric channel from the radar measurements themselves, using objects with known locations (such as islands or transponders) as “fiducials” or reference points. This approach helps break the degeneracy present in the backpropagation problem and provides support for a generalized representation of corrections that could work for any perturbed ionosphere without explicit retraining or human operators in the loop once deployed. Finally, once trained, the inference speed of the DL solution can be orders of magnitude faster than typical forward-model driven iterative solutions.