Autonomous Storm Detection and Tracking Using Random Finite Sets

Lightning storm observations are used in advanced weather models to predict and provide warning of severe weather. Current lightning observations gathered from geostationary orbit are static and not well resolved. Observations made from Low Earth Orbit (LEO) offer improved resolution and could take advantage of decreasing LEO construction and launch costs, however challenges are introduced by rapidly changing observing location and smaller sensor field of view. Therefore, to achieve autonomous extended storm target tracking from LEO, ASTER Labs will develop innovative Random Finite Set (RFS)-theory-based software using measurement filtering methods that include the Gamma-Gaussian-Inverse Wishart (GGIW) with Cardinalized Probability Hypothesis Density (CPHD) and GGIW with Joint Generalized Labeled Multi-Bernoulli (JGLMB) filters. Using sequential measurements, these methods enable identification and tracking of extended targets, such as thunderstorms producing lightning flashes. ASTER Labs’ team will develop RFS-based algorithms that identify a storm, quantify and narrow a region of interest (ROI), and output the global location of that ROI. The newly developed STORM Module software tool will provide storm tracking on-orbit with RFS filtering for CMOS sensors. Predictive estimation of the storm dynamics will be provided via a near-constant velocity and unknown turning rate model, inferred by a Reinforcement Learning (RL) agent. Archival data from current lightning mappers, e.g. GOES-16 and GOES-17, will be processed for RL agent training and simulation, as well as evaluation of the STORM algorithms’ ability to properly identify, track, and report a thunderstorm ROI. Phase I will focus on developing the STORM Module and associated algorithms, along with executing storm tracking simulations to assess the algorithms’ performance in select scenarios. The target market for the STORM Module is weather satellites operated by government or private entities.