Autonomous Storm Tracking and Control for Space Based Lightning Sensors

NeuraLightning is a Neural Network based framework to generate regions of interest for lightning storm clusters from the lightning event stream (i.e., L1b data generated by the real time event processor) that can be used to adjust the sensor/satellite attitude and as input to the CIS. NeuraLightning will use nowcasting to predict future lighting event frames from recent aggregated LIS/GLM/LXM data output using a Convolutional LSTM (C-LSTM). Lightning storms in each predicted frame will then be detected using a "one shot" CNN based object detection model (e.g., YOLO). The chosen NN architectures will also be optimized to execute efficiently on a space qualified hardware. A fundamental constraint is that the autonomous processing of lightning event data must be performed on the processing hardware available on a space-borne satellite, which must satisfy multiple constraints such as weight, power, radiation hardness etc. Since the achievable throughput for the various algorithmic components/options on appropriate space-borne hardware are not known a priori, our approach will be to develop a family of algorithmic variants that represent a trade-off between achievable throughput and model accuracy. Once hardware platform choices are identified, the achieved throughput for the various variants can be evaluated and the version with highest accuracy and sufficient throughput can be selected.