Machine Learning

You Only Look Once (YOLO): Real-Time Object Detection

Ships and other objects of interest can be detected and tracked in satellite optical or thermal imagery using existing machine learning algorithms. Once an object of interest has been detected, the object can be “fingerprinted” to be re-acquired later or be picked up by another satellite. Object tracking tasks can be performed autonomously by multiple satellites with the last known location and object fingerprint being passed to the satellite that currently has the best view of the target area.

 

YOLO is an effective single shot object detection (SSD) architecture that is computationally cheap enough to run in real time on existing satellite hardware. The speed of the YOLO architecture comes from processing the entire image in a single pass, rather than a two-shot model such as Faster-RCNN that processes portions of the image repeatedly. Two-shot models are generally more accurate at the cost of increased processing time, but YOLO has experimentally been able to complete all tracking tasks with one-fifth the processing time of two-shot models.

Once the objects in the image have been detected, the localized image of the object is used to generate a fingerprint by first removing the background then separately analyzing each channel of the image. The reflectance of each channel is analyzed, unique characteristics of the object class are identified, in the case of ships these are length, breadth and ellipticity. All distinguishing features along with heading and speed are then used to positively identify the ship in the future.

Citations:

Heiselberg, “A Direct and Fast Methodology for Ship Recognition in Sentinel-2 Multispectral Imagery,” Remote Sensing, vol. 8, no. 12, p. 1033, 2016.

Joseph Redmon, Ali Farhadi: “YOLOv3: An Incremental Improvement”, 2018; [http://arxiv.org/abs/1804.02767 arXiv:1804.02767].

The Fermi Power Margin Model (PMM)

NASA Goddard’s Fermi spacecraft suffered a solar array drive failure in 2018 that resulted in less solar power to operate the spacecraft. Subsequently, new modes of operation and close monitoring of the power system are now required on a daily basis. The Fermi Power Margin Model (PMM) is a support vector regression model trained to gauge the stress level placed on the Fermi spacecraft’s power system. The PMM model was trained on ~17GB of telemetry data using labels generated from a manually tuned heuristic and input from the flight operations team (FOT). The PMM model is able to achieve an accuracy of 97.5% on a reduced telemetry set and allows the FOT to gauge the health of the power system once per orbit rather than once per day.

The PMM also contains a univariate and multivariate anomaly detection component that identifies significant power events and anomalies through telemetry. Univariate analysis performs a historical z-score analysis per telemetry item to detect events signifying a drop in power relative to the beta angle of the spacecraft. Multivariate analysis applies a X2 test of the mahalanobis distance for combined telemetry inputs to detect global power events such as solar eclipses or changes in spacecraft attitude affecting solar power.