Kristen Haynes, Applied Optimization Inc.; Jeffrey Hollon, Applied Optimization, Inc.; Kimberly Kinateder, Applied Optimization; Victoria Carone, Applied Optimization; Tamara Payne, Applied Optimization Inc.; Phan Dao, Applied Optimization, Inc.; Cory Hufford, Applied Optimization, Inc.; Stephen Gregory, Stephen A. Gregory LLC; Scott Milster, AFRL/RV
Keywords: data fusion, multi-sensor, photometry, non-resolved object characterization
Abstract:
With the development of the Unified Data Library (UDL), large amounts of commercial electro-optical data have become available for analysis. This data is collected by multiple ground-based sensors spread across the Earth and provided by multiple commercial sources. As a result, highly diversified multi-sensor data may be available for a particular satellite on any given date. Ideally, data from all the sensors collected on a satellite would be used for non-resolved object characterization. However, blindly combining data from multiple sensors can result in the joining of different noise profiles or multiple overlapping, yet different, trends in the light curve as a result of different locations, observation conditions, and data processing methods potentially used by different sensors. Since noise or multiple trends in the light curve can cause issues in characterizing the satellite, data from different sensors must either be kept separate or combined in a more selective manner. While separating data from each sensor into their own separate light curves is an easy option, the resulting light curves may be too sparse in number of observations or phase angle span to use for effective characterization.
Therefore, we explored methods to handle multi-sensor data that would combine subsets of the data to avoid both the issue of noisy, multi-trend light curves from all available sensors and the issue of light curves being too sparse for use in characterization from separate sensors. This paper presents a method we developed to selectively combine subsets of data by analyzing the mathematical similarity of data trends between sensors that make up a light curve in combination with the physical distance between sensor pairs to determine whether the multi-sensor data can be combined. The similarity of data trends from sensor pairs is analyzed by fitting simple statistical models to each sensor independently and to the union of each sensor pairs data. The residuals of each fit are then compared to determine if fusion of the multi-sensor data is possible. If the residuals from the fits to the individual sensors data are considered similar to the residuals for the fit to the union of the sensors data, then we consider the two sensors trends to be similar and eligible for fusion. This process is performed iteratively across all possible pairs of sensors present in the data until the statistical tests indicate no more fusion may occur. Before a sensor pairs data can be fused, we also consider the geographic distance between the sensors to help guide the statistical fusion method.
Since different facets and payloads of a satellite may be visible for sensors separated geographically from one another, we leveraged simulations that considered both north-south and east-west differences in sensor locations when observing the same satellite. These simulations were used to analyze the effects of directional differences on the resulting light curves and aided in the determination of when the distance between sensors causes their light curves to no longer have visually similar trends. With the simulation results, we can determine distance thresholds that will restrict certain pairs of sensors being fused if their distance surpasses the threshold.
If a sensor pair has similar light curve trends and a sufficiently small distance, the data from those sensors are fused. This process is performed iteratively until either all sensors are fused together or no remaining sensors fulfill both the trend similarity and distance requirements. The result of this process is one or more light curves that can then be used for enhanced object characterization, monitoring, and change detection procedures.
This paper will present the statistical method used to analyze trends in multi-sensor light curves, the simulations used to determine a distance threshold, and results of our sensor fusion method.
Date of Conference: September 14-17, 2021
Track: Non-Resolved Object Characterization