Embedded GPU implementation of sensor correction for on-board real-time stream computing of high-resolution optical satellite imagery

摘要

In recent years, the data generation rate of high-resolution optical satellites is increasing rapidly, and this situation brings massive stress to the data downloading link and data processing system. To relieve downloading link stress and enhance the time efficiency of information acquisition, on-board processing needs to be introduced. Current on-board solutions are mostly based on digital signal processing and field-programmable gate array, which are obturated, lack flexibility, and are highly costly to implement. This paper takes sensor correction, the prerequisite geometric step of high-resolution optical satellite data processing, as an example, using a simulation prototype made of a double-module data parallel pipeline system based on NVIDIA embedded graphics processing unit platform, and proposes a feasible stream computing approach for low power consumption, flexible, and expandable on-board real-time processing. The experiments use a short strip of GaoFen-9 satellite data that contain 4.5 standard scenes to validate the performance and correctness of this approach. Compared to the same algorithms on the Dell PowerEdge T630 Server, the on-board version demonstrates an obvious performance advantage. When considering the contrast of power consumption for each platform, the advantage becomes even more significant. With statistics and analysis of experimental results, the timeline of processing demonstrates that this approach could meet the on-board real-time sensor correction requirement. And correctness is also verified by the root-mean-square error of pixel-by-pixel image comparison experiments.