2021 IEEE Space Computing Conference (SCC) | 2021
Onboard Multi-Scale Tile Classification for Satellites and Other Spacecraft
Abstract
As space-based sensors for Earth observation continue to increase in resolution, extreme computational capability is required to process, compress, and interpret the exorbitant volumes of data collected. Historically, computation has been performed on ground systems. However, this research proposes and benchmarks a flexible onboard tile-classification system for high-resolution Earth observation on satellites and spacecraft. Modern computer-vision techniques for classification and segmentation have progressed to orbital platforms. However, for many applications, the granularity of data analysis has tradeoffs: classification of the entire image is too coarse to yield useful scientific products, and segmentation at the pixel-level is too computationally expensive. Tile classification serves as a middle ground between these two paradigms that can be applied to high-resolution imagery. Transfer learning is conducted on large and small variants of MobileNetV2 deep-learning models using Earth-observation imagery at different ground-resolved distances (GRDs). A case study then compares the inference accuracy for fine-tuned models tested on different GRDs. This process demonstrates the effectiveness of applying one model to a variety of image scales. The inference performance for these models is evaluated in terms of execution time, parallel performance, memory use, and energy consumption. This research showcases increased capacity for onboard remote sensing with tile classification that can foster more versatile space situational awareness.