Managing the Data Flood - Analyzing Images Onboard Small Earth Observation Satellites

Satellites as small as cubesats and other Nanosatellites can generate vast amounts of optical data using highresolution and multispectral image sensors. Coupled with the growing popularity of proliferated earth observation constellations consisting of many small satellites instead of larger, more traditional earth observation satellites, this presents new challenges. Firstly, this data has to be downlinked from the satellites to ground stations, a task which is made more complicated with increasing numbers of satellites in a constellation. Additionally, smaller satellites tend to have less power available for transmitting, as well as smaller antennas, reducing downlink speeds. Secondly, the data must be analyzed to extract mission-specific information depending on the current use-case: For example, wildfires have to be identified or the location and type of ships detected. For many applications, only these extracted meta-information are relevant, not the actual satellite image itself. These necessary steps of downlinking and image analysis can lead to large delays between image acquisition and actionable decisions being made on the basis of this data. In this paper we describe the integration of an onboard image analysis payload into the UWE-5 satellite mission, an educational communications satellite mission consisting of two 3U CubeSats. This processing unit features two boards, each containing a Microchip PolarFire MPFS250T FPGA SoC. In this architecture, one FPGA SoC controls the camera payload, while the second one synthesizes a Machine Learning (ML) accelerator to analyze camera data onboard the satellite using Convolutional Neural Networks (CNNs). This approach enables the satellite to evaluate the usefulness of specific images by analyzing the degree of cloud cover or to detect features deemed important by the operator, accelerating the path from data acquisition to actionable insight and reducing data volume.