FilmDrop Edge On-Orbit is a reconfigurable data thinning solution that quickens access to insights and data security, while lowering downlink costs and bandwidth usage.

Increasingly, satellite providers are faced with the challenge of transmitting very large amounts of data from earth observation (EO) satellites to ground stations. As new EO instruments with even greater spatial and spectral resolution come online, this issue will only be compounded. FilmDrop Edge On-Orbit addresses this problem by using on-satellite processing to reduce the volume of data that must be downlinked from the satellite.

Why use edge processing? 

Previous FilmDrop solutions from Element 84 provided a pipeline that processed data from the ground station through to actionable, analysis-ready data.  Specifically, our work with AWS snowcone hardware facilitated FilmDrop Edge processing from drone data. FilmDrop Edge On-Orbit adds on-satellite processing to improve the quality of the data and reduce latency for downlink. This solution offers accelerated GPU-based processing on-orbit, or can run lighter-weight models on a CPU.  FilmDrop Edge reduces the quantity of data for clients to downlink by, for example, detecting and discarding images that are not useful for data insights (such as cloudy images or certain terrain types), or by downlinking only the results of the analysis rather than the raw imagery. It also provides clients with greater resilience—if less data needs to be downlinked, missions can better deal with communication outages.

Cloud detection as one use case for on-orbit technology

Earth observation satellite imagery consumers often need relatively cloud-free data for their uses. To address this common need, FilmDrop Edge On-Orbit offers cloud detection as one of its capabilities. The cloud detection machine learning algorithm was pre-trained on Sentinel-2 imagery, and then trained using the Azavea Cloud Dataset. Pre-training gives the model a head-start, making it easier and quicker to fine-tune, and ultimately requiring less training data. The algorithm itself consists of a fusion of two architectures—an FPN architecture with a Convolutional Neural Network backbone, and CheapLab, a much lighter-weight architecture. CheapLap in particular is ideal for edge computing applications, because it is highly effective at segmenting “natural classes” (such as vegetation, water, and clouds), and can run on CPUs. By combining a deep learning architecture with a shallower architecture that produces high-resolution output (CheapLab), we arrive at a highly accurate cloud detection algorithm that can run on restricted compute resources.

Additional FilmDrop Edge On-Orbit applications 

So far, we have focused heavily on cloud detection to demonstrate the potential of FilmDrop Edge On-Orbit technology. Although cloud detection represents a clear-cut application of this technology that will benefit satellite imagery providers, there are countless additional use-cases in which we are looking forward to seeing this technology implemented. 

For more information about our progress on this work, and to connect with our team to learn how FilmDrop Edge On-Orbit can work for your organization, find more about FilmDrop and our machine learning offerings on our website.