With clouds covering an average of 65% of Earth's surface, valuable data collected by Earth Observation satellites is often obscured. Typically, 30% to 40% of downlinked images are discarded due to excessive cloud cover, leading to significant financial losses.
Satellites equipped with edge computing capabilities are revolutionizing how we process and analyze this data. As the volume of data collected by the hundreds of EO satellites continues to grow, traditional methods of downloading raw binary data to ground stations are becoming inefficient and costly. This blog will explore satellite edge devices, computing on such devices, existing methodologies for cloud detection, and how onboard cloud detection can improve satellite operations.
Edge computing in satellite operations involves deploying software on satellite-based hardware, which is often constrained by limited resources such as disk space, RAM, and processing power. Edge devices, like Raspberry Pi, Jetson Nano/Orin, and similar System-on-chip (SoC) hardware, operate in remote and harsh environments. Optimizing applications for these edge devices requires innovative approaches to ensure efficient data processing and transmission. In the context of satellites, edge computing enables real-time data analysis and decision-making, reducing the need to transmit vast amounts of raw data to ground stations.
On Board Cloud Detection Models
Clouds can significantly interfere with data collection for Multispectral (MSI) and Hyperspectral (HSI) data. Due to the nature of passive sensors and their wavelengths, clouds mask out swathes of land in an image. Hence, current methodologies require downlinking all captured data from the satellite to a ground station and then checking for clouds either via thresholds or using cloud detection algorithms. These methods can become bottlenecks as data volume increases, exacerbated by slow downlink speeds reminiscent of the early 90s/2000s internet speeds. Hence, some basic processing on the satellite is necessary to filter out images containing clouds over a certain percentage. Advances in embedded systems and chips now allow software to run complex deep-learning algorithms to mask out cloudy pixels effectively.
This method is still restricted to individual sensors as there is very little standardization between various sensors and satellites. Even the same sensor on two different satellites can result in varying images and constraints. To bypass this, a product that can run on all data collected from various sensors and satellites is critical for deployment at scale. The resolution, number of bands, and wavelengths will make very little difference to such a product. A key requirement is compressing the model to minimize space on both the hard disk and RAM, optimizing resource use. A decent benchmark is reducing the model size to at least 50-100 Megabytes.
Reduce Downlink Costs and Enable Rapid Insights
Integrating cloud detection models into the edge-computing framework of satellites can significantly reduce the volume of data that needs to be downlinked. This selective transmission approach, which filters out cloud-contaminated images, can lead to cost savings of at least 30% as resources are no longer wasted on transmitting unusable data.
By removing cloud-polluted pixels, data size can be reduced by up to 40%. This efficiency enables faster decision-making in critical applications such as disaster response, infrastructure monitoring, and national security.
The Market
The market for on-board cloud detection models is vast and rapidly expanding, driven by the increasing deployment of hyperspectral and multispectral imaging satellites. Currently, there are approximately 980 Optical, Hyperspectral & Multispectral satellites in orbit, with deployment plans for many more in the near future, as indicated by the NewSpace Index.
The market potential underscores the strategic advantage for satellite data providers to incorporate these models, ensuring they can offer their customers affordable, cloud-free imagery. As the demand for timely and precise satellite data grows from downstream users, integrating on-board cloud detection models will become increasingly critical in maintaining pricing advantage, and further passing on the cost-saving benefits to the user.
About LPL
Little Place Labs specializes in near-real-time space analytics for terrestrial and space applications, leveraging machine learning solutions designed for satellites and space infrastructures. We remove the data bottleneck by processing raw data in orbit. Our applications service commercial and national security sectors by significantly enhancing response times and decision-making for threats against civilians, national assets, and our guardians.
Orbitfy Prep is LPL’s suite of in-orbit software modules designed to enhance satellite operations and increase utilization rate by enabling real-time intelligence and optimizing data delivery. The software includes cloud detection and thresholding functions, processing raw binary inputs into analysis-ready data before cloud detection. Our model is agnostic to the number of bands, resolutions, and wavelengths, making it versatile across various sensors.
If you want to learn more about Orbitfy Prep and how it can optimize your satellite operations, book a call with us using Calendly.
Commenti