Title

AWS re:Invent 2023 - Reducing EO data volume to optimize satellite operations (AES206)

Summary

• Quentin Morris, a solutions architect on the aerospace and satellite team, introduces Alexander D'Souza, the image processing and machine learning lead at Skywatch.
• Skywatch, founded after winning the NASA Space Apps Challenge in 2014, pivoted to Earth observation in 2016 to democratize data access and reduce costs.
• The company received Series B funding in 2021 and provides high-resolution imagery, such as 50 cm data from Airbus.
• A significant challenge in Earth observation is cloud occlusion, which affects around 70% of the Earth at any given time, leading to unusable data and wasted resources.
• Skywatch developed a machine learning model using a UNet architecture with a ResNet encoder-decoder pre-trained on ImageNet to identify and reject cloudy images before downlinking, saving costs for satellite operators.
• The model is sensor and resolution agnostic, and Skywatch has created a large dataset for manually annotated cloud imagery.
• AWS serverless components, including ECR, step functions, lambdas, and SageMaker, are used for model training and deployment.
• Skywatch aims to push critical classification tasks onto satellites to avoid downlinking unusable images, requiring model scaling.
• A demonstration shows the model's ability to detect clouds over snow-covered fields, with the model infrastructure contained in a 3 GB Docker container on the spacecraft and model artifacts of only 22 MB.
• The model produces accurate pixel maps of cloud coverage, enabling cost reductions and operational efficiency for satellite imagery providers.

Insights

• Skywatch's pivot to Earth observation and subsequent funding success highlight the growing commercial interest and investment in space-related data services.
• The cloud occlusion problem is a significant issue for Earth observation, and Skywatch's machine learning solution represents a valuable innovation in the field, potentially saving satellite operators substantial costs.
• The use of AWS serverless architecture for model training and deployment demonstrates the scalability and flexibility of cloud services in supporting complex machine learning tasks.
• The ability to deploy sophisticated machine learning models directly onto satellites could revolutionize satellite operations by reducing the need for ground-based processing and enabling real-time decision-making in orbit.
• The demonstration of the model's performance, particularly in challenging conditions like cloud detection over snow, showcases the practical application of machine learning in enhancing the quality and usability of satellite imagery.
• The size of the model artifacts and the efficiency of the inference process (2-3 seconds) are critical for space-based operations, where computational resources and communication bandwidth are limited.