Astrophysics is the ultimate remote sensing challenge. We figured that the problem of inferring the properties of distant galaxies from a limited set of (very) remote observations is no different to inferring the surface properties of the Earth using data taken from orbit. Complex data, weak signals and buried information is our bread and butter. So we are applying our experience in astrophysics-based hyper-spectral image and time series analysis to EO, bringing to bear some of the latest AI techniques to boot. We are particularly interested in applications of Synthetic Aperture Radar observations. Our first major project is ClearSky, but we are also working on other exciting things that you can read about below.
We are working closely with satellite communications innovator and space gateway Goonhilly Earth Station Ltd., a member of NVIDIA's Inception Programme. Goonhilly is a legendary satellite data hub with outstanding data connectivity. Recently, GES has established a powerful centre for AI-accelerated data processing: Goonhilly's new green Tier 3 data centre hosts an AI & Deep Learning optimised data platform, which includes the NVIDIA® DGX-1™ supercomputer. This is enabling us to perform truly deep and innovative Earth Observation analysis at lightning speed.
The NVIDIA® DGX-1™ is an integrated system for AI and deep learning. DGX-1 features 8 NVIDIA® Tesla® V100 Tensor Core GPU accelerators, interconnected with NVIDIA® NVLink™ in a hybrid cube-mesh network. DGX-1 delivers over 2 petaFLOPS of AI training performance. Moreover, the DGX-1 system is powered by NVIDIA-optimized software, including the latest AI frameworks and libraries, tuned for maximized performance and scaling for multi-GPU and multi-system workload. DGX-1 is the first AI system purpose-built for AI and deep learning, providing developers and researchers plug-in, power-up deployment simplicity, effortless data science workflow and revolutionary performance that scales in both production and research settings.
This image shows an ESA Sentinel composite of the Cornish peninsula observed on 18th May 2019. The lower part shows the direct Sentinel-2 RGB observation, affected by cloud cover. The upper part shows our ClearSky prediction of the RGB response using Synthetic Aperture Radar (SAR) imagery observed by Sentinel-1. SAR offers the capability to penetrate cloud, but interpretation of this data is not straightfoward. ClearSky remedies this by recovering imagery across the full visible-to-infrared spectral window, dramatically enriching the utility of SAR data.
We are developing a technique to evaluate the health of pasture below the resolution of individual pixels. Our approach allows us to map things like bare soil fraction and predict dry matter yields within a field. This information can be used by farmers to maximise the efficiency of land use.
Our coast is dynamic. The ebb and flow of the tide, storms and coastal erosion result in a complex mixture of factors that affect the coastal environment. We are working on a new method to monitor the land-sea interface with an aim of providing a hyper-localised tidal map of the UK.
Radar backscatter is dependent on the surface geometry at the point of interest. While there are sophisticated approaches for measuring surface elevation, we have developed a 'cheat' algorithm that uses simple SAR imagery to detect changes in elevation, for example due to landslips or mining activity.
Water is a precious resource, and proper management of the supply will become even more vital as our climate changes. We are developing an algorithm that can simultaneously measure the surface area of every reservoir in the UK. This provides a more sensitive indicator of stresses on the water supply.