Researchers at the School of Informatics have developed DeepSpace, a new system that overcomes the major challenges of satellite image data transfer and storage. Using deep learning-based super-resolution, DeepSpace can compress images by more than 100 times in real time on board satellites, without losing accuracy. The breakthrough reduces transfer delays from days to minutes, cuts the cost of storing vast Earth observation datasets, and enables faster, more reliable applications such as wildfire detection, plastic pollution monitoring, and land-use mapping. Pelican-2 rendering. [Image credit: Planet Labs PBC] Large groups of low-earth orbit satellites allow for frequent high-resolution earth imaging for numerous geospatial applications, such as environmental monitoring, navigation (GPS), weather forecasting, disaster management and mapping. They generate hundreds of Terabytes of data per day in space, which must then be transferred to Earth through constrained intermittent connections to ground stations. These large volumes lead to large day-level delay in data download and exorbitant cloud storage costs. What are the challenges for earth imagery and storage? Efficient high-quality satellite image data storage and processing is incredibly challenging due to the large volumes of data, downlink bottlenecks, increasing cloud storage costs and the limited capabilities of satellites to run computations on board. Constellations of satellites generate hundreds of Terabytes (TB) of data per day. Due to their low orbit, they have intermittent connectivity to ground stations, with limited bandwidth connections. This causes transfer delays of several hours to a few days, heavily impacting time-sensitive applications such as natural disaster monitoring. Further to this, the growth in satellite deployments continues to outpace space-earth data capacity, implying that in future only a small fraction of the data can be downloaded. The cost of storing Earth observation data is increasingly expensive. The study estimates that it costs millions of dollars each month to store and make available a few years’ worth of data. The longer a satellite constellation operates, and with the increase in the number of constellations and image sizes increasing over time, this cost will continue to rise. There are also limited onboard compute resources available on satellites. Satellites have small GPUs which are further limited by power availability which prioritises mission-critical satellite operations. It is non-trivial to accommodate compute-intensive compression algorithms on the satellite. The study evaluated existing end-to-end compression systems and found that there were no existing methods which successfully addressed all these challenges. DeepSpace, a novel satellite image data acquisition system Researchers have proposed DeepSpace, a new system designed to meet each of these three challenges. For the first time in satellite image data collection, DeepSpace uses the deep learning-based super-resolution (SR) approach. DeepSpace works by compressing satellite images by hundreds of times in real-time onboard satellites, whilst ensuring the accuracy and fidelity of the images. DeepSpace can perform lightweight computation onboard the satellite by shifting the compute burden to the cloud, where images can then be decompressed on demand. The study extensively tested and evaluated DeepSpace against a wide range of state-of-the-art baselines and considered multiple satellite image datasets to both support and evidence its use and application. They were able to demonstrate that reconstructed images with DeepSpace, even after more than 100x compression, yield similar application performance as with the ‘ground truth’ raw uncompressed images. This means far less bandwidth is needed to move data from space to Earth, reducing data transfer delays and lowering the cost of storing and maintaining data archives. Practical impact Using the reconstructed imagery from using DeepSpace, rather than the raw ‘ground truth’ images has many practical advantages. The researchers tested DeepSpace against a range of real-world scenarios to demonstrate its potential to transform Earth observation: Wildfire detection – DeepSpace outperformed existing methods on both response time and accuracy of detection, cutting response times from days to less than 30 minutes. Plastic detection in oceans – Detecting plastic in oceans requires a very high level of image reconstruction. DeepSpace maintained high accuracy while achieving compression ratios still 50 times greater than existing methods, where other methods failed at preserving or recovering detailed pixel-level information. Land use and cropland classification – Land use measurement prioritises accuracy and cost over response speed. DeepSpace achieved a cropland classification performance like that achieved using ground truth raw data. Towards sustainable planetary intelligence As satellite deployments continue to increase, Space-to-Earth data capacity and latency will remain a limiting factor. Systems like DeepSpace demonstrate how machine learning can reshape satellite operations, enabling more sustainable, efficient, and accessible Earth observation. The research team also plan to make the DeepSpace implementation publicly available, accelerating innovation across the global Earth Observation community. Related links DeepSpace on GitHub Link to paper 'DeepSpace: Super Resolution Powered Efficient and Reliable Satellite Image Data Acquisition' Link to Professor Mahesh Marina's personal website Link to Chuanhao Sun's IEEE Xplore website Publication date 06 Oct, 2025
