Towards global flood mapping onboard low cost satellites with machine learning

Gonzalo Mateo‐García, Joshua Veitch-Michaelis, Lewis Smith, Silviu Oprea, Guy Schumann, Yarin Gal, Atılım Güneş Baydin, Dietmar BackesScientific 2021 · DOI

Abstract

Spaceborne Earth observation is a key technology for flood response, offering valuable information to decision makers on the ground. Very large constellations of small, nano satellites—'CubeSats'—are a promising solution to reduce revisit time in disaster areas from days to hours. However, data transmission to ground receivers is limited by constraints on power and bandwidth of CubeSats. Onboard processing offers a solution to decrease the amount of data to transmit by reducing large sensor images to smaller data products. The ESA's recent PhiSat-1 mission aims to facilitate the demonstration of this concept, providing the hardware capability to perform onboard processing by including a power-constrained machine learning accelerator and the software to run custom applications. This work demonstrates a flood segmentation algorithm that produces flood masks to be transmitted instead of the raw images, while running efficiently on the accelerator aboard the PhiSat-1. Our models are trained on WorldFloods: a newly compiled dataset of 119 globally verified flooding events from disaster response organizations, which we make available in a common format. We test the system on independent locations, demonstrating that it produces fast and accurate segmentation masks on the hardware accelerator, acting as a proof of concept for this approach.

When a major flood hits, every hour counts for disaster response. While sophisticated satellites like the ESA’s Sentinel-2 provide high-quality data, they often have "revisit times" of several days, meaning a flooded region might not be imaged until the water has already receded. Large constellations of tiny, inexpensive "CubeSats" could solve this by providing hourly coverage.

However, CubeSats face a major physical constraint: downlink bandwidth. Sending high-resolution, multi-spectral images back to Earth requires significant power and time that these small satellites lack. Currently, it can take days to downlink, process, and deliver actionable flood maps to workers on the ground. This research proposes moving the "brain" to the satellite—using onboard machine learning to process images in orbit and sending back only the essential "flood masks," reducing data downlink requirements by two orders of magnitude.

Methods and Hardware Constraints

The researchers focused on semantic segmentation, a computer vision task where every pixel in an image is classified as land, water/flood, or cloud. To prove feasibility, they targeted the Intel Movidius Myriad 2 VPU, a low-power machine learning accelerator designed for edge devices, part of the ESA’s PhiSat-1 mission.

Comparison of spatial resolution and spectral response between Sentinel-2 and the CubeSat-compatible Hyperscout-2 sensor.

A significant challenge is the sensor discrepancy: the target CubeSat sensor (HyperScout-2) has a much lower spatial resolution (80m) than Sentinel-2 (10m). To account for this, the team intentionally degraded training data with bilinear interpolation, Gaussian noise, and motion blur to ensure the models would be robust to real-world CubeSat imagery. They compared standard remote sensing indices (like NDWI) against a Simple CNN (SCNN) and a more complex U-Net architecture.

The WorldFloods Dataset

A major contribution of this work is the release of WorldFloods, a "machine-learning ready" dataset designed to train global flood mapping models. It comprises 119 globally verified flooding events (2015–2019) totaling 422 flood extent maps. Data was collated from the Copernicus Emergency Management Service (EMS), UNOSAT, and the Global Flood Inundation Map Repository (GLOFIMR).

Global distribution of the 119 verified flooding events included in the WorldFloods dataset.

The dataset includes over 12 Gigapixels of labeled data (~266 GB), pairing 13-band Sentinel-2 imagery with rasterized labels for land, water, and clouds. The researchers found that training on "noisy" labels (derived from different satellites or acquired a few days apart from the imagery) actually improved the model's ability to generalize to new, unseen locations compared to training on a smaller set of "perfect" labels.

Key Results and Efficiency

  • The 100x Compression Advantage: By converting a 12-bit multi-channel hyperspectral image into a 2-bit flood mask onboard, the amount of data needing to be downlinked drops by a factor of 100. This turns a 13-minute transmission into a few seconds.
  • Speed: The SCNN model successfully processed a 12-megapixel image (a typical CubeSat frame) in less than one minute on the Myriad 2 hardware.
  • Accuracy: Deep learning models significantly outperformed traditional threshold-based methods. The U-Net achieved an Intersection over Union (IoU) of ~72% at 10m resolution, maintaining a high recall (over 94%) critical for disaster response.
  • Generalization: The models proved robust across diverse geographies, from the Philippines to Sicily, acting as a successful proof-of-concept for a global, automated flood-monitoring constellation.

Discussion and Future Work

While the approach offers massive speed gains, there is a trade-off: downlinking only the mask loses raw hyperspectral data that could be used for secondary analysis, such as detecting water pollutants. Current models also occasionally struggle with "dark" features, misclassifying cloud shadows or dense urban shadows as water.

Future work suggests that adding Digital Elevation Models (DEMs) as an extra input channel could help the model understand that water is unlikely to be found on steep slopes, further reducing false positives. Additionally, a hybrid approach where raw images are periodically downlinked to "fine-tune" the onboard models could maintain long-term accuracy.

Summary updated

Abstract

Spaceborne Earth observation is a key technology for flood response, offering valuable information to decision makers on the ground. Very large constellations of small, nano satellites- 'CubeSats' are a promising solution to reduce revisit time in disaster areas from days to hours. However, data transmission to ground receivers is limited by constraints on power and bandwidth of CubeSats. Onboard processing offers a solution to decrease the amount of data to transmit by reducing large sensor images to smaller data products. The ESA's recent PhiSat-1 mission aims to facilitate the demonstration of this concept, providing the hardware capability to perform onboard processing by including a power-constrained machine learning accelerator and the software to run custom applications. This work demonstrates a flood segmentation algorithm that produces flood masks to be transmitted instead of the raw images, while running efficiently on the accelerator aboard the PhiSat-1. Our models are trained on WorldFloods: a newly compiled dataset of 119 globally verified flooding events from disaster response organizations, which we make available in a common format. We test the system on independent locations, demonstrating that it produces fast and accurate segmentation masks on the hardware accelerator, acting as a proof of concept for this approach.
BibTeX
@inproceedings{mateogarcia2021towards,
  author = {Gonzalo Mateo‐García and Joshua Veitch-Michaelis and Lewis Smith and Silviu Oprea and Guy Schumann and Yarin Gal and Atılım Güneş Baydin and Dietmar Backes},
  title = {{Towards global flood mapping onboard low cost satellites with machine learning}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-86650-z},
}
Figure