ML4Floods

ML4Floods looks at a satellite photo taken during a flood and marks which pixels are water. It works on optical images, the ordinary colour and infrared pictures taken by Sentinel-2 and Landsat.^1,2 The package covers the whole chain in Python. It fetches images and existing flood maps, trains a neural network, runs it on a new scene and shows the result on an interactive map.²

The work began with the 2019 Frontier Development Lab Europe team on disaster response, and went on with funding from the UK Space Agency, led by Trillium Technologies.^1,2 The current models and dataset come from a paper by Enrique Portalés-Julià, Gonzalo Mateo-García, Cormac Purcell and Luis Gómez-Chova, published in Scientific Reports in November 2023,^2,6 and are hosted on Hugging Face by the Image and Signal Processing group (ISP).^2,4,5 The models have also been run on board a D-Orbit satellite in a set of experiments.²

The training set is called WorldFloods. Version 2 holds 509 Sentinel-2 images of floods between 2016 and 2023, each paired with a mask drawn mostly from Copernicus emergency flood maps and corrected by hand.³ Each mask has two layers: one says whether a pixel was clear or cloudy, the other whether it was land or water.³ The trained models follow the same split, so a map shows where cloud hid the ground rather than calling it dry.^2,4 The Hugging Face dataset card puts the download at about 76 GB.^3,5 Tutorials walk through a full flood map of the Kherson dam break and a run over a time series of Sentinel-2 images.^1,2

For a model of a place, a map of where water stood in a past flood is a check on a flood model's predictions. It also gives a quick picture of a new flood when the next clear image comes in. Optical satellites cannot see through cloud, and floods tend to come with it; the radar-based AI4G Flood works through cloud and at night.⁷

The code is under the LGPL-3.0 licence.¹ The WorldFloods dataset and all the pre-trained models are under Creative Commons Attribution-NonCommercial 4.0, which does not allow commercial use.^1,4,5