ML4CC - An ML-ops Toolbox for Flooding

Applying machine learning (ML) to Earth observation (EO) data gives us the ability to better make predictions about how to adapt and mitigate our changing climate. ML4CC is a new public ML Toolbox for flood segmentation and mapping comprising an open repository of artificial intelligence tools such as enhanced, simulated and labelled geospatial data and advanced machine learning modules. ML4CC is dedicated to simplifying ML production and validation and ultimately improving climate related decision-making within the UK.

In 2020, with the support of the UK Space Agency, Trillium worked with its partners to create the first toolbox for floods: ML4Floods.

UNOSAT were the first organization to test and use the tools.

All of the tools are available on:

Flood extent segmentation over a time series of satellite images from Albania.

Understanding and using the tools

The figure below presents an overview of the ML4Floods toolkit alongside the users of each component (click for a larger version). The toolkit is structured as an end-to-end pipeline with components that 1) ingest, sort and organise satellite data, integrating ground-truth masks, 2) tile, augment and normalise the data, 3) train new models on the data, or run existing models on new data, and display uncertainty maps generated by the models, and 4) query and visualise the results via a web-based mapping application.

Each of the components 1-3 can be accessed via an application programming interface (API) so that technical users can fine-tune their workflows, or adopt the components in their own tools. However, the teams also developed a graphical interface (4) that can run the toolkit through a simple point-and-click interface. This last component places the power of ML-enhanced flood segmentation models in the hands of ordinary users, like disaster relief coordinators and urban planners. Finally, the graphical tool is incredibly useful for machine-learning researchers, allowing them to quickly compare and contrast model results on the same interface, greatly speeding up the model development process.