Many co-risk participants expressed an interest in the machine learning session, so the session was split over the final two weeks of the conference, and a number of sub-projects were proposed. The broad range of knowledge and expertise at the conference ensured that everyone was able to contribute and learn something new from the other participants. The sub-projects focused mainly on methods to provide general information and tools for practitioners coming from either machine learning or flood modelling disciplines. However, the development of data-driven models are also being planned, aided by the acquisition of a high-spec Tensor Notebook. In general, the group was interested, enthusiastic, engaged and keen to contribute and learn. Summaries of the separate sub-sessions are given below;
Application of deep learning for flood extent detection:
This group is attempting to develop deep learning architecture to estimate flood extents using microwave remote sensing in an automated fashion. Flood labels are required to run deep learning models to estimate flood extents. Flood labels is hard to generate and is a perfect candidate for community mapping. The project involves digitizing 4 to 5 land cover classes including flooded areas. The input data is a Digital Globe imagery for a flood in Phitsanulok in 2017. This flood event is selected because it has both microwave and optical data with one-day difference. Microwave remote sensing data is used as input to deep learning models for which labels are created using optical data in a QGIS environment.
Unsupervised algorithms for hazard consistent flood maps selection:
Probabilistic flood risk analysis often requires multiple simulations of complex systems, modelling all possible potential events. This is usually computationally infeasible, thus we propose to extract a reduced set of all potential floodmaps without losing statistical information about the risk. One benefit of this would be to reduce the required calculations in the damage analysis. This can hopefully be achieved using well-known machine learning techniques such as clustering. The project will kick-off in the final week.
Clustering of cities for spatial transfer of damage models
This project was a general ambition to help in the selection of damage models, thus making them more applicable to the local situation. We are planning to do this by introducing a distance measure between regions/cities that reflects the difference in urban physical characteristics. The mixture of flooding, machine learning and urban growth modelling expertise in the group greatly helps this project. We are hopeful of being able to produce a tool to help flood impact practitioners in choosing appropriate damage models in data-scarce regions around the world.
Global flood damage data collection
Disaster damage data is often not available for damage modelling. This session is about how damage data could be collected and made available to damage modellers. This could include new damage collection procedures or a way to get already collected damage data to modellers. In this session, we will brainstorm ways to organize this and then feed the conclusions towards the vision paper.
SLEUTH modeling for Chiang Mai
SLEUTH is an urban growth model which allows future urban land use maps to be predicted. The model is based on the concept of cellular automata, and uses a number of inputs such as land-use, DEM data and road networks. In the context of flooding, it can be used to estimate exposure and therefore damage for the assessment of potential future damage losses. This sub-group are attempting to build such a model for Chiang Mai, and are being helped by a professor from the local university. Thanks to the GIS expertise and contributions from local partners, we are now hopeful of producing urban growth predictions for Chiang Mai by the end of the conference. Vision paper: Machine learning for flood risk in the next 5 to 10 years
The vision paper is an ambitious project to put together an opinion/vision paper about challenges and opportunities of machine learning for flood assessment. Prior to the conference, the scientific journal NHESS (Natural Hazards and Earth Systems Science) was approached, and they invited the group to submit an ‘Invited perspectives’ article type on the future of machine learning for flood assessment’. This focused the group on a scope for the paper, and they were further split into looking into applications, opportunities and challenges for the different components of flood risk analysis; Hazard, exposure and Impact. We are still hopeful of putting together a draft by the final week. Watch this space!
