Towards urban flood susceptibility mapping using machine and deep learning models (part1)

In this series of articles, we will show the application of data-driven models in urban pluvial flood mitigation. Urban pluvial flood is caused by short intensive rainfall events that exceed the capacity of the storm drainage system. It could occur anywhere even in areas without flooding history. There are three types of flood maps as shown in the figure below:

1. Flood susceptibility maps: they show the probability of flooding at a given location based on its physical characteristics.
2. Flood inundation maps: they identify the flood extent after or during a flood event. They are obtained from satellite images or drones.
3. Flood hazard maps: they show the spatial distribution of flood variables such as velocity, water depth, time of flooding, etc. They are obtained from hydrodynamic models.

Figure. Different types of flood maps (Bentivoglio et al.,2022)

While performing hydrologic modelling followed by 1D-2D hydrodynamic modelling is the best representation of the physical process of runoff generation and concentration in a complex urban watershed. These models are computationally expensive to be applied on a city scale and hence their application is limited to small study areas.

Data-driven models try to find a relationship between the input and output datasets (see Figure 2). They are raising as a surrogate to the complex hydrodynamic models. However, their major challenge is to generalise to a study area that was not included in the training dataset. Furthermore, they are considered as a black-box, they have a good performance but we are not able to understand how the model is working.

Work flow of data-driven model

In the next articles, we will address the following:

1. We will compare deep learning models with traditional machine learning algorithms for flood susceptibility and flood hazard mapping.
2. We will evaluate the models' transferability in space.
3. We will use some methods to try to explain these black-box models.

References:

Bentivoglio, R., Isufi, E., Jonkman, S. N., and Taormina, R.: Deep Learning Methods for Flood Mapping: A Review of Existing Applications and Future Research Directions, Hydrol. Earth Syst. Sci. Discuss. [preprint], https://doi.org/10.5194/hess-2021-614, 2021.