How much data do we need to assess future flood risk in data-scarce mega-cities?

Flood risk in urban areas is rising due to climate change and urbanization. Rapidly developing megacities located in the Global South are oftentimes referred to as risk hotspots with a particularly strong rise in risk. Therefore, it is essential to adapt these cities to rising risks and to minimize negative impacts for society, economy and the environment. Scenarios and simulation are popular tools to assess future trajectories of risk and the effectiveness of potential adaptation strategies. Yet, while state of the art simulation approaches are well suited to assess quantifiable flood risk processes in data-rich cities of the Global North, they often cannot be applied in data-scarce urban contexts of the Global South. Nevertheless, it is important to inform adaptation processes in these megacities, not just with respect to infrastructural development but also in terms of social adaptation measures. Here we therefore analyse how much data is actually needed in order to develop valid simulations of future flood risk and potential adaptation strategies in mega-cities with limited data availability. We apply a novel approach linking Bayesian networks with parameterized urban structure types. Bayesian networks are well suited as they allow the consideration of various data types and they are able to handle missing data. Urban structure types are used to make estimates of parameters where no data is available. Thus, we manage to represent different patterns of flood hydrology, exposure values, social vulnerability etc. We use Mumbai as a case study. The Bayesian network is built based on process understanding. For parameter learning we used three levels of data availability and validate it against observations to evaluate how much data we need to get reasonable results. In the first level we use only open-source data and expert knowledge to learn the parameters. Within the second level we further incorporate data obtained during a structured household survey and within the third level we also include data we obtained from local authorities in Mumbai. Results show that the quantity of data in general is oftentimes not decisive for the quality of representing large scale flood risk in Mumbai. Rather it is important to consider all site-specific parameters which are forming flood risk even if they are just quantified by expert estimates. We advocate for the exploration and implementation of adaptable and versatile approaches that enable the integration of both local insights and expert knowledge while moving away from models which are limited to measured data. These approaches would prove advantageous in conducting risk assessments within contexts where data availability is limited.