Getting the most out of satellite-based precipitation data for forecasting river floods with deep neural networks

Abstract:

Precipitation, being a major forcing of hydrological systems, is the main input to flood forecasting models. In recent years, machine learning models are increasingly utilized for flood prediction, as they can potentially accommodate complex relationships between precipitation and the resulting floods. The presented study focuses on space and time aspects related to precipitation input into deep neural networks. Specifically, we look at a spectrum of forms of precipitation data that can be served into the model. For example, precipitation data can be kept at their most detailed form (i.e., the highest resolution and the maximal extent in space and time), letting the model learn to predict floods with a maximal flexibility. Alternatively, pre-processing, for example by space-time averaging, could make the features more hydrological-relevant, less prone to overfitting errors and learning can be more effective and robust. Uncertainty, which can be substantial for precipitation data, is also reduced with space-time averaging. On the downside, in this process important information may be lost, and a shift towards the pre-processing efforts can eliminate machine learning advantages. We explore this trade-off from several points of view including prediction performance and interpretability.

The research is done as a part of Google’s flood forecasting initiative, focusing on large river watersheds in India. Near real-time satellite-precipitation datasets with a base resolution of up to 0.1 deg and 1 h and hourly stream gauge measurements are used for training and validation. Sensitivity of different performance metrics to space-time precipitation properties and to some network properties is examined. The generalization of the modeling process and the interpretability of the learned models are also discussed. Finally, we propose a procedure that can be followed to determine appropriate precipitation feature structure in modelling settings similar to the discussed above.