Latent Neural Mapping for Hydrological Data Assimilation in Flood Prediction

Floods are one of the most frequent and severe natural disasters, and it is important to be prepared to predict them. Accurate prediction of floods requires the provision of accurate estimates of river discharge. Data assimilation (DA) as a technique for integrating background fields and observations can be a helpful solution to improve the accuracy of the river discharge prediction. DA can be a highly effective technology, however, when DA is performed on a large amount of data or high dimensional data, it results to be computationally very expensive, which is inappropriate for flood prediction, where timely results are required. Also, DA is used to merge data from diverse sources of information and, when the background fields and the observations are not from the same place, e.g. the observations are sparse, data must be interpolated on different grinds which increase the errors’ accumulation. In this work, latent neural mapping is designed to mitigate problems related to errors propagation and computational cost. We integrated DA with neural network (NN) and the resulting model helps on saving computational cost and solve the problem of sparse observation. Convolutional NN are employed to build a mapping function which converts data from the background space to the observation space (and vice versa). We tested the model with real data and flooding events in the UK. Data provided by the National River Flow Archive (NRFA) served as observations and the data provided by the European Flood Awareness System (EFAS) served as background fields. The Result shows that the accuracy is improved by 54.4% in MSE and the runtime of the model in 50s for 300 iterations.