Latent data assimilation with non‐explicit observation operator in hydrology

Abstract Natural hazards can cause significant damage to human life and property. Among them, floods are one of the most severe and frequent natural disasters, making flood prediction crucial. River discharge is an essential factor in causing floods, so accurate and fast prediction of river discharge is crucial for flood mitigation. Data assimilation (DA) as a method of combining different sources of data (e.g., state field and observations) has the ability to estimate the possible states of river discharge. However, DA on high‐dimensional data such as river discharge can be computationally expensive. Furthermore, when the DA process lacks explicit mappings from the state field to the observations, DA cannot be conducted effectively. In this work, we design a latent neural mapping (LNM) in the form of a neural network (NN) as the observation operator and integrate this within a three‐dimensional variational data assimilation (3D‐Var) framework. By operating within latent space, the resulting approach helps mitigate computational costs and allows us to run DA within seconds despite the high‐dimensional data. In addition, several alternative NNs are employed to build mapping functions, which map data from the state space to the observation space (and vice versa), and benchmarked against the latent‐space‐based LNM approach. We test the LNM with real river discharge data from the UK and Ireland. The National River Flow Archive (NRFA) dataset provides the observations, and the data provided by a surrogate model from the European Flood Awareness System (EFAS) dataset serves as the state field. LNM outperforms the alternative methods in terms of accuracy and efficiency. The LNM developed can be applied to areas other than hydrology to integrate data efficiently with models.