Long-Short Term Memory networks as observation operator for the states in a conceptual hydrological model

During the last decades, data assimilation has demonstrated its merit for updating hydrological models with remotely sensed observations. Generally, physically-based models are used as these contain model states that can effectively be observed. Yet, remote sensing data, such as microwave backscattering, often needs to be converted to these model states using an observation operator, which often is a physically-based retrieval algorithm. When using conceptual models, the problem becomes more complicated as the model states cannot be related to actual physical properties in the field. Because of this, the observation operator is often an empirical relation between a hydrological (state) variable and a model state. In this poster presentation we demonstrate the use of a Long-Short Term Memory (LSTM) network as alternative observation operator that allows to convert remotely sensed observations into model state estimations of the Probability Distributed Model (PDM). Therefore, Sentinel-1 observations averaged at the catchment scale and for each land use type within the catchment, along with other data sources (such as LAI, precipitation, …)  are fed to an LSTM in order to estimate a critical capacity of the probability distributed soil moisture reservoir. These data are then used to update the PDM through a classical data assimilation method (i.e. the ensemble Kalman Filter).