Source code: Learning the Kalman Update End-to-End for Hydrological Forecasting

This resource contains the source code and analysis artifacts accompanying the manuscript 'Learning the Kalman Update End-to-End for Hydrological Forecasting: A Unified Differentiable Framework' submitted to Water Resources Research. The framework couples the WALRUS lumped hydrological model with the Unscented Kalman Filter and learns the Kalman update end-to-end via two complementary approaches: (i) a learned-noise UKF that estimates diagonal process and observation noise covariances by backpropagation (6 parameters), and (ii) an LSTM-based Neural Gain Filter (LSTM-NGF) that learns the Kalman gain directly from data. The archive includes configuration files, training and evaluation scripts, and analysis utilities required to understand and reproduce the methodology. Experiments use hourly hydrometeorological data from the Regge catchment (Netherlands, 1999-2011), publicly available via 4TU.ResearchData (DOI: 10.4121/12717392.v2). Trained model checkpoints and processed windowed datasets are not included in this archive.