Abstract

Recent deep learning approaches for river discharge forecasting have improved the accuracy and efficiency in flood forecasting, enabling more reliable early warning systems for risk management. Nevertheless, existing deep learning approaches in hydrology remain largely confined to local-scale applications and do not leverage the inherent spatial connections of bodies of water. Thus, there is a strong need for new deep learning methodologies that are capable of modeling spatio-temporal relations to improve river discharge and flood forecasting for scientific and operational applications. To address this, we present RiverMamba, a novel deep learning model that is pretrained with long-term reanalysis data and that can forecast global river discharge and floods on a 0.05° grid up to 7 days lead time, which is of high relevance in early warning. To achieve this, RiverMamba leverages efficient Mamba blocks that enable the model to capture global-scale channel network routing and enhance its forecast capability for longer lead times. The forecast blocks integrate ECMWF HRES meteorological forecasts, while accounting for their inaccuracies through spatio-temporal modeling. Our analysis demonstrates that RiverMamba delivers reliable predictions of river discharge, including extreme floods across return periods and lead times, surpassing both operational AI- and physics-based models.

RiverMamba at NeurIPS 2025 [in preparation...]

Video

RiverMamba at AI for Good

-->

Overeview

Overview of the task

Overview of the objective of this work. We are interested in medium-range river discharge and flood forecasting. For this we aim to build a vision model to consider the spatial relations between points at and near bodies of water. To address uncertainty in the meteorological forcing, we aim to build forecast layers so that they can, for each catchment point, incorporate information about meteorological forcing from the neighboring points and throughout the temporal dimension.

Model architecture

model design

An overview of the proposed RiverMamba model for river discharge forecasting. The model forecasts at time t, high-resolution river discharge maps (\(\mathbf{X}^{t+1:t+L}_{dis24}\)) from initial conditions (\(\mathbf{X}^{t−T:t−1}_{ERA5}, \mathbf{X}^{t−T:t−1}_{GloFAS}, \mathbf{X}^{t−T−1:t−2}_{CPC}\)), static river attributes (\(\mathbf{X}_{static}\)), and meteorological forecasts (\(\mathbf{X}^{t+1:t+L}_{HRES}\)).

model design

The structure of the hindcast block and forecast block. Both use a Bidirectional Mamba block and the forecast block has the same structure as the hindcast block, but it additionally incorporates meteorological forecasts HRES by concatenation. The forecast block also includes LOAN layers although it is not shown in the previous figure.

Serialization

We use Space-filling Curves to connect the points or pixels into one sequence. A space filling curve can be defined as a bijective function \(\Phi: \mathbb{Z}^3 \rightarrow \mathbb{N} \) where every point in the discrete space corresponds to a unique index within one sequence. By altering the curves sequentially through the netowk's blocks, the points or pixels will be connected and scanned from diverse spatial perspectives, enabling RiverMamba to capture different contextual features.

Comparison to the baselines 1
Comparison to the baselines 2

Comparison to the baselines on GloFAS-Reanalysis

Here we compare on 3366 gauged stations (2021-2024) using reanalysis data. LSTM is based on the encoder-decoder model developed by Google.

Comparison to the baselines 1

Results on GloFAS-Reanalysis

We compare daily flood severity maps from GloFAS reanalysis as a ground truth and RiverMamba model at extreme flood events with different causes in 2024 and from different places around the Earth. These maps are usually used in the operational flood forecast service like GloFAS to provide a quick overview of the ongoing and upcoming flood events.

Results on GRDC observations

Here we compare on 3366 gauged stations (2021-2023) using GRDC observational data. LSTM is based on the encoder-decoder model developed by Google and GloFAS is the operational system operated by ECMWF.

Comparison to the baselines 2

Results on GRDC observations

River discharge of the Sauer river in 2021. The grey shaded area highlights the Germany flood between July 10 to July 20 in 2021.

Comparison to the baselines 2

Poster

Poster

BibTeX



			@article{rivermamba,
					  title={RiverMamba: A State Space Model for Global River Discharge and Flood Forecasting},
					  author={Shams Eddin, Mohamad Hakam and Zhang, Yikui and Kollet, Stefan and Gall, Juergen},
					  journal={arXiv preprint arXiv:2505.22535},
					  year={2025}
					  }

				  

Dataset, reforecasts and pretrained models

-->