Research map / Time series / Long-term forecasting
RiverMamba: A State Space Model for Global River Discharge and Flood Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RiverMamba, a bidirectional Mamba-based spatio-temporal model that pretrains on GloFAS reanalysis and fine-tunes on GRDC observations to forecast global daily river discharge and floods at 0.05° up to 7 days lead time.
Problem
Global river discharge and flood forecasting at high spatial resolution requires modeling spatio-temporal river network relations, which existing deep learning methods (mostly locally lumped LSTMs) do not do, while physics-based systems are expensive and need calibration.
Approach
Serializes sampled global river points into 1D sequences via space-filling curves (Sweep and Gilbert) so bidirectional Mamba blocks can capture long spatio-temporal receptive fields over river networks; forecast blocks fuse ECMWF HRES forecasts sequentially, with LOAN conditioning on static attributes and a return-period-weighted MSE loss.
Key results
- On GloFAS reanalysis test (2021-2024), RiverMamba reaches R2 0.8728, KGE 0.9125, F1 0.4589 versus LSTM at 0.8485, 0.8924, 0.3582.
- On GRDC gauged stations test (2021-2023), RiverMamba reaches R2 0.5057, KGE 0.6612, F1 0.2427, outperforming LSTM (0.4615, 0.6141, 0.1475) and GloFAS reforecast (0.2892, 0.4944, 0.2044).
- Pretraining took about 3 days on 16 GPUs and fine-tuning on GRDC about 4 hours on 16 GPUs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RiverMamba instead of LSTM | GloFAS reanalysis forecasting · GloFAS reanalysis test 2021-2024 | R2: 0.8485 ±0.0021 -> 0.8728 ±0.0013 | Helped |
| use RiverMamba instead of LSTM | GloFAS reanalysis forecasting · GloFAS reanalysis test 2021-2024 | KGE: 0.8924 ±0.0029 -> 0.9125 ±0.0008 | Helped |
| use RiverMamba instead of LSTM | GloFAS reanalysis forecasting · GloFAS reanalysis test 2021-2024 | F1: 0.3582 ±0.0058 -> 0.4589 ±0.0080 | Helped |
| fine-tune and use RiverMamba instead of LSTM | GRDC observational discharge forecasting · GRDC gauged stations test 2021-2023 | R2: 0.4615 ±0.0039 -> 0.5057 ±0.0028 | Helped |
| fine-tune and use RiverMamba instead of LSTM | GRDC observational discharge forecasting · GRDC gauged stations test 2021-2023 | KGE: 0.6141 ±0.0018 -> 0.6612 ±0.0010 | Helped |
| fine-tune and use RiverMamba instead of LSTM | GRDC observational discharge forecasting · GRDC gauged stations test 2021-2023 | F1: 0.1475 ±0.0014 -> 0.2427 ±0.0111 | Helped |
| add flood weighting (w-hat) (ablation) | RiverMamba loss · Europe validation | F1: 0.2236 -> 0.2859 | Helped |
| add lead-time weighting (u-hat) (ablation) | RiverMamba loss · Europe validation | KGE: 0.9086 -> 0.9136 | Helped |
| add lead-time weighting (u-hat) (ablation) | RiverMamba loss · Europe validation | F1: 0.2236 -> 0.2593 | Helped |
| add LOAN in hindcast blocks (ablation) | RiverMamba location embedding · Europe validation | KGE: 0.9183 -> 0.9160 | Hurt |
| add LOAN in hindcast blocks (ablation) | RiverMamba location embedding · Europe validation | F1: 0.2790 -> 0.2827 | Helped |
| add LOAN in forecast blocks (ablation) | RiverMamba location embedding · Europe validation | KGE: 0.9183 -> 0.9166 | Hurt |
Surprising findings
- Ablation shows that using LOAN in both hindcast and forecast blocks is needed to balance KGE and F1; hindcast-only or forecast-only LOAN increases F1 but lowers KGE.
- Reported that GRDC performance drops for all models relative to GloFAS reanalysis, attributed to unmodeled human water management such as dams and irrigation.
Limitations
- Assumes ERA5-Land is available at t-1, whereas operationally it is released with about 5-day delay.
- Observed discharge is affected by human interventions (dams, irrigation) that the model does not represent.
- No uncertainty estimation for forecasts; errors from weather forecasts, rare floods and data bias remain.
Experimental setup
Datasets: GloFAS-ERA5 reanalysis GRDC observational discharge ERA5-Land CPC precipitation ECMWF HRES LISFLOOD static maps
Benchmarks: GloFAS-ERA5 reanalysis test 2021-2024 GRDC gauged stations test 2021-2023
Metrics: R2 KGE F1 for flood return periods 1.5-20 years
Baselines: Persistence Climatology Encoder-Decoder LSTM GloFAS reforecast
Builds on: Mamba S4 LOAN Encoder-Decoder LSTM VMamba
Closest papers by meaning
- The Merit of River Network Topology for Neural Flood Forecasting (ICML 2024 · similarity 0.88)
- FIDLAR: Forecast-Informed Deep Learning Architecture for Flood Mitigation (AAAI 2025 · similarity 0.88)
- Pluvial Flood Emulation with Hydraulics-informed Message Passing (ICML 2024 · similarity 0.86)
- Learning from Polar Representation: An Extreme-Adaptive Model for Long-Term Time Series Forecasting (AAAI 2024 · similarity 0.84)
- Data-driven Precipitation Nowcasting Using Satellite Imagery (AAAI 2025 · similarity 0.83)
- Physics-Guided Fair Graph Sampling for Water Temperature Prediction in River Networks (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.