FloodWatch Beta — data may be periodically unavailable.
AI Research

Training towards best-in-class flood prediction

We're building deep learning models that forecast river levels across multiple time horizons, trained on gauge, rainfall, and temperature data from across Great Britain.

National-scale training data

Our models learn from multi-agency flood data spanning England, Scotland, and Wales, combining river gauge readings, radar rainfall, catchment boundaries, and river network topology.

1,600+
River gauge stations

Environment Agency, SEPA, and Natural Resources Wales

1km
Radar rainfall

Met Office NIMROD composite, catchment-aggregated

GB
Temperature data

Station-level temperature as a model input feature

2,103
Catchment boundaries

NRFA topographic drainage polygons covering Great Britain

192,000
River network segments

OS Open Rivers topology with flow direction

NWP
Rainfall nowcasting

Forecast rainfall as a future input for longer-horizon prediction

Model architectures

We train and evaluate multiple deep learning architectures to find the best approach for each forecast horizon, from 1-hour nowcasts to 24-hour outlooks.

Recurrent models

GRU and LSTM networks trained on sequential gauge and rainfall time series with multi-horizon forecasting from 1 hour to 24 hours ahead.

Production target Q2 2026

Graph neural networks

GNNs that model flood propagation along real river topology. Message-passing between upstream and downstream stations captures spatial flood dynamics.

Active R&D

Next-generation architectures

Temporal Fusion Transformers, N-BEATS, physics-informed hybrid models, and GNN-GRU ensembles currently in evaluation for multi-horizon forecasting.

Planned

Research directions

Beyond core flood prediction, we're expanding into complementary areas that will enhance forecast accuracy and enable new capabilities.

Satellite flood extent detection

Computer vision models analysing Sentinel-1 SAR and Sentinel-2 optical imagery to detect and map flood extents in near real-time, independent of cloud cover.

AI-accelerated hydraulic simulation

Training neural network surrogates to approximate 1D/2D hydraulic models at a fraction of the computational cost, enabling real-time flood inundation mapping.

National terrain processing

Assembling high-resolution DTM and DSM data from LiDAR surveys across England, Scotland, and Wales for terrain-aware flood modelling and surface water routing.

Ensemble forecasting

Combining predictions from multiple model architectures to produce calibrated probabilistic forecasts with uncertainty bounds for each station and horizon.

Built on GPU-accelerated infrastructure

Our training and inference pipeline is built on the NVIDIA AI platform, from data preprocessing through to production model serving.

NVIDIA Inception Program member badge

FloodWatch.UK is a member of the NVIDIA Inception program, which supports startups building AI-driven solutions. Membership accelerates our access to GPU infrastructure, technical expertise, and co-development opportunities.

Training

Active
  • CUDA & cuDNN — GPU-accelerated training across A100 and H100 hardware
  • PyTorch with AMP — Automatic mixed precision for faster training and lower memory usage
  • PyTorch Geometric — GNN training on river network graph topology

Data processing

Integrating
  • RAPIDS cuGraph — GPU-accelerated river network graph analytics
  • RAPIDS cuDF & cuSpatial — GPU dataframes for rainfall data and catchment spatial processing
  • NVIDIA DALI — GPU data loading pipeline for satellite imagery and radar data

Production inference

Planning
  • NVIDIA Triton Inference Server — Serving multiple concurrent models with batched GPU inference across 1,600+ stations
  • NVIDIA TensorRT — Model optimisation for low-latency nationwide inference every 15 minutes

Future research

Vision
  • NVIDIA Modulus — Physics-informed neural networks for hydraulic simulation surrogates
  • NVIDIA TAO Toolkit — Transfer learning for satellite flood extent detection

© 2025 NVIDIA, the NVIDIA logo, CUDA, cuDNN, DALI, DGX, Modulus, RAPIDS, TAO Toolkit, TensorRT, and Triton Inference Server are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. PyTorch is a trademark of The Linux Foundation.

Where we are

We're actively training and evaluating models across multiple architectures and forecast horizons. Results will be published when our evaluation programme is complete.

21
Training runs completed
3+5
Architectures in R&D

GRU, LSTM, GNN — plus TFT, N-BEATS, Transformer, hybrid ensemble, and more in the pipeline

5
Forecast horizons

1h, 2h, 4h, 8h, 24h

Split
Archetype models

Specialist sub-models per catchment type: peat, urban, clay, chalk, mixed

Published results

Coming soon

Collaboration

We're interested in research partnerships with universities, government agencies, and organisations working on flood resilience in the UK. If you're working in hydrology, climate science, or environmental AI, we'd like to hear from you.

Get in touch