climate-weather/ai produced the result/Nature Geoscience 2026 · v2
AI fills gaps in river temperature records to track overlapping heatwaves
Researchers used a neural network to reconstruct daily river temperature and oxygen levels for 796 basins, then counted how often river heatwaves coincided with air heatwaves, and projected how that overlap changes to 2100.
spectrum · one line per step, placed by what the step does · bright lines used AI
Threefold increase in atmospheric–riverine compound heatwaves under climate change
Nature Geoscience, 2026
doi:10.1038/s41561-026-02040-y · record aix-00286 v2 · checked 2026-11-05
- AI was for
- Property prediction, Forecasting
- Model family
- Recurrent neural network, Gradient-boosted trees, Random forest
- Checked by
- Held-out796 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
A heatwave is usually thought of as a spell of unusually hot air. Rivers have them too: stretches of days when the water is far warmer than normal for the time of year. Warm water holds less dissolved oxygen, the oxygen that fish and other river life breathe, so hot spells in a river matter for what lives in it. The difficulty is in the measurements. Air temperature is recorded every day at thousands of stations, but river water temperature and oxygen are often sampled only now and then, perhaps when someone visits a gauging site. Without a continuous daily record you cannot say when a river heatwave started, how long it lasted or whether it lined up with a hot spell in the air above.
The researchers set out to build those continuous records and use them to study compound events, meaning days when a river heatwave and an atmospheric heatwave happen together. They covered 796 river basins in the United States and Central Europe over 1981 to 2019, looked at which features of a basin go with more frequent compound events, compared water temperature and oxygen under different conditions, and extended the picture forward to 2100 under two future emissions scenarios.
Where AI came in
The central tool was a long short-term memory network, or LSTM, a type of neural network built to learn patterns in sequences such as a time series. It was trained on the patchy existing samples of water temperature and dissolved oxygen, paired with daily weather and river data (air temperature, precipitation, discharge, snow water equivalent, surface pressure and wind speed) and with fixed characteristics of each basin. Three quarters of each basin's record was used for training and the remaining quarter held back for testing. The trained network was then run across all 796 basins to produce an unbroken day-by-day series. In effect it stood in for measurements that were never taken.
Everything that follows rests on those reconstructed series. The heatwave events themselves were picked out by fixed rules rather than by learning, using thresholds and a five-day window for deciding that two events coincided. Separate machine learning came in for the question of which basin features matter: a random forest procedure narrowed 197 candidate basin attributes down to 12, and a gradient boosted tree model, XGBoost, was fitted to annual compound event counts and inspected to see which predictors carried the most weight. The same LSTM was applied to downscaled output from ten climate models to carry the water temperature projections to 2100.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors trained a long short-term memory network on irregularly sampled water quality observations, hydrometeorological forcings and basin attributes to reconstruct continuous daily river water temperature and dissolved oxygen for 796 basins in the USA and Central Europe over 1981–2019. From these reconstructions they identified riverine heatwaves and their co-occurrence with atmospheric heatwaves, reporting an increase in compound events of about 0.40 events per decade, from 0.76 to about 2.3 events per year, and relative increases in riverine heatwave frequency, duration and intensity of 113.6%, 147.8% and 94.8% between 1981–1990 and 2010–2019. An XGBoost model attributed 59.2% of the importance for annual compound frequency to climatic predictors, 22.4% to topographic and 18.3% to hydrological ones. Compound events coincided with an additional 16% rise in water temperature and a further 2.9% decline in dissolved oxygen relative to isolated riverine heatwaves, and under SSP5-8.5 the proportion of riverine heatwave days occurring as compound events is projected to reach 98.5% by 2091–2100.
How AI was used
A cross-continental LSTM was fitted to each basin's record using a 75%/25% train/test split of the time series, taking daily forcings (air temperature, precipitation, discharge, snow water equivalent, surface pressure, wind speed) and static basin attributes as predictors of daily water temperature and dissolved oxygen, and was then run over all 796 basins to reconstruct continuous daily series for 1981–2019. Heatwave events were extracted from the reconstructed water temperature and the air temperature forcings by a rule-based percentile-threshold procedure in the heatwaveR package, and compound events were identified by an event-synchronisation rule using a 5-day lag window, with sensitivity tests at 1, 3, 5, 10, 15 and 20 days. A VSURF random-forest procedure reduced 197 HydroATLAS attributes to 12 predictors, and a grid-searched XGBoost model was fitted to annual compound event frequency, with gain importance, SHAP values, individual conditional expectation curves and partial dependence plots used to characterise driver contributions. Downscaled forcings from ten NEX-GDDP-CMIP6 models under SSP2-4.5 and SSP5-8.5 were processed at basin scale to drive daily projections of air and river water temperature from 2020 to 2100.
The shape of the work
Structural · the record, drawn
Assemble basin-scale observations, forcings and attributes
Obtaining raw data, whether by measurement, download or retrieval.
We assembled a basin-scale dataset for 796 river basins across the USA and CEwhere the paper describes this · verbatim
Train cross-continental LSTM on observed water quality
Fitting model parameters, including fine-tuning an existing model.
partitioning each basin’s time series into 75% for training and 25% for testingwhere the paper describes this · verbatim
Reconstruct continuous daily water temperature and dissolved oxygen
Running a trained model over new data to predict, classify or score. The AI stood in for unresolved measurement.
reconstructing continuous daily records (1981–2019) across 796 rivers in the USA and Central Europewhere the paper describes this · verbatim
Identify atmospheric, riverine and compound heatwave events
Cleaning, filtering, normalising or labelling data already obtained.
AHWs and RHWs were identified using the heatwaveR packagewhere the paper describes this · verbatim
Select predictors and attribute drivers of compound frequency
Extracting understanding from model behaviour.
we applied the XGBoost model for feature importance analysiswhere the paper describes this · verbatim
Project future air and water temperature under SSP scenarios
Running a trained model over new data to predict, classify or score.
Compare water quality and trends across heatwave conditions
Extracting understanding from model behaviour.
we examined WT and DO changes under three conditionswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The riverine and compound heatwave statistics that the paper reports are computed from daily water temperature and dissolved oxygen series reconstructed by the LSTM model, so the central findings rest on model output rather than on direct observation.
We used a trained cross-continental long short-term memory (LSTM) model to reconstruct daily WT and DO for 796 riverswhere the paper describes this · verbatim
partitioning each basin’s time series into 75% for training and 25% for testingwhere the paper describes this · verbatim
Code used to analyse data and develop the figures is available via GitHubwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- Version of Cross-continental LSTM for water temperature and dissolved oxygenWhich version of the model was used is not stated.
- Version of XGBoostWhich version of the model was used is not stated.
- Version of VSURF random-forest variable selectionWhich version of the model was used is not stated.
- What step 2 replacedThe paper gives no basis for what the AI stood in for.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00286, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). AI fills gaps in river temperature records to track overlapping heatwaves. Record aix-00286, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00286/