How can hydrological models extrapolate to changing conditions?

Most hydrological models are tuned on past records and then asked to predict a future with a different climate, vegetation and land use. How can models be built and tested so that they stay trustworthy outside the conditions they were tuned to?

Question

How can hydrological models be adapted so that they extrapolate reliably to changing conditions, including changing vegetation dynamics, rather than only reproducing the period they were calibrated on? Changing conditions include a warmer or drier climate and altered land use.

Background

The authors of the twenty-three unsolved problems in hydrology note that most hydrologists would probably agree this requires a more process-based rather than calibration-based approach, because calibrated conceptual models do not usually extrapolate well, and probably also abandoning potential evapotranspiration in modelling evaporation [1].

What is known

• Klemeš proposed a hierarchy of operational tests, generalising the split-sample test, to check whether a model can be transposed to other climates and land uses [2].
• In 86 Australian catchments, many apparent failures of conceptual models in multi-year drought were false negatives of the split-sample test (155 of 282 cases under one criterion), so models may be more capable under change than thought [3].
• Long short-term memory (LSTM) networks trained on 531 US basins predicted streamflow in basins left out of training with a higher median Nash–Sutcliffe efficiency (0.69) than a conceptual model calibrated basin by basin (0.64) [4].
• Trained on 1960–1990 and tested on the warmer 2000–2023, an LSTM was the most accurate model overall but its performance shifted more between periods than that of a conceptual model or a hybrid of the two, which generalised better [5].

What would count as progress

No single result settles a question like this; contributions like these would move it forward:

• A benchmark of differential split-sample tests on open large-sample datasets such as CAMELS and Caravan: calibrate in one climate regime, evaluate in a contrasting one, with a fixed protocol and scores.
• Models with explicit vegetation dynamics (such as rooting depth or stomatal response to CO₂), checked against catchments that went through drought, fire or land-cover change.
• Machine-learning and hybrid models evaluated for physical realism under warming, such as their runoff-ratio and evaporation response, not only for accuracy in the training climate.

References

1. Blöschl, G. et al. (2019). Twenty-three unsolved problems in hydrology (UPH) – a community perspective. Hydrological Sciences Journal 64(10), 1141–1158.
2. Klemeš, V. (1986). Operational testing of hydrological simulation models. Hydrological Sciences Journal 31(1), 13–24.
3. Fowler, K. J. A. et al. (2016). Simulating runoff under changing climatic conditions: revisiting an apparent deficiency of conceptual rainfall-runoff models. Water Resources Research 52(3), 1820–1846.
4. Kratzert, F. et al. (2019). Toward improved predictions in ungauged basins: exploiting the power of machine learning. Water Resources Research 55(12), 11344–11354.
5. Bohl, J. P. et al. (2026). Hybrid models generalize better to warmer climate conditions than process-based and purely data-driven models. Hydrology and Earth System Sciences 30(14), 4667–4698.

Source

Imported from Twenty-three unsolved problems in hydrology, question 19, curated by Blöschl et al. Listed as open in the source (2019) and still an active research question as of 2026-10-20. The statement above is written in our own words; see the source for the original wording, history and discussion.