An official website of the United States government
Abstract This study examines whether deep learning models can produce reliable future projections of streamflow under warming. We train a regional long short‐term memory network (LSTM) to daily streamflow in 15 watersheds in California and develop three process models (HYMOD, SAC‐SMA, and VIC) as benchmarks. We force all models with scenarios of warming and assess their hydrologic response, including shifts in the hydrograph and total runoff ratio. All process models show a shift to more winter runoff, reduced summer runoff, and a decline in the runoff ratio due to increased evapotranspiration. The LSTM predicts similar hydrograph shifts but in some watersheds predicts an unrealistic increase in the runoff ratio. We then test two alternative versions of the LSTM in which process model outputs are used as either additional training targets (i.e., multi‐output LSTM) or input features. Results indicate that the multi‐output LSTM does not correct the unrealistic streamflow projections under warming. The hybrid LSTM using estimates of evapotranspiration from SAC‐SMA as an additional input feature produces more realistic streamflow projections, but this does not hold for VIC or HYMOD. This suggests that the hybrid method depends on the fidelity of the process model. Finally, we test climate change responses under an LSTM trained to over 500 watersheds across the United States and find more realistic streamflow projections under warming. Ultimately, this work suggests that hybrid modeling may support the use of LSTMs for hydrologic projections under climate change, but so may training LSTMs to a large, diverse set of watersheds.
more »
« less
On the need for physical constraints in deep learning rainfall–runoff projections under climate change: a sensitivity analysis to warming and shifts in potential evapotranspiration
Abstract. Deep learning (DL) rainfall–runoff models outperform conceptual, process-based models in a range of applications. However, it remains unclear whether DL models can produce physically plausible projections of streamflow under climate change. We investigate this question through a sensitivity analysis of modeled responses to increases in temperature and potential evapotranspiration (PET), with other meteorological variables left unchanged. Previous research has shown that temperature-based PET methods overestimate evaporative water loss under warming compared with energy budget-based PET methods. We therefore assume that reliable streamflow responses to warming should exhibit less evaporative water loss when forced with smaller, energy-budget-based PET compared with temperature-based PET. We conduct this assessment using three conceptual, process-based rainfall–runoff models and three DL models, trained and tested across 212 watersheds in the Great Lakes basin. The DL models include a Long Short-Term Memory network (LSTM), a mass-conserving LSTM (MC-LSTM), and a novel variant of the MC-LSTM that also respects the relationship between PET and evaporative water loss (MC-LSTM-PET). After validating models against historical streamflow and actual evapotranspiration, we force all models with scenarios of warming, historical precipitation, and both temperature-based (Hamon) and energy-budget-based (Priestley–Taylor) PET, and compare their responses in long-term mean daily flow, low flows, high flows, and seasonal streamflow timing. We also explore similar responses using a national LSTM fit to 531 watersheds across the United States to assess how the inclusion of a larger and more diverse set of basins influences signals of hydrological response under warming. The main results of this study are as follows: The three Great Lakes DL models substantially outperform all process-based models in streamflow estimation. The MC-LSTM-PET also matches the best process-based models and outperforms the MC-LSTM in estimating actual evapotranspiration. All process-based models show a downward shift in long-term mean daily flows under warming, but median shifts are considerably larger under temperature-based PET (−17 % to −25 %) than energy-budget-based PET (−6 % to −9 %). The MC-LSTM-PET model exhibits similar differences in water loss across the different PET forcings. Conversely, the LSTM exhibits unrealistically large water losses under warming using Priestley–Taylor PET (−20 %), while the MC-LSTM is relatively insensitive to the PET method. DL models exhibit smaller changes in high flows and seasonal timing of flows as compared with the process-based models, while DL estimates of low flows are within the range estimated by the process-based models. Like the Great Lakes LSTM, the national LSTM also shows unrealistically large water losses under warming (−25 %), but it is more stable when many inputs are changed under warming and better aligns with process-based model responses for seasonal timing of flows. Ultimately, the results of this sensitivity analysis suggest that physical considerations regarding model architecture and input variables may be necessary to promote the physical realism of deep-learning-based hydrological projections under climate change.
more »
« less
- Award ID(s):
- 2144332
- PAR ID:
- 10533032
- Publisher / Repository:
- European Geosciences Union
- Date Published:
- Journal Name:
- Hydrology and Earth System Sciences
- Volume:
- 28
- Issue:
- 3
- ISSN:
- 1607-7938
- Page Range / eLocation ID:
- 479 to 503
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Abstract In dryland ecosystems, vegetation within different plant functional groups exhibits distinct seasonal phenologies that are affected by the prevailing hydroclimatic forcing. The seasonal variability of precipitation, atmospheric evaporative demand, and streamflow influences root-zone water availability to plants in water-limited environments. Increasing interannual variations in climate forcing of the local water balance and uncertainty regarding climate change projections have raised the potential for phenological shifts and changes to vegetation dynamics. This poses significant risks to plant functional types across large areas, especially in drylands and within riparian ecosystems. Due to the complex interactions between climate, water availability, and seasonal plant water use, the timing and amplitude of phenological responses to specific hydroclimate forcing cannot be determined a priori , thus limiting efforts to dynamically predict vegetation greenness under future climate change. Here, we analyze two decades (1994–2021) of remote sensing data (soil adjusted vegetation index (SAVI)) as well as contemporaneous hydroclimate data (precipitation, potential evapotranspiration, depth to groundwater, and air temperature), to identify and quantify the key hydroclimatic controls on the timing and amplitude of seasonal greenness. We focus on key phenological events across four different plant functional groups occupying distinct locations and rooting depths in dryland SE Arizona: semi-arid grasses and shrubs, xeric riparian terrace and hydric riparian floodplain trees. We find that key phenological events such as spring and summer greenness peaks in grass and shrubs are strongly driven by contributions from antecedent spring and monsoonal precipitation, respectively. Meanwhile seasonal canopy greenness in floodplain and terrace vegetation showed strong response to groundwater depth as well as antecedent available precipitation (aaP = P − PET) throughout reaches of perennial and intermediate streamflow permanence. The timings of spring green-up and autumn senescence were driven by seasonal changes in air temperature for all plant functional groups. Based on these findings, we develop and test a simple, empirical phenology model, that predicts the timing and amplitude of greenness based on hydroclimate forcing. We demonstrate the feasibility of the model by exploring simple, plausible climate change scenarios, which may inform our understanding of phenological shifts in dryland plant communities and may ultimately improve our predictive capability of investigating and predicting climate-phenology interactions in the future.more » « less
-
Abstract Process-based models of tree-ring width are used both for reconstructing past climates and for projecting changes in growth due to climate change. Since soil moisture observations are unavailable at appropriate spatial and temporal scales, these models generally rely on simple water budgets driven in part by temperature-based potential evapotranspiration (PET) estimates, but the choice of PET model could have large effects on simulated soil moisture, moisture stress, and radial growth. Here, I use four different PET models to drive the VS-Lite model and evaluate the extent to which they differ in both their ability to replicate observed growth variability and their simulated responses to projected 21st century warming. Across more than 1200 tree-ring width chronologies in the conterminous United States, there were no significant differences among the four PET models in their ability to replicate observed radial growth, but the models differed in their responses to 21st century warming. The temperature-driven empirical PET models (Thornthwaite and Hargreaves) simulated much larger warming-induced increases in PET and decreases in soil moisture than the more physically realistic PET models (Priestley–Taylor and Penman–Monteith). In cooler and more mesic regions with relatively minimal moisture constraints to growth, the models simulated similarly small reductions in growth with increased warming. However, in dry regions, the Thornthwaite- and Hargreaves-driven VS-Lite models simulated an increase in moisture stress roughly double that of the Priestley–Taylor and Penman–Monteith models, which also translated to larger simulated declines in radial growth under warming. While the lack of difference in the models’ ability to replicate observed radial growth variability is an encouraging sign for some applications (e.g. attributing changes in growth to specific climatic drivers), the large differences in model responses to warming suggest that caution is needed when applying the temperature-driven PET models to climatic conditions with large trends in temperature.more » « less
-
Abstract Water temperatures in mountain streams are likely to rise under future climate change, with negative impacts on ecosystems and water quality. However, it is difficult to predict which streams are most vulnerable due to sparse historical records of mountain stream temperatures as well as complex interactions between snowpack, groundwater, streamflow and water temperature. Minimum flow volumes are a potentially useful proxy for stream temperature, since daily streamflow records are much more common. We confirmed that there is a strong inverse relationship between annual low flows and peak water temperature using observed data from unimpaired streams throughout the montane regions of the United States' west coast. We then used linear models to explore the relationships between snowpack, potential evapotranspiration and other climate‐related variables with annual low flow volumes and peak water temperatures. We also incorporated previous years' flow volumes into these models to account for groundwater carryover from year to year. We found that annual peak snowpack water storage is a strong predictor of summer low flows in the more arid watersheds studied. This relationship is mediated by atmospheric water demand and carryover subsurface water storage from previous years, such that multi‐year droughts with high evapotranspiration lead to especially low flow volumes. We conclude that watershed management to help retain snow and increase baseflows may help counteract some of the streamflow temperature rises expected from a warming climate, especially in arid watersheds.more » « less
-
Hourly potential evapotranspiration at 0.1° resolution for the global land surface from 1981-presentAbstract Challenges exist for assessing the impacts of climate and climate change on the hydrological cycle on local and regional scales, and in turn on water resources, food, energy, and natural hazards. Potential evapotranspiration (PET) represents atmospheric demand for water, which is required at high spatial and temporal resolutions to compute actual evapotranspiration and thus close the water balance near the land surface for many such applications, but there are currently no available high-resolution datasets of PET. Here we develop an hourly PET dataset (hPET) for the global land surface at 0.1° spatial resolution, based on output from the recently developed ERA5-Land reanalysis dataset, over the period 1981 to present. We show how hPET compares to other available global PET datasets, over common spatiotemporal resolutions and time frames, with respect to spatial patterns of climatology and seasonal variations for selected humid and arid locations across the globe. We provide the data for users to employ for multiple applications to explore diurnal and seasonal variations in evaporative demand for water.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.5194/hess-28-479-2024
