Laboratoire d’hydrologie, climat et changements climatiques

Abstract. This study investigates the ability of long short-term memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A set of state-of-the-art, hydrological model-dependent regionalization methods are applied to 148 catchments in northeast North America and compared to an LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 % to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model's structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

It has become apparent in recent decades that river water temperature can have immediate and lasting impacts on aquatic organisms and their lotic habitat. In rivers that are dammed, there is an opportunity and a responsibility to regulate flows in order to control these temperatures to ensure the survival of the fish and other aquatic life. This paper uses a physically based hydraulic model (HEC-RAS) to run a water temperature component, allowing the thermal model to simulate water temperatures at the same hourly time step as the hydraulic model in a data-sparse region using two meteorological reanalysis datasets (ERA5 and ERA5-Land) as inputs allowing for a full representation of the diurnal cycle. This was achieved by making use of the HEC-RAS controller to automate the calibration and subsequent simulation processes. Results show that these products are able to provide high-quality thermal simulations on a 200 km river system in British Columbia, Canada, obtaining mean absolute errors in validation of 0.66 °C and a root mean square error of 0.84 °C. Some of the boundary conditions seemed to have little effect on downstream water temperatures. This is due to the measured point of interest being far enough downstream of the dam that a thermal equilibrium is reached well before. Simulations using shorter river reaches confirm that long lakes in the study region contribute to the thermal equilibrium being attained. There also seems to be a limit to the advantage conveyed by increased spatial density of the data, as results indicate a form of skill plateau after a certain input data density is attained.

This paper evaluates the impact of climate change on the water temperature of the Nechako River near the town of Vanderhoof (British Columbia, Canada). To do so, the Hydrologic Engineering Center’s River Analysis System (HEC-RAS) hydraulic and water temperature model was used with data from 10 climate models representing two Shared Socioeconomic Pathways (SSP 2-4.5 and SSP 5-8.5) over two future time periods (2041-2070 and 2071-2100). The results showed an upward trend in projected water temperatures for all tested discharge rates from the impounding reservoir during the warmest periods of the year. The study found that water temperatures are expected to increase by up to 2.57 °C for the near future (2041-2070) and up to 3.56 °C on average for all flow scenarios studied for a far future (2071-2100) when using SSP5-8.5. The timing of the peak water temperature during the summer is also expected to shift, with maximum water temperatures occurring up to 10 days later than in the reference period. In 10.3% of the far future SSP5-8.5 scenarios, at least one day per summer had a mean daily temperature of at least 24 °C, which exceeds limits of 20 °C for sockeye salmon and 21 °C for white sturgeon which are considered detrimental for the fish. It has been shown that over 50% of sockeye salmon will stop their sustained swimming at water temperatures of 24 °C due to cardiac limitations.

Abstract. Efficient adaptation strategies to climate change require estimating future impacts and the uncertainty surrounding this estimation. Over- or under-estimating future uncertainty may lead to maladaptation. Hydrological impact studies typically use a top-down approach in which multiple climate models are used to assess the uncertainty related to climate model structure and climate sensitivity. Despite ongoing debate, impact modelers have typically embraced the concept of "model democracy" in which each climate model is considered equally fit. The newer CMIP6 simulations, with several models showing a climate sensitivity larger than that of CMIP5 and larger than the likely range based on past climate information and understanding of planetary physics, have reignited the model democracy debate. Some have suggested that hot models be removed from impact studies to avoid skewing impact results toward unlikely futures. This large-sample study looks at the impact of removing hot models on the projections of future streamflow over 3,107 North American catchments. More precisely, the variability of future projections of mean, high, and low flows is evaluated using an ensemble of 19 CMIP6 GCMs, 5 of which are deemed "hot" based on their global equilibrium climate sensitivity (ECS). The results show that the reduced ensemble of 14 climate models provides streamflow projections with reduced future variability for Canada, Alaska, the Southwest US, and along the Pacific coast. Elsewhere, the reduced ensemble has either no impact or results in increased variability of future streamflow, indicating that outlier climate models do not necessarily provide outlier projections of future impacts. These results emphasize the delicate nature of climate model selection, especially based on global fitness metrics that may not be appropriate for local and regional assessments.

<strong class="journal-contentHeaderColor">Abstract.</strong> Efficient adaptation strategies to climate change require estimating future impacts and the uncertainty surrounding this estimation. Over- or under-estimating future uncertainty may lead to maladaptation. Hydrological impact studies typically use a top-down approach in which multiple climate models are used to assess the uncertainty related to climate model structure and climate sensitivity. Despite ongoing debate, impact modelers have typically embraced the concept of "model democracy" in which each climate model is considered equally fit. The newer CMIP6 simulations, with several models showing a climate sensitivity larger than that of CMIP5 and larger than the likely range based on past climate information and understanding of planetary physics, have reignited the model democracy debate. Some have suggested that hot models be removed from impact studies to avoid skewing impact results toward unlikely futures. This large-sample study looks at the impact of removing hot models on the projections of future streamflow over 3,107 North American catchments. More precisely, the variability of future projections of mean, high, and low flows is evaluated using an ensemble of 19 CMIP6 GCMs, 5 of which are deemed "hot" based on their global equilibrium climate sensitivity (ECS). The results show that the reduced ensemble of 14 climate models provides streamflow projections with reduced future variability for Canada, Alaska, the Southwest US, and along the Pacific coast. Elsewhere, the reduced ensemble has either no impact or results in increased variability of future streamflow, indicating that outlier climate models do not necessarily provide outlier projections of future impacts. These results emphasize the delicate nature of climate model selection, especially based on global fitness metrics that may not be appropriate for local and regional assessments. &emsp;

This study investigates the ability of Long Short-Term Memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A series of state-of-the-art, hydrological model-dependent regionalization methods is applied to 148 catchments in Northeast North America and compared to a LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model's structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

<strong class="journal-contentHeaderColor">Abstract.</strong> This study investigates the ability of Long Short-Term Memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A series of state-of-the-art, hydrological model-dependent regionalization methods is applied to 148 catchments in Northeast North America and compared to a LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model&rsquo;s structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

<strong class="journal-contentHeaderColor">Abstract.</strong> This study investigates the ability of long short-term memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A set of state-of-the-art, hydrological model-dependent regionalization methods are applied to 148 catchments in northeast North America and compared to an LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 % to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model's structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

<strong class="journal-contentHeaderColor">Abstract.</strong> Efficient adaptation strategies to climate change require estimating future impacts and the uncertainty surrounding this estimation. Over- or under-estimating future uncertainty may lead to maladaptation. Hydrological impact studies typically use a top-down approach in which multiple climate models are used to assess the uncertainty related to climate model structure and climate sensitivity. Despite ongoing debate, impact modelers have typically embraced the concept of "model democracy" in which each climate model is considered equally fit. The newer CMIP6 simulations, with several models showing a climate sensitivity larger than that of CMIP5 and larger than the likely range based on past climate information and understanding of planetary physics, have reignited the model democracy debate. Some have suggested that hot models be removed from impact studies to avoid skewing impact results toward unlikely futures. This large-sample study looks at the impact of removing hot models on the projections of future streamflow over 3,107 North American catchments. More precisely, the variability of future projections of mean, high, and low flows is evaluated using an ensemble of 19 CMIP6 GCMs, 5 of which are deemed "hot" based on their global equilibrium climate sensitivity (ECS). The results show that the reduced ensemble of 14 climate models provides streamflow projections with reduced future variability for Canada, Alaska, the Southwest US, and along the Pacific coast. Elsewhere, the reduced ensemble has either no impact or results in increased variability of future streamflow, indicating that outlier climate models do not necessarily provide outlier projections of future impacts. These results emphasize the delicate nature of climate model selection, especially based on global fitness metrics that may not be appropriate for local and regional assessments. &emsp;

This study investigates the ability of Long Short-Term Memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A series of state-of-the-art, hydrological model-dependent regionalization methods is applied to 148 catchments in Northeast North America and compared to a LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model’s structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

<strong class="journal-contentHeaderColor">Abstract.</strong> Efficient adaptation strategies to climate change require estimating future impacts and the uncertainty surrounding this estimation. Over- or under-estimating future uncertainty may lead to maladaptation. Hydrological impact studies typically use a top-down approach in which multiple climate models are used to assess the uncertainty related to climate model structure and climate sensitivity. Despite ongoing debate, impact modelers have typically embraced the concept of "model democracy" in which each climate model is considered equally fit. The newer CMIP6 simulations, with several models showing a climate sensitivity larger than that of CMIP5 and larger than the likely range based on past climate information and understanding of planetary physics, have reignited the model democracy debate. Some have suggested that hot models be removed from impact studies to avoid skewing impact results toward unlikely futures. This large-sample study looks at the impact of removing hot models on the projections of future streamflow over 3,107 North American catchments. More precisely, the variability of future projections of mean, high, and low flows is evaluated using an ensemble of 19 CMIP6 GCMs, 5 of which are deemed "hot" based on their global equilibrium climate sensitivity (ECS). The results show that the reduced ensemble of 14 climate models provides streamflow projections with reduced future variability for Canada, Alaska, the Southwest US, and along the Pacific coast. Elsewhere, the reduced ensemble has either no impact or results in increased variability of future streamflow, indicating that outlier climate models do not necessarily provide outlier projections of future impacts. These results emphasize the delicate nature of climate model selection, especially based on global fitness metrics that may not be appropriate for local and regional assessments. &emsp;

<strong class="journal-contentHeaderColor">Abstract.</strong> This study investigates the ability of long short-term memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A set of state-of-the-art, hydrological model-dependent regionalization methods are applied to 148 catchments in northeast North America and compared to an LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 % to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model's structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

<strong class="journal-contentHeaderColor">Abstract.</strong> Efficient adaptation strategies to climate change require estimating future impacts and the uncertainty surrounding this estimation. Over- or under-estimating future uncertainty may lead to maladaptation. Hydrological impact studies typically use a top-down approach in which multiple climate models are used to assess the uncertainty related to climate model structure and climate sensitivity. Despite ongoing debate, impact modelers have typically embraced the concept of "model democracy" in which each climate model is considered equally fit. The newer CMIP6 simulations, with several models showing a climate sensitivity larger than that of CMIP5 and larger than the likely range based on past climate information and understanding of planetary physics, have reignited the model democracy debate. Some have suggested that hot models be removed from impact studies to avoid skewing impact results toward unlikely futures. This large-sample study looks at the impact of removing hot models on the projections of future streamflow over 3,107 North American catchments. More precisely, the variability of future projections of mean, high, and low flows is evaluated using an ensemble of 19 CMIP6 GCMs, 5 of which are deemed "hot" based on their global equilibrium climate sensitivity (ECS). The results show that the reduced ensemble of 14 climate models provides streamflow projections with reduced future variability for Canada, Alaska, the Southwest US, and along the Pacific coast. Elsewhere, the reduced ensemble has either no impact or results in increased variability of future streamflow, indicating that outlier climate models do not necessarily provide outlier projections of future impacts. These results emphasize the delicate nature of climate model selection, especially based on global fitness metrics that may not be appropriate for local and regional assessments. &emsp;

<strong class="journal-contentHeaderColor">Abstract.</strong> This study investigates the ability of long short-term memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A set of state-of-the-art, hydrological model-dependent regionalization methods are applied to 148 catchments in northeast North America and compared to an LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 % to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model's structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.

<strong class="journal-contentHeaderColor">Abstract.</strong> This study investigates the ability of long short-term memory (LSTM) neural networks to perform streamflow prediction at ungauged basins. A set of state-of-the-art, hydrological model-dependent regionalization methods are applied to 148 catchments in northeast North America and compared to an LSTM model that uses the exact same available data as the hydrological models. While conceptual model-based methods attempt to derive parameterizations at ungauged sites from other similar or nearby catchments, the LSTM model uses all available data in the region to maximize the information content and increase its robustness. Furthermore, by design, the LSTM does not require explicit definition of hydrological processes and derives its own structure from the provided data. The LSTM networks were able to clearly outperform the hydrological models in a leave-one-out cross-validation regionalization setting on most catchments in the study area, with the LSTM model outperforming the hydrological models in 93 % to 97 % of catchments depending on the hydrological model. Furthermore, for up to 78 % of the catchments, the LSTM model was able to predict streamflow more accurately on pseudo-ungauged catchments than hydrological models calibrated on the target data, showing that the LSTM model's structure was better suited to convert the meteorological data and geophysical descriptors into streamflow than the hydrological models even calibrated to those sites in these cases. Furthermore, the LSTM model robustness was tested by varying its hyperparameters, and still outperformed hydrological models in regionalization in almost all cases. Overall, LSTM networks have the potential to change the regionalization research landscape by providing clear improvement pathways over traditional methods in the field of streamflow prediction in ungauged catchments.