Improved streamflow simulations in hydrologically diverse basins using physically-informed deep learning models

Improved streamflow simulations in hydrologically diverse basins using physically-informed deep learning models. Magotra, B., Saharia, M., & Dhanya, C. T. Hydrological Sciences Journal, February, 2025.

Paper doi abstract bibtex 2 downloads

Physically-informed deep learning models, especially long short-term memory (LSTM) networks, have shown promise in large-scale streamflow simulations. However, an in-depth understanding of the relative contribution of physical information in deep learning models has been missing. Using a large-sample testbed of 220 catchments in hydrologically diverse regions of the Indian sub-continent, we quantify the impact of incremental addition of physical information on model performance using multiple variants of the LSTM model based on various combinations of static catchment attributes and simulated land surface states. We found that LSTM models trained with catchment geophysics as additional input outperformed the base LSTM model in terms of the nationwide median Kling-Gupta efficiency (KGE) of in-sample catchments, increasing KGE from 0.32 to 0.60. Additionally, the model retained significant prediction skill in out-of-sample catchments, demonstrating that a pre-trained LSTM model can be a powerful tool to predict streamflow in data-scarce regions.

@article{magotra_improved_2025,
	title = {Improved streamflow simulations in hydrologically diverse basins using physically-informed deep learning models},
	issn = {0262-6667, 2150-3435},
	url = {https://www.tandfonline.com/doi/full/10.1080/02626667.2025.2458545},
	doi = {10.1080/02626667.2025.2458545},
	abstract = {Physically-informed deep learning models, especially long short-term memory (LSTM) networks, have shown promise in large-scale streamflow simulations. However, an in-depth understanding of the relative contribution of physical information in deep learning models has been missing. Using a large-sample testbed of 220 catchments in hydrologically diverse regions of the Indian sub-continent, we quantify the impact of incremental addition of physical information on model performance using multiple variants of the LSTM model based on various combinations of static catchment attributes and simulated land surface states. We found that LSTM models trained with catchment geophysics as additional input outperformed the base LSTM model in terms of the nationwide median Kling-Gupta efficiency (KGE) of in-sample catchments, increasing KGE from 0.32 to 0.60. Additionally, the model retained significant prediction skill in out-of-sample catchments, demonstrating that a pre-trained LSTM model can be a powerful tool to predict streamflow in data-scarce regions.},
	language = {en},
	urldate = {2025-03-04},
	journal = {Hydrological Sciences Journal},
	author = {Magotra, Bhanu and Saharia, Manabendra and Dhanya, C. T.},
	month = feb,
	year = {2025},
	pages = {1--14},
}

Downloads: 2

{"_id":"yCrTvuxiWqLfmvGQ7","bibbaseid":"magotra-saharia-dhanya-improvedstreamflowsimulationsinhydrologicallydiversebasinsusingphysicallyinformeddeeplearningmodels-2025","author_short":["Magotra, B.","Saharia, M.","Dhanya, C. T."],"bibdata":{"bibtype":"article","type":"article","title":"Improved streamflow simulations in hydrologically diverse basins using physically-informed deep learning models","issn":"0262-6667, 2150-3435","url":"https://www.tandfonline.com/doi/full/10.1080/02626667.2025.2458545","doi":"10.1080/02626667.2025.2458545","abstract":"Physically-informed deep learning models, especially long short-term memory (LSTM) networks, have shown promise in large-scale streamflow simulations. However, an in-depth understanding of the relative contribution of physical information in deep learning models has been missing. Using a large-sample testbed of 220 catchments in hydrologically diverse regions of the Indian sub-continent, we quantify the impact of incremental addition of physical information on model performance using multiple variants of the LSTM model based on various combinations of static catchment attributes and simulated land surface states. We found that LSTM models trained with catchment geophysics as additional input outperformed the base LSTM model in terms of the nationwide median Kling-Gupta efficiency (KGE) of in-sample catchments, increasing KGE from 0.32 to 0.60. Additionally, the model retained significant prediction skill in out-of-sample catchments, demonstrating that a pre-trained LSTM model can be a powerful tool to predict streamflow in data-scarce regions.","language":"en","urldate":"2025-03-04","journal":"Hydrological Sciences Journal","author":[{"propositions":[],"lastnames":["Magotra"],"firstnames":["Bhanu"],"suffixes":[]},{"propositions":[],"lastnames":["Saharia"],"firstnames":["Manabendra"],"suffixes":[]},{"propositions":[],"lastnames":["Dhanya"],"firstnames":["C.","T."],"suffixes":[]}],"month":"February","year":"2025","pages":"1–14","bibtex":"@article{magotra_improved_2025,\n\ttitle = {Improved streamflow simulations in hydrologically diverse basins using physically-informed deep learning models},\n\tissn = {0262-6667, 2150-3435},\n\turl = {https://www.tandfonline.com/doi/full/10.1080/02626667.2025.2458545},\n\tdoi = {10.1080/02626667.2025.2458545},\n\tabstract = {Physically-informed deep learning models, especially long short-term memory (LSTM) networks, have shown promise in large-scale streamflow simulations. However, an in-depth understanding of the relative contribution of physical information in deep learning models has been missing. Using a large-sample testbed of 220 catchments in hydrologically diverse regions of the Indian sub-continent, we quantify the impact of incremental addition of physical information on model performance using multiple variants of the LSTM model based on various combinations of static catchment attributes and simulated land surface states. We found that LSTM models trained with catchment geophysics as additional input outperformed the base LSTM model in terms of the nationwide median Kling-Gupta efficiency (KGE) of in-sample catchments, increasing KGE from 0.32 to 0.60. Additionally, the model retained significant prediction skill in out-of-sample catchments, demonstrating that a pre-trained LSTM model can be a powerful tool to predict streamflow in data-scarce regions.},\n\tlanguage = {en},\n\turldate = {2025-03-04},\n\tjournal = {Hydrological Sciences Journal},\n\tauthor = {Magotra, Bhanu and Saharia, Manabendra and Dhanya, C. T.},\n\tmonth = feb,\n\tyear = {2025},\n\tpages = {1--14},\n}\n\n","author_short":["Magotra, B.","Saharia, M.","Dhanya, C. T."],"key":"magotra_improved_2025","id":"magotra_improved_2025","bibbaseid":"magotra-saharia-dhanya-improvedstreamflowsimulationsinhydrologicallydiversebasinsusingphysicallyinformeddeeplearningmodels-2025","role":"author","urls":{"Paper":"https://www.tandfonline.com/doi/full/10.1080/02626667.2025.2458545"},"metadata":{"authorlinks":{}},"downloads":2},"bibtype":"article","biburl":"https://api.zotero.org/users/605836/collections/98W6NZ9F/items?key=Py0EWgS20Ou10SSxCEyfvr4L&format=bibtex&limit=100","dataSources":["Yph7rNjSLqDTJjbD8"],"keywords":[],"search_terms":["improved","streamflow","simulations","hydrologically","diverse","basins","using","physically","informed","deep","learning","models","magotra","saharia","dhanya"],"title":"Improved streamflow simulations in hydrologically diverse basins using physically-informed deep learning models","year":2025,"downloads":2}