Never, never, (really never!?) train an LSTM on a single basin. Or should you?
Abstract
Machine learning (ML) has played an increasing role in the hydrological sciences. In particular, Long Short-Term Memory (LSTM) networks are popular for rainfall–runoff modeling. A large majority of studies that use this type of model do not follow best practices, and there is one mistake in particular that is common: training deep learning models on small, homogeneous data sets, typically data from only a single hydrological basin. In this position paper, we show that LSTM rainfall–runoff models are best when trained with data from a large number of basins.
Links
The run directories of all experiments, including model weights, simulations, and pre-computed metrics, are available at https://doi.org/10.5281/zenodo.10139248. The code that was used for analyzing all the experiments and to create all the figures, based on the run directories, can be found at https://doi.org/10.5281/zenodo.13691802.
Citation
@Article{kratzert2024never,
author = {Kratzert, F. and Gauch, M. and Klotz, D. and Nearing, G.},
title = {HESS Opinions: Never train a Long Short-Term Memory (LSTM) network on a single basin},
journal = {Hydrology and Earth System Sciences},
volume = {28},
year = {2024},
number = {17},
pages = {4187--4201},
doi = {10.5194/hess-28-4187-2024}
}