Prediction and Inference of Flow Duration Curves Using Multioutput Neural Networks

Worland, Scott C.; Asquith, William H.; Knight, Rodney R.; Wieczorek, Michael E.

doi:10.1029/2018wr024463

Cited by 45 publications

(27 citation statements)

References 68 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In most applications considered here, 95% CIs based on fitted copula models provided adequate coverage of the observations, except in the tails of the distribution ( Figures 5 and 11). It is possible that some of these tail uncertainty estimates could be improved if uncertainty from other sources was propagated forward, for instance uncertainty in the FDC estimate (see Worland et al, 2019aWorland et al, , 2019b or even measurement error at the donor sites. In addition, when the NSE values were decomposed for each model, copula-based approaches tended to have less variance error in the NSE (Figure 10).…”

Section: Benefits and Limitations To Prediction Using Copulasmentioning

confidence: 99%

Copula Theory as a Generalized Framework for Flow‐Duration Curve Based Streamflow Estimates in Ungaged and Partially Gaged Catchments

Worland

Farmer

Asquith

et al. 2019

Water Resources Research

Self Cite

View full text Add to dashboard Cite

Flow‐duration curve (FDC) based streamflow estimation methods involve estimating an FDC at an ungaged or partially gaged location and using the time series of nonexceedance probabilities estimated from donor streamgage sites to generate estimates of streamflow. We develop a mathematical framework to illustrate the connection between copulas and prior FDC‐based approaches. The performance of copula methods is compared to several other streamflow estimation methods using a decade of daily streamflow data from 74 sites located within two river basins in the southeast United States with different climate characteristics and physiographic properties. We show that copula approaches: (1) outperform other methods in the limiting case of perfect information with regard to the rank‐based correlation structure and FDCs across the gaging network; (2) provide a hedge against poor performance when donor information becomes sparser and less informative; (3) outperform other methods when used for partially gaged sites with several years of available data; and (4) remain a competitive albeit nondominating method for ungaged sites and partially gaged sites with limited data when realistic error is introduced in the estimation of FDCs and correlations across the gaging network.

show abstract

Section: Benefits and Limitations To Prediction Using Copulasmentioning

confidence: 99%

Copula Theory as a Generalized Framework for Flow‐Duration Curve Based Streamflow Estimates in Ungaged and Partially Gaged Catchments

Worland

Farmer

Asquith

et al. 2019

Water Resources Research

Self Cite

View full text Add to dashboard Cite

show abstract

“…Dropout is a computationally efficient way to combine many network structures and prevent over-fitting; it adds noise and limits co-dependencies between neurons during DNN training (Srivastava et al, 2014;Goodfellow et al 2016). It involves temporarily removing randomly selected neurons during DNN training (Srivastava et al, 2014;Goodfellow et al, 2016;Worland et al 2019).…”

Section: Deep Neural Network Development Architecture and Testingmentioning

confidence: 99%

“…However, we observed that the DNN overpredicted monthly runoff in California, Texas, and Florida as indicated by large negative (i.e., < -25%) median watershed residuals (Figure 3). Future studies may use local interpretable model-agnostic explanations (LIME; Ribeiro et al, 2018) and other machine learning interpretation techniques to further explain these patterns in model residuals (e.g., Worland et al, 2019).…”

Section: Deep Neural Network Testingmentioning

confidence: 99%

“…While this study does not attempt to characterize the filtering processes of watersheds across the CONUS, future studies may apply tools such as partial response functions (Rice et al, 2016;Rice & Emanuel, 2017) and local interpretable model-agnostic explanations (Worland et al, 2019) to explore the impact of GCM inputs and watershed characteristics on watershed-scale runoff. Also, these statistical model interpretation methods may be used to open up the machine learning "black box" by generating hypotheses that can be tested using physically-based hydrology models and field experiments (Shen et al, 2018;Rice and Emanuel 2019).…”

Section: Implications and Future Directionsmentioning

confidence: 99%

“…Recent advances in optimization algorithms, computer hardware (e.g., GPUs), computer software (e.g., Google TensorFlow), and cloud computing services (e.g., Amazon Web Services' Sage Maker) have also made the utilization of machine learning methods readily feasible for applications in hydrology and other areas (LeCun et al, 2015;Schmidhuber et al, 2015;Shen et al, 2018). Furthermore, explanation techniques such as local interpretable model-agnostic explanations (LIME; Ribeiro et al, 2018;Worland et al, 2019), can help model developers assess the trustworthiness of their machine learning algorithm results.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Improved Accuracy of Watershed-Scale General Circulation Model Runoff Using Deep Neural Networks

Rice¹,

Saia²,

Emanuel³

2019

Preprint

View full text Add to dashboard Cite

Projecting impacts of climate change on water resources is a vital research task, and general circulation models (GCMs) are important tools for this work. However, the spatial resolution of downscaled GCMs makes them difficult to apply to non-grid conforming scales relevant to water resources management: individual watersheds. Machine learning techniques like deep neural networks (DNNs) may address this issue. Here we use a DNN to predict monthly watershed-scale runoff (i.e., stream discharge divided by watershed area) from monthly gridded and downscaled Coupled Model Intercomparison Project Phase 5 (CMIP5) GCM hydroclimatic fluxes (i.e., precipitation, evapotranspiration, and temperature). We used hydroclimatic fluxes, biotic, and abiotic characteristics from 2,731 watersheds in the conterminous United States to train and test a DNN that can predict watershed-scale runoff. The DNN described 93% (Pearson’s correlation coefficient = 0.962) of the variability in observed runoff and was temporally and spatially robust. The median absolute error (MAE) of DNN predictions was approximately 25 percentage points lower than that of gridded, downscaled GCM runoff or monthly normal runoff (i.e., 30-year average of runoff observations at the watershed-outlet). DNN monthly runoff predictions had the lowest MAE of all the grid-to-watershed-scale conversion approaches we tested, including: linear ridge regression, support vector machines, extreme gradient boosting, and artificial neural networks. We demonstrated why using DNNs to convert gridded GCM hydroclimatic fluxes to watershed-scales is relevant to water resources research and management. We also provided a methods guide for hydrologists interested in implementing machine learning techniques.

show abstract

Mind the information gap: How sampling and clustering impact the predictability of reach‐scale channel types in California (USA)

Guillon,

Lane,

Byrne

et al. 2024

Earth Surf Processes Landf

View full text Add to dashboard Cite

Clustering and machine learning‐based predictions are increasingly used for environmental data analysis and management. In fluvial geomorphology, examples include predicting channel types throughout a river network and segmenting river networks into a series of channel types, or groups of channel forms. However, when relevant information is unevenly distributed throughout a river network, the discrepancy between data‐rich and data‐poor locations creates an information gap. Combining clustering and predictions addresses this information gap, but challenges and limitations remain poorly documented. This is especially true when considering that predictions are often achieved with two approaches that are meaningfully different in terms of information processing: decision trees (e.g., RF: random forest) and deep learning (e.g., DNNs: deep neural networks). This presents challenges for downstream management decisions and when comparing clusters and predictions within or across study areas. To address this, we investigate the performance of RF and DNN with respect to the information gap between clustering data and prediction data. We use nine regional examples of clustering and predicting river channel types, stemming from a single clustering methodology applied in California, USA. Our results show that prediction performance decreases when the information gap between field‐measured data and geospatial predictors increases. Furthermore, RF outperforms DNN, and their difference in performance decreases when the information gap between field‐measured and geospatial data decreases. This suggests that mismatched scales between field‐derived channel types and geospatial predictors hinder sequential information processing in DNN. Finally, our results highlight a sampling trade‐off between uniformly capturing geomorphic variability and ensuring robust generalisation.

show abstract

Prediction and Inference of Flow Duration Curves Using Multioutput Neural Networks

Cited by 45 publications

References 68 publications

Copula Theory as a Generalized Framework for Flow‐Duration Curve Based Streamflow Estimates in Ungaged and Partially Gaged Catchments

Copula Theory as a Generalized Framework for Flow‐Duration Curve Based Streamflow Estimates in Ungaged and Partially Gaged Catchments

Improved Accuracy of Watershed-Scale General Circulation Model Runoff Using Deep Neural Networks

Mind the information gap: How sampling and clustering impact the predictability of reach‐scale channel types in California (USA)

Contact Info

Product

Resources

About