Understanding the Information Content in the Hierarchy of Model Development Decisions: Learning From Data

Gharari, Shervan; Gupta, Hoshin; Clark, Martyn P.; Hrachowitz, Markus; Fenicia, Fabrizio; Matgen, Patrick; Savenije, Hubert

doi:10.1029/2020wr027948

Cited by 29 publications

(33 citation statements)

References 156 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These include the observation of inputs to the model (typically precipitation, potential evapotranspiration), the specification of inputs to the model (e.g., time and space resolution), the model structure (e.g., physics-based or lumped), and the observed discharge data used for model validation (or calibration). It is of course not straightforward to treat these sources independently, as there are many possible interactions and analysis of one source must be considered conditional on other choices of the modeling procedure (see Gharari et al (2021) for a comprehensive discussion). Precipitation uncertainty seldom considered.…”

Section: Introductionmentioning

confidence: 99%

Is Precipitation Responsible for the Most Hydrological Model Uncertainty?

et al. 2022

View full text Add to dashboard Cite

Rainfall-runoff modeling is highly uncertain for a number of different reasons. Hydrological processes are quite complex, and their simplifications in the models lead to inaccuracies. Model parameters themselves are uncertain—physical parameters because of their observations and conceptual parameters due to their limited identifiability. Furthermore, the main model input—precipitation is uncertain due to the limited number of available observations and the high spatio-temporal variability. The quantification of model output uncertainty is essential for their use. Most approaches used for the quantification of uncertainty in rainfall-runoff modeling assign the uncertainty to the model parameters. In this contribution, the role of precipitation uncertainty is investigated. Instead of a standard sensitivity analysis of the model output with respect to the input variations, it is investigated to what extent realistic precipitation fields could improve model performance. Realistic precipitation fields are defined as gridded realizations of precipitation which reproduce the observed values at the observation locations, with values which reproduce the distribution of the observed values and with spatial variability the same as the spatial variability of the observations. The above conditions apply to each observation time step. Through an inverse modeling approach based on Random Mixing precipitation fields fulfilling the above conditions and reproducing the discharge output better than using traditional interpolated observations can be obtained. These realizations show how much rainfall runoff models may profit from better precipitation input and how much remains for the parameter and model concept uncertainty. The methodology is applied using two hydrological models with a contrasting basis, SHETRAN and HBV, for three different mesoscale sub-catchments of the Neckar basin in Germany. Results show that up to 50% of the model error can be attributed to precipitation uncertainty. Further, inverting precipitation using hydrological models can improve model performance even in neighboring catchments which are not considered explicitly.

show abstract

Section: Introductionmentioning

confidence: 99%

Is Precipitation Responsible for the Most Hydrological Model Uncertainty?

et al. 2022

View full text Add to dashboard Cite

show abstract

“…However, when the system forcing is much reduced in magnitude (for instance, dry period or minimal amount of rainfall), the system's internal processes (such as infiltration rates which can easily be affected by human activities like deforestation) will have a more pronounced effect on the output. Some of the relevant literature on concepts regarding analysis of the effect of forcing on the systems in the context of hydrological modeling can be obtained from [49][50][51][52].…”

Section: Discussionmentioning

confidence: 99%

Contributions of Human Activities and Climatic Variability to Changes in River Rwizi Flows in Uganda, East Africa

2021

View full text Add to dashboard Cite

This study employed Soil and Water Assessment Tool (SWAT) to analyze the impacts of climate variability and human activities on River Rwizi flows. Changes in land use and land cover (LULC) types from 1997 to 2019 were characterized using remotely sensed images retrieved from Landsat ETM/TM satellites. SWAT was calibrated and validated over the periods 2002–2008 and 2009–2013, respectively. Correlation between rainfall and river flow was analyzed. By keeping the optimal values of model parameters fixed while varying the LULC maps, differences in the modeled flows were taken to reflect the impacts of LULC changes on rainfall–runoff generation. Impacts due to human activities included contributions from changes in LULC types and the rates of water abstracted from the river as a percentage of the observed flow. Climate variability was considered in terms of changes in climatic variables such as rainfall and evapotranspiration, among others. Variability of rainfall was analyzed with respect to changes in large-scale ocean-atmosphere conditions. From 2000 to 2014, the portion of River Rwizi catchment area covered by cropland increased from 23.0% to 51.6%, grassland reduced from 63.3% to 37.8%, and wetland decreased from 8.1% to 4.7%. Nash–Sutcliffe Efficiency values for calibration and validation were 0.60 and 0.71, respectively. Contributions of human activities to monthly river flow changes varied from 2.3% to 23.5%. Impacts of human activities on the river flow were on average found to be larger during the dry (14.7%) than wet (5.8%) season. Using rainfall, 20.9% of the total river flow variance was explained. However, climate variability contributed 73% of the river flow changes. Rainfall was positively and negatively correlated with Indian Ocean Dipole (IOD) and Niño 3, respectively. The largest percentages of the total rainfall variance explained by IOD and Niño 3 were 12.7% and 9.8%, respectively. The magnitude of the correlation between rainfall and IOD decreased with increasing lag in time. These findings are relevant for developing River Rwizi catchment management plans.

show abstract

“…Finally, we note that the LSTM architecture exhibits better regional transferability than the SN17 model structure (Figure S10 in Supporting Information S1). This points to a fundamental difference between what is achieved when training an LSTM network as opposed to calibrating the SN17 model, where the former corresponds more closely to a structure learning problem (Gharari et al, 2021), while the latter is restricted to only parameter learning given a predefined model structure.…”

Section: Evaluation Of Regional Tl Network Against Three Benchmarksmentioning

confidence: 99%

Exploring the Potential of Long Short‐Term Memory Networks for Improving Understanding of Continental‐ and Regional‐Scale Snowpack Dynamics

Wang

Gupta

Zeng

et al. 2022

Water Resources Research

Self Cite

View full text Add to dashboard Cite

Accurate estimation of the spatio‐temporal distribution of snow water equivalent is essential given its global importance for understanding climate dynamics and climate change, and as a source of fresh water. Here, we explore the potential of using the Long Short‐Term Memory (LSTM) network for continental and regional scale modeling of daily snow accumulation and melt dynamics at 4‐km pixel resolution across the conterminous US (CONUS). To reduce training costs (data are available for ∼0.31 million snowy pixels), we combine spatial sampling with stagewise model development, whereby the network is first pretrained across the entire CONUS and then subjected to regional fine‐tuning. Accordingly, model evaluation is focused on out‐of‐sample predictive performance across space (analogous to the prediction in ungauged basins problem). We find that, given identical inputs (precipitation, temperature, and elevation), a single CONUS‐wide LSTM provides significantly better spatio‐temporal generalization than a regionally calibrated version of the physical‐conceptual temperature‐index‐based SNOW17 model. Adding more meteorological information (dew point temperature, vapor pressure deficit, longwave radiation, and shortwave radiation) further improves model performance, while rendering redundant the local information provided by elevation. Overall, the LSTM exhibits better transferability than SNOW17 to locations that were not included in the training data set, reinforcing the advantages of structure learning over parameter learning. Our results suggest that an LSTM‐based approach could be used to develop continental/global‐scale systems for modeling snow dynamics.

show abstract

Understanding the Information Content in the Hierarchy of Model Development Decisions: Learning From Data

Abstract:  We present a strategy for characterizing and quantifying the information added at each model building step  Model building steps are interdependent in a hierarchical manner  We call for the focus of model calibration to shift from "parameter spaces" to "function spaces"

Cited by 29 publications

References 156 publications

Is Precipitation Responsible for the Most Hydrological Model Uncertainty?

Is Precipitation Responsible for the Most Hydrological Model Uncertainty?

Contributions of Human Activities and Climatic Variability to Changes in River Rwizi Flows in Uganda, East Africa

Exploring the Potential of Long Short‐Term Memory Networks for Improving Understanding of Continental‐ and Regional‐Scale Snowpack Dynamics

Contact Info

Product

Resources

About