Conceptual Rainfall-runoff Model Research Articles

Satellite-derived spatiotemporal data on imperviousness for improved hydrological modelling of urbanised catchments

• Performance of both lumped and distributed conceptual rainfall-runoff models were improved dramatically when temporal variations in impervious surface fraction (ISF) were used as input to these models. • Both SIMHYD and AWBM could be used for assessing urbanisation effect on streamflow of urbanized catchments. • Both lumped and distributed conceptual rainfall-runoff models showed similar performance in simulation of daily streamflows for the small urban catchment tested. • Given the observed rainfall for the study catchment, one percent increase in impervious surface fraction would lead to 6.9–7.3 mm (2.9–3%) increase in the annual streamflow during 1991–2018. Conceptual rainfall-runoff models have been used extensively to simulate streamflow in urban catchments. However, spatiotemporal changes in the impervious surface area, as an important land use indicator for urban environment, has not been used explicitly as input into conceptual rainfall-runoff models. In this study, spatiotemporal impervious surface fraction (ISF) data were incorporated into two lumped and distributed (1 km) models, namely SIMHYD and AWBM in an urbanized catchment in southeast Queensland, Australia. A sub-pixel classification technique was used to derive ISF from 6 LandSat images between 1988 and 2015. Daily rainfall-runoff models were calibrated with ISF as an input variable. To evaluate model performance for different simulation periods, four model runs including M1 (lumped models with constant ISF), M2 (lumped models with temporally varying ISF), M3 (distributed models with fixed spatial ISF) and M4 (distributed models with spatiotemporal ISF and climate data). Results showed that (1) the sub-pixel classification algorithm can be used to accurately derive ISF data with a mean absolute error (MAE) ranging from 0.07 to 0.12 for 6 images, (2) ISF in the catchment has increased gradually from 12.2% to 34.6% from 1988 to 2015, (3) the performance of SIMHYD and AWBM were not improved noticeably when distributed climate and ISF data were used, (4) simulation capability of lumped and distributed SIMHYD and AWBM were improved (e.g., increase in NSE from 0.68 to 0.73, and decrease in PBIAS from −30.2% to 1.2% for lumped SIMHYD) when temporal (M2) and/or spatiotemporal (M4) ISF data (5-year) were used, (5) difference in model performance is small and insignificantly whether annual or 5-yearly ISF data were used and (6) annual streamflow would increase by 6.9–7.3 mm (equal to 2.9–3%) in response to every 1% increase in ISF for the catchment tested. The methodology developed in this study can be applied for assessment of the effect of urbanisation on regional water balance.

Comparison of sequential and variational assimilation methods to improve hydrological predictions in snow dominated mountainous catchments

• Assimilation of streamflow and snow data using variational and sequential techniques in a hydrological model. • Satellite snow coverage products are assimilated together with streamflow data. • Variational techniques outperform sequential techniques in two mountainous basins. • Average relative gains in variational assimilation are 49% for streamflow and 28% for snow data. Snowmelt modeling is vital issue for proper operational flow forecasting systems but still challenging in upstream mountainous basins due to limited observations and associated uncertainties. Data assimilation is a valuable tool to improve model state estimates to improve hydrological forecast. While assimilation of snow has been increasingly implemented in research, particularly using sequential data assimilation methods, less attention has been dedicated to compare different types of assimilation techniques. This study aims at comparing the well-known Ensemble Kalman Filter (EnKF) sequential data assimilation and the less known Moving Horizon Estimation (MHE) variational assimilation. Initial conditions of the hydrological model states are improved by assimilating in-situ streamflow and remotely sensed snow covered area data into the Hydrologiska Byråns Vattenbalansavdelning (HBV) conceptual rainfall-runoff model. The study is applied to improve runoff predictions over two mountainous basins: one in the uppermost catchment of the Karasu river, in Eastern Turkey, and the second in an alpine basin located in the headwaters of the Genil River in the northern flank of the Sierra Nevada Mountain in Southern Spain. The forecast models are tested in a hindcasting condition, i.e. closed loop environment, which simulates real-time forecasting systems. The results are analyzed using forecast verification metrics. Results showed that streamflow and snow state predictions using MHE outperform the commonly used EnKF based on the continuous ranked probability score (CRPS).

Integrating Meteorological Forcing from Ground Observations and MSWX Dataset for Streamflow Prediction under Multiple Parameterization Scenarios

Precipitation and near-surface air temperatures are significant meteorological forcing for streamflow prediction where most basins are partially or fully data-scarce in many parts of the world. This study aims to evaluate the consistency of MSWXv100-based precipitation, temperatures, and estimated potential evapotranspiration (PET) by direct comparison with observed measurements and by utilizing an independent combination of MSWXv100 dataset and observed data for streamflow prediction under four distinct scenarios considering model parameter and output uncertainties. Initially, the model is calibrated/validated entirely based on observed data (Scenario 1), where for the second calibration/validation, the observed precipitation is replaced by MSWXv100 precipitation and the daily observed temperature and PET remained unchanged (Scenario 2). Furthermore, the model calibration/validation is done by considering observed precipitation and MSWXv100-based temperature and PET (Scenario 3), and finally, the model is calibrated/validated entirely based on the MSWXv100 dataset (Scenario 4). The Kling–Gupta Efficiency (KGE) and its components (correlation, ratio of bias, and variability ratio) are utilized for direct comparison, and the Hanssen–Kuiper (HK) skill score is employed to evaluate the detectability strength of MSWXv100 precipitation for different precipitation intensities. Moreover, the hydrologic utility of MSWXv100 dataset under four distinct scenarios is tested by exploiting a conceptual rainfall-runoff model under KGE and Nash–Sutcliffe Efficiency (NSE) metrics. The results indicate that each scenario depicts high streamflow reproducibility where, regardless of other meteorological forcing, utilizing observed precipitation (Scenario 1 and 3) as one of the model inputs, shows better model performance (KGE = 0.85) than MSWXv100-based precipitation, such as Scenario 2 and 4 (KGE = 0.78–0.80).

Land-Use-Based Runoff Yield Method to Modify Hydrological Model for Flood Management: A Case in the Basin of Simple Underlying Surface

The study of runoff under the influence of human activities is a research hot spot in the field of water science. Land-use change is one of the main forms of human activities and it is also the major driver of changes to the runoff process. As for the relationship between land use and the runoff process, runoff yield theories pointed out that the runoff yield capacity is spatially heterogeneous. The present work hypothesizes that the distribution of the runoff yield can be divided by land use, which is, areas with the same land-use type are similar in runoff yield, while areas of different land uses are significantly different. To prove it, we proposed a land-use-based framework for runoff yield calculations based on a conceptual rainfall–runoff model, the Xin’anjiang (XAJ) model. Based on the framework, the modified land-use-based Xin’anjiang (L-XAJ) model was constructed by replacing the yielding area (f/F) in the water storage capacity curve of the XAJ model with the area ratio of different land-use types (L/F; L is the area of specific land-use types, F is the whole basin area). The L-XAJ model was then applied to the typical cultivated–urban binary land-use-type basin (Taipingchi basin) to evaluate its performance. Results showed great success of the L-XAJ model, which demonstrated the area ratio of different land-use types can represent the corresponding yielding area in the XAJ model. The L-XAJ model enhanced the physical meaning of the runoff generation in the XAJ model and was expected to be used in the sustainable development of basin water resources.

Trends in reconstructed monthly, seasonal and annual flows for Irish catchments (1900–2016)

Trends in reconstructed monthly, seasonal and annual flows for Irish catchments (1900–2016)

A Comparative Evaluation of Conceptual Rainfall–Runoff Models for a Catchment in Victoria Australia Using eWater Source

Hydrological modelling at a catchment scale was conducted to investigate the impact of climate change and land-use change individually and in combination with the available streamflow in the Painkalac catchment using an eWater Source hydrological model. This study compares the performance of three inbuilt conceptual models within eWater Source, such as the Australian water balance model (AWBM), Sacramento and GR4J for streamflow simulation. The three-model performance was predicted by bivariate statistics (Nash–Sutcliff efficiency) and univariate (mean, standard deviation) to evaluate the efficiency of model runoff predictions. Potential evapotranspiration (PET) data, daily rainfall data and observed streamflow measured from this catchment are the major inputs to these models. These models were calibrated and validated using eight objective functions while further comparisons of these models were made using objective functions of a Nash–Sutcliffe efficiency (NSE) log daily and an NSE log daily bias penalty. The observed streamflow data were split into three sections. Two-thirds of the data were used for calibration while the remaining one-third of the data was used for validation of the model. Based on the results, it was observed that the performance of the GR4J model is more suitable for the Painkalac catchment in respect of prediction and computational efficiency compared to the Sacramento and AWBM models. Further, the impact of climate change, land-use change and combined scenarios (land-use and climate change) were evaluated using the GR4J model. The results of this study suggest that the higher climate change for the year 2065 will result in approximately 45.67% less streamflow in the reservoir. In addition, the land-use change resulted in approximately 42.26% less flow while combined land-use and higher climate change will produce 48.06% less streamflow compared to the observed flow under the existing conditions.

Lithologic control on parameters of conceptual rainfall-runoff model and runoff characteristics: a case study of the Xinanjiang model

Lithologic control on parameters of conceptual rainfall-runoff model and runoff characteristics: a case study of the Xinanjiang model

Lithologic control on parameters of conceptual rainfall-runoff model and runoff characteristics: a case study of the Xinanjiang model

Lithologic control on parameters of conceptual rainfall-runoff model and runoff characteristics: a case study of the Xinanjiang model

Identifying most relevant controls on catchment hydrological similarity using model transferability – A comprehensive study in Iran

A key goal of hydrological science is to understand the climatic controls on catchment hydrological response behavior. Progress toward this goal would result in improved model transferability from gauged to gauged catchments. In this sense, the effectiveness of the model's transferability is contingent on the proper selection of donor and target catchment pairs. Thus, using a rainfall-runoff model, we evaluate two distinct types of hydrological similarity in this study: (i) the apparent similarity measured by similarity distance based on observable catchments descriptors (CDs) and a Euclidean distance based on physical similarity (PS) method and (ii) behavioral similarity, which is determined by highest-performance of transferred model parameters between gauged donor catchment and ungauged target catchment (best-donor case (BD)). It is believed that catchments that apparently to be similar in terms of CDs, have a similar hydrological behavior. We wish to see if that assumption is valid in this paper. Spatial proximity (SP) is also implemented to see if it might be used as an alternative for PS where there is no apparent physical similarity between catchments. To test the study's assumptions, the HBV conceptual rainfall-runoff model is used in 576 catchments across four climate regions in Iran. The results indicate that: (1) as expected, the best-donor (BD) case performs the best, and the more than 75 % of our physically similar catchments have a hydrological similarity (the overlap was ≥ 70 %), (2) the superiority of PS over SP demonstrates that the CDs exert a great influence on transferability within each climate region than geographical distance. However, we demonstrated that the SP is superior, when spatial distance between donor and target catchments is reduced (nearest neighbor ≤ 20 km), (3) consistent with CDs, when utilizing SP method, geographical distance has a varying effect on model transferability within wetter and drier regions, such that SP performs better in wetter regions than it does in dry interior regions, (4) throughout Iran, the dominant controls on model transferability differ by region. Thus, the climatic (aridity index or PET/P), topographic (mean elevation), and physiographic (catchment area) properties exert a greater influence on parameter transfer to ungauged catchments than do other CDs, and (5) the runoff ratio (streamflow signature) confirmed the superiority of the wetter regions over the drier regions in terms of control on the parameters transfer.

On constraining a lumped hydrological model with both piezometry and streamflow: results of a large sample evaluation

Abstract. The role of aquifers in the seasonal and multiyear dynamics of streamflow is undisputed: in many temperate catchments, aquifers store water during the wet periods and release it all year long, making a major contribution to low flows. The complexity of groundwater modelling has long prevented surface hydrological modellers from including groundwater level data, especially in lumped conceptual rainfall–runoff models. In this article, we investigate whether using groundwater level data in the daily GR6J model, through a composite calibration framework, can improve the performance of streamflow simulation. We tested the new calibration process on 107 French catchments. Our results show that these additional data are superfluous if we look only at model performance for streamflow simulation. However, parameter stability is improved, and the model shows a surprising ability to simulate groundwater levels with a satisfying level of performance in a wide variety of hydrogeological and hydroclimatic contexts. Finally, we make several recommendations regarding the model calibration process to be used according to the hydrogeological context of the modelled catchment.

Multimodel quantification of green and blue water components and its error propagations through parameter transferability approach across input choices

A robust quantitative assessment of water security through efficient water accounting concepts like blue and green water footprints has gained a lot of attention for water resources management at river basin scale. However, such water accounting tools and models have dependencies to the choice of precipitation data used to drive the model, as the input bias propagate into simulated streamflow through interaction with different hydrological processes. This study performs a holistic investigation of sensitivity of water accounting fractions like blue water flow (BWF), green water flow (GWF), green water storage (GWS), and their dependencies on secondary precipitation datasets (SPDs) choices using a lumped conceptual rainfall-runoff model [Hydrological Simulation model (HYSIM)] and a physically based semi-distributed hydrological model [Soil and Water Assessment Tool (SWAT)] in the Damodar River basin, India. These models were driven by seven gridded precipitation datasets, including India Meteorological Department (IMD) data, Indian Monsoon Data Assimilation and Analysis (IMDAA) data, Asian Precipitation-Highly-Resolved Observational Data Integration Towards Evaluation (APHRODITE) data, Watch forcing ERA-Interim data, Climate Hazard Group Infrared Precipitation with Station data (CHIRPS), PRINCETON precipitation data and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Network-Climate Data Records (PERSIANN-CDR) data. The reference IMD based multimodel results revealed that the annual average blue and green water components of the Damoder River were in the range of 558–656 mm/year, 574–728 mm/year and 34–112 mm/year for BWF, GWF, and GWS respectively during 1994–2010 period. The results obtained after applying parameter transferability approach across the input space have revealed exact quantification of choice dependent annual variations in spatial and temporal characteristics of BWF, GWF, and GWS. A further evaluation based on error propagation ratios (γCC, γrRMSE, γB) has revealed a quantification of input choice induced dampening/magnification effects on flow estimations of both the HYSIM and SWAT and these quantifications were free from the compensation effects of recalibration. The findings in this study demonstrate that the change in annual average BWF varies between −46.97% to +41.38% in comparison to benchmark IMD simulations, and the corresponding changes for GWF is between −19.63% and +17.06% and for GWS is between −29.97% to +35.5% with different choices of SPD scenarios. The systematic (γB) and random (γrRMSE) error propagation factors during monsoon season are larger for SWAT model ([0.21–0.78], [0.69–0.89] respectively) than that of HYSIM model ([0.18–0.74], [0.66–0.81], respectively) and it is varying with choice of different SPDs.

Towards improving the calibration practice of conceptual hydrological models of extensive green roofs

Conceptual rainfall-runoff models (CRRMs) can be used as a design tool for green roofs due to their simplicity and acceptable accuracy. This study showed how the uncertainty of CRRM parameters could be reduced by changing the calibration practice, which can enhance the interpretation and identifiability of CRRM parameters. A CRRM was developed and tested on a dataset of 14 extensive green roofs located in four Norwegian cities with different climatic conditions. Two calibration schemes were compared: a common scheme using runoff data as a basis for calibration (single-objective), and a scheme combining runoff and soil moisture data for the calibration (multi-objective scheme). The results confirmed the ability of the CRRM to simulate runoff from extensive green roofs across multiple climatic zones and different roof configurations (Kling Gupta Efficiency > 0.75). The multi-objective calibration scheme was found to reduce the uncertainty of the CRRM parameters, especially the storage parameters, enhancing the physical interpretation of parameter values. The study attempted to give guidelines to estimate parameters of the CRRM which can be used by practitioners for new roof configurations under different climatic conditions.

Intercomparison of joint bias correction methods for precipitation and flow from a hydrological perspective

The typical framework of the climate change impact assessment on water resources relies on plausible scenarios obtained from global climate models (GCMs) and hydrological models (HMs). Although regional climate models (RCMs) can better simulate local climate at a high-resolution grid, the direct use of model outputs from RCMs is not recommended as inputs for HMs due to systematic error. Existing studies have focused on the bias correction (BC) of climate model outputs without considering uncertainties/biases in hydrological modeling. In this regard, this study proposed an integrated framework that combines the BC of RCM precipitation and the simulated flow from the rainfall-runoff model, considering the underlying uncertainty in the parameters of the distribution function. The regional climate model, HadRM3, and the conceptual rainfall-runoff model, HYMOD, are employed. Observed daily precipitation, evapotranspiration, and discharge time series over the Thorverton catchment are compiled from the UK Meteorological Office. To examine the effectiveness of the combined strategy, four different BC approaches have been explored to reduce systematic biases in the flow simulated through the HMs using the RCM precipitation as input. Here, BCs of RCM and HM outputs have been applied under the condition that the bias-corrected ensembles should be within the range of the observed climate variability. The four BC models are considered: aathe RCM precipitation and flow are corrected by preserving their natural variabilities (Case-4). From a hydrological perspective, the Case-4 model showed the best performance among the four cases in terms of correcting the bias and the spread of the flow ensemble.

Assessing climate change impact on the hydropower potential of the Bamboi catchment (Black Volta, West Africa)

Abstract. This study evaluates the impact of future climate change (CC) on the hydropower generation potential of the Bamboi catchment (Black Volta) in West Africa using a conceptual rainfall-runoff model (HBV light) and regional climate models (RCMs)–global climate models (GCMs). Two climate simulation datasets MPI-ESM-REMO (CORDEX) and GFDL-ESM2M-WRF (WASCAL) under RCP4.5 were applied to the validated hydrological model to simulate the catchment runoff. Based on reference and future simulated discharges, a theoretical 1.3 MW run of river hydro power plant was designed to evaluate the hydropower generation. Hydrological and hydropower generation changes were expressed as the relative difference between two future periods (2020–2049 and 2070–2099) and a reference period (1983–2005). The climate models' ensemble projected a mean annual precipitation increase by 8.8 % and 7.3 % and discharge increase by 11.4 % and 9.735 % for the 2020–2049 and 2070–2099 periods respectively (for bias corrected data). On the contrary an overall decrease of hydropower generation by −9.1 % and −8.4% for the 2020–2049 and 2070–2099 periods was projected respectively. The results indicate that projected increases in discharge should not solely be considered as leading to an increase in hydropower potential when prospecting climate change impact on hydropower.

The role of land use and morphology representation in the setup and calibration of the conceptual TUW model

The processes of the transformation of rainfall to runoff are highly complicated, and the proper characterisation of these processes with conceptual hydrological models is a very challenging task. Morphology and land cover have a significant influence on a river basin’s hydrologic response. Thus, catchment characteristics of the topography and land use play an essential role in parametrising the runoff concentration processes in hydrological models. In the study, our goal was to detect which characteristics and their spatial distribution influence the efficiency of a conceptual rainfall-runoff model efficiency most. The spatially lumped and semi-distributed versions of the TUW conceptual rainfall model, which is an HBV type model, were compared. Both models use the concept of lumped storages associated with the surface and subsurface, interconnected by thresholds and links to simulate the runoff transformation. We focused on two land-use characteristics, the percentage cover of the agricultural land and percentage cover of the forests, and the mean slope of the terrain as a topography characteristic. The differences between runoff model efficiencies both in the calibration and validation periods were evaluated. Based on which version of the model was more effective in the simulation of the runoff, it was detected which types of catchment land use, and morphology were better represented by using the lumped or semi-distributed version of the TUW model, respectively. The analysis aimed to improve the understanding of the influence of spatial representation morphology and land cover in conceptual models on model efficiency and may help to improve model setup and calibration.

Comparing spatial and temporal scales of hydrologic model parameter transfer: A guide to four climates of Iran

Simulating streamflow in ungauged catchments remains a challenging task in hydrology and increases the demand for regionalization studies worldwide. Here, we investigate the effect of three modes of parameter transfer, including temporal (transferring across different periods), spatial (transferring between same calibration periods but different sites), and spatiotemporal (transferring across both different periods and sites) on simulating streamflow using HBV conceptual rainfall-runoff model at 576 unregulated catchments throughout Iran (407,000 Km2). Our main conclusions are: (1) temporal mode shows the best performance, with the lowest decline in performance (median decline of 5.8%) as measured using the NSE efficiency metric, (2) difference between spatial and spatiotemporal options was negligible (median decline of 13.7% and 15.1% respectively), (3) all parameters are associated with some uncertainties and those related to runoff and snow components of the model are associated with the highest and lowest uncertainties, respectively, (4) overall, the model performance in arid regions is not as good as humid regions which confirmed that elevation and climate play a major role in parameter estimation in these areas, and (5) aridity and catchment elevation are two major controls on model transferability at regional (climate classes) and local (the whole country) scales. We also show that the superiority of the temporal mode is maintained with: (i) increasing spatial distance between gauged (donor) and ungauged (target) catchments, (ii) increasing time lag (10 years) between calibration and validation, and (iii) gradually increased time lags between calibration and validation. Our study suggest that spatiotemporal parameter transfer can be a reliable option for PUB studies and climate change-related studies, at least in wetter catchments. However, further research is needed to explore the complicated relationship between temporal and spatial aspects of hydrological variability.

Parameter Optimization of a Conceptual Rainfall-Runoff Model in the Coastal Urban Region

Sunwoo, W.; Lee, G.; Nguyen, H.H.; Shin, H., and Jun, K.S., 2021. Parameter optimization of a conceptual rainfall-runoff model in the coastal urban region. In: Lee, J.L.; Suh, K.-S.; Lee, B.; Shin, S., and Lee, J. (eds.), Crisis and Integrated Management for Coastal and Marine Safety. Journal of Coastal Research, Special Issue No. 114, pp. 261–265. Coconut Creek (Florida), ISSN 0749-0208. The rainfall-runoff simulation based on local conditions is crucial for flood monitoring and mitigation. Although numerous rainfall-runoff models were widely used, only few studies have been applied in coastal regions in South Korea. The major reasons are that model parameters are incorrectly determined, and the peak flooding is overestimated due to extreme rainfall events. In this study, a conceptual rainfall-runoff model, the Probability Distributed Model (PDM), was employed to simulate discharge with an incorporation of combined satellite-, model-, and ground-based soil moisture as an input data. The combined soil moisture data was generated using two renormalization strategies. In this study, we aimed to compare the performances of two optimization techniques associated in the PDM model, including a traditional non-linear and a newly metaheuristics algorithms, for the discharge prediction. The simulation was conducted in a coastal urban region, Hongseong, using the optimized parameter sets for flood events. The simulated results from the PDM model revealed that parameters obtained with the metaheuristics algorithm indicated a superior performance when comparing against the observed runoff data. Moreover, parameter sets obtained by the metaheuristics algorithm and non-linear technique showed significant differences. The method of this study can provide more improvements in rainfall-runoff simulation results for flood management.

Parameterization of the rainfall-runoff model in changing climate

Abstract A conceptual rainfall-runoff model was used for estimating the impact of climate change on the runoff regime in the Myjava River basin. Changes in climatic characteristics for future decades were expressed by a regional climate model using the A1B emission scenario. The model was calibrated for 1981–1990, 1991–2000, 2001–2010, 2011–2019. The best set of model parameters selected from the recent calibration period was used to simulate runoff for three periods, which should reflect the level of future climate change. The results show that the runoff should increase in the winter months (December and January) and decrease in the summer months (June to August). An evaluation of the long-term mean monthly runoff for the future climate scenario indicates that the highest runoff will occur in March.

Evaluating the impact of rainfall–runoff model structural uncertainty on the hydrological rating of regional climate model simulations

Abstract We propose to evaluate the impact of rainfall–runoff model (RRM) structural uncertainty on climate model evaluation, performed within a process-oriented framework using the RRM. Structural uncertainty is assessed with an ensemble approach using three conceptual RRMs (HBV, IHACRES and GR4J). We evaluate daily precipitation and temperature from 11 regional climate models forced by five general circulation models (GCM–RCMs), issued from EURO-CORDEX. The assessment was performed over the reference period (1970–2000) for five catchments situated in northern Tunisia. Seventeen discharge performance indexes were used to explore the representation of hydrological processes. The three RRMs performed well over the reference period, with Nash–Sutcliffe efficiency values ranging from 0.70 to 0.90 and bias close to 0%. The ranking of GCM–RCMs according to hydrological performance indexes is more meaningful before the bias correction, which considerably reduces the differences between GCM- and RCM-driven hydrological simulations. Our results illustrate a strong similarity between the different RRMs in terms of raw GCM–RCM performances over the reference period for the majority of performance indexes, in spite of their different model structures. This proves that the structural uncertainty induced by RRMs does not affect GCM–RCM evaluation and ranking, which contributes to consolidate the RRM as a standard tool for climate model evaluation.

Hydrologically informed machine learning for rainfall–runoff modelling: towards distributed modelling

Abstract. Despite showing great success of applications in many commercial fields, machine learning and data science models generally show limited success in many scientific fields, including hydrology (Karpatne et al., 2017). The approach is often criticized for its lack of interpretability and physical consistency. This has led to the emergence of new modelling paradigms, such as theory-guided data science (TGDS) and physics-informed machine learning. The motivation behind such approaches is to improve the physical meaningfulness of machine learning models by blending existing scientific knowledge with learning algorithms. Following the same principles in our prior work (Chadalawada et al., 2020), a new model induction framework was founded on genetic programming (GP), namely the Machine Learning Rainfall–Runoff Model Induction (ML-RR-MI) toolkit. ML-RR-MI is capable of developing fully fledged lumped conceptual rainfall–runoff models for a watershed of interest using the building blocks of two flexible rainfall–runoff modelling frameworks. In this study, we extend ML-RR-MI towards inducing semi-distributed rainfall–runoff models. The meaningfulness and reliability of hydrological inferences gained from lumped models may tend to deteriorate within large catchments where the spatial heterogeneity of forcing variables and watershed properties is significant. This was the motivation behind developing our machine learning approach for distributed rainfall–runoff modelling titled Machine Induction Knowledge Augmented – System Hydrologique Asiatique (MIKA-SHA). MIKA-SHA captures spatial variabilities and automatically induces rainfall–runoff models for the catchment of interest without any explicit user selections. Currently, MIKA-SHA learns models utilizing the model building components of two flexible modelling frameworks. However, the proposed framework can be coupled with any internally coherent collection of building blocks. MIKA-SHA's model induction capabilities have been tested on the Rappahannock River basin near Fredericksburg, Virginia, USA. MIKA-SHA builds and tests many model configurations using the model building components of the two flexible modelling frameworks and quantitatively identifies the optimal model for the watershed of concern. In this study, MIKA-SHA is utilized to identify two optimal models (one from each flexible modelling framework) to capture the runoff dynamics of the Rappahannock River basin. Both optimal models achieve high-efficiency values in hydrograph predictions (both at catchment and subcatchment outlets) and good visual matches with the observed runoff response of the catchment. Furthermore, the resulting model architectures are compatible with previously reported research findings and fieldwork insights of the watershed and are readily interpretable by hydrologists. MIKA-SHA-induced semi-distributed model performances were compared against existing lumped model performances for the same basin. MIKA-SHA-induced optimal models outperform the lumped models used in this study in terms of efficiency values while benefitting hydrologists with more meaningful hydrological inferences about the runoff dynamics of the Rappahannock River basin.