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

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.

Effects of snowmelt on the runoff dynamics in two catchments with different forest stands

<p>The forest stand can significantly affect the snow deposition and consequently the runoff during the melt period. This study focuses on water and element fluxes from snowpack in two Czech boreal headwater lake catchments with different forest stands (mature vs. regenerating after bark beetle tree dieback) using isotopic and hydrochemical tools. Sampling and analysis of the surface water, precipitation and snowpack throughout one  hydrological year enabled us to estimate the isotopic balance and chemical snowpack evolution, but also the snowmelt contribution in lakes inlets and outlets.</p><p>Isotopic signatures of the snowpack were seasonal, with δ<sup>2</sup>H amplitudes of -25‰ in the mature and -17‰ in the regenerating forest catchments. The mature forest had a ~1 month longer duration of snow cover and higher concentration of solutes in the precipitation and snowpack. In both catchments, heavier isotopes (<sup>18</sup>O and <sup>2</sup>H) preferentially left the snowpack, which was saturated with rainwater. This resulted in the final spring snowmelt being enriched with lighter isotopes (<sup>16</sup>O and <sup>1</sup>H). Ions were also eluted from the snowpack during rain-on-snow events and partial snow melting throughout the winter, causing fluxes of diluted water at the end of the snowmelt. Our results demonstrate the hydrological and hydrochemical variability of the snowpack, which in the future may even increase with rising temperatures and changes of precipitation patterns.</p>

Streamflow Predictions in a Small Urban–Rural Watershed: The Effects of Radar Rainfall Resolution and Urban Rainfall–Runoff Dynamics

The authors predicted streamflow in an urban–rural watershed using a nested regional–local modeling approach for the community of Manchester, Iowa, which is downstream of a largely rural watershed. The nested model coupled the hillslope-link model (HLM), used to simulate the upstream rural basins, and XPSWMM, which was used to simulate the more complex rainfall–runoff dynamics and surface and subsurface drainage in the urban areas, making it capable of producing flood maps at the street level. By integrating these models built for different purposes, we enabled fast and accurate simulation of hydrological processes in the rural basins while also modeling the flows in an urban environment. Using the model, we investigated how the spatial and temporal resolution of radar rainfall inputs can affect the modeled streamflow. We used a combination of three radar rainfall products to capture the uncertainty of rainfall estimation in the model results. Our nested model was able to simulate the hydrographs and timing and duration above the threshold known to result in nuisance flooding in Manchester. The spatiotemporal resolution the radar rainfall input to the model impacted the streamflow outputs of the regional, local, and nested models differently depending on the storm event.

Commercial microwave links for urban drainage modelling: The effect of link characteristics and their position on runoff simulations

Commercial microwave links (CMLs), radio connections widely used in telecommunication networks, can provide path-integrated quantitative precipitation estimates (QPEs) which could complement traditional precipitation observations. This paper assesses the ability of individual CMLs to provide relevant QPEs for urban rainfall-runoff simulations and specifically investigates the influence of CML characteristics and position on the predicted runoff. The analysis is based on a 3-year-long experimental data set from a small (1.3 km2) urban catchment located in Prague, Czech Republic. QPEs from real world CMLs are used as inputs for urban rainfall-runoff predictions and subsequent modelling performance is assessed by comparing simulated runoffs with measured stormwater discharges. The results show that model performance is related to both the sensitivity of CML to rainfall and CML position. The bias propagated into the runoff predictions is inversely proportional to CML path length. The effect of CML position is especially pronounced during heavy rainfalls, when QPEs from shorter CMLs, located within or close to catchment boundaries, better reproduce runoff dynamics than QPEs from longer CMLs extending far beyond the catchment boundaries. Interestingly, QPEs averaged from all available CMLs best reproduce the runoff temporal dynamics. Adjusting CML QPEs to three rain gauges located 2-3 km outside of the catchment substantially reduces the bias in CML QPEs. Unfortunately, this compromises the ability of the CML QPEs to reproduce runoff dynamics during heavy rainfalls. More experimental case studies are necessary to provide specific recommendations on CML preprocessing methods tailored to different water management tasks, catchments and CML networks.

Modelling historical variability of phosphorus and organic carbon fluxes to the Mackenzie River, Canada

This study provides an improved statistical modelling framework for understanding historical variability and trends in water constituent fluxes in subarctic western Canada. We evaluated total phosphorus (TP) and dissolved organic carbon (DOC) fluxes for the Hay, Liard and Peel tributaries of the Mackenzie River. The TP and DOC concentrations primarily exhibit chemodynamic relationships with discharge, with the exception of the chemostatic relationship between DOC and discharge for the Hay River. With this understanding, we explored a number of enhancements in the load estimation model that included the use of (i) linear regression and logarithmic models, (ii) air-temperature as an alternate input variable and (iii) quantile mapping for bias-correction. Further, we evaluated uncertainties in the simulation of fluxes and trends by using a bootstrapping method. The modelled TP and DOC fluxes show considerable seasonal and interannual variability that generally follow the runoff dynamics. The annual and seasonal trends are mostly small and insignificant, with the largest significant increases occurring in the winter months. These trends are amplified compared with discharge, suggesting the possibility of pronounced changes with large changes in discharge. Additionally, the results provide evidence that directly using limited water constituent samples for trend analysis can be problematic.