Hydrologic Performance Simulation of Green Infrastructures: Why Data-Driven Modelling Can Be Useful?

Benchmarking data-driven rainfall-runoff modeling across 54 catchments in the Yellow River Basin: Overfitting, calibration length, dry frequency

Much of the prairie region in North America is characterized by relatively flat terrain with many depressions on the landscape. The hydrological response (runoff) is a combination of the conventional runoff from the contributing areas and the occasional overflow from the non-contributing areas (depressions). In this study, we promote the use of a hybrid modelling structure to predict runoff generation from prairie landscapes. More specifically, the Soil and Water Assessment Tool (SWAT) is fused with artificial neural networks (ANNs), so that SWAT and the ANN module deal with the contributing and non-contributing areas, respectively. A detailed experimental study is performed to select the best set of inputs, training algorithms and hidden neurons. The results obtained in this study suggest that the fusion of process-based and data-driven models can provide improved modelling capabilities for representing the highly nonlinear nature of the hydrological processes in prairie landscapes.Editor D. Koutsoyiannis; Associate editor L. See

Future climate projections require Earth system models to simulate conditions outside their calibration range. It is therefore crucial to understand the applicability of such models and their modules under transient conditions. This study assesses the robustness of different types of models in terms of rainfall–runoff modeling under changing conditions. In particular, two process-based models and one data-driven model are considered: 1) the physically based land surface model of the European Centre for Medium-Range Weather Forecasts, 2) the conceptual Simple Water Balance Model, and 3) the Long Short-Term Memory-Based Runoff model. Using streamflow data from 161 catchments across Europe, a differential split-sample test is performed, i.e., models are calibrated within a reference period (e.g., wet years) and then evaluated during a climatically contrasting period (e.g., drier years). Models show overall performance loss, which generally increases the more conditions deviate from the reference climate. Further analysis reveals that the models have difficulties in capturing temporal shifts in the hydroclimate of the catchments, e.g., between energy- and water-limited conditions. Overall, relatively high robustness is demonstrated by the physically based model. This suggests that improvements of physics-based parameterizations can be a promising avenue toward reliable climate change simulations. Further, our study illustrates that comparison across process-based and data-driven models is challenging due to their different nature. While we find rather low robustness of the data-driven model in our particular split-sample setup, this must not apply generally; by contrast, such model schemes have great potential as they can learn diverse conditions from observed spatial and temporal variability both at the same time to yield robust performance.

Process-based forest models generally have many parameters, multiple outputs of interest and a small underlying empirical database. These characteristics hamper parameterization. Bayesian calibration offers a solution to the calibration problem because it applies to models of any type or size. It provides parameter estimates, with measures of uncertainty and correlation among the parameters. The procedure begins by quantifying the uncertainty about parameter values in the form of a prior probability distribution. Then data on the output variables are used to update the parameter distribution by means of Bayes' Theorem. This yields a posterior calibrated distribution for the parameters, which can be summarized in the form of a mean vector and variance matrix. The predictive uncertainty of the model can be quantified by running it with different parameter settings, sampled from the posterior distribution. In a further step, one may evaluate the posterior probability of the model itself (rather than that of the parameters) and compare that against the probability of other models, to aid in model selection or improvement. Bayesian calibration of process-based models cannot be performed analytically, so the posterior parameter distribution must be approximated in the form of a representative sample of parameter values. This can be achieved by means of Markov Chain Monte Carlo simulation, which is suitable for process-based models because of its simplicity and because it does not require advance knowledge of the shape of the posterior distribution. Despite the suitability of Bayesian calibration, the technique has rarely been used in forestry research. We introduce the method, using the example of a typical forest model. Further, we show that reductions in parameter uncertainty, and thus in output uncertainty, can be effected by increasing the variety of data, increasing the accuracy of measurements and increasing the length of time series.

The impacts on ecosystems of global and local shifts in the environment due to e.g. climate change, land use change, and urbanisation present us with complex and interacting societal challenges. Solving a single problem often aggravates another. Scientific insights needed to face these challenges are based on interdisciplinary multi-scale research. It requires that we work together across (sub)disciplinary boundaries and combine our insights and knowledge to understand the complexities and interrelatedness within ecosystems. Making this combination is currently hampered by the availability and scatteredness of datasets and the lack of methods and tools to combine knowledge, data and models. Developing digital twins of ecosystems could provide a paradigm to achieve this combination of knowledge, data and models, and generate the insight and knowledge that we require to face the challenges of global and local change. A digital twin of an ecosystem is a digital representation of an ecosystem tailored to the user’s research objectives and perspective of that ecosystem. They bring together (long-term) data collected of multiple ecosystem facets and across multiple scales, process-based models, and data science and computational modelling tools, making them a versatile instrument. Digital twins can be used to understand ecosystem functioning, by aiding the discovery of intricate ecological relationships not yet understood or measured. Furthermore, they are ideal for exploring the effects of in situ management strategies, or for the detection of early-warning signals of global change impacts on ecosystem dynamics. LTER-LIFE is a scientific infrastructure, embedded in LTER-NL (Long-Term Ecosystem Research Netherlands) and LifeWatch ERIC, aimed at supporting shared and integrated ecological research, providing scientists the platform and tools to build digital twins of ecosystems. We do this by supporting the sharing of data and models and bringing these together into a site-focused virtual lab where people can work together. LTER sites provide an ideal opportunity for this as they often come with a long history of extensive data collection and research across multiple spheres. Drawing on experience of building digital twins in the Veluwe LTSER platform, this keynote will take you through the journey of digital twinning parts of an ecosystem with a long legacy of ecological and environmental research and data collection. As these data exist in many different formats, we built data mobilisation and FAIRification pipelines to augment the focal experimental sites in the Veluwe with relevant historical, privately-owned or national datasets. Subsequently, we explored the potential of the Notebook as a Virtual Research Environment (NaaVRE), which provided us with a virtual lab in which we developed workflows for processing and linking FAIRified data with process-based and data-driven models. Ultimately, we show that this way of working can bring together researchers to generate insights on the impacts of climate change, land use change, and urbanisation on the functioning of the ecosystem, providing scientific support for management and policy decisions.

The purpose of this study was to characterize mountain river regimes in the Spanish Pyrenees and to assess the importance of snow accumulation and snowmelt on the timing of river flows. Daily streamflow data from 9 gauging stations in the Pyrenees were used to characterize river regimes. These data were analyzed by hydrological indices, with a focus on periods when snow accumulation and snowmelt occurred. These results were combined with data on Snow Water Equivalent (SWE) (from measurements of depth and density of snow in the main river basins and also simulated by a process-based hydrological model), snowmelting (simulated by a process-based hydrological model), precipitation (from observations), and temperature (from observations). Longitude and elevation gradients in the Pyrenees explain the transition of river regimes from those that mostly had low nival signals (in the west and at low elevations) to those that mostly had high nival signals (low winter runoff and late spring peakflow, in the east and at high elevations). Although trend analyses indicated no statistically significant changes, there was a trend of decreased nival signal over time in most of the analyzed rivers. Our results also demonstrated that snow processes cannot explain all of the interannual variability of river regimes, because the temporal distribution of liquid precipitation and temperature play key roles in hydrography.

Study regionAthabasca River basin, Alberta, Canada (156,000 km2). Study focusHydrology often relies upon hydrologic models in data-sparse regions; however, it is unclear if such models are reliably accurate, or if internal process simulations are reasonable representations of watershed function. Standard model evaluation and calibration approaches often prioritize accurate reproduction of recorded streamflow, ignoring process simulation fidelity, regardless of the intended model application. This study evaluates whether combined use of streamflow and isotope tracer performance metrics can improve representation of simulated streamflow-generating processes within a large river basin, the Athabasca watershed, to inform calibration of a process-based, distributed hydrologic model. New hydrological insights for the regionFlow-based performance metrics were found to be sensitive to processes influencing streamflow volume and timing, but insensitive to internal flow paths and storage volumes. Although somewhat less reliable than flow metrics, isotope tracer performance metrics are found to be most sensitive to processes influencing mixing and water age, and appreciably responsive to many other processes. We demonstrate that process-based hydrologic models for rivers such as the Athabasca River cannot be optimally calibrated using streamflow metrics alone, as such optimizations cannot tune parameters or process representations to which the objective function is insensitive. Importantly, isotope tracers have demonstrable value for informing process-based hydrologic model optimization by providing a window into the sub-surface black box within complex regional-scale simulations.

Enhancing process-based hydrological models with embedded neural networks: A hybrid approach

The utilization of data-driven models in hydrology has witnessed a significant increase in recent years. The open-source philosophy underpinning much of the code developed and research being conducted has facilitated widespread access to the the hydrological community to sophisticated machine learning models and technology (Reichstein 2019). These data driven approaches to hydrological modelling has witnessed growing interest after multiple studies has shown how machine-learning models were able to outperform nationwide traditional physics-based hydrological models (Kratzerts et al. 2019). The latter often demands substantial man-hours for development, calibration and fine-tuning to accurately represent relevant hydrological processes. In this national-scale explorative study we undertake an in-depth examination of Danish catchment hydrology. Our objective is to understand what processes and dynamics are well captured by a purely data driven model without physical constraints, namely the Entity-Aware Long Short-Term Model (EA-LSTM). The model code was developed by Kratzerts et al. (2019) and the analysis build on top of a newly published national CAMELS data set covering 301 catchments in Denmark (Koch and Schneider, 2022), with an average resolution of 130 km2. Denmark, spanning an area of around 43 000 km2, demonstrates a relatively high data coverage. Presently more than 400 stations record water level measurements in the Danish stream network, while a network of 243 stations have collected meteorological data since 2011. These datasets maintained by the Danish Environmental Protection Agency and the Danish Meteorological Institute, respectively, and are publicly available. Despite Denmark’s data abundance, Koch and Schneider (2022) demonstrated that the data-driven EA-LSTM model, trained with the CAMELS dataset for Denmark (from now on referred to as the DK-LSTM) were not able to outperform the traditional physics-based hydrological model, against which it was benchmarked. Consequently, performance of the DK-LSTM model could be increased by pre-training it with simulations from a national physics-based model indicating that dominating hydrological processes are not described by the readily available input data in the CAMELS dataset. This study conducts a comprehensive analysis of Danish catchment hydrology aiming to explore three aspects: 1) the common characteristics of the catchments where the DK-LSTM performs well or encounters challenges, 2) the identification of hydrological characteristics, that exhibit improvement when informing the data-driven model with physics-based model simulations, and 3) an exploration of whether the aforementioned findings can guide us in determining necessary physical constraints and/or input variables that explains the hydrological processes for the data-driven model approach at a national scale, using the example of DK-LSTM.   Koch, J., and Schneider, R. Long short-term memory networks enhance rainfall-runoff modelling at the national scale of Denmark. GEUS Bulletin, 49. https://doi.org/10.34194/geusb.v49.8292, 2022. Kratzert, F., Klotz, D., Shalev, G., Klambauer, G., Hochreiter, S., and Nearing, G.: Towards learning universal, regional, and local hydrological behaviors via machine learning applied to large-sample datasets, Hydrol. Earth Syst. Sci., 23, 5089-5110, https://doi.org/10.5194/hess-23-5089-2019, 2019. Reichstein, M., Camps-Valls, G., Stevens, B. et al. Deep learning and process understanding for data-driven Earth system science. Nature 566, 195–204. https://doi.org/10.1038/s41586-019-0912-1, 2019.

<p>The recent boom in machine learning and data science has led to a number of new opportunities in the environmental sciences. In particular, climate models represent the best tools we have to predict, understand and potentially mitigate climate change, however these process-based models are incredibly complex and require huge amounts of high-performance computing resources. Machine learning offers opportunities to greatly improve the computational efficiency of these models.</p><p>Here we discuss our recent efforts to reduce the computational cost associated with running a process-based model of the physical ocean by developing an analogous data-driven model. We train statistical and machine learning algorithms using the outputs from a highly idealised sector configuration of general circulation model (MITgcm). Our aim is to develop an algorithm which is able to predict the future state of the general circulation model to a similar level of accuracy in a more computationally efficient manner.</p><p>We first develop a linear regression model to investigate the sensitivity of data-driven approaches to various inputs, e.g. temperature on different spatial and temporal scales, and meta-variables such as location information. Following this, we develop a neural network model to replicate the general circulation model, as in the work of Dueben and Bauer 2018, and Scher 2018.</p><p>We present a discussion on the sensitivity of data-driven models and preliminary results from the neural network based model.</p><p> </p><p><em>Dueben, P. D., & Bauer, P. (2018). Challenges and design choices for global weather and climate models based on machine learning. Geoscientific Model Development, 11(10), 3999-4009.</em></p><p><em>Scher, S. (2018). Toward Data‐Driven Weather and Climate Forecasting: Approximating a Simple General Circulation Model With Deep Learning. Geophysical Research Letters, 45(22), 12-616.</em></p>

Forested peatlands are globally significant carbon pools, important forest resources and areas for other bioproduction. Management of drained peatland forests has partly contradictory targets, such as economic profit, climate change mitigation and adaptation, and water protection. Balancing between these targets by comparing different management options requires a thorough understanding of the ecosystem processes, as well as modelling tools that are able to represent their complexity. This thesis presents three such tools, which are based on process-based models, and their applications to concrete water table management scenarios in drained peatlands. All three works are built upon a common peat hydrological model. Two of the studies analyze the effect of canal-blocking restoration practices on the water table and carbon dioxide emissions in tropical peatlands. The first work shows that using optimization algorithms to choose the location of a fixed number of canal blocks can lead to sizeable improvements on the amount of peat they rewet. The second work systematically analyzes the impact of canal blocks on tropical peatland water tables, and provides insights about their performance for different weather conditions and peat types. The third study presents a peatland ecosystem model focused on the effect of drainage on nutrient dynamics and forest growth in boreal peatlands. By connecting the relevant hydrological and biogeochemical processes, this ecosystem model enables the study of interconnected phenomena such as the identification of the stand growth limiting factor, and the impact of typical ditch network management operations on the nutrient balance and forest productivity.

Process-based forest models are important tools for predicting forest growth and their vulnerability to factors such as climate change or responses to management. One of the most widely used stand-level process-based models is the 3-PG model (Physiological Processes Predicting Growth), which is used for applications including estimating wood production, carbon budgets, water balance and susceptibility to climate change. Few 3-PG parameter sets are available for central European species and even fewer are appropriate for mixed-species forests. Here we estimated 3-PG parameters for twelve major central European tree species using 1418 long-term permanent forest monitoring plots from managed forests, 297 from un-managed forest reserves and 784 Swiss National Forest Inventory plots. A literature review of tree physiological characteristics, as well as regression analyses and Bayesian inference, were used to calculate the 3-PG parameters.The Swiss-wide calibration, based on monospecific plots, showed a robust performance in predicting forest stocks such as stem, foliage and root biomass. The plots used to inform the Bayesian calibration resulted in posterior ranges of the calibrated parameters that were, on average, 69% of the prior range. The bias of stem, foliage and root biomass predictions was generally less than 20%, and less than 10% for several species. The parameter sets also provided reliable predictions of biomass and mean tree sizes in mixed-species forests. Given that the information sources used to develop the parameters included a wide range of climatic, edaphic and management conditions and long time spans (from 1930 to present), these species parameters for 3-PG are likely to be appropriate for most central European forests and conditions.

We present a Bayesian geostatistical model for mean annual runoff that incorporates simulations from a process-based hydrological model by treating the simulations as a covariate in the statistical model. The regression coefficient of the covariate is modeled as a spatial field such that the relationship between the covariate (simulations from a hydrological model) and the response variable (observed mean annual runoff) is allowed to vary within the study area. Hence, it is a spatially varying coefficient. A preprocessing step for including short records in the modeling is also suggested and we obtain a model that can exploit several data sources by using state of the art statistical methods. The geostatistical model is evaluated by predicting mean annual runoff for 1981–2010 for 127 catchments in Norway based on observations from 411 catchments. Simulations from the process-based HBV model on a 1 km × 1 km grid are used as input. We found that on average the proposed approach outperformed a purely process-based approach (HBV) when predicting runoff for ungauged and partially gauged catchments: The reduction in RMSE compared to the HBV model was 20 % for ungauged catchments and 58 % for partially gauged catchments, where the latter is due to the preprocessing step. For ungauged catchments the proposed framework also outperformed a purely geostatistical method with a 10 % reduction in RMSE compared to the geostatistical method. For partially gauged catchments however, purely geostatistical methods performed equally well or slightly better than the proposed combination approach. It is not surprising that purely geostatistical methods perform well in areas where we have data. In general, we expect the proposed approach to outperform geostatistics in areas where the data availability is low to moderate.

Flow, low flow statistics, and variability of Darbandikhan and Dukan inflow time series have been analyzed. Darbandikhan and Dukan dams’ inflow time series were best fitted with three parameter inverse Gaussian distribution. Long term, annual, and monthly flow duration curves (FDCs) are constructed. Two parameters logarithmic function appeared to be the most appropriate functions fitted to FDCs. Low flow percentiles are extracted from flow duration curves for 95, 90, 75, 70, 50, and 25 percentiles. Annual minimum 1-, 3-, 7-, 14-, 30-, 60-, and 90-day mean inflow with recurrence intervals of 2, 5, 10, 20, 30, and 50 years were estimated.Base flows were separated from long-term inflow time series. Results indicated that a large amount of long-term inflow was supported by base flow for both Darbandikhan and Dukan dams’ inflow time series. The discharges were generally less than their mean inflow with no large fluctuations. The trend of monthly mean inflow discharge versus time have been analyzed using different methods. Results are indicating a significant negative monthly trends of 0.488 m3/s and 0.593 m3/s with reduction percentages of 0.0127 % and 0.0120% for Darbandikhan and Dukan dams’ inflow time series; respectively. The results of current study were compared with USGS Data Series 540 results. All month’s inflow defined a reduction that maximized in August. Comparison of low flow statistics are also referred to maximum reduction percentage in Q95% which is occurring commonly in summer. The larger reduction during summer months is indicating artificial water withdrawals from feeding streams of the dams during periods of no rain.

Comparison of deep learning models and a typical process-based model in glacio-hydrology simulation