HS1.3.2
Bridging physical, analytical, information-theoretic and machine learning approaches to system dynamics and predictability in Hydrology and Earth System Sciences

HS1.3.2

EDI
Bridging physical, analytical, information-theoretic and machine learning approaches to system dynamics and predictability in Hydrology and Earth System Sciences
Co-organized by NP2
Convener: Rui A. P. Perdigão | Co-conveners: Julia HallECSECS, Cristina Prieto, Maria KireevaECSECS, Shaun HarriganECSECS
Presentations
| Tue, 24 May, 10:20–11:50 (CEST)
 
Room 2.31

Presentations: Tue, 24 May | Room 2.31

Chairpersons: Rui A. P. Perdigão, Julia Hall, Cristina Prieto
10:20–10:23
Welcome and Opening Remarks
10:23–10:29
|
EGU22-2993
|
Highlight
|
Virtual presentation
|
Spatio-temporal synchronization of heavy rainfall events triggered by atmospheric rivers in North America 
Frederik Wolf, Sara M. Vallejo-Bernal, Niklas Boers, Norbert Marwan, Dominik Traxl, and Jürgen Kurths

Atmospheric rivers (ARs) are filaments of extensive water vapor transport in the lower troposphere. They are important triggers of heavy rainfall events, contributing to more than 50% of the rainfall sums in some regions along the western coast of North America. ARs play a crucial role in the distribution of water, but can also cause natural and economical damage by facilitating heavy rainfall. Here, we investigate the large-scale spatio-temporal synchronization patterns of heavy rainfall triggered by ARs over the western coast and the continental regions of North America.

For our work, we employ daily ERA5 rainfall estimates at a spatial resolution of 0.25°x0.25° latitude and longitude which we threshold at the 95th percentile to obtain binary time series indicating the absence or presence of heavy rainfall. Subsequently, we separate periods with ARs and periods without ARs and investigate the differing spatial synchronization pattern of heavy rainfall. To establish that our results are not dependent on the chosen AR catalog, this is conducted in two different ways: first based on a recently published catalog by Gershunov et al. (2017) , and second based on a catalog constructed using the IPART algorithm (Xu et al, 2020). For both approaches, we subsequently utilize event synchronization and a complex network framework to reveal distinct spatial patterns of heavy rainfall events for periods with and without active ARs. Using composites of upper-level meridional wind, we attribute the formation of the rainfall synchronization patterns to well-known atmospheric circulation configurations, whose intensity scales with the strength of the ARs. Furthermore, we demonstrate that enhanced AR activity is going in hand with a suppressed seasonal shift of the characteristic meridional wind pattern. To verify and illustrate how small changes of the high-level meridional wind affect the distribution of heavy rainfall, we, additionally, perform a case study focusing on the boreal winter.

Our results indicate the strong sensitivity of the intensity, location, frequency, and pattern of synchronized heavy rainfall events related to ARs to small changes in the large-scale circulation.

How to cite: Wolf, F., Vallejo-Bernal, S. M., Boers, N., Marwan, N., Traxl, D., and Kurths, J.: Spatio-temporal synchronization of heavy rainfall events triggered by atmospheric rivers in North America, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-2993, https://doi.org/10.5194/egusphere-egu22-2993, 2022.

10:29–10:35
|
EGU22-3694
|
ECS
|
Virtual presentation
|
Prediction of drain flow fraction at high spatial resolution by combining physically based models and machine learning 
Raphael Schneider, Hafsa Mahmood, Rasmus Rumph Frederiksen, Anker Lajer Højberg, and Simon Stisen

In Denmark, about half of the agricultural land is artificially drained. These drainage systems have a significant effect on the hydrological system. Knowledge about the spatio-temporal distribution of drain flow is crucial to understand aspects such as groundwater recharge, streamflow partitioning and nutrient transport. Still, quantification of drain flow at regional and large scale remains a major challenge: Data on the distribution of the installed subsurface drainage system are scarce, as are measurements of drain flow. Large-scale simulations of drain with physically-based hydrological models are challenged by scale, as drain flow is controlled by small-scale variations in groundwater depth often beyond the model resolution. Purely data-driven models can struggle representing the complex controls behind drain flow.

Here, we suggest a metamodel approach to obtain a more accurate estimate of generated drain flow at high spatial resolution of 10 m, combining physically-based with data-driven models. Our variable of interest is drain fraction, defined as the ratio between drain flow and recharge per grid cell, which is an indicator for flow partitioning between drain and recharge to deeper groundwater.

First, we setup distributed, integrated groundwater models at 10 m grid resolution for 28 Danish field-scale drain catchments with observations of drain flow timeseries. A joint calibration of these field-scale models against observed drain flow resulted in an average KGE of above 0.5. Subsequently, the simulated drain fractions from the field-scale models were used to train a decision tree machine learning algorithm. This metamodel uses various mappable covariates (topography and geology-related) available at high resolution for all of Denmark. The metamodel then is used to predict drain fractions, within its limits of applicability, across relevant areas of Denmark with significant drain flow outside of the field-scale models.

Eventually, the predicted drain fractions are intended to inform national, large-scale physically based hydrological models: An improved representation of drain can, for example, make those models more fit to improve national targeted nitrate regulation.

How to cite: Schneider, R., Mahmood, H., Frederiksen, R. R., Højberg, A. L., and Stisen, S.: Prediction of drain flow fraction at high spatial resolution by combining physically based models and machine learning, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-3694, https://doi.org/10.5194/egusphere-egu22-3694, 2022.

10:35–10:41
|
EGU22-4076
|
ECS
|
On-site presentation
|
Information theory approach for enhancing time series analysis and predictability of soil environments 
Luiza Cristina Corpaci, Sebastian Raubitzek, and Kevin Mallinger

Soil environments are naturally governed by a multitude of interdependent chemical, biological, and physical processes that define their macro-state. In the context of farming these features are further complemented and affected by anthropogenic activities (ploughing, fertilizing, use of pesticides, etc.) that systematically aim to change soil and plant environments to enhance yield, but often with unforeseen detrimental effects (biodiversity loss, erosion, etc.). Assessing strategies for sustainable environmental management is therefore a highly challenging task that is often accompanied by incomplete knowledge of systemic feedback mechanisms and a lack of continuous and reliable data. 

To address this issue, we investigate the use of complexity metrics from information theory to gain insights about underlying patterns of multivariate soil systems and their potential implications for time series analysis. Here we apply existing methods for the processing and analysis of similar systems, we verify current theories about the dynamics and mechanisms of ecological processes in time and study innate interactions between separate components. Thereby, we will use available agricultural datasets that display a wide range of soil properties and explore several notions of complexity approaches, such as entropy measures (e.g., Permutation entropy, transfer entropy, Shannon entropy) and the Hurst exponent. Characteristic features will be highlighted that can be used to enhance time series prediction accuracy and systemic soil functions understanding.

How to cite: Corpaci, L. C., Raubitzek, S., and Mallinger, K.: Information theory approach for enhancing time series analysis and predictability of soil environments, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4076, https://doi.org/10.5194/egusphere-egu22-4076, 2022.

10:41–10:47
|
EGU22-4933
|
ECS
|
Virtual presentation
Integrating historical information, systematic data, and rainfall-runoff modelling to improve flood frequency estimates 
Miguel Angel Fernandez-Palomares and Luis Mediero

Abstract

Flood frequency curves are usually fitted to short time series of observations, leading to great uncertainties mainly for high return periods. However, reliable estimates are required for designing and assessing safety of hydraulic infrastructure, such as bridges and dams. Therefore, flood frequency analyses based on instrumental data collected at gauging stations can be improved by incorporating available information about historical floods before the beginning of the systematic period. This study presents how to identify and integrate all the information available, in order to improve flood frequency curve estimates. The Cuevas de Almanzora Dam located in southeast Spain is selected as case study.

The Cuevas de Almanzora Dam catchment has an area of 2122 km2 with a mean annual precipitation of 316 mm. However, daily precipitation can be higher, such as 600 mm for the 1973 flood event. Flood data are available at a gauging station located in the River Almanzora upstream of the dam, with a draining catchment of 1850 km2. The systematic period is 1963-2008 with information about 36 annual maximum floods. The largest flood in the 20th century was recorded at the gauging station in 1973. A two-dimensional (2D) hydrodynamic model of the River Almanzora was calibrated with such information.

Historical information about floods has been collated from local newspapers, books, chronicles, research papers, photographs, national archives of historical floods, and municipal archives. The three largest floods in the River Almanzora between 1830 and 1963 were identified, extending the systematic period to a total period of 191 years. Information about water depths and flood extensions at different cross sections of the River Almanzora were collected. The 2D hydrodynamic model was used to estimate the peak discharges in such historical flood events.

After the end of the systematic period, the hydrograph of the great 2012 flood event was estimated from the data recorded at the Cuevas de Almanzora reservoir. A rainfall-runoff model was calibrated in the catchment with 1-h precipitation data to estimate the flood hydrograph at the gauging station.

The five historical floods that exceed the perception threshold in the period 1830-2020 were integrated with the annual maximum floods extracted from the systematic data, using five techniques to incorporate historical information in the flood frequency curve. The Generalized Extreme Value (GEV) and the Two-Component Extreme Value (TCEV) distribution functions were considered. The best fit was selected considering the accuracy and the uncertainty of estimates by a stochastic procedure. Flood quantiles for the highest return periods triple the estimates obtained by using only the systematic data.

The methodology proposed can improve the reliability of flood quantile estimates, mainly in arid regions where the lack of information about the rare greatest flood, which can exceed several times the mean magnitude of floods in the systematic period, can lead to strong underestimates for the highest return periods that are needed to design and assess the safety of hydraulic infrastructure.

Acknowledgments: This research has been supported by the project SAFERDAMS (PID2019-107027RB-I00) funded by the Spanish Ministry of Science and Innovation.

How to cite: Fernandez-Palomares, M. A. and Mediero, L.: Integrating historical information, systematic data, and rainfall-runoff modelling to improve flood frequency estimates, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-4933, https://doi.org/10.5194/egusphere-egu22-4933, 2022.

10:47–10:53
|
EGU22-5690
|
ECS
|
Presentation form not yet defined
Analysis of wind wave statistics preceding beach cast events on the southeastern shore of the Baltic Sea (Kaliningrad region): preliminary results 
Igor Isachenko and Elena Esiukova

Beach cast is a material deposited on beaches after being washed up by storm (or tidal movement). The composition of beach cast usually includes seagrass or algae fragments, wracks of land plants and other materials of natural origin, (anthropogenic) marine litter, including plastic debris and microplastics. Beach casts monitoring is of current interest for beach management and maintenance of the sandy shores for recreational purposes, tracing marine litter transport and dispersion, evaluating environmental contamination by microplastiсs.

Large patches of marine debris appear on beaches after stormy weather. However, little is known about the sea state that precedes the formation of beach casts. From an observer's point of view, beach casts occur at random locations along the coast at unpredictable times. They may even be washed back to the sea at some time later. This work aims to disclose characteristic features of temporal variations of surface wave field parameters, which lead to beach cast formation.

Results of incidental surveys of the northern coast of the Sambia Peninsula, stretching from west to east in the southeastern part of the Baltic Sea, were analyzed. The presence of beach cast (at one or more locations) was observed during 234 days of 2011-2021. Some of the observations were performed during or shortly after the ending of the beaching process. Field information was collated with a freely available re-analysis database on surface waves (http://marine.copernicus.eu). Surface wave spectrum parameters were picked up from the database at the geographical point corresponding to the coastal zone's open-sea limit. Elements of Bayesian analysis were applied to overcome the lack of information on the very time of the beach casts formation and/or the unknown duration of the beaching process.

The analysis shows the values of significant wave height, peak period, and wave direction, which occurred before the beach cast appearance more often than follows from the overall time statistics ("climate"). A separate analysis of only recently formed beach casts made it possible to determine the evolution of wave spectrum parameters during the beaching process. Data suggests that most of the beach cast events on this coast are preceded by waves caused by cyclone passages from the northern direction.

Data analysis is carried out by I.I., with the support of the Russian Science Foundation, grant No 21-77-00027. Beach surveys are carried out by E.E. voluntarily and with partial support from IO RAS state assignment.

How to cite: Isachenko, I. and Esiukova, E.: Analysis of wind wave statistics preceding beach cast events on the southeastern shore of the Baltic Sea (Kaliningrad region): preliminary results, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-5690, https://doi.org/10.5194/egusphere-egu22-5690, 2022.

10:53–10:59
|
EGU22-6761
|
ECS
|
Presentation form not yet defined
|
Understanding the Information Content in the Hierarchy of Model Development Decisions: Learning From Data 
Shervan Gharari, Hoshin Gupta, Martyn Clark, Markus Hrachowitz, Fabrizio Fenicia, Patrick Matgen, and Hubert Savenije

Process-based hydrological models seek to represent the dominant hydrological processes in a catchment. However, due to unavoidable incompleteness of knowledge, the construction of “fidelius” process-based models depends largely on expert judgment. We present a systematic approach that treats models as hierarchical assemblages of hypotheses (conservation principles, system architecture, process parameterization equations, and parameter specification), which enables investigating how the hierarchy of model development decisions impacts model fidelity. Each model development step provides information that progressively changes our uncertainty (increases, decreases, or alters) regarding the input-state-output behavior of the system. Following the principle of maximum entropy, we introduce the concept of “minimally restrictive process parameterization equations—MR-PPEs,” which enables us to enhance the flexibility with which system processes can be represented, and to thereby investigate the important role that the system architectural hypothesis (discretization of the system into subsystem elements) plays in determining model behavior. We illustrate and explore these concepts with synthetic and real-data studies, using models constructed from simple generic buckets as building blocks, thereby paving the way for more-detailed investigations using sophisticated process-based hydrological models. We also discuss how proposed MR-PPEs can bridge the gap between current process-based modeling and machine learning. Finally, we suggest the need for model calibration to evolve from a search over “parameter spaces” to a search over “function spaces.”

How to cite: Gharari, S., Gupta, H., Clark, M., Hrachowitz, M., Fenicia, F., Matgen, P., and Savenije, H.: Understanding the Information Content in the Hierarchy of Model Development Decisions: Learning From Data, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-6761, https://doi.org/10.5194/egusphere-egu22-6761, 2022.

Vimeo: EGU22-6761-material-r0
10:59–11:05
|
EGU22-8321
|
ECS
|
Highlight
|
On-site presentation
|
Partitioning of green-blue water fluxes around the world: ML model explainability and predictability 
Daniel Althoff and Georgia Destouni

The consequences of ever-increasing human interference with freshwater systems, e.g., through land-use and climate changes, are already felt in many regions of the world, e.g., by shifts in freshwater availability and partitioning between green (evapotranspiration) and blue (runoff) water fluxes around the world. In this study, we have developed a machine learning (ML) model for the possible prediction of green-blue water flux partitioning (WFP) under different climate, land-use, and other landscape and hydrological catchment conditions around the world. ML models have shown relatively high predictive performance compared to more traditional modelling methods for several tasks in geosciences. However, ML is also rightly criticized for providing theory-free “black-box” models that may fail in predictions under forthcoming non-stationary conditions. We here address the ML model interpretability gap using Shapley values, an explainable artificial intelligence technique. We also assess ML model predictability using a dissimilarity index (DI). For ML model training and testing, we use different parts of a total database compiled for 3482 hydrological catchments with available data for daily runoff over at least 25 years. The target variable of the ML model is the blue-water partitioning ratio between average runoff and average precipitation (and the complementary, water-balance determined green water partitioning ratio) for each catchment. The predictor variables are hydro-climatic, land-cover/use, and other catchment indices derived from precipitation and temperature time series, land cover maps, and topography data. As a basis for the ML modelling, we also investigate and quantify (through data averaging over moving sub-periods of different time lengths) a minimum temporal aggregation scale for water flux averaging (referred to as the flux equilibration time, Teq) required to reach a stable temporal average runoff (and evapotranspiration) fraction of precipitation in each catchment; for 99% of catchments, Teq is found to be ≤2 years, with longer Teq emerging for catchments estimated to have higher ratio Rgw/Ravg, i.e., higher groundwater flow contribution (Rgw) to total average runoff (Ravg). The cubist model used for the ML modelling yields a Kling-Gupta efficiency of 0.86, while the Shapley values analysis indicates mean annual precipitation and temperature as the most important variables in determining the WFP, followed by average slope in each catchment. A DI threshold is further used to label new data points as inside or outside the ML model area of applicability (AoA). Comparison between test data points outside and inside the AoA reveals which catchment characteristics are mostly responsible for ML model loss of predictability. Predictability is lower for catchments with: larger Teq and Rgw/Ravg; higher phase lag between peak precipitation and peak temperature over the year; lower forest and agricultural land fractions; and aridity index much higher or much lower than 1 (implying major water or energy limitation, respectively). Identifying such predictability limits is crucial for understanding, and facilitating user awareness of the applicability and forecasting ability of such data-driven ML modelling under different prevailing and changing future hydro-climatic, land-use, and groundwater conditions.

How to cite: Althoff, D. and Destouni, G.: Partitioning of green-blue water fluxes around the world: ML model explainability and predictability, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8321, https://doi.org/10.5194/egusphere-egu22-8321, 2022.

11:05–11:11
|
EGU22-8372
|
Highlight
|
Virtual presentation
Deciphering Hydroclimatic Complexity with Information Physics and Quantum Technologies 
Rui A. P. Perdigão and Julia Hall

Discerning the dynamics of complex systems in a mathematically rigorous and physically consistent manner is as fascinating as intimidating of a challenge, stirring deeply and intrinsically with the most fundamental Physics, while at the same time percolating through the deepest meanders of everyday life.

The socio-natural coevolution in hydroclimate dynamics is an example of that, exhibiting a striking articulation between governing principles and free will, in a stochastic-dynamic resonance that goes way beyond a reductionist dichotomy between deterministic and probabilistic approaches and between physical principles and information technologies.

Subjacent to the conceptual and operational interdisciplinarity of that challenge, lies the simple formal elegance of a “lingua franca” for communication with Nature. This emerges from the innermost mathematical core of Information Physics articulating the wealth of insights and flavours from frontier natural, social and technical sciences in a coherent, integrated manner.

Communicating thus with Nature, we equip ourselves by developing formal innovative methodologies and technologies to better appreciate and discern complexity in articulation with expert knowledge. Thereby opening new pathways to assess and predict elusive non-recurrent phenomena such as irreversible geophysical transformations and extreme hydro-meteorological events in a coevolutionary climate.

Our novel advances will be shared across the formal, structural and functional theory of the Information Physics of Coevolutionary Complex Systems, along with the analysis, modelling and decision support in crucial matters afflicting our environment and society, with special emphasis onto hydroclimatic problems.

In an optic of operational empowerment, some of our flagship initiatives will be addressed such as Earth System Dynamic Intelligence and Quantum Information Technologies in the Earth Sciences (QITES) on a synergy among our information physical and quantum technological developments.

The articulation between these flagships leverages our proprietary synergistic quantum gravitational and electrodynamic QITES constellation from deep undersea to outer space to take the pulse of our planet, ranging from high resolution 4D sensing and computation to unveiling early warning signs of critical transitions and extreme events.

How to cite: Perdigão, R. A. P. and Hall, J.: Deciphering Hydroclimatic Complexity with Information Physics and Quantum Technologies, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-8372, https://doi.org/10.5194/egusphere-egu22-8372, 2022.

11:11–11:17
|
EGU22-10119
|
ECS
|
Virtual presentation
|
Life cycles of glacial lakes in Norway: Insights from machine learning algorithms on Landsat series and Sentinel-2 
Ghazal Moghaddam, Liss Marie Andreassen, and Irina Rogozhina

The observed retreat of mountain glaciers on a global scale promotes the formation and growth of glacial lakes across newly exposed ice-free areas. In mainland Norway, this process drives the rise in glacial lake outburst floods (GLOFs), posing a considerable threat to people and infrastructure  downstream. Moreover, many glacial lakes are used as reservoirs for hydropower production and thus represent an important energy source, emphasizing the need for continuous monitoring of glacial lake life cycles.

Remote sensing is currently the most efficient technique for tracking changes in glacial lakes, understanding their responses to climate change and observing lakes prone to GLOFs. Recent advances in machine learning techniques have presented new opportunities to automatize glacial lake mapping over large areas. For the first time, this study presents a Norway-wide reconstruction of glacial lake changes through the last three decades using  machine learning algorithms and long-term satellite observations. It contrasts the performance of two classification methods - maximum likelihood  classification (MLC) and support vector machine (SVM) - to outline glacial lakes and study their evolution using the Landsat series and Sentinel-2 images.

This study zooms into the pros and cons of each classification method and satellite product through the prism of glacial lake processes occurring over  disparate temporal and spatial scales - from lake formation, growth and dissociation from the proximal glaciers to the aftermath of rapid GLOF events. Based on this analysis, I conclude that the recognition skills of supervised classification methods largely depend on the quality of satellite images and careful selection of training samples. Some of the factors that adversely affect the classification results are unfavourable weather conditions such as  cloud, snow and ice cover, image disturbances through atmospheric corrections and shadows on slopes that lead to misclassifications. Regardless of higher spatial and temporal resolution, Sentinel imagery has not revealed significant advantages over Landsat but has shown a potential for their  complementary use to continue glacial lake observations in the future. The performance of SVM is clearly superior to MLC, but it is difficult to use over  large spatial scales, at least in the form it is currently implemented in ENVI.

How to cite: Moghaddam, G., Andreassen, L. M., and Rogozhina, I.: Life cycles of glacial lakes in Norway: Insights from machine learning algorithms on Landsat series and Sentinel-2, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10119, https://doi.org/10.5194/egusphere-egu22-10119, 2022.

11:17–11:23
|
EGU22-10890
|
Virtual presentation
|
One Saddle Point and Two Types of Sensitivities Within the Lorenz 1963 and 1969 Models 
Bo-Wen Shen, Roger Pielke, Sr., and Xubin Zeng

The fact that both the Lorenz 1963 and 1969 models suggest finite predictability is well-known. However, it is less known that mechanisms (i.e., sensitivities) within both models that lead to finite predictability are different. Additionally, the mathematical and physical relationship between these two models has not been fully documented. New analyses along with literature review are performed here to provide insights on the similarities and differences for these two models. The models represent different physical systems, one for convection and the other for barotropic vorticity. From theperspective of mathematical complexities, the Lorenz 1963 (L63) model is limited-scale and nonlinear; and the Lorenz 1969 (L69) model is closure-based, physically multiscale, mathematically linear, and numerically ill-conditioned. The former possesses a sensitive dependence of solutions on initial conditions, known as the butterfly effect, and the latter contains numerical sensitivities due to an ill-conditioned matrix with a large condition number (i.e., a large variance of growth rates).

Here, we illustrate that the existence of a saddle point at the origin is a common feature that produces instability in both systems. Within the chaotic regime of the L63 nonlinear model, unstable growth is constrained by nonlinearity, as well as dissipation, yielding time varying growth rates along an orbit, and, thus, a dependence of (finite) predictability on initial conditions. Within the L69 linear model, multiple unstable modes at various growth rates appear, and the growth of a specific unstable mode (i.e., the most unstable mode during a finite time interval) is constrained by imposing a saturation assumption, thereby yielding a time varying system growth rate. Both models have been interchangeably applied for qualitatively revealing the nature of finite predictability in weather and climate. However, only single type solutions were examined (i.e., chaotic and linearly unstable solutions for the L63 and L69 models, respectively), and the L69 system is ill-conditioned and easily captures numerical instability. Thus, an estimate of the predictability limit using either of the above models, with or without additional assumptions (e.g., saturation), should be interpreted with caution and should not be generalized as an upper limit for predictability of the atmosphere.

How to cite: Shen, B.-W., Pielke, Sr., R., and Zeng, X.: One Saddle Point and Two Types of Sensitivities Within the Lorenz 1963 and 1969 Models, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-10890, https://doi.org/10.5194/egusphere-egu22-10890, 2022.

11:23–11:29
|
EGU22-11148
|
ECS
|
Highlight
|
On-site presentation
Disentangling direct and indirect soil moisture effects onecosystem carbon uptake with Causal Modeling 
Christian Reimers, Alexander Winkler, Vincent Humphrey, and Markus Reichstein

Soil moisture affects gross primary production through two pathways. First, directly through
drought stress and second, indirectly through temperature via evaporative cooling of the near-
surface atmospheric layer. Because it is not possible to disentangle these effects experimentally
at a biome level, Humphrey et al. (2021) has used Earth system model experiments in which soil
moisture is fixed to its seasonal cycle and evaluated the effects on gross primary production. In
contrast, we aim to use causal modeling to infer impacts directly from observation. To predict the
effects of soil moisture anomalies on gross primary production, we extend existing causal mod-
eling frameworks to cover situations where two variables influence one other. A major challenge
in applying causal modeling here lies in the bidirectional relationship between soil moisture and
temperature via evapotranspiration. On one hand, higher temperature leads to higher evapotran-
spiration and thus lower soil moisture. On the other hand, lower soil moisture leads to lower evap-
otranspiration and thus higher temperatures. Therefore, neither soil moisture nor temperature can
be adequately modeled as a function of the other. To address this challenge, we extend existing
causal modeling frameworks to account for these situations where the variables are not functions
of each other but are determined by equilibrium. We show that our method identifies the correct
links between variables in synthetic data. We further evaluate whether our new approach is con-
sistent with the results of Humphrey et al. (2021) based on idealized counterfactual experiments
using Earth system models. To this end, we use the control runs of the models to directly predict
the results of the idealized counterfactual experiment as proof-of-concept. Finally, we apply our
method to observations and determine the direct and indirect effect of soil moisture anomalies on
gross primary production.

References:
Vincent Humphrey, Alexis Berg, Philippe Ciais, Pierre Gentine, Martin Jung, Markus Reichstein,
Sonia I Seneviratne, and Christian Frankenberg. Soil moisture–atmosphere feedback dominates
land carbon uptake variability. Nature, 592(7852):65–69, 2021.

How to cite: Reimers, C., Winkler, A., Humphrey, V., and Reichstein, M.: Disentangling direct and indirect soil moisture effects onecosystem carbon uptake with Causal Modeling, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-11148, https://doi.org/10.5194/egusphere-egu22-11148, 2022.

11:29–11:35
|
EGU22-12346
|
ECS
|
Virtual presentation
Configuration entropy analysis of river water quality dynamics under fine time resolution and network topology 
Tianrui Pang, Jiping Jiang, Peng Wang, Yi Zheng, and Tong Zheng

The water environment is an important carrier of material processes, in which a large number of biochemical reactions and energy transmission processes occur. High-frequency water quality observation can help us understand the dynamics of solute transport in the water environment. The information-theoretic approaches to system dynamics are receiving more and more attention that it reveals the new laws and support board applications. Configuration entropy (H*) is one of the derivative indexes that originated from information entropy, which was first introduced in 1994 to describe the disorder in random morphologies. It can reflect the complexity of the system under different space or time resolutions. Researchers have analyzed the characteristics of configuration entropy in some of the environment scenarios, such as spatial arrangement of rainfall. In this paper, we analyzed the space structure of river basin water quality dynamic system under the network topology of rivers, together with the time structure of water quality dynamic system. We calculated the configuration entropy of six water quality parameter data from four monitoring stations at Potomac River in two dimensions of time and space with topological treatment of river water system map. We arranged the high-frequency water quality time series according to different time slices to form a two-dimensional pixel image for calculating configuration entropy and the variation under different time resolutions. Results show that with the increasing length of time slice (from 1 day to 9 days), except pH and turbidity, the configuration entropy curve of other parameters has only one peak (1 day, 1.5 days, 2 days) to the valley (2.5 days and later), which confirms a hypothesis that the configuration entropy will not have a valley when the length of time grid is significantly greater than the width. When the length of the time slice is more than 2.5 days, even if the length of the time slice is increased, the overall shape of configuration entropy curve does not change significantly, suggesting that the configuration entropy of specific water quality parameters did not show temporal heterogeneity in a long-time period observation. We also assumed that temporal fractal phenomena exist in some water quality parameters consistent with previous studies. More analysis is in progress.

How to cite: Pang, T., Jiang, J., Wang, P., Zheng, Y., and Zheng, T.: Configuration entropy analysis of river water quality dynamics under fine time resolution and network topology, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12346, https://doi.org/10.5194/egusphere-egu22-12346, 2022.

11:35–11:41
|
EGU22-12588
|
ECS
|
Virtual presentation
A complex network perspective on catchment water quality dynamics: characteristics, pattern, and drivers 
Qingzhi Wen, Jiping Jiang, and Bellie Sivakumar

Understanding the connections in the dynamics of water quality at different locations in a catchment is important for catchment studies and watershed management. Complex network science provides effective ways to uncover connections and patterns in catchment water quality dynamics. This study  investigates the spatial connections in each of five water quality indexes (Chloride, Dissolved oxygen, pH, Total nitrogen, and Total organic carbon) and flow rate in the Chesapeake Bay basin, USA.High-resolution data (five minutes) from 120 water quality monitoring stations are analyzed. 1) The clustering coefficient (CC) and degree distribution methods are employed to examine the connections and identify the type of the water quality networks. The results indicate that the networks of water quality parameters are  scale-free. The power-law (γ) values of for the networks of Chl, DO, flow rate, pH, TN, TOC are 0.74, 0.67, 0.37, 2.0, 0.57 and 1.2, respectively. 2) Monte Carlo simulation of degree distributions and clustering coefficients (CC) shown that all water quality parameters present a decrease in the CC along with the turn down of the threshold of correlation coefficient (R), but the R threshold for DO and flow rate was 0.9. Other water quality parameters showed a sharp decline in the range of correlation coefficient (R) of 0.3-0.6, show a gentle decrease, and then decrease sharply, with an inverse s-curve. 3) All the WQ parameters show stable patterns of CC versus R, for different sizes of networks, arrived by randomly reducing the number of nodes (i.e. stations) of the networks. This seems to indicate that the pattern is an internal systemically feature of the networks, regardless of the node selected for analysis. The variations of CC values for the different stations in the networks  with different R values also help identify the “heat area” of the whole catchment, which  has some nodes with stable large CC. For the different water quality parameters, the heat area is basically the same, except for pH and TN for which the area is much smaller. The present findings on the characteristics, patterns, and drivers of water quality dynamics in catchments have important implications for water quality studies, especially in large networks of monitoring stations.

How to cite: Wen, Q., Jiang, J., and Sivakumar, B.: A complex network perspective on catchment water quality dynamics: characteristics, pattern, and drivers, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12588, https://doi.org/10.5194/egusphere-egu22-12588, 2022.

11:41–11:47
|
EGU22-13100
|
Highlight
|
Virtual presentation
Machine Learning Methods with the Standardized VPD Drought Index to Identify and Assess Drought in the United States 
Brandi Gamelin, Vishwas Rao, Julie Bessac, and Mustafa Altinakar

Extreme drought has a strong socio-economic impact on the human environment, especially where surface and ground water supplies are significantly reduced due to reduced stream flow, reduced hydroelectric generation, and increased ground water pumping for agricultural and human consumption. This reduction will likely increase in the future as drought is expected to increase in the United States due to global warming and climate change. However, identifying drought is problematic due to the lack of standardized classification or reliable methods for drought prediction. Recently, machine learning techniques have been applied to drought indices to identify drought features and for risk assessment. For this work, we utilize unsupervised machine learning (ML) computational algorithms to identify drought characteristics with a new drought index based on vapor pressure deficit (VPD). The Standardized VPD Drought Index (SVDI) is used to cluster points with common features to characterize spatial and temporal drought characteristics. The SVDI is calculated with the NASA’s Land Surface Assimilation System (NLDAS) data from 1990-2010. Several ML cluster techniques (e.g. HMM, k_means, BIRCH, DBSCAN) are applied to the SVDI to identify known short and long term drought events. Optimized techniques will be applied to downscaled global climate models (e.g. CCSM4, GFDL-ESM2G, and HadGEM2-ES) based on the 8.5 Representative Concentration Pathway (RCP8.5). From the space-time clustering algorithm, we will extract the spatiotemporal information for each identified event as a means of determining the probability of each type of event under global warming in the future.

How to cite: Gamelin, B., Rao, V., Bessac, J., and Altinakar, M.: Machine Learning Methods with the Standardized VPD Drought Index to Identify and Assess Drought in the United States, EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-13100, https://doi.org/10.5194/egusphere-egu22-13100, 2022.

11:47–11:50
Closing Remarks