SSS10.6 | Quantifying and communicating uncertain information in earth and environmental sciences
PICO
Quantifying and communicating uncertain information in earth and environmental sciences
Convener: Gerard Heuvelink | Co-conveners: Lorenzo Menichetti, Alice Milne, Madlene Nussbaum, Nadezda Vasilyeva
PICO
| Mon, 24 Apr, 16:15–18:00 (CEST)
 
PICO spot 3b
Mon, 16:15
In complex earth systems, uncertain information (whether in measurements, maps or models) is the norm, and this impinges on most knowledge that earth scientists generate. It is important to quantify and account for this uncertainty, otherwise results can be misleading. This is particularly important when predictions are used in a decision-making process where the end user needs to be able to properly evaluate the risks involved.

Quantifying uncertainty is a difficult challenge, that continually calls for the development of more refined tools. Many diverse methods have been developed, such as for spatial prediction using kriging and machine learning, stochastic simulation, uncertainty propagation and in expert elicitation, but many challenges still remain. A second and often overlooked challenge with uncertainty is how to communicate and visualize it effectively to end users such as scientists, engineers, policy makers, regulators and the general public.

In this session, we will examine the state of the art of both uncertainty quantification and communication in earth and environmental sciences. We welcome submissions on three components of the problem: 1) new methods and applications of uncertainty quantification; 2) use of uncertainty information in decision-making and for risk assessment; and 3) efficient and effective communication and visualization of uncertainty to end-users. Dealing with uncertainty across all these three components is a truly multidisciplinary task, requiring input from diverse disciplines (such as earth and environmental science, statistics, economics and psychology) to ensure that it is successful. The main aim of this session is to bring these disciplines together so that we can learn from each other. Previous topics discussed in this session include (but are not limited to) quantifying uncertainty in carbon budgets for climate change research, advising farmers on fertilizer application and liming given uncertainties in soil nutrients and weather forecasts, statistical modelling of laboratory measurement errors, and communication of uncertainty in micronutrient concentrations in staple crops to decision makers.

PICO: Mon, 24 Apr | PICO spot 3b

Chairpersons: Gerard Heuvelink, Madlene Nussbaum, Alice Milne
16:15–16:20
16:20–16:30
|
PICO3b.1
|
EGU23-15802
|
solicited
|
On-site presentation
Alice Milne, Bader Oulaid, imane el fartassi, Joanna Zawadzka, Rafiq el Alami, Khouloud Tabiti, Helen Metcalfe, Alhousseine Diarra, Vasthi Alonso chavez, Toby Wayne, and Ron Corstanje

Policy makes and other stakeholders working to improve the sustainability of agriculture need access to information on soil spatial variation, and the impacts of management strategies that policy might promote. Models and data can provide such information but typically such products are developed in isolation and so do not allow for an integrated trade-off analysis. To overcome this limitation, we have developed a GIS- modelling framework (GIS-MF) that allows users to interrogate integrated models and data layers. This GIS-MF prototype has been developed for the Tensift watershed which is in the region of Marakesh-Safi, Morocco. The alpha version of the software contains four main components: (i) a field to watershed-scale Digital Soil Mapping viewer, (ii) a watershed scale Ecosystems Services Report viewer (iii) an Interactive Ecosystem Service and Environmental Impacts viewer that allows trade-offs to be explored and (iv) a field-scale yield prediction tool.  For each component it is essential to communicate the uncertainties associated with predictions in a way that is both informative and intuitive to the end users. We have explored several ways for communicating uncertainty that we will present. For the DSM viewer we consider both methods that show uncertainty distributions as well as the probability of exceeding agronomically relevant thresholds. We take a similar approach for the Ecosystems Services Report viewer, and the yield prediction tool. For the uncertainties related to our integrated trade-offs we consider approaches to communicate the changes in multiple objectives. We draw on previous analyses to make conclusions and we will ask the EGU audience their views on each of the methods. 

How to cite: Milne, A., Oulaid, B., el fartassi, I., Zawadzka, J., el Alami, R., Tabiti, K., Metcalfe, H., Diarra, A., Alonso chavez, V., Wayne, T., and Corstanje, R.: Communicating the Uncertainty in Predictions from a GIS-Modelling Framework, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15802, https://doi.org/10.5194/egusphere-egu23-15802, 2023.

16:30–16:32
|
PICO3b.2
|
EGU23-2353
|
ECS
|
On-site presentation
Laura Dawkins

Climate change adaptation decisions often require the consideration of risk rather than the environmental hazard alone. One approach for quantifying risk is to use a risk assessment framework which combines information about hazard, exposure and vulnerability to estimate risk in a spatially consistent way. In recent years, publicly available, open-source risk assessment frameworks have been made available, including the CLIMADA tool. Such tools are increasingly being used in combination with ensembles of climate model projections to quantify risk on climate timescales, presenting the ensemble spread as a measure of climate uncertainty. As climate models are computationally expensive to run, this quantification of uncertainty derived from the ensemble of projections is often limited by the number of members available.

We present a novel framework involving the application and extension of the CLIMADA open-source climate risk assessment tool, demonstrating an approach for providing a richer quantification of uncertainty. We show how a statistical Generalised Additive Model, involving an `ensemble member' random effect term, can be used to statistically represent the climate model ensemble summary of risk and be used to simulate many more realisations of risk, representative of a larger collection of plausible ensemble members. We present an application of the framework to an idealised example related to heat-stress and the associated risk of reduced outdoor physical working capacity in the UK.

How to cite: Dawkins, L.: A framework for uncertainty quantification in climate risk assessments (POSTER), EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-2353, https://doi.org/10.5194/egusphere-egu23-2353, 2023.

16:32–16:34
|
PICO3b.3
|
EGU23-11306
|
ECS
|
On-site presentation
Florian Darmann, Monika Kumpan, Irene Winkler, Peter Strauss, and Thomas Weninger

As the measurement of soil hydraulic properties is time-consuming and expensive, they are often computed from easily measurable soil properties via pedotransfer functions (PTFs). There are plenty of existing PTFs which were mainly derived for specific regions or catchments.

In our study, new PTF’s for Austrian soils were developed to estimate soil hydraulic properties such as field capacity or permanent wilting point. The PTFs were built by applying the random forest method, including prediction of uncertainty measures.

The used database includes soil physical data from different project and monitoring campaigns all over Austria and represents many different landscapes and soil types within the country. It consists of 2300 samples from 518 agricultural sampling locations.

For the derivation of new PTF’s, several inputs were available. Since the model is to be applied to standard data from official Austrian soil mapping, we focused on the therein existing data as input variables. We tested four possible combinations for each target property, with particle size distribution and depth as fixed variables. Bulk density and soil organic matter were added in different compositions.

To ensure the applicability of the newly derived PTF’s, the available data set was randomly divided into two subsets. A fraction of 80 % of the samples were contained in training data set for the derivation of the prediction models, the validation set includes 20 % of the data for determining statistical parameters and for comparison with other PTFs. This comparison showed good applicability of the new PTFs in relation to the previous models with R² ranging from 0,73 to 0,83 per target variable.

Due to the fact that each prediction contains uncertainties, uncertainty maps for all agricultural areas in Austria were created. Therefore the 25% and 75% percentile and the respective interquartile range (IQR) of our predicted properties were determined. The results show mean values for the IQR between 6,8 % for field capacity at 60 hPa and 5,7 % for permanent wilting point.

As main outcome, new models for hydraulic properties of Austrian soils were produced. The comparison with established PTFs showed higher accuracy for the prediction of soil properties in Austrian soils than other established PTFs. With the implemented prediction of uncertainty, areas with good predictions can be identified as well as areas with high uncertainties.

How to cite: Darmann, F., Kumpan, M., Winkler, I., Strauss, P., and Weninger, T.: Progressing pedotransfer functions for nation-wide mapping of soil hydraulic properties, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11306, https://doi.org/10.5194/egusphere-egu23-11306, 2023.

16:34–16:36
|
PICO3b.4
|
EGU23-7145
|
On-site presentation
Carl Bolster and Peter Vadas

The long-term application of inorganic fertilizer and animal manure to agricultural fields has resulted in soil test P (STP) concentrations that greatly exceed the agronomic critical level in many areas of the world. This build-up of soil P, often referred to as legacy P, can increase the risk of P being mobilized during runoff and leaching events, even years or decades after P inputs have decreased or ceased. If this mobilized P reaches P-limited surface waters, eutrophication can result, leading to serious water quality problems that can adversely impact the environment, human health, and recreational activities. Due to the cost and effort involved in soil, manure, and runoff sampling and testing, as well as implementation of best management practices, nutrient management policies and strategies are often guided by computer model predictions. However, computer model predictions are inherently uncertain, thus it is important to account for this uncertainty when interpreting modeling data and using modeling results to guide decision making

In this study we conducted a sensitivity and uncertainty analysis using the Annual P Loss Estimator (APLE) model focusing on model predictions of STP. We calculated and evaluated the sensitivity coefficients of predicted STP and changes in STP using 1- and 10-yr simulations with and without P application. We also compared two methods for estimating prediction uncertainties: first-order variance approximation (FOVA) and Monte Carlo simulation (MCS). Finally, we compared uncertainties in APLE-predicted STP to uncertainties in measured STP collected from multiple sites in Maryland under different manuring and cropping treatments. Results from our sensitivity analysis showed that predicted STP and changes in STP for 1-yr simulations without P inputs were most sensitive to initial STP whereas model STP predictions were most sensitive to manure and fertilizer application rates when sensitivity analyses included P inputs. For the 10-yr simulations without P application inputs, the range in sensitivity coefficients for crop uptake and precipitation were much greater than for the 1-yr simulations. Prediction uncertainties from FOVA were comparable to those from MCS for model input uncertainties up to 50%. Using FOVA to calculate APLE STP prediction uncertainties using the Maryland data set, the mean measured STP for nearly all site years fell within the 95% confidence intervals of the STP prediction uncertainties. Our results provide users of APLE insight into what model inputs require the most careful measurement when using the model to predict changes in STP under conditions of P drawdown (i.e., no P application) or P buildup. Our results also demonstrate the importance of including model prediction uncertainties when estimating long-term drawdown of STP in agricultural fields.

How to cite: Bolster, C. and Vadas, P.: Sensitivity and uncertainty analysis for predicted soil test phosphorus using the APLE model, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7145, https://doi.org/10.5194/egusphere-egu23-7145, 2023.

16:36–16:38
|
PICO3b.5
|
EGU23-6954
|
ECS
|
On-site presentation
Assessment of the topsoil organic carbon saturation in Hungary using machine learning-based pedotransfer function with uncertainty propagation
(withdrawn)
Gábor Szatmári, Gergely Jakab, Annamária Laborczi, Zoltán Szalai, and László Pásztor
16:38–16:40
|
PICO3b.6
|
EGU23-4250
|
ECS
|
On-site presentation
|
Franca Giannini Kurina, Johannes Lund Jensen, Bent Tolstrup Christensen, and Jørgen Eriksen

C-TOOL is a simple flexible SOM turnover model competitive with other alternatives that often demands more input information. As most SOC turnover models, C-TOOL rely on plant C inputs derived from measured agricultural yields using simple allometric equations that establish the relation between C inputs and crop yields. The main sources of uncertainty in C turnover models rely on C input calculations and parametrization of initial pool distribution. Since the main output is the temporal dynamic SOC stock, we should refer to long-term data that can be challenging to obtain to appreciate meaningful changes. Nevertheless, in Denmark an experiment from 1981 to 2019 examined the effect of annual additions of different levels of C inputs on SOC storage. This experiment produced a robust validation platform for soil C modelling, permitting to account not only the temporal variability but also the variability beneath each treatment. In this work we implemented the C-TOOL model at a plot level, based on this precise data from spring barley straw disposal at plot level in order to explore SOC turnover modeling uncertainties and validate it performance. We performed a variance-based sensitivity analysis to evaluate how sensitive are C inputs calculations to allometric parametrization using the distributional information on spring barley allometrics (grain yield, harvest index and root biomass and root exudates). After exploring alternatives on the model parametrization related to allometric, initial soil C conditions and initialization period, we arranged a simulation design to test all the possible combinations to assess how accurate is the model in predicting the temporal plot variability of SOC. To be able to perform the numerous scenarios we worked on a feasible computational implementation trough R. Finally, we studied how dependent is the lack of fit to the alternative parametrization and the spatiotemporal variability of field conditions trough a variance component analysis. From the C input variance-based sensitivity analysis we conclude that root exudates and root biomass are the most sensitive parameters. Validation results show that C-TOOL was able to accurately describe the temporal dynamic of SOC in the topsoil due to non-significant differences between simulated and observed data. All the alternative parametrizations register a prediction error below 15 % related to the mean showing differences between observed and predicted between 3.60 and 6.46 Mg C/ha in the top 20 cm depth. This lack of fit was mainly explained by the spatiotemporal (year and block) variability rather than the parametrization alternatives tested. Nevertheless, we conclude that is relevant to focus on the initial soil C condition parametrization but not the in-situ measurements of harvest index. Besides, using a fixed amount of root biomass for spring barley presented better than using the standard allometrics. Further studies dive into a global sensitivity analysis on the multivariate variability distribution of all the inputs involved to get to a robust uncertainty estimation.

How to cite: Giannini Kurina, F., Lund Jensen, J., Tolstrup Christensen, B., and Eriksen, J.: Long term plot scale variability to explore Soil Carbon turnover modeling uncertainties: a C-TOOL implementation, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-4250, https://doi.org/10.5194/egusphere-egu23-4250, 2023.

16:40–16:42
|
PICO3b.7
|
EGU23-1500
|
ECS
|
On-site presentation
|
Yi Xiao, Jie Xue, Xianglin Zhang, Nan Wang, Emanuele Lugato, Dominique Arrouays, Zhou Shi, and Songchao Chen

Soil organic carbon (SOC) sequestration is a promising natural climate solution for capturing atmospheric CO2, and it provides crucial co-benefits in improving soil functions and services at the same time. Given that SOC is not a single, uniform entity, further knowledge of SOC fractions with distinctive properties, such as particulate organic carbon (POC) and mineral associated organic carbon (MAOC), is necessary in order to fully comprehend how SOC responds to environmental changes. Despite their enormous significance, POC and MAOC information is still scarce in the soil databases, especially on a large scale. The pedotransfer function (PTF) is a useful method for estimating missing soil parameters, but its application in SOC fractions has not received much attention. We assessed the potential of MAOC prediction using machine learning-based PTF (random forest (RF), Cubist, and gradient boosted machine (GBM)) along with predictor selection methods (recursive feature elimination (RFE), and forward recursive feature selection (FRFS)) on 352 representative mineral topsoil samples (0-20 cm) from across Europe. The repeated validation (100 times) revealed that machine learning-based PTFs were capable of accurately predicting MAOC. RFE can effectively reduce the number of predictors from 21 to 12 with comparable performance to the models using all predictors. With only 6 predictors (SOC, silt + clay, nitrogen, nitrogen deposition, soil erosion, and sand), the suggested FRFS algorithm outperformed RFE and had the best model parsimony. Of the three machine learning models, Cubist performed the best when combined with FRFS. Our results also showed that, when compared to a single machine learning model, five model ensemble approaches can increase model accuracy and robustness. This study offers a valuable reference for coupling PTF and legacy soil databases, in order to improve the spatial coverage and effectiveness of SOC fraction forecasts based on machine learning.

How to cite: Xiao, Y., Xue, J., Zhang, X., Wang, N., Lugato, E., Arrouays, D., Shi, Z., and Chen, S.: Ensemble machine learning improves pedotransfer functions for predicting soil mineral associated organic carbon, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1500, https://doi.org/10.5194/egusphere-egu23-1500, 2023.

16:42–16:44
|
PICO3b.8
|
EGU23-689
|
ECS
|
On-site presentation
|
Kerstin Rau, Thomas Gläßle, Philipp Hennig, and Thomas Scholten

Artificial neural networks (ANN), which are mainly used in pattern and image recognition, have now found a wide range of applications in soil science and geoscience. They have proven to be a useful tool for complex questions that also involve a large amount of data, for example prediction of soil classes or soil properties on various scales. However, we face two main challenges when applying ANN: In their basic form, deep-learning algorithms do not provide interpretable predictive uncertainty. Thus, in geosciences and in particular in soil science, they have been used more as black-box models and properties of a machine learning model such as the certainty and plausibility of the predicted variables, for example soil classes, were interpretation by experts rather than quantified by metrics validating the ANN. In most cases regression coefficients or comparable statistical measure are reported for the overall performance of the model. This leads to the second challenge, that is that these algorithms have high confidence of their predictions in areas far away from the training area or in areas where they receive only little information from a small number of data points.
In order to gain a better understanding of these aforementioned properties, we implement in our explorative study on soil classification a Bayesian deep learning approach (i.e., a method to add uncertainty to deep networks) known as last layer Laplace approximation. This is a technique that can be applied as a post-hoc "add-on" without destroying the otherwise good performance of deep classifiers. It helps us to correct the overconfident areas without reducing the accuracy of our prediction, giving us a more realistic uncertainty expression of the model's prediction.  
Our predictor variable soil type provides us with a large amount of complex information about soil processes and properties, which is a great advantage since it would take a lot of time and money to collect all this information individually. At the same time, soil maps are in high demand by authorities, construction companies or farmers. In our study area around Tübingen in southern Germany, there are 41 different soil types, determined according to the German soil classification, sub divisible into typical soils of the Neckar and Ammer valleys, the Swabian Jura and Black Forest, and non-area related soil. In addition to the underlying soil map, remotely sensed variables, a digital elevation model and its derivatives are used as input to the ANN, which is designed to learn the relationship between these and the soil type. As a test case, we then explicitly exclude the Swabian Jura and Black Forest in the training area but include them as prediction regions. Both regions are characterized by very different soil types compared to the rest of the study area due to their considerably different geology, climate, and terrain. Our goal is then to enrich soil type maps with a structured uncertainty to better understand the causality of machine learning models in soil science and their transferability to regions other than the training and validation area.

How to cite: Rau, K., Gläßle, T., Hennig, P., and Scholten, T.: How can we quantify, explain and apply the uncertainty of complex soil maps predicted with neural networks?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-689, https://doi.org/10.5194/egusphere-egu23-689, 2023.

16:44–16:46
|
PICO3b.9
|
EGU23-3313
|
ECS
|
On-site presentation
|
Jonas Schmidinger and Gerard B.M. Heuvelink

There is growing interest in the field of digital soil mapping (DSM) to quantify the underlying uncertainty of point predictions through different probabilistic prediction models. Uncertainty in DSM is often described in the format of a prediction interval (PI). Yet, PIs or uncertainty estimates in general are only of value for end users if they have good quality, i.e. when they are reliable. Reliability refers to the consistency between predicted conditional probabilities and observed frequencies of independent validation data. Ideally, PIs are also sharp, which refers to the concentration of probabilistic information, i.e. the narrowness of conditional probability distributions. The prediction interval coverage probability (PICP) is currently used in DSM to assess the reliability of PIs but it is ignorant to a potential one-sided bias of its bounding quantiles. Therefore, we propose to complement the current validation procedure with new metrics suggested in the broader probabilistic literature. This includes metrics that do not only evaluate uncertainty estimates in PI format but also quantiles or full conditional probability distributions. The newly proposed metrics are the quantile coverage probability (QCP), the probability integral transform (PIT) and so-called proper scoring rules for relative comparisons. Examples of scoring rules are the continuous ranked probability score (CRPS), which can be decomposed into a reliability part (RELI) and the interval score (IS). Sharpness can be evaluated through the prediction interval width (PIW). We illustrated the use of the various metrics in a case-study using soil pH data from The Land Use and Coverage Area Frame Survey (LUCAS). Thereby, uncertainty estimates of five different models were compared: Kriging with external drift (KED), quantile regression forest (QRF), quantile regression post-processing of a random forest (QRPP RF), quantile regression neural network (QRNN) and a reference null-model (NM).  KED, NM and QRPP RF showed very good reliability according to QCP, PICP, PIT and RELI. QRF was slightly pessimistic in the centre and QRNN very overoptimistic at the edges of the conditional probability distributions. Despite this, QRF performed best according to mean CRPS and mean IS because it produced fewer outliers at the edges. As expected, NM had the lowest sharpness, i.e. the largest PIW values. Sharpness of the other models was overall similar but QRNN had sharper predictions at the edges and QRF was less sharp in the centre of the conditional probability distributions. Lastly, we also generated PIW maps to indicate the spatial uncertainty of the five prediction models. The spatial variability of PIW was larger for QRF, QRNN and QRPP RF in comparison to KED. Whereas with NM, PIW was completely uniform. 

How to cite: Schmidinger, J. and Heuvelink, G. B. M.: Validation of uncertainty estimates in digital soil mapping, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-3313, https://doi.org/10.5194/egusphere-egu23-3313, 2023.

16:46–16:48
|
PICO3b.10
|
EGU23-17408
|
On-site presentation
Stéphane Burgos, Simon Tanner, and Madlene Nussbaum

Detailed soil maps are crucial to balance interests of urban planning and soil protection. In Switzerland, high-quality arable land is protected by an inventory and surfaces inside the inventory can only exceptionally be developed. Recently, legal frameworks were modified to require detailed soil maps to change the inventory. So far, conventional soil maps at scale of 1:5'000 – directly linked to a prescribed sampling density – are acknowledged to be sufficiently accurate. As conventional mapping is very time-consuming digital soil mapping is currently considered by regional governments to accelerate data collection.

To test digital soil mapping to generate such detailed maps we selected a typical area on the Swiss Plateau of 800 hectares. We sampled 1'120 locations by feature space coverage design. Spatial predictions were computed for soil properties and soil suitability classes were derived. Additional independent validation data was sampled by a stratified random design at 120 locations and used to evaluate overall prediction accuracy. For rootable soil depth, the main soil attribute decisions are based on, model uncertainty was quantified by quantile regression forest.  

Various representations of uncertainty at selected point locations and for a map excerpt were prepared. Following one of the recently formulated ten Pedometrics challenges, we evaluated (mis)understanding thereof by in-depth guided expert interviews with six inventory decision makers. Level of acceptable uncertainty for the inventory and whether end users rather trusted certain sampling densities as opposed to statistical accuracy measures was further discussed in the interviews.

How to cite: Burgos, S., Tanner, S., and Nussbaum, M.: What is the accepted level of erroneous decision for the Swiss arable land inventory? Uncertainty perception of digital soil maps., EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-17408, https://doi.org/10.5194/egusphere-egu23-17408, 2023.

16:48–16:50
|
PICO3b.11
|
EGU23-17024
|
On-site presentation
Linda Lilburne, Gerard Heuvelink, and Anatol Helfenstein

There is an implicit quality associated with all soil maps which depends on a range of factors including the mapping algorithm, the extent and quality of the calibration data, and the quality and relevance of the co-variates. Spatially explicit prediction uncertainty information is increasingly provided when digital soil mapping methods are used. This may take the form of a 90% prediction interval such as that supplied by ISRIC in their global SoilGrids product. In this study we used independent data sets of particle size measurements from New Zealand and the Netherlands to investigate the accuracy of the prediction interval information in the SoilGrids clay, sand and silt layers. While prediction intervals were wide for both countries, we had expected that these would be narrower for the Netherlands as SoilGrids has more calibration points in the Netherlands than in New Zealand. Spatially, there was much more variation in the prediction interval width in the Netherlands than in New Zealand, with some areas being less uncertain and some highly uncertain. New Zealand had uncertain predictions for the entire country, although the prediction intervals were not as wide as in the most uncertain areas in the Netherlands.

Independent validation showed that the clay prediction intervals were too wide: for New Zealand between 95 and 98% of the validation data were within the 90% prediction interval, for the Netherlands this was between 95 and 97%. For sand we found the opposite, with only between 60 and 70% of the data falling in the 90% prediction interval for New Zealand and between 77 and 88% for the Netherlands. Comparison of prediction errors with prediction interval widths showed that the prediction errors tended to be larger at locations with wide prediction intervals, although this relationship was clearer for the Netherlands than for New Zealand. Estimates that fell outside the sand prediction interval were associated with narrow as well as wide prediction interval widths. Our analyses highlight the importance of users considering soil uncertainty information before using the soil data. It is also important that producers of soil information document the accuracy limitations to help guide potential users and that they evaluate the validity of the uncertainty information prior to release.

How to cite: Lilburne, L., Heuvelink, G., and Helfenstein, A.: Interpreting and evaluating digital soil mapping prediction uncertainties, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-17024, https://doi.org/10.5194/egusphere-egu23-17024, 2023.

16:50–16:52
|
PICO3b.12
|
EGU23-8081
|
Virtual presentation
Andrés Almeida-Ñauñay, Ernesto Sanz, Miguel Quemada, Juan José Martin-Sotoca, and Ana María Tarquis

Grasslands are one of the most important and complex ecological systems due to their characteristic dynamics influenced by meteorological and climate patterns. In this sense, drought is one of the most challenging obstacles to overcome. Especially in semi-arid areas, where biomass production is greatly limited by the amount of precipitation. In this line, remote sensing methods have been demonstrated to be a valuable instrument for monitoring vegetation in wide areas, and vegetation indices (VIs) have shown a high sensitivity to vegetation variations. In this line, the soil water content has been shown to be a key factor in vegetation growth. In this work, we compare the temporal dynamics of two semi-arid grassland areas based on a soil water content index estimated in each area.

We selected two semi-arid areas in Spain and time series of VIs are built based on multispectral images of MODIS TERRA product with a temporal resolution of 8 days in each area. Red (620-670 nm) and Near Infrared (841-876 nm) reflectance channels were extracted and filtered by the quality of the pixel. Then, a soil water content index (WCI) is calculated based on the water balance of the soil over time. Recurrence plots (RP) and recurrence quantification analysis (RQA) were calculated to characterize the influence of soil water content on vegetation index dynamics. The characterisation was based on various RQA complexity measurements, including Determinism (DET), among others.

In general, our results revealed that WCI was able to distinguish between areas. RPs revealed a different temporal pattern in each area using WCI and VIs. Furthermore, RQA measurements revealed that the dry area presented a different dynamic in contrast to the wetter area. In general, WCI was shown to be a useful index in characterizing soil water content, and recurrence plots were able to describe and characterise the dynamics of each area.

Acknowledgements: The authors acknowledge the support of Clasificación de Pastizales Mediante Métodos Supervisados - SANTO, from Universidad Politécnica de Madrid (project number: RP220220C024).

 

References

Almeida-Ñauñay, A.F., Benito, R.M., Quemada, M., Losada, J.C., Tarquis, A.M., 2022. Recurrence plots for quantifying the vegetation indices dynamics in a semi-arid grassland. Geoderma 406, 115488. https://doi.org/10.1016/j.geoderma.2021.115488

Almeida-Ñauñay, A.F., Benito, R.M., Quemada, M., Losada, J.C., Tarquis, A.M., 2021. The Vegetation–Climate System Complexity through Recurrence Analysis. Entropy 23, 559. https://doi.org/10.3390/e23050559

Martín-Sotoca, J.J., Saa-Requejo, A., Moratiel, R., Dalezios, N., Faraslis, I., Tarquis, A.M., 2019. Statistical analysis for satellite-index-based insurance to define damaged pasture thresholds. Nat. Hazards Earth Syst. Sci. 19, 1685–1702. https://doi.org/10.5194/nhess-19-1685-2019

Sanz, E., Saa-Requejo, A., Díaz-Ambrona, C.H., Ruiz-Ramos, M., Rodríguez, A., Iglesias, E., Esteve, P., Soriano, B., Tarquis, A.M., 2021. Normalized Difference Vegetation Index Temporal Responses to Temperature and Precipitation in Arid Rangelands. Remote Sens. 13, 840. https://doi.org/10.3390/rs13050840

How to cite: Almeida-Ñauñay, A., Sanz, E., Quemada, M., Martin-Sotoca, J. J., and Tarquis, A. M.: Soil water content in vegetation indices dynamics through a recurrence plots approach, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-8081, https://doi.org/10.5194/egusphere-egu23-8081, 2023.

16:52–16:54
|
PICO3b.13
|
EGU23-15495
|
Virtual presentation
Ruhollah Taghizadeh-Mehrjardi, Gerard B.M. Heuvelink, and Thomas Scholten

Machine learning models have been widely used in digital soil mapping to predict the spatial distribution of soil properties across the landscape. However, due to the unpredictable behavior of soil properties, insufficient training sample size, input data error, uncertainty in model parameters and structure, uncertainty inherently exists in machine learning predictions. Therefore, knowledge of prediction uncertainty is vital for decision-makers and end-users. This study quantifies the overall uncertainty of digital soil maps of Germany for soil acidity, organic carbon, cation exchange capacity, and clay using quantile regression forest (QRF) and artificial neural network (ANN) models. Here, we propose the use of a novel ANN model that directly estimates the lower and upper bounds of a prediction interval by using an architecture with two output neurons. A multi-objective evolutionary algorithm (i.e., non-dominated sorting genetic algorithm II) was employed to parameterize the ANN weights. The results of the modeling indicated that ANN performed better than the QRF for predicting soil properties. Additionally, the ANNs produced narrower prediction intervals in comparison to the QRF. Most importantly, ANN yielded prediction interval coverage probabilities that were more closely aligned to their associated confidence levels in comparison to the QRF. In general, the ANNs were not only effective in predicting soil properties, but they were also effective in constructing reasonable prediction intervals for soil characteristics; and therefore, it is recommended to be used for predicting soil properties and quantifying their uncertainty in digital soil mapping.

Keywords: uncertainty estimation, prediction interval, multi-objective optimization, artificial neural networks, machine learning, Germany

How to cite: Taghizadeh-Mehrjardi, R., B.M. Heuvelink, G., and Scholten, T.: Quantification of uncertainty using artificial neural networks for mapping of soil properties in Germany, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15495, https://doi.org/10.5194/egusphere-egu23-15495, 2023.

16:54–16:56
|
PICO3b.14
|
EGU23-17330
|
Virtual presentation
Formalization of experimental protocols for automated data processing and uncertainty estimation
(withdrawn)
Nadezda Vasilyeva, Artem Vladimirov, and Taras Vasiliev
16:56–18:00