Proper use of land for various purposes requires digital soil mapping. Nowadays, the use of machine learning techniques in digital soil mapping has been a major breakthrough. The resulting maps are accurate, objective and easily reproducible. Furthermore, the process is less expensive than traditional methods. A supervised machine learning technique needs soil samples and environmental covariates for the creation of a map. The lack of soil samples in some regions is a major issue in digital soil mapping. In the case of acid sulfate (AS) soils, the absence of maps can be a high risk for the environment. This is due to AS soils can lead to environmental damage when they are oxidized during the drainage of the land. Therefore, in the unavailability of maps, AS soils may be accidentally drained by external activities related to agriculture, forestry or urban activities. A possible solution for mapping areas with few soil samples is to use soil samples from other regions. In a previous work, we showed that a machine learning model is able to correctly classify soil samples from a region where it had not been trained if the composition of the soils of the region where it has been trained is the same [1]. In this study, we have analyzed whether a machine learning technique is capable of predicting the AS soils of a region when the model has been trained in a region with a very different soil composition. Four different regions located in the coastal areas of Finland have been considered. The machine learning method used is Random Forest, which has shown very high predicted abilities for the classification and prediction of AS soils [2-5]. The results show that the model is able to correctly predict AS soils when the model is trained with soil samples from other regions in most cases. This is a significant advancement in the field because it permits the first recognition of regions with a limited number of soil samples.
[1] V. Estévez et al. 2024. “A First Approximation for Acid Sulfate Soil Mapping in Areas with Few Soil Samples”. Environ. Sci. Proc. 2024, 29, 4. https://doi.org/ 10.3390/ECRS2023-15831
[2] V. Estévez et al. 2022. “Machine learning techniques for acid sulfate soil mapping in southeastern Finland”. Geoderma 406 (2022) 115446.
[3] V. Estévez et al. 2023. “Improving prediction accuracy for acid sulfate soil mapping by means of variable selection”. Front. Environ. Sci. 11:1213069 (2023).
[4] V. Estévez et al. 2024. “Acid sulfate soil mapping in western Finland: How to work with imbalanced datasets and machine learning”. Geoderma 447 (2024) 116916.
[5] V. Estévez et al. 2025. “Mapping of acid sulfate soil types in Laihianjoki River catchment: A multiclass classification” . European Journal of Soil Science 76, no. 5: e70204.
How to cite: Estévez, V., Mattbäck, S., and Boman, A.: Cross-regional transfer learning to predict acid sulfate soils in Finland using Random Forest, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-550, https://doi.org/10.5194/egusphere-egu26-550, 2026.
Accurate and spatially explicit soil information is a fundamental prerequisite for Earth system modelling, land surface simulations, and global environmental assessments. Although existing global soil datasets, such as GSDEv1 (The first version of this study, DOI:10.1002/2013ms000293), HWSD 2.0 and SoilGrids 2.0, have substantially advanced large-scale soil representation, they still exhibit limitations in spatial resolution, vertical consistency, and physical realism. Here we present GSDEv2, a next-generation global soil physical and chemical property dataset developed to meet the increasing demand for high-resolution Earth system modelling. GSDEv2 provides seamless global predictions at 90 m spatial resolution for nearly 30 static soil properties, including soil organic carbon, texture fractions, bulk density, porosity, and related variables, across six standard depth intervals (0–200 cm). The dataset is built upon an unprecedented compilation of approximately 23 million soil profiles, primarily sourced from the World Soil Information Service (WoSIS) and complemented by high-quality regional and national datasets. All profiles were subjected to rigorous, pedologically informed quality control procedures to remove implausible or inconsistent observations that can bias machine-learning predictions. To better capture pedological heterogeneity, soil profiles and environmental covariates were stratified into desert, non-desert mineral, and organic soil domains. Separate Random Forest models were trained for each domain using a comprehensive set of covariates representing climate, topography, vegetation, and parent material, including AlphaEarth Foundations data. Model predictions were validated using both internal cross-validation and independent reference datasets, demonstrating clear improvements in spatial detail and physical realism compared with GSDEv1, SoilGrids 2.0, and HWSD-based products. In addition, GSDEv2 adopts a data fusion framework, allowing high-quality regional soil maps to be integrated into the global predictions while preserving global consistency. GSDEv2 represents a substantial step forward in global digital soil mapping, providing a physically consistent, high-resolution soil dataset that is better suited for hydrological, biogeochemical, and land–atmosphere modelling applications. The dataset is intended to support the GlobalSoilMap initiative and next-generation Earth system simulations.
How to cite: Shangguan, W., Shi, G., and Dai, Y.: A Global Soil Dataset for Earth System Modeling (Version 2, GSDEv2), EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6831, https://doi.org/10.5194/egusphere-egu26-6831, 2026.
Spatial predictive mapping is widely used in geoscience to generate spatially explicit maps from limited field observations and holds particular significance for soil mapping. In this approach, point-based observations are linked to spatially continuous predictor variables, and (due to expected nonlinearity) machine learning algorithms are often employed to learn their relationships and produce spatial predictions. A key challenge, however, is assessing the quality and reliability of the resulting maps.
While there is consensus that map accuracy is ideally assessed using an independent probability sample from the prediction area, such data are often unavailable. Consequently, practitioners commonly rely on splitting the available observations into training and testing sets or on repeated data partitioning via cross-validation. The resulting performance statistics are used to obtain a proxy for the final map accuracy, where cross-validation additionally supports model tuning and selection. In recent years, a considerable debate has emerged regarding how data should be partitioned into training and test sets during cross-validation. Studies have shown that estimated performance metrics can differ substantially depending on the chosen data-splitting strategy, for example, whether observations are split randomly or according to spatial structures, such as in spatial cross-validation approaches that partition data by spatial units (e.g., spatial block cross-validation or leave-region-out schemes). While some researchers argue that random cross-validation is inappropriate because it yields overly optimistic performance estimates, others contend that spatial cross-validation can be overly pessimistic and therefore advocate for random validation instead.
We argue that both spatial and random validation approaches can provide appropriate proxies for map accuracy, but their suitability depends on how well they align with the specific prediction context. Many spatial prediction tasks involve a combination of interpolation and extrapolation in geographic space, feature space, or both, and the chosen cross-validation strategy should explicitly account for this. To address this, we propose a new category of cross-validation methods, termed prediction-domain adaptive validation. Methods in this category flexibly adapt data partitioning to reflect the underlying prediction task, ensuring that validation data resemble the intended prediction scenario. To illustrate the potential of these new methods, we reproduce a simulation study, compare different validation methods, and discuss their purpose.
We show that random cross-validation methods are suitable when training samples are randomly distributed across the prediction area, whereas spatial cross-validation is better suited for extrapolation-dominated scenarios. In practice, however, most applications fall between these two extremes. In such cases, prediction-domain adaptive cross-validation can provide more reliable estimates of map accuracy, as it explicitly adapts to the underlying prediction situation. We believe that the proposed prediction-domain adaptive validation approach helps consolidate the ongoing discussion on various strategies by providing a balanced approach that yields more suitable estimates of map accuracy during cross-validation. This, in turn, supports model tuning and enhances the quality of the resulting maps, such as those generated during digital soil mapping.
How to cite: Meyer, H., Linnenbrink, J., and Nowosad, J.: A call for prediction-domain adaptive validation for assessing the performance of spatial prediction models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6965, https://doi.org/10.5194/egusphere-egu26-6965, 2026.
Marginal agricultural lands, characterised by limited productivity and constrained suitability for conventional farming, represent a growing challenge for sustainable land management and agricultural planning. While several review studies have proposed generalised threshold values to identify marginal conditions, these broad definitions often fail to capture regional heterogeneity. Consequently, robust delineation requires approaches that derive thresholds directly from the characteristics of the study area, combining data-driven analysis with decision-relevant thresholds. In this study, we developed a geospatial framework to delineate marginal croplands across Hungary using high-resolution (100 m) spatial data. Soil properties, topographic variables, and climatic indicators were analysed at the pixel level. Marginality was assessed without relying on a single dependent variable, instead applying multiple threshold-based approaches to identify unfavourable conditions across individual variables. Thresholds were derived using complementary data-driven methods, including univariate unsupervised techniques such as Gaussian Mixture Models, one-dimensional clustering, and percentile-based classification. In parallel, expert-based threshold definitions were applied using distribution-based rescaling with predefined class boundaries. Each variable was evaluated independently to derive marginality scores, which were subsequently analysed by thematic categories (soil and environmental) and combined into an integrated marginality assessment. The results show consistent spatial patterns of agricultural marginality across Hungary, driven mainly by soil-related limitations, topographic constraints, and climate. Spatial agreement among different threshold-based classifications identifies areas of stable marginality, whereas disagreement delineates transitional zones influenced by methodological choices. The framework allows systematic comparison of threshold-based delineations of marginal agricultural lands.
How to cite: Csikós, N., Laborczi, A., Takács, K., Szatmári, G., Pásztor, L., and Tóth, G.: Beyond general thresholds: A soil-based, data-driven identification of marginal agricultural lands, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13469, https://doi.org/10.5194/egusphere-egu26-13469, 2026.
Soil organic carbon (SOC) plays a central role in soil fertility, carbon sequestration, and greenhouse gas flux regulation. In agricultural landscapes, water erosion (Ew) and tillage erosion (Et) processes alter the spatial and vertical distribution of SOC, leading to significant losses in erosional areas and accumulations in depositional areas (Baert et al. 2024) 1. Understanding and quantifying these erosion-driven controls on SOC distribution is essential for the sustainable management of soils and their carbon stocks. This study aims to quantify and model these processes in an agricultural catchment of 92 ha, located in the central Belgian loess belt, by developing a novel fully explicit 3D model prediction SOC as a function of depth, Ew and Et. The performance of this 3D model has been evaluated by comparing SOC stocks obtained through (i) its integrated application, versus (ii) the more classical approach based on a fixed stock calculations per site, considering a reference depth of 0.3m and 1m. The comparison shows that both approaches have very similar performances, both in terms of random (RMSE) and systematic error (%Bias). However, the 3D model has the advantage, over the more classical approach, that it is depth-explicit, and can therefore predict SOC values at any given depth.
The study combined the use of WaTEM/SEDEM model (Notebaert et al. 2006) 2 with the sampling of 45 soil profiles until a depth of 1m and a sampling interval of 0.1 m, across different topographical positions (i.e. plateau, convexity, slope, concavity and footslope) covering a wide range of both Ew and Et values in erosional and depositional contexts. In addition, as the proposed novel DSM approach is based on the general depth distribution of Meersmans et al. (2009) 3 , it provides a set of specific parameters related to the vertical heterogeneity of SOC (i.e. SOC at the surface (SOCsurf), SOC at the bottom of the profile(SOCinf), tillage depth (td), and the rate of the exponential decline underneath the plough layer (α)), which on its turn are all expressed as a function of Ew and/or Et rates .
Our research shows that SOC stocks is two-fold higher in depositional areas as compared to eroded sites. Moreover, the present study highlights that Et and Ew are having a different impact on the redistribution, and hence both the vertical and horizontal heterogeneity, of SOC within agricultural landscapes. In this respect, Et mainly affects surface SOC stocks, whereas Ew has a greater impact on deeper stocks. Given its good performance, the presented novel spatially explicit 3D model can be considered as a promising tool for refining the quantification of SOC stocks and associated Digital Soil Mapping-products at the catchment scale.
1 Baert et al. (2024). Assessing the 3D distribution of soil organic carbon by integrating predictions of water and tillage erosion into a digital soil mapping-approach: A case study for silt loam cropland (Belgium).
2 Notebaert et al. (2006). WaTEM / SEDEM version 2006 Manual.
3 Meersmans et al. (2009). Modelling the three dimensional spatial distribution of soil organic carbon (SOC) at the regional scale (Flanders, Belgium).
How to cite: Maindiaux, M., Degré, A., Baert, P., Li, X., Colinet, G., Paternostre, L., Pissard, A., and Meersmans, J.: A Novel Depth-Explicit Model to Map the 3D Distribution of Soil Organic Carbon by Water and Tillage Erosion at the Catchment Scale., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16917, https://doi.org/10.5194/egusphere-egu26-16917, 2026.
Pedotransfer functions (PTFs) are widely used to estimate soil hydraulic properties from easily measurable soil attributes and are essential in hydrological modelling and soil evaluation. However, most established PTFs are derived from datasets dominated by agricultural soils or mineral forest soils with moderate to high bulk density (>0.9 g cm-3). As a result, forest soils characterized by low bulk density and elevated organic carbon (Corg) contents are insufficiently represented, despite their abundance in mountainous and temperate forest regions. The hydraulic characteristics of such soils, intermediate between mineral and organic soils, differ markedly and therefore the application of existing PTFs fails to adequately represent the hydraulic complexity of these soils.
In the present study, we developed and evaluated new PTFs specifically tailored to forest soils with low bulk density (0.1-0.9 g cm-3) also including organic horizons and litter layers (Corg >20%). The analysis is based on an extensive dataset of undisturbed soil samples collected from forest sites across the Eastern Alps from different depth intervals. Soil water retention curves were determined over a wide suction range using the evaporation method (HYPROP) combined with dew-point-method (WP4C). The dataset was further extended with a comparable, recently published dataset for soil hydraulic properties of forest soils from a global compilation to enable independent validation.
In a first step we investigated the dependence of derived Mualem van Genuchten parameters of the selected soil horizons on soil texture, bulk density, and Corg across different bulk-density classes. Results show a clear shift in controlling factors below a bulk density threshold of 0.9 g cm⁻³. While soil texture remains the dominant predictor in higher bulk-density classes, its influence diminishes at lower bulk densities. In contrast, bulk density and Corg content increasingly control the shape of the water retention curve, indicating that structural and organic-matter-related effects outweigh textural controls in these soils.
Based on these findings, we derived PTFs for forest soils grouped in bulk density classes. Validation against an independent dataset demonstrates that the proposed PTFs significantly reduce root mean squared errors compared to established PTFs developed for mineral soils, forest soils with higher bulk density, or pure organic horizons. Improvements are particularly pronounced in the wet and intermediate suction ranges that are most relevant for plant available water.
Our results highlight the need for specialized PTFs for forest soils with low bulk density and high Corg content. The new PTFs contribute to reducing the knowledge gap in soil hydraulic properties within the complex transition space between mineral soils and purely organic soils (Corg > 20%) and support improved representation in hydrological and ecosystem models.
How to cite: Gadermaier, J., Zieher, T., Behringer, M., Klebinder, K., Keßler, D., Hohenbrink, T., Englisch, M., and Katzensteiner, K.: Pedotransfer functions for forest soils with low bulk density and high organic carbon content: Insights from the Eastern Alps., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18576, https://doi.org/10.5194/egusphere-egu26-18576, 2026.
Digital soil mapping at the field-scale faces a fundamental challenge of building accurate predictive models from small, high-dimensional tabular datasets where training sample sizes are limited by cost and labor constraints. Traditional machine learning methods like Random Forest have long dominated pedometrics, but recent advances in artificial neural network architectures challenge this status. To investigate this, we develop a comprehensive evaluation framework built upon LimeSoDa, our diverse and fully open-access collection of field-scale digital soil mapping datasets. This allows us to assess the application of modern neural networks in pedometrics under realistic conditions of data scarcity. Our results demonstrate that contemporary architectures consistently outperform classical methods when coupled with specific methodological enhancements that address training instability. In-context learning tabular foundation models, such as TabPFN, show particular promise and surpass established baselines even on very small datasets. We go further and investigate the application of tabular foundation models on datasets with unfavorable feature-to-sample ratios typical in soil spectroscopy. Building upon principal component analysis and partial least squares, we propose hybrid strategies that effectively address the challenges posed by soil spectroscopy datasets. Going beyond purely tabular regression modeling, we extend our framework to incorporate spatial information through Kriging prior Regression, integrating geostatistical features into tabular machine learning predictions and further improving accuracy when sensor data alone provide limited information. Our findings establish a new baseline for field-scale digital soil mapping and offer methodological insights applicable to any precision agriculture domain constrained by small tabular datasets.
How to cite: Barkov, V., Schmidinger, J., Gebbers, R., and Atzmueller, M.: Field-scale digital soil mapping in the era of tabular foundation models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-14583, https://doi.org/10.5194/egusphere-egu26-14583, 2026.
Accurate, high-resolution mapping of soil total carbon (TC) stocks on a global scale is fundamental to global carbon cycle modeling, international climate policy (e.g., IPCC inventories), and sustainable land management. Current Digital Soil Mapping (DSM) efforts often rely on monolithic global machine learning models that frequently fail to capture fine-scale local variability and are prone to significant regional biases. These biases stemmed from spatial non-stationarity, disjointed calibration datasets from varied sources, and instrumentation mismatches, leading to poor predictive performance and high uncertainty in under-sampled regions. To address this critical challenge, the Mixture of Localised Experts (MoLE) framework was introduced as a novel deep learning architecture designed for robust and responsible soil property prediction. The MoLE framework overcame the limitations of traditional GLOBAL–LOCAL approaches by employing a dynamic gating network (router) that learned to partition the problem space. This router intelligently directed input data comprising multiple proximal soil sensor features from a multinational dataset to one of several specialised “expert” sub-models. Each “localised expert” was trained to become highly proficient within a specific geographical or data-driven domain, effectively creating a single, cohesive model that “thinks globally but acts locally.” This framework was developed using a large, harmonised proximal sensor dataset (n = 1443) from six countries across five continents to predict TC. When assessed against an independent hold-out validation set, the MoLE framework demonstrated outstanding precision for TC prediction, achieving a coefficient of determination (R²) of 0.98 and a root mean squared error (RMSE) of 0.06%. Crucially, the results indicated that the MoLE architecture substantially reduced regional prediction bias. The interpretable routing mechanism offered fresh perspectives on model decision-making, revealing the experts activated for various ecoregions and boosting the transparency of the model. The MoLE framework offered a scalable, resilient, and comprehensible framework for the advancement of next-generation global soil information systems. By adeptly addressing spatial heterogeneity and reducing regional bias, this methodology represented a substantial advancement in the precise quantification of global TC stocks.
How to cite: Nayak, A. and Chakraborty, S.: Tackling Spatial Heterogeneity in Global Soil Total Carbon Mapping using a Mixture of Localised Experts, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-291, https://doi.org/10.5194/egusphere-egu26-291, 2026.
High-resolution soil property maps (SPMs) are of high relevance on multiple spatial scales. At field-scale, highly resolved knowledge of soil properties can inform soil management zone delineation for precision agriculture applications like irrigation, fertilisation or compaction risk management. On regional to national scales, high-resolution SPMs can aid in formulating soil policy with accurate baselines and realistic improvement goals. They may assist in soil unit definition, aiding in the balancing act between scientific precision and geographic cohesion, i.e. administrative effectiveness. On a continental, European scale, these SPMs can inform on the adequacy of Soil Monitoring Law implementation and can increase the representativity of continental soil sampling campaigns.
Contributing to this effort, we produced 10 m-resolution maps of soil pH, bulk density, coarse fragments, and texture (sand, silt, clay). This represents a large increase in spatial resolution compared to previously published maps, whose cells cover at least 250 x 250 m. We generated our maps by training shallow property-specific artificial neural networks (ANNs) on soil sample data across the EU-27 + the United Kingdom. We used soil data from the Land Use/Coverage Area frame statistical Survey (LUCAS): texture and coarse fragments data from the 2009 campaign, and pH and bulk density data from 2018. Instead of collecting a large number of spatial covariate data layers, we used the AlphaEarth Foundations satellite embeddings, produced by Google DeepMind. The satellite embeddings are the result of a representation learning model trained to efficiently compress Earth observation data. In this way, data from various sources is represented as a 64-dimensional vector per 10 m pixel. We extracted the 2018 embeddings as predictors for all soil properties. After training the models, we produced the full maps on Google Earth Engine (GEE) using a hand-implemented ANN with five neurons in the hidden layer, followed by a dropout layer. The dropout layer gave us the opportunity to additionally provide a prediction uncertainty map per soil property. A single output model was used to produce pH, bulk density and coarse fragments, while a multioutput NN generated the three texture components. Model performances varied, with R² ranging from 0.21 (coarse fragments) to 0.69 (pH). We validated our maps using a variety of field- and regional-scale soil datasets, furthermore by comparison of the value distributions with LUCAS data on European and national scales, and by visually contrasting with published soil property maps. We look forward to testing the temporal stability of our models once LUCAS 2022 data are available. Our soil maps will be necessary and useful for the scientific community across scales, from the field to the continent.
How to cite: Smit, E., Bernardini, L. G., Moreno Martínez, Á., Muñoz-Marí, J., Vuolo, F., and Izquierdo-Verdiguier, E.: High-resolution European soil property maps leveraging foundation-model Earth observation embeddings, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18609, https://doi.org/10.5194/egusphere-egu26-18609, 2026.
Rising temperatures and increasing climatic stress will push forest ecosystems in mountain regions towards their ecological limits, intensifying the need for informed decisions on forest management and tree species suitability. Such assessments rely on spatially explicit ecological models that require area-wide, depth-resolved soil information as a key input. Digital soil mapping (DSM) provides a framework to generate such information; however, the reliability and interpretability of model outputs strongly depend on how soil parent material information is represented at the input stage. Soil parent material, defined by bedrock and/or overlying material derived from geomorphological processes, is commonly described using geological maps and related datasets that encode properties as categorical units. While this representation is widespread across many landscapes, it becomes particularly limiting in complex mountain terrain, where fine-scale spatial variability and vertical heterogeneity arise from interacting geological and geomorphological processes. These processes strongly control the physical and chemical characteristics of the soil parent material but are difficult to resolve using class-based representations. In addition, soil parent material information is typically available at coarser and inconsistent spatial resolutions compared to terrain-derived predictors used in DSM. This limits its suitability for data-driven mapping and ecological modelling, where spatial consistency, depth discreteness and plausibility of predictors are essential.
We address the preparation of soil parent material information as machine-learning-compatible predictors. Geological and geomorphological information differentiated by genetic process types is derived from project-specific geological and geomorphological field mapping and encoded as a set of categorical chemical and physical property classes, including rare but pedologically relevant types. These class-based descriptions are transformed into continuous representations of parent material composition, expressed by five mineral component layers and physical fractions describing grain-size distribution and proportions of consolidated (bedrock-derived) versus unconsolidated (deposit-derived) material. This preserves pedological meaning, reduces the dominance of unevenly represented classes in data-driven modelling, and enables proportional mixing across geometric soil depth intervals, resulting in more stable and interpretable learning than purely categorical predictors.
Using a rule-based allocation scheme, polygon-based information on soil parent material genetics, layer thickness, and areal extent of unconsolidated cover is used to derive depth-discrete parent material layers for four geometric soil depth levels. Bedrock is represented as the basal parent material, while overlying unconsolidated material may be present with a defined areal coverage fraction within a homogeneous geological polygon. Vertical mixing is handled proportionally based on depth contribution. For parent material types associated with gravitational processes, an additional standalone two-dimensional distribution model, independent of mapped areal coverage information, is used to resolve pixel-scale presence or absence of overlying unconsolidated parent material. Parent material types associated with other genetic processes (e.g. aeolian, fluvial or glacial) are assumed to exhibit spatially continuous coverage and are mixed vertically according to their thickness.
By providing vertically consistent and physically interpretable predictors, including the systematic transformation of class-based soil parent material descriptions into continuous representations, the proposed depth-aware approach enables the generation of area-wide, spatially coherent soil information. These products are suitable as input for downstream ecological applications, such as tree species suitability and soil hydrological assessments.
How to cite: Huber, T., Simon, A., Klebinder, K., Englisch, M., Kessler, D., Ganser, C., Gruber, J., Wilhelmy, M., Szentiványi, J., Brandstätter, J., Wagner, T., Vremec, M., and Winkler, G.: Depth-discrete, machine-learning-interpretable soil parent material representation for robust soil mapping in complex mountain terrain, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20486, https://doi.org/10.5194/egusphere-egu26-20486, 2026.
Accurate spatial prediction of soil organic carbon (SOC) is a key component of digital soil mapping and is essential for environmental modelling and land management. The increasing availability of multi-source satellite observations offers new opportunities to improve SOC mapping. However, the relative contribution of different optical and radar data sources remains still not fully clear. In this study, we systematically evaluate the influence of multi-temporal optical and synthetic aperture radar (SAR) observations on SOC prediction at the continental scale. A total of 18,984 topsoil samples (0–20 cm) from the LUCAS soil survey were combined with historical satellite-derived predictors. SOC prediction models were developed using geostatistical and machine-learning approaches and evaluated using cross-validation. Different combinations of optical and SAR predictors were tested to assess their impact on model performance and spatial patterns of SOC maps. Results showed that SOC prediction accuracy was strongly dependent on the type of satellite information used. Sentinel-3 long-term optical observations provided the highest predictive performance, explaining up to 70% of SOC spatial variability. SAR co- and cross-polarization contributed similarly to SOC prediction, while their combination further improved model performance. Predicted SOC maps exhibited pronounced spatial heterogeneity, with consistent large-scale patterns but varying local details across data sources. These findings highlight the complementary value of optical and radar observations and provide practical guidance for data selection in large-scale digital soil mapping of SOC.
How to cite: lyu, W. and Maerker, M.: Digital mapping of soil organic carbon across Europe using multi-source optical and radar observations, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21110, https://doi.org/10.5194/egusphere-egu26-21110, 2026.
Agricultural soils are increasingly under pressure due to land use intensification and the ongoing effects of climate change. Current EU policies, such as the "Carbon Removals and Carbon Farming Regulation" (CRCF), aim to improve the resilience of cropland soils by carbon sequestration through climate-smart management. However, such regulations significantly increase the demand for spatiotemporal soil data to monitor and verify the effectiveness of carbon farming measures.
In recent years, the analysis of remote sensing-based bare soil observations has been increasingly used to generate accurate, high-resolution maps of cropland properties, such as soil organic carbon (SOC). However, due to the lack of robust reference data and the slow-changing nature of SOC, validating temporal model performance remains challenging. In this study, we tested the extent to which spatiotemporal models based on satellite data can support wall-to-wall soil monitoring and provide information on the temporal variability of SOC in cropland soils.
To achieve this, bare soil composites were derived from Landsat and Sentinel-2 data using a moving window approach and compared to Bavarian long-term soil monitoring data from 1986 to 2022. The results showed that while overall model performance was high, the validation of measured SOC trends yielded significantly lower accuracy, underlining the high uncertainty in predicting temporal soil carbon dynamics. While long-term analyses of 25+ years were necessary to detect significant SOC changes in most cases, the classification of the results revealed a low confusion rate between sites with increasing or decreasing SOC trends across the observation period.
These findings are supported by recent results based on the repetition of the German agricultural soil inventory, currently being conducted at the Thünen Institute of Climate-Smart Agriculture. Although significant uncertainties remain in quantifying SOC dynamics within 10-year intervals, results can be improved by taking the plot-scale SOC variability into account. This preprocessing step not only improves the significance of spatiotemporal SOC models ("model-then-derive") but also allows for the direct prediction of SOC changes based on a "derive-then-model" approach.
In summary, these results provide a first step toward an integrated soil monitoring system based on remote sensing and repeated soil sampling. While the findings demonstrate that it is possible to validate spatiotemporal SOC models using long-term sampling data, they also highlight the necessity of further improving the accuracy and applicability of the models. Based on our studies, we will further discuss the opportunities and challenges to independently validate SOC trends claimed by carbon farming schemes using remote sensing data.
How to cite: Broeg, T., Erasmi, S., and Don, A.: Quantifying the uncertainty of remote sensing-based soil carbon monitoring, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9102, https://doi.org/10.5194/egusphere-egu26-9102, 2026.
Agricultural soils in South Korea are prone to acidification due to the leaching of exchangeable bases caused by intensive rainfall. To mitigate this issue, the government subsidizes lime fertilizers based on soil test data. However, for untested fields lacking analytical data, a standardized application rate is uniformly applied, which limits the precision of soil management. This study aims to develop an optimal prediction model and establish environmental and geographical covariates for the precise estimation of soil pH in untested fields using Machine Learning (ML) techniques. A dataset comprising 203,941 soil pH measurements collected between 2022 and 2024 was utilized for model training. Based on the SCORPAN framework, we constructed environmental and geographical covariates across a grid covering South Korea, incorporating Soil (s), Climate (c), Organisms (o), Relief (r), Parent material (p), and Spatial position (n) factors. The Random Forest (RF) algorithm was employed as the primary prediction model. Furthermore, to account for spatial autocorrelation, a hybrid model combining RF with Ordinary Kriging (OK) of the prediction residuals (RF+OK) was developed, and its performance was compared with the standalone RF model. Variable importance analysis indicated that geographical variables (distance information) had the most significant influence on pH prediction, followed by organism variables (NDVI and crop cultivation density) and climatic variables. In the model performance evaluation, the standalone RF model achieved an R² of 0.509 and an RMSE of 0.550. However, the hybrid RF+OK model demonstrated significantly improved prediction precision, achieving an R² of 0.590 and an RMSE of 0.492. The big data-driven pH prediction model established in this study is expected to eliminate policy blind spots in areas lacking soil test information and enable the precise calculation of lime requirements for sustainable soil management.
How to cite: Lee, E., Kim, M.-S., Jung, H., and Lee, T.-G.: Development of a Soil pH Prediction Model for South Korean Agricultural Fields Using Digital Soil Mapping, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16339, https://doi.org/10.5194/egusphere-egu26-16339, 2026.
Soil acidification in China's croplands, driven by intensive agriculture, threatens both agricultural sustainability and ecological security. However, its large-scale spatiotemporal dynamics and underlying drivers remain poorly understood. We developed a meta model coupling of process model and machine-learning, integrating national soil survey information from the Chinese Soil Series Records and the VSD+ model to simulate and map soil pH trajectories from 1980 to 2100. We found a pronounced national acidification trend from 1980 to 2020, with average cropland soil pH declining by 0.37 units (from 7.05 to 6.68). This trend varied regionally, with the most severe acidification in Northeast China (ΔpH = -0.67) and the slowest decline on the Qinghai-Tibet Plateau (ΔpH = -0.11). Projections show that even under a nitrogen fertilizer zero-growth scenario, soil pH will continue to decrease to 6.49 by 2100. Our model reveals that fertilizer management exerts a stronger influence on acidification than climate change, as evidenced by minor differences among climate scenarios. Mechanistically, nitrogen transformation was the dominant acidifying process in upland systems, while both HCO3 leaching and nitrogen transformation were primary drivers in paddy systems. This study provides a quantitative basis for sustainable nutrient management and offers valuable insights for mitigating soil acidification in global agricultural systems.
How to cite: wang, X., chen, S., shi, Z., and duan, S.: Spatiotemporal divergence of soil acidification in China’s cropland, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4730, https://doi.org/10.5194/egusphere-egu26-4730, 2026.
Mapping fine-grained soil particle size distributions (PSDs) in complex forest ecosystems remains a significant challenge in pedometrics. Traditional pixel-based machine learning approaches often struggle to capture the spatial heterogeneity and dependencies inherent in forest soils, particularly when ground-truth sampling is limited by cost and accessibility. This study presents a novel, holistic GeoAI framework that integrates geostatistical augmentation with Graph Neural Networks (GNNs) to map fine-grained soils (<60 µm) using LiDAR and Sentinel-2 data.
Our methodology addresses the "small data" problem through a two-stage process. First, we employed CoKriging (geostatistics) to locally upscale point-based soil samples within measured forest stands. This geostatistical interpolation generated a dense set of annotated training data, effectively augmenting the dataset to train GNN models. Second, we shifted from varying pixel resolutions to object-based analysis by segmenting forests into homogeneous polygonal zones based on tree species and canopy structure, which served as nodes in a graph structure.
We evaluated five GNN architectures (GAT, RGCN, GCN, EGNN, and MPNN) to predict PSDs using ~60 covariates derived from high-resolution LiDAR (geomorphometry, hydrology) and Sentinel-2 time-series (vegetation/soil indices). The graph attention network (GAT) emerged as the superior model, demonstrating remarkable stability and predictive accuracy. By utilizing multi-head attention mechanisms, the GAT model successfully learned the importance of neighboring nodes and complex spatial dependencies that standard convolutional models often miss. The GAT model achieved R2 values exceeding 0.98 across all soil particle groups. Feature importance analysis revealed that LiDAR-derived geomorphometry (specifically elevation and downslope) and Sentinel-2 derived organisms (e.g., WDVI) were the dominant covariates driving predictions. This approach demonstrates that combining geostatistical data augmentation with the relational learning capabilities of GATs offers a scalable, accurate solution for digital soil mapping in data-sparse environments, with significant implications for forest management and hydrological modelling.
Reference: Abdi, O., Laamanen, V., & Uusitalo, J. (2025). Mapping forest fine-grained soil particle size distributions: a holistic GeoAI approach via graph neural networks, LiDAR, and Sentinel-2. International Journal of Applied Earth Observation and Geoinformation, 143, 104807. https://doi.org/10.1016/j.jag.2025.104807.
How to cite: Abdi, O., Laamanen, V., and Uusitalo, J.: From Geostatistics to Graph Attention Networks: A Holistic GeoAI Approach for Mapping Forest Soil Particle Size Distributions with Limited Samples, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20902, https://doi.org/10.5194/egusphere-egu26-20902, 2026.
Successful tea cultivation is dependent on careful soil management practices that are underpinned by information on soil properties, which tends to be sparse in tea growing regions. Such information is often periodically captured through field sampling, however, may only be available for selected plantations or fields within larger tea estates. Consequently, soil management decisions on some plantations are made in the absence of soil information.
In this study, digital soil mapping techniques were used to create 30 m resolution maps of selected soil properties that were captured within Kericho, Kimugu and Cheymen tea plantations on a large tea estate managed by Browns East Africa Plantations Kenya Ltd in Western Kenya. Preliminary results, obtained from relating soil properties to topographic and climatic SCORPAN factors using random forests revealed differing importance of climatic and topographic predictors for different soil properties, suggesting different drivers behind variation in these properties. The accuracy of the predictions, measured with the root mean square error, was 0.96% for soil organic carbon, 0.31 for pH, 0.097 mg/kg for nitrogen, 9.77% for sand, 3.65% for silt, and 10.56% for clay. Maps for plantations with no validation data available were then sense-checked against the predictive soil maps for Africa (AfSIS).
Further improvements in accuracies are expected from inclusion of NDVI image composites to aid soil carbon modelling as well as data on fertiliser applications within tea plots for nitrogen predictions, coupled with the XGBoost algorithm. Finalised maps are expected to be used within the digital platform for tea crop management called “Internet of TeaTM” or “IoTeaTM” that incorporates a model of tea growth and development called “CUPPA-Tea”. Alongside, the underpinning soil data will help us understand the fundamental processes in the soil that influence greenhouse gas emissions, and using advanced genomic technologies to accelerate the tea breeding process.
How to cite: Zawadzka, J., Singh, M., Oulaid, B., Holden, A., Tuwei, G., Wallace, A., Waine, T., and Corstanje, R.: Digital soil mapping of soil properties for enhanced management of Kenyan tea estates., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21123, https://doi.org/10.5194/egusphere-egu26-21123, 2026.
Accurate characterization of soil properties is fundamental for quantifying terrestrial carbon stocks, land–atmosphere interactions, and assessing agroecosystem functioning and ecosystem services. Launched in August 2024, Tanager-1 is the first satellite in Planet’s hyperspectral constellation, delivering over 400+ contiguous spectral bands across the 400–2,500 nm range at a spatial resolution of 30 m. Such data is vital for monitoring vegetation and soil health. However, retrieving soil properties from satellite data in diverse agricultural landscapes remains challenging in heterogeneous croplands where soil, vegetation, and moisture vary strongly within and between fields. This is a particular issue in perennial cropping systems, such as vineyards. This research, conducted as part of the Horizon Europe AI4SoilHealth project, explores the potential of Tanager data to derive key soil properties, such as soil organic carbon (SOC) and soil texture, while addressing the confounding effects of different scenes' radiometric signatures.
The methodology leverages the European Soil Data Center's (ESDAC) Land Use/Cover Area Frame Statistical Survey (LUCAS), which includes around 20,000 soil samples, as a foundation for retrieving soil properties from Tanager surface reflectance, using machine learning approaches such as partial least squares regression (PLSR). The study investigates strategies for decoupling the soil and vegetation components. For example, drawing on the PROSAIL radiative transfer model to explicitly simulate and account for the vegetation contribution in the training dataset and using an autoencoder-based spectral unmixing model to mitigate vegetation effects. and estimate bare soil reflectance from Tanager observations in a pre-processing step. Validation is supported by field-based spectrometers and laboratory analysis of physical soil samples to provide the ground truth for distinct scene endmembers.
Preliminary findings suggest that the high signal-to-noise ratio of the Tanager hyperspectral sensor, when combined with artificial intelligence models, shows promising improvements in soil property retrieval accuracy.
How to cite: Grady, M., Ampas, H., Guillevic, P., Karyotis, K., Roth, K., and Wania, A.: Planet Tanager Hyperspectral Data to Retrieve Soil Properties of Heterogeneous Agricultural Landscapes , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12559, https://doi.org/10.5194/egusphere-egu26-12559, 2026.
The organic carbon content of soil is of increasing global interest. Soil is the largest terrestrial store of carbon on the planet, yet many agricultural soils are highly degraded, and a significant source of greenhouse gas emissions. However, appropriate management can mitigate carbon loss and even turn agricultural soils into carbon sinks. Soil Organic Carbon (SOC) content is also a good indicator of soil health: higher SOC levels are typically associated with improved soil structure, moisture retention, and plant nutrient availability. Increasing soil carbon stocks is therefore important from both food security and climate perspectives.
Understanding the spatial variability of soil organic carbon (SOC) is critical for accurately monitoring and managing soil carbon stocks. Field-scale soil sampling is widely used to estimate SOC and infer its spatial distribution; however, SOC can vary substantially over short distances, leading to challenges in characterising within-field heterogeneity. When short-range variability is high, low-density sampling and spatial interpolation may fail to capture meaningful structure, resulting in uncertain predictions and misleading estimates of field-scale means.
We present a case study wherein we compared hyperlocal (≤10 m) and field-scale soil sampling results. Statistical analyses were applied to evaluate SOC variability across different spatial scales and to assess the significance of hyperlocal soil carbon variability.
Results show that hyperlocal variability can closely reflect that observed at the field scale, helping to explain why spatial patterns may not be effectively captured by coarse-scale sampling. Meanwhile, where hyperlocal variance was small, field-scale spatial structure was clearer, making interpolation more defensible. This demonstrates that hyperlocal sampling provides a useful diagnostic for assessing whether field-scale SOC data can meaningfully support spatial modelling.
This study highlights the value of incorporating hyperlocal sampling into soil carbon assessments to better capture spatial heterogeneity and improve the reliability of carbon monitoring.
How to cite: Lakey, C., Knox, N., and McGlade, J.: Soil Organic Carbon Heterogeneity: Insights from Hyperlocal and Field-Scale Sampling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22141, https://doi.org/10.5194/egusphere-egu26-22141, 2026.
Abstract
Reliable estimates of the root zone plant-available water holding capacity (RZ-PAWHC) are essential for assessing crop water availability and supporting climate-resilient agricultural planning. RZ-PAWHC is calculated by summing the plant-available water holding capacity (PAWHC), adjusted for the soil fine earth fraction (SFEF), over the root zone depth (RZD). Existing Sub-Saharan Africa (SSA) maps of RZ-PAWHC rely on coarse datasets and lack validation. We generated 100 m-resolution RZ-PAWHC maps for maize and wheat across central Ethiopia by integrating national soil datasets, digital soil mapping, pedotransfer function (PTF), and rule-based rootability indices. Thirteen primary soil properties were mapped at 5 cm-thick intervals from 0 to 150 cm depth using Random Forest models. Two additional primary soil property maps—depth to bedrock and drainage class, each providing a single value per pixel—were obtained from Africa-SoilGrids. These properties served as inputs for PTF to estimate volumetric moisture content at field capacity (VMC-FC) and at permanent wilting point (VMC-PWP), and for rootability rules to derive RZD. A coarse fragment map was also used to derive SFEF. RZD was defined as the shallowest of the depth to bedrock, to oxygen shortage (derived from drainage class), to a restrictive soil layer, or to a crop-specific maximum rooting depth. Prediction accuracy of the primary soil property maps ranged from a model efficiency coefficient of 0.17 for coarse fragments to 0.79 for pH-H₂O. Derived maps showed that PAWHC over the 0150 cm depth ranged from 1 to 237 mm (mean = 122 mm), SFEF averaged 89%, reducing total PAWHC by 11%, and mean maize RZD was 124 cm, primarily constrained by bedrock (covering 45% of the mapped area) and oxygen shortage (22%). RZ-PAWHC maps had mean values of 91 mm for maize and 83 mm for wheat, with the lowest values observed in poorly drained, clay-rich soils. Compared with SSA maps for maize, the generated maps indicated higher RZ-PAWHC due to higher PAWHC, higher SFEF (89 vs 83%) and deeper rooting (124 vs 107 cm), with only minimal chemical root restrictions (salinity, sodicity, toxicity; 1 vs 17%). Validation against independent soil and root observations from 50 soil pits (up to 2 m depth) showed that the generated maps outperformed SSA for VMC-FC and VMC-PWP, while the SSA map better estimated PAWHC and SFEF. RZD and RZ-PAWHC were overestimated in the generated maps, whereas SSA underestimated them. Deriving PAWHC, SFEF, RZD, and RZ-PAWHC from measured- instead of from mapped soil properties revealed that errors were mainly driven by inaccuracies in the mapped soil properties rather than by the PTF or RZD derivation rules. These findings highlight the importance of improving the accuracy of critical soil property maps —particularly drainage and depth to bedrock— which most strongly constrained RZD.
Keywords: root zone depth, plant-available water holding capacity, pedotransfer function, validation, central Ethiopia
How to cite: Abegaz, M. R., Heuvelink, G. B. M., and Leenaars, J. G. B.: Mapping and validating the root zone plant-available water holding capacity in central Ethiopia, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20473, https://doi.org/10.5194/egusphere-egu26-20473, 2026.
Roads in Sub-Saharan Africa provide essential transportation functions; however, they often cause adverse environmental impacts, such as enhanced erosion, that are frequently underestimated or overlooked during planning and implementation. Global digital soil mapping (DSM) products such as OpenLandMap and SoilGrids provide open-source soil information for large-scale ecosystem service assessment and monitoring. However, the accuracy of these modelled datasets varies spatially because the input data used for model development are unevenly distributed. Thus, their reliability and implications for erosion modelling in data-scarce semi-arid regions remain insufficiently understood.
In this study, we investigated the spatial variability of soil properties (soil texture fractions, soil organic carbon, bulk density, and pH) in Baringo County, Kenya and examined whether paved and unpaved roads influence this variability. We evaluated the accuracy of SoilGrids250m 2.0 and OpenLandMap (30 m) in representing field-measured soil properties and assessed how DSM-related uncertainties propagate into erosion estimates. Using pedotransfer functions based on soil texture fractions and soil organic carbon, we derived soil erodibility factors from both field data and OpenLandMap. The factors were subsequently used in the Integrated Valuation of Ecosystem Services and Trade-offs (InVEST) Sediment Delivery Ratio model to estimate spatial patterns of soil loss.
Both DSM products underestimated topsoil silt content (RMSE = 35.6% and 36.1% for OpenLandMap and SoilGrids, respectively) and overestimated clay (RMSE = 17.8% and 22.3%) and sand contents (21.4% and 17.9%, respectively), with accuracy decreasing with depth. Field data revealed significantly lower silt content and higher sand content in topsoils near roads compared to further away with a moderate effect size. Sediment deposition and export computed using the parametrized OpenLandMap factor showed high correlation with results parametrized from field data across varying distances for both paved roads and unpaved roads (R2 > 0.87). Despite high correlations, modelling results parametrized with OpenLandMap underestimated sediment deposition and export by factors of approximately 2.2 and 2.7 for paved and unpaved roads, respectively. Unpaved roads showed greater sediment export near the road corridor compared to paved roads, while paved roads were associated with greater sediment deposition.
Our results demonstrate that while global DSM products can reproduce relative spatial patterns of road-related erosion, systematic biases in soil property predictions an affect erosion estimate. This highlights the need to explicitly consider DSM uncertainty when using open soil data for erosion modelling and infrastructure, especially in data-scarce regions.
How to cite: Nyamari, N., Stollenwerk, C., Kienzler, L., van der Meij, M., Ochuodho Otieno, D., and Bogner, C.: Evaluating digital soil mapping products for modelling road-related soil erosion in Baringo County, Kenya, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10455, https://doi.org/10.5194/egusphere-egu26-10455, 2026.
Volcanic soils are characterized by andic properties such as organic carbon (OC), bulk density (Bd), phosphate retention (Pret), and the sum of ammonium oxalate-extractable aluminum and half iron (Alo + 0.5Feo), and play an important role in agricultural production, global carbon cycling, and ecological functions. However, conventional determination of andic properties relies on destructive, labor-intensive, and time-consuming wet chemistry analyses. Soil spectroscopic techniques such as visible and near-infrared (Vis-NIR) spectroscopy and portable X-ray fluorescence (pXRF) provide rapid and non-destructive alternatives. Previous studies have shown that soil properties can be well predicted by integrating spectroscopic data with machine learning algorithms such as partial least squares regression (PLSR) and Cubist. However, no study has investigated the data fusion of Vis-NIR and pXRF for predicting andic properties. Therefore, this study aimed to elucidate the relationships between andic properties and signals from Vis-NIR and pXRF, and to evaluate the accuracy of sensor-based models for predicting andic properties and soil classification. A total of 93 soil samples were collected from 24 pedons of volcanic soils (0–60 cm depth) in northern Taiwan, including Andisols and Inceptisols. Soil samples were measured by Vis-NIR and pXRF, and predictive models were developed using individual sensors and a data fusion approach calibrated with PLSR and Cubist algorithms. Laboratory analyses were conducted to quantify andic properties as reference values. Both Vis-NIR and pXRF signals demonstrated associations with andic properties. Data fusion of these two sensors markedly improved model performance compared with single-sensor approaches. In particular, the Vis-NIR + pXRF-based model calibrated with Cubist yielded good predictive performance for all andic properties, achieving R2 and LCCC values higher than 0.90 for OC, Pret, and Alo + 0.5Feo, and R2 = 0.83 and LCCC = 0.89 for Bd. Moreover, 23 out of the 24 studied pedons were correctly classified by this model. Integrating Vis-NIR and pXRF through data fusion provides an efficient approach for assessing andic properties, improving management and resource-use efficiency in volcanic soils and supporting sustainable smart agriculture. Further studies incorporating additional spectroscopic sensors may further broaden applicability across diverse soil types.
How to cite: Wu, P.-H., Minasny, B., Huang, Y.-C., Demattê, J. A. M., and Hseu, Z.-Y.: Data fusion of Vis-NIR and pXRF with machine learning for predicting andic properties in volcanic soils, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3261, https://doi.org/10.5194/egusphere-egu26-3261, 2026.
Proximal sensing techniques play an increasingly important role in pedometrics and digital soil mapping, yet methodological challenges remain in achieving pedologically consistent, transferable predictions. Portable gamma-ray spectrometry (pGRS) offers a physically grounded sensing approach, but its sensitivity to environmental conditions and soil type complicates the development of robust soil property conversion algorithms. This study addresses key session themes by evaluating sampling design, depth consistency, and physical interpretability in pGRS-based soil prediction.
Using a structured grid-based sampling framework, we combined in-situ pGRS measurements with laboratory analyses across contrasting arable and grassland systems at the North Wyke Farm Platform (UK), an experimental research farm, and a geologically distinct external site. Relationships between gamma-derived variables (40K, 238U, 232Th) and soil texture, SOC, and soil moisture were examined alongside depth-profile data to assess vertical consistency.
Results show that ⁴⁰K provides a pedologically meaningful predictor of texture and SOC, while soil moisture exerts a physically interpretable attenuation effect on gamma signals. Depth analyses demonstrate that pGRS sensitivity is heavily influenced by the near-surface (0-10 cm), with soil property gradients, rather than radionuclide redistribution, controlling signal response. These findings demonstrate the potential for pGRS-derived predictions and inform the development of depth-aware pedo-transfer functions.
Building on this work, the data collected will be used to develop predictive models for soil properties from pGRS measurements through the integration of machine learning approaches informed by soil process knowledge. This includes the development of an open-source conversion algorithm to translate pGRS signals into soil property estimates with quantified uncertainty, supporting reproducible, explainable, and transferable soil mapping frameworks for soil health assessment.
How to cite: Trust, B., Blackwell, M., Ansell, L., Collins, A., Mansfield, N., Rowntree, J., and Blake, W.: Environmental Controls on Portable Gamma Spectrometry for Soil Property Assessment: Influence of Land Use and Soil Type on a Farm Level, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22985, https://doi.org/10.5194/egusphere-egu26-22985, 2026.
Soil organic carbon (SOC) is a critical indicator of soil health and central to land management initiatives with climate change mitigation co-benefits. Conventional sampling and laboratory analysis however remain labour-intensive and costly, creating significant gaps in spatial and temporal coverage. These limitations hinder the development of explicit, quantitative, and spatially realistic SOC maps needed for environmental modelling, land management, and impact verification. Portable proximal sensing technologies, such as gamma ray spectrometry (GRS), offer a promising solution by enabling rapid, in situ measurements for high-resolution digital soil mapping.
We explore how radionuclide proxy measurements with portable gamma ray spectrometry can provide scalable, spatially explicit SOC estimates that can be integrated into pedometric frameworks. Soil samples were collected from two fields, one arable and one permanent pasture, on an estate in southwest UK. Radionuclide activity concentrations were measured at sample locations, and continuous walking surveys were conducted to generate spatial maps of measured radionuclide activity across both fields.
Correlation analysis and principal component analysis (PCA) were used to explore relationships between radionuclides and SOC. Results show that radionuclide activity concentrations are consistently and negatively associated with SOC, particularly thorium-232 (-0.77) and potassium-40 (-0.69). Elastic net regression and partial least squares regression (PLSR) identified thorium-232, potassium-40/uranium-238, potassium-40, and thorium-232/uranium-238 as consistently important variables (PLSR VIP scores 1.21, 1.12, 1.08, 1.07 respectively), emerging as strong indicators of SOC variation. These findings highlight the potential for radionuclide proxies to explain SOC distribution and offer insight into broader soil health dynamics across contrasting land uses.
The results support the adoption of portable gamma ray spectrometry as a transparent digital soil mapping tool addressing current gaps in spatial SOC representation. To ensure trust and reproducibility, protocols must be validated across a range of soil conditions, and conversion approaches from radionuclide proxies to SOC must be standardized (work currently advancing in the Joint FAO/IAEA Coordinated Research Programme D12015 on “Combining Gamma-Ray Sensing and Digital Technology for Soil Moisture and Soil Property Mapping”). Transparency is essential, from raw proxy measurements through to SOC map products, so that stakeholders can confidently use these data for decision-making. When applied with rigour and data traceability, this approach offers meaningful support for Climate-Smart Agriculture and sustainable land management strategies while reducing uncertainty in soil property mapping.
How to cite: Mansfield, N., Taylor, A., Jackson, R., Goddard, R., Keenor, S., Reid, B., and Blake, W.: Radionuclide-based prediction of Soil Organic Carbon: a proximal sensing approach for high-resolution soil mapping, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19548, https://doi.org/10.5194/egusphere-egu26-19548, 2026.
- Nitrous oxide (N₂O) emissions from rice paddies represent an important but highly variable pathway of nitrogen loss, with strong dependence on local hydrological conditions, soil properties, climate regimes and management practices. This pronounced variability poses major challenges for process-based models (PBMs), which often rely on fixed functional structures and site-specific parameterization, limiting their ability to generalize across heterogeneous regions. In this study, we develop a hybrid modelling framework that integrates machine learning (ML) with PBMs to improve predictive generalization while retaining mechanistic interpretability. Within this framework, PBMs are used to explicitly describe N₂O responses to key environmental drivers, whereas the ML component is employed to capture, distill and generalize data-driven response relationships from multi-site observational datasets compiled at the global scale. Beyond methodological development, the hybrid approach is used to explore the spatial heterogeneity and dynamic responses of N₂O emissions across contrasting rice-growing regions. By jointly analysing climatic, soil and management drivers, we assess how response behaviours may differ between regions and under varying water management regimes. Our results highlight the potential of hybrid modelling as both a predictive and diagnostic tool for understanding N₂O variability in rice paddy systems. This framework provides a flexible foundation for future scenario analysis and supports the development of region-specific mitigation strategies for more sustainable rice production.
How to cite: Cao, Y., Pan, B., Chen, D., and Lam, S. K.: A Hybrid Machine Learning–Process-Based Modelling Approach to Explore Dynamic Responses and Spatial Heterogeneity of N₂O Emissions in Rice Paddies, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16639, https://doi.org/10.5194/egusphere-egu26-16639, 2026.
Accurate estimation of soil organic carbon (SOC) content at large scales is very important for sustainable agriculture, climate change mitigation, and land management. This study was performed to consider the effect of using different soil sampling algorithms (SSA) for selecting optimal soil samples from the legacy soil datasets for predicting SOC content in the bare soil areas of the State of Bavaria, Germany. For this purpose, the matrix of soil samples alongside the corresponding values of covariates for each sample point was provided under three different scenarios. In the first scenario, which is the most commonly used scenario in digital soil mapping (DSM) studies, the values of each covariate at each soil sample location were captured from the exact pixel corresponding to the soil sample location (sample pixel). However, in the second and third scenarios, based on a filter-based parameterization, the covariate values at each soil sample location were calculated using 3 × 3 and 5 × 5 pixel windows, respectively. After providing the dataset for each scenario, three SSA including simple random sampling (SRS), conditioned Latin hypercube sampling (cLHS), and feature space coverage sampling (FSCS) were applied for selecting the optimal numbers of soil samples from each scenario to be used as the calibration dataset. In addition, those soil samples that were not selected as the calibration dataset were considered as the validation dataset. In fact, these SSA were applied to create four splitting ratios of calibration and validation (cal/val) datasets including 50–50, 60–40, 70–30, and 80–20. For each scenario, the splitting ratio of cal/val datasets using each SSA was provided 50 times to consider the deterministic ability of SSA to select the same soil samples across multiple repetitions. The random forest (RF) model was trained using the calibration datasets to predict the SOC content in the validation datasets for each scenario. The results of the performance analysis showed that the cLHS method with a splitting ratio of 80–20 from the second scenario outperformed other SSA and scenarios for predicting SOC content. The median of three statistical indices including root mean square error (RMSE (%)), coefficient of determination (R²), and mean error (ME) were 1.13, 0.73, and –0.07, respectively, for this selected SSA and the cal/val datasets from the second scenario. Therefore, the results of this study demonstrated that the type of SSA, the splitting ratio of cal/val datasets, and the parameterization of covariate values for the sample pixel could influence the prediction performance of the machine learning model for predicting SOC content. However, before generalizing these findings, more studies would be required using other SSA and different pixel windows around the sample pixel in different conditions (climate, soil types, and geology, etc.).
How to cite: Garosi, Y., Nussbaum, M., Htitiou, A., Gabriel, D., Rode, M., and Möller, M.: Selecting the best combination of different parameterizations of covariates and sampling algorithms for the spatial prediction of the soil organic carbon contents, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-1458, https://doi.org/10.5194/egusphere-egu26-1458, 2026.
Forest soils represent the largest carbon reservoir in terrestrial ecosystems, yet decadal scale changes in soil organic carbon (SOC) across Europe remain insufficiently quantified. Here, we analyze harmonized topsoil data from 0 to 20 cm in the LUCAS Soil Surveys conducted between 2009 and 2018 to investigate spatial patterns, environmental controls, and national level changes in forest SOC across the European Union and the United Kingdom. Using spatiotemporal machine learning models, we estimate a net SOC increase of approximately 1.31 Pg C over the decade, accompanied by pronounced regional heterogeneity. Climatic factors, particularly the aridity index, mean annual temperature, and available water capacity, emerged as the dominant drivers of SOC distribution, while forest structure, topography, and land cover change provided additional contributions. France, Sweden, and Germany accounted for the largest shares of total SOC gains and associated CO₂ equivalent reductions. These results highlight forest soils as dynamic and spatially heterogeneous carbon sinks and underscore their importance for national carbon accounting and climate mitigation strategies.
How to cite: Chen, Z., Lu, R., Schillaci, C., Shi, Z., and Chen, S.: Soil Organic Carbon Dynamics in European Forests over the Period 2009–2018, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-3793, https://doi.org/10.5194/egusphere-egu26-3793, 2026.
Agronomic optimization is essential in developing countries, particularly in regions where soil resources are limited and spatial variability is poorly characterized. This study, the first of its kind in Haiti, applied predictive modeling to link laboratory-derived physical and chemical soil properties with proximal and remotely sensed data collected from 32,949 georeferenced surface soil (0–20 cm) samples across the Arcahaie region. A representative subset of samples (n = 300) was analyzed using multiple machine-learning models, including Random Forest, Gradient Boosting, Stacking Ensemble, and XGBoost, to predict soil pH, texture components (sand, silt, clay), soil organic carbon, soil organic matter, cation exchange capacity, and plant-available P, K, Si, Fe, and Cu from proximal sensing data. Strong predictive performance was achieved for sand, silt, clay, soil organic carbon, soil organic matter, and cation exchange capacity (R² ≥ 0.80), with particularly robust results for soil texture and carbon-related properties, while predictions for other parameters were statistically significant but less accurate. The optimized models were subsequently applied to the full dataset, and spatial interpolation was performed to generate high-resolution maps of soil physical and chemical variability across the region. These outputs provide a decision-support framework for site-specific agronomic management. The methodology demonstrated here is readily transferable to other agriculturally important regions of Haiti and comparable developing-country contexts and could be further extended to three-dimensional modeling and mapping of subsoil properties to better characterize fertility within the root zone.
How to cite: Chakraborty, S., Nayak, A., Weindorf, D., Cean, R., and Bakr, N.: Integrating Proximal and Remote Sensing for Regional Soil Characterization and Mapping in Rural Haiti, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2767, https://doi.org/10.5194/egusphere-egu26-2767, 2026.
Bare-soil mapping is essential for agricultural monitoring, land-surface characterization, and environmental modelling, supporting applications such as soil organic carbon (SOC) estimation, evapotranspiration retrieval, erosion assessment, and land degradation monitoring. However, accurate detection of exposed soil remains challenging due to spectral confusion with sparse vegetation and crop residues, strong seasonal variability, and the heterogeneous structure of agricultural landscapes. The Harmonized Landsat–Sentinel (HLS) dataset, providing 30 m spatial resolution and a 5-day revisit cycle, offers new opportunities for multi-temporal bare-soil mapping. This study develops an automated workflow combining spectral indices and a machine-learning algorithm (Maximum Entropy, MaxEnt) to map bare soil across agricultural regions using HLS surface reflectance imagery. Multiple indices capturing vegetation–soil contrasts were employed, including NDVI, BSI, NDMI, SAVI/GSAVI, NBR, EVI, and DBSI. High-confidence bare-soil pixels were first identified using a rule-based approach with strict thresholds (e.g., NDVI < 0.2 and BSI > 0.4–0.7), which minimized commission errors and generated reliable presence samples for model calibration. To improve generalization across different day-of-year (DOY) mosaics, these samples were integrated into a presence-background modelling framework using the MaxEnt algorithm (maxnet). Background samples were constrained to non-bare conditions (e.g., NDVI ≥ 0.3). Model performance was evaluated using AUC, Kappa, sensitivity, and specificity, while permutation importance and jackknife analyses quantified predictor contributions. The model achieved an AUC of 0.9 and a Kappa value of 0.7, indicating strong discriminative ability and substantial agreement. NDVI and BSI were identified as the most influential predictors. The resulting products include DOY-specific bare-soil probability maps, binary masks, and aggregated bare-soil frequency maps, providing a robust and scalable framework for long-term agricultural and soil-related applications.
How to cite: Syu, C.-H., Valdez, M. C., Chen, C.-F., Yang, J.-H., Yen, C.-C., and Chang, Y.-H.: Multi-temporal Bare-Soil Mapping in Agricultural Landscapes Using HLS Imagery and MaxEnt Modelling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-4575, https://doi.org/10.5194/egusphere-egu26-4575, 2026.
Soil systems provision many essential ecosystem services including food resources, water filtration and climate regulation. The accumulation and storage of soil organic carbon (SOC) is central to soil health and delivers a climate mitigation service of global significance. However, widespread monitoring of SOC is constrained by the costs, expertise and labour demands associated with traditional quantification methodologies, such as wet oxidation and dry combustion techniques. Diffuse reflectance spectroscopy (DRS) within the visible-near infrared (VNIR) range (400-1100 nm) offers a rapid and low cost means of SOC quantification. Specifically, recent developments to portable/ handheld VNIR spectrophotometer instruments are successfully facilitating cost-effective and accurate in-situ measurements, expanding global access to SOC quantification tools.
This research investigates the use of portable VNIR spectrophotometers for predicting SOC content in soils across parkland and links-style golf courses. Our aim is to develop novel machine learning-based pedometric SOC prediction models based on VNIR spectra and SOC reference values, and to integrate models into an open access software for non-specialists to easily quantify SOC at their own golf courses. To determine the optimal spectrum acquisition strategy, spectrophotometer scans were conducted in-situ using sub-sections of fresh soils cores (to 15cm depth) and ex-situ using processed samples. Processed samples underwent drying and milling to increase homogenisation and reduce the impacts of spectrum disturbing factors, such as water content and surface roughness. Furthermore, in-situ scans of surface vegetation were conducted to understand the link between the VNIR spectra of vegetation and SOC content. Following preliminary investigations, we will compare the accuracy of SOC quantification by three VNIR spectrophotometers, which vary in wavelength range and cost. Overall, this research aims to support greenskeepers to monitor SOC sequestration and storage, and in-turn soil health and quality, across courses. In doing so, this research will empower greenskeepers to use evidence-based soil management practices. Additionally, understanding the carbon sequestration and storage abilities of golf course soil systems will ameliorate the accuracy and transparency of the golfing industry’s net climate impact.
Our findings will contribute to the development of in-situ VNIR spectroscopy as a scalable, cost-effective and environmentally friendly analytical tool for SOC quantification by expanding its applications to turfgrass systems. Overall, this research will advance machine learning-based pedometric approaches, their application in practical land management and the global accessibility of SOC quantification tools.
How to cite: Risbey, C., Smeaton, C., and McCarthy, W.: Developing portable VNIR spectroscopy for soil organic carbon quantification, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5070, https://doi.org/10.5194/egusphere-egu26-5070, 2026.
Variable-rate nitrogen (N) fertilization has the potential to improve nitrogen-use efficiency and reduce losses in rainfed cereal systems, yet practical frameworks linking ground-based management zones with satellite time series remain limited. Here, we assess variable-rate N fertilization in a commercial winter barley field in Catalonia using Sentinel-2 optical and Sentinel-1 C-band SAR observations. Two variable-rate strategies were evaluated against the farmer’s uniform-rate (UR) practice. Spatio-temporal clustering of Sentinel-1 and Sentinel-2 time series was used to evaluate satellite sensitivity to within-field heterogeneity identified by proximal soil sensing, complemented by point-level time-series analysis to track site-specific crop development. For two seasons of winter barley, spatio-temporal clusterings revealed that GNDVI and NDWI produced the most heterogeneous spatial clustering, indicating potential sensitivity to spatio-temporal variability throughout the crop growth stages.
This work promotes to implementation of freely available optical and radar satellite data for within-field farming practices integrated with ground based proximal sensing to support operational, site-specific nitrogen management in Mediterranean cereal systems.
How to cite: Amangeldy, N., Chourak, Y., A. Garcia-Braga, E., Luo, H., Portal, G., Bardají, I., Vall-llossera, M., Vilaplana Ventura, R., Morellos, A., Vázquez-Arellano, M., Meers, E., and M. Mouazen, A.: Spatio-Temporal Variability Mapping for Variable-Rate Nitrogen Fertilization, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-5576, https://doi.org/10.5194/egusphere-egu26-5576, 2026.
While topsoil layers (0–30 cm) are routinely sampled in most soil surveys, subsoil data (30–60 cm) are collected less consistently, and deeper soil layers (>60 cm) are only rarely investigated, typically in detailed or site-specific studies. This is to be expected, as deeper sampling increases costs and soil surveys are typically performed to inform short-term agronomic decisions, where the specific characteristics of the soil below the plough layer are largely irrelevant. Consequently, large-scale soil databases exhibit a pronounced vertical data gap, with dense information available for topsoil layers but sparse or missing measurements at depth. This depth bias introduces significant limitations for applications that depend on the full soil profile, such as hydrological modeling, groundwater recharge estimation, and nutrient leaching assessments, where subsoil and deeper soil properties play a critical role. To combat this limitation, we are examining Machine Learning and geostatistical frameworks of predicting subsoil textural composition on the heterogeneous landscape of Greece. More specifically, (i) the prediction based on raw compositional data versus isometric log-ratio (ilr)–transformed coordinates, (ii) the integration of spatial information within machine-learning frameworks, (iii) univariate per-component regression versus multivariate regression approaches, and (iv) the inclusion and exclusion of predictor variables are being examined. Additionally, all machine-learning models are benchmarked against equivalent ordinary least squares (OLS) regressions, which serve as baseline models. This comparison enables the assessment of potential relationships between the performance of simple, interpretable regression models and that of more complex, traditionally less-interpretable machine-learning approaches. Preliminary results are encouraging, with R² values of 0.77, 0.75, and 0.61 for the prediction of subsoil clay, sand, and silt contents, respectively, on the most robust univariate raw compositional spatially aware Random Forest model. Previous studies suggest that predictive performance may be further improved through compositional data pre-processing using isometric log-ratio (ilr) transformation and multivariate Random Forest modeling.
How to cite: Soulis, K., Nikitakis, E., Gerontidis, S., Stavropoulos, A., and Kalivas, D.: Comparative examination of methods for the prediction of subsoil texture composition., EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6486, https://doi.org/10.5194/egusphere-egu26-6486, 2026.
ABSTRACT
The Northeast China black soil region is one of the country’s most important grain-producing areas. However, long-term conventional tillage characterized by deep plowing and intensive soil disturbance has caused severe soil erosion and continuous declines in surface soil organic matter (SOM). In response, conservation tillage has been widely promoted, with pre-sowing crop residue cover (CRC) regarded as a key indicator of its effectiveness. Nevertheless, residue cover substantially weakens soil signals in optical remote sensing, reducing sensitivity to SOM and making accurate SOM mapping under high-coverage conditions particularly challenging. Consequently, achieving robust SOM inversion across CRC conditions has become a critical bottleneck for long-term soil quality monitoring.
To address this challenge, this study identifies the pre-sowing period as the optimal temporal window for SOM remote sensing inversion and develops a SOM estimation framework integrating Sentinel-2 optical imagery, Sentinel-1 SAR data, and multi-source environmental covariates. A total of 585 surface soil samples (0–10 cm) and 117 UAV observations were collected from representative black soil areas in Northeast China. Continuous CRC maps were first generated and used as prior information to partition the study area into bare-soil and residue-covered zones, for which independent random forest regression models were constructed. In bare-soil areas, surface SOM was directly estimated using spectral indices and environmental variables. In residue-covered areas, spectral unmixing was applied to separate soil and residue components, which were combined with SAR penetration features to supplement surface soil dielectric information. In addition, long-term CRC indicators represented by multi-year cumulative values were incorporated to characterize cumulative residue return effects on surface SOM accumulation.
The results demonstrate that the proposed framework significantly improves SOM estimation accuracy in residue-covered areas. Compared with a CRC-agnostic baseline model, R² increased from 0.72 to 0.86 and RMSE decreased from 0.58 to 0.42, corresponding to an approximate 27.6 % reduction in estimation error. High-resolution SOM maps for 2016—2025 reveal a stable northeast-high to southwest-low spatial gradient across the black soil region. High SOM contents (>25 g·kg⁻¹) occur in the Sanjiang Plain and in regions extending from the Lesser Khingan Mountains to the Changbai Mountains, where humus-rich dark brown forest soils predominate. Moderate SOM levels (12–25 g·kg⁻¹) dominate the Songnen and Liaohe Plains, while low SOM contents (0–12 g·kg⁻¹) persist in the southern and western Songnen Plain and aeolian sandy regions of eastern Inner Mongolia. Spatial statistical analysis further indicates that SOM accumulation rates in high-CRC areas are approximately 20 % higher than those in low-coverage regions.
Overall, the proposed multi-source remote sensing framework integrates spectral unmixing, SAR penetration information, and conservation tillage-related features to achieve accurate SOM estimation under CRC conditions. This framework provides a transferable and operational approach for soil quality monitoring under conservation tillage, supporting soil improvement assessment, policy evaluation, and sustainable agricultural management in the Northeast China black soil region.
Keywords: Soil Organic Matter; Conservation Tillage; Multi-source Remote Sensing; Machine Learning; Black Soil Region
How to cite: Liu, H., Li, J., Liu, S., Wan, W., and Liu, Z.: Digital mapping of soil organic matter under conservation tillage in Northeast China based on multi-source remote sensing, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9103, https://doi.org/10.5194/egusphere-egu26-9103, 2026.
Accurate estimation of soil organic carbon (SOC) is crucial for soil health, agricultural productivity, and climate change mitigation. Digital SOC mapping often lacks multi-sensor integration, improved bare soil compositing, and robust uncertainty assessment. We used 6-year multi-temporal Synthetic Aperture Radar (SAR) from Sentinel-1 and optical data from Sentinel-2. This study also utilized confidence interval (CI)–based bare soil compositing for SOC prediction in an agricultural landscape in northeast Germany to enhance SOC estimation.
Four Random Forest models were developed to isolate and compare the independent and combined contributions of optical and SAR data. Local soil samples collected between 2013 and 2022 were divided into training and testing datasets. Independent validation was conducted using samples collected in 2024. Pixel-wise uncertainty was quantified through 100 repeated model runs with different training and testing splits, resulting in a spatially explicit SOC uncertainty map.
Combining SAR and optical data improved in model calibration, while CI-based compositing further enhanced prediction accuracy. Using important features, the model achieved a coefficient of determination (R²) of 0.79 and a ratio of performance to deviation (RPD) of 2.23 in independent validation. The models incorporating SAR data showed higher uncertainty due to its sensitivity to soil conditions; however, standalone SAR data still yielded acceptable SOC mapping performance (R² = 0.57, RPD = 1.54). These results show that combining multi-temporal optical and SAR data with explicit uncertainty assessment enhances the robustness and reliability of SOC mapping across agricultural landscapes.
How to cite: Mahmood, T., Conrad, C., Wenzel, J. L., and Pöhlitz, J.: Improving soil organic carbon estimation and uncertainty assessment using multi-temporal optical and SAR data , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-10160, https://doi.org/10.5194/egusphere-egu26-10160, 2026.
High resolution (HR) soil information is critical for a broad range of applications, including agriculture and sustainable land practice, environmental modelling, nature conservation and resource management. Its importance is particularly pronounced at local scales and in low-relief landscapes where subtle topographic variations can strongly influence soil properties.
Our aim was to evaluate the contribution of LiDAR-derived topographic information to HR soil property mapping and to assess its added value compared to conventional digital elevation models (DEMs) on a low-relief, alluvial plain in Hungary. Soil pH was modelled using a hybrid, machine learning and geostatistical approach that integrates LiDAR-derived DEM and its derivatives, Sentinel-2 imagery, geological map and land cover information, achieving acceptable predictive performance (RMSE = 0.55, ME = 0.01). The results indicate that the LiDAR-derived topographic covariates were the most important predictors. Comparisons with existing large-scale soil pH maps revealed very weak spatial agreement in both spatial patterns and value distributions, which can be largely attributed to the limited capability of conventional DEMs to represent microtopography and subtle elevation changes in low-relief areas, particularly under forest cover.
These findings demonstrate the substantial added value of LiDAR-derived data for HR soil mapping. When used as environmental covariates, LiDAR-DEM and its derivatives, effectively represent topographic features and capture soil moisture and drainage effects that influence soil formation even in low-relief areas, regardless of vegetation cover. Since topographic parameters are decisive factors in soil property modelling, the quality of the applied DEM directly determines the quality of the resulting soil maps.
How to cite: Takács, K., Árvai, M., Szatmári, G., and Pásztor, L.: LiDAR-based high resolution soil mapping in a Hungarian lowland area, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11524, https://doi.org/10.5194/egusphere-egu26-11524, 2026.
Bulk density (BD) is a key soil property due to its strong influence on the physical, chemical, and biological properties of soil, serving as an important indicator of soil health, compaction, and physical quality. An earlier study revealed that BD observations in the Hungarian Soil Information and Monitoring System (SIMS) might have been biased and therefore require correction. SIMS is a nationwide soil monitoring system that collects information on the temporal variability of Hungarian soils across Hungary at 1236 soil profiles. Accordingly, the objective of this research was to bias-correct BD measurements in SIMS using advanced pedotransfer functions (PTFs).
Various PTFs, including multiple linear regression, generalized additive model, cubist, random forest, and artificial neural networks, were developed based on the Hungarian Detailed Soil Hydrophysical Database and an extensive set of environmental covariates (e.g., long-term climatic data, topography and its derivatives), which serve as proxies for soil-forming factors. The developed PTFs were evaluated and compared using five times repeated 10-fold cross-validation, revealing that the random forest-based (RF) PTF outperformed other techniques, with RMSE and model efficiency coefficient values of 0.099 g cm−3 and 0.539, respectively. Consequently, the RF-based PTF was used to correct BD measurements in SIMS and to provide quantitative information on the uncertainty associated with the corrected BD values. The latter is essential to support end users in the proper interpretation and application of the corrected BD values. Subsequently, a dataset was compiled containing information on the SIMS profile and layer identifiers, the upper and lower depth boundaries of each soil genetic horizon, the corrected BD values, and their associated uncertainties. The dataset is publicly available on Zenodo (https://doi.org/10.5281/zenodo.16926945).
The aim of our presentation is to introduce the applied methodology, the RF-based PTF developed for bias correction of BD measurements in SIMS, and, most importantly, the resulting BD dataset compiled within the framework of this study.
Reference:
Sohrab S., Szabó B., Pásztor L., Makó A., Szatmári G., 2025: Adjusting bulk density observations in the Hungarian Soil Information and Monitoring System using pedotransfer functions. European Journal of Soil Science 76(6): e70245.
How to cite: Szatmári, G., Sohrab, S., Szabó, B., Makó, A., and Pásztor, L.: Pedotransfer functions for bias-correcting bulk density observations in the Hungarian Soil Information and Monitoring System, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12098, https://doi.org/10.5194/egusphere-egu26-12098, 2026.
A wide range of disciplines and many national and international initiatives require spatially, and increasingly, spatiotemporally explicit information on soil properties, functions, processes and services at various scales. These emerging demands are mostly due to the recognition of the multifunctionality of soils. During the past decades a general demand for quantitative digital soil information for environmental modelling and management has compelled soil scientist to address many soil-related questions from a quantitative point of view. Digital soil mapping (DSM) together with pedometrics and proximal soil sensing will keep on playing crucial role in the forecasted near future referred as digital pedology as the upcoming era of the fundamental understanding of soil in space and time.
Hungarian soil spatial data infrastructure has recently been renewed within the DOSoReMI.hu initiative. Primary soil property maps compiled according to GlobalSoilMap.net specifications provide an operational and widely accepted source of information on soil at national level. Nevertheless, more recently new challenges have arisen from temporal, thematic and spatial point of view: higher resolution, up-to-date DSM products are expected, which describe soil features (properties, functions, processes) in more details, from various aspects. Our overall aim is to address these challenges by (i) developing approaches that extend and/or rethink the three pillars of DSM, (ii) linking DSM products of different scales and (iii) bridging potential gaps between them.
Based on the above background, it is both rational and socially beneficial to identify opportunities to expand the potential and improve the performance of DSM in the support of the multifaceted concept of soil security. Our objectives are twofold: (i) to continue developing and modernizing the national spatial soil data infrastructure capable of meeting the emerging needs for soil information for various purposes and at different scales, (ii) by addressing (at least some of) the ‘pedometrics ten challenges’, methodological issues that are currently at the forefront of soil research.
During our forecasted activity we are going to focus on some highlighted topics keeping the possibility to extend them by identifying new challenges and solutions:
- Introduction of spatially dense observations together with proper pedotransfer function development into DSM to increase thematic versatility and spatial resolution.
- Application of ancillary data originating from high-resolution (both spatial and spectral) imagery and geophysical sensors to support e.g. pedometric zoning of agricultural plots and increase prediction accuracy on flat areas.
- Introducing spatially distributed, physics-based models into digital mapping of specific soil functions and processes through pedological interpretation of spatiotemporal model outputs.
- Testing the impact of iterative mapping and ensemble modelling on prediction performance based on detailed and refined accuracy assessment.
According to our experiences, working on approaches to solve certain mapping challenges may easily conclude to newer methodological questions and/or ideas. With the present paper and the topics listed above, our aim is to stimulate DSM stakeholders (both users and producers) to identify and/or formulate new challenges and solutions, which might be worth for overthinking and/or solving within the spatial branch of the upcoming era of digital pedology.
How to cite: Pásztor, L., Szatmári, G., Laborczi, A., Takács, K., Mészáros, J., Árvai, M., Takáts, T., Kocsis, M., Benő, A., and Csikós, N.: Some thoughts on the opportunities to expand the potential and improve the performance of digital soil mapping in Hungary, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12463, https://doi.org/10.5194/egusphere-egu26-12463, 2026.
Spatiotemporal variation of soil organic matter (SOM) contents was significant to research on global warming, greenhouse effect, and ecosystem health and quality. However, the spatiotemporal modeling for soil properties in most studies focused on a discrete-time calibration and validation, and faced the problems of missing observations. Integrated Nested Laplace Approximation with the Stochastic Partial Differential Equation (INLA-SPDE) model that is robust to missing data and unbalanced sampling design was proposed as a potential model for spatiotemporal soil modeling. This study presented an application of INLA-SPDE for spatiotemporal modeling of SOM (2006, 2010, 2018) using 924 samples and 8 environmental covariates in Lishu County, Northeast China. The results demonstrated that the INLA-SPDE model incorporating spatiotemporal information generally outperformed the two-phase methods based on Cubist and Random Forest, particularly in years requiring greater temporal extrapolation (2006 and 2018), while achieving comparable performance in 2010. This superiority can be attributed to its comprehensive consideration of various sources of uncertainty. Furthermore, the posterior distributions derived from the model provided valuable insights into the effects of environmental covariates on SOM spatiotemporal variation, with clay content showing the strongest positive influence and annual precipitation exhibiting a notable negative effect. The spatial pattern of SOM consistently exhibited higher values in the east and lower values in the west. After an overall decline from 2006 to 2010, mean SOM content increased from 17.97 g kg-1 to 20.85 g kg-1 between 2010 and 2018 (a total increase of 2.88 g kg-1 at an annual rate of 0.36 g kg-1 yr-1), with notable recovery in central and eastern areas, likely associated with the implementation of straw returning practices. In addition to prediction accuracy, computational complexity, and uncertainty analysis, the study evaluated the model from new perspectives, including covariate interpretability and flexibility. This research provides a promising spatiotemporal modeling framework for digital soil mapping.
How to cite: Ji, W., Yu, Q., Li, B., Huang, Y., and Yan, Y.: Spatiotemporal modeling of soil organic matter in the black soil area of Northeast China with INLA-SPDE and remote sensing data, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15979, https://doi.org/10.5194/egusphere-egu26-15979, 2026.
Digital Soil mapping (DSM) provides standardised information layers. The recent availability of global and continental remote sensing-derived products coupled with the ease-of-access to computational resources has made the production of such layers easier. It is ever to characterise and evaluate such DSM-derived products, in particular the type of actual information they can provide to users.
DSM studies commonly assess prediction uncertainty using various approaches, including multiple simulations or quantile random forests. These studies provide measures of accuracy derived from statistical (cross-)validation, often based on non-probability and non-representative observations. However, these accuracy metrics and uncertainty assessments do not encompass all the potential elements that could be used to characterise a DSM product, and they do not directly address the needs of the users. We assessed maps based on area of applicability (i.e., the area in covariate space where the model learns about relationships based on the training data), the landscape heterogeneity both in the landscape itself and in covariate space, and the local influence of the covariates on the final products.
We present examples of continental and global mapping products, highlighting main accuracy, uncertainty and interpretability aspects and how these influence their suitability for intended use by stakeholders, decision makers and users in general at the given resolution. The results permit some practical reflections on how to integrate all the above elements to identify regions where the confidence in the predictions is highest and the associated uncertainty lowest, but also, where the product is not considered fit for the intended use.
How to cite: Poggio, L., Rossiter, D., Batjes, N., and Kempen, B.: Towards user needs assessment for DSM products, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-17489, https://doi.org/10.5194/egusphere-egu26-17489, 2026.
Machine-learning (ML) models are increasingly used to predict environmental process parameters from literature-derived datasets. A common but rarely scrutinized practice is random data splitting for model training and evaluation, which implicitly assumes independence among samples. However, environmental datasets often contain strong group structures arising from shared soil sources. Samples originating from the same soil may share substantial, unquantified microbial information, including community composition, functional potential, and legacy effects, which cannot be fully represented by standard physicochemical descriptors. Ignoring such group structure may therefore induce group-wise information leakage and lead to overoptimistic assessments of model performance.
Here, we systematically examine the consequences of random versus group-wise data splitting for ML-based prediction of distribution coefficient (Kd) and first-order degradation rate constants (μ) of atrazine in soils. A dataset was compiled from published batch experiments and incubation studies, comprising 306 datasets from 205 distinct soils (adsorption) and 329 datasets derived from 77 distinct soil sources (degradation); grouping was defined exclusively based on shared soil sources. This grouping strategy explicitly reflects the presence of latent microbial controls that remain unobservable to the model. ML models were trained using identical algorithms but evaluated under two contrasting strategies: (i) conventional random splitting that ignores soil-based group structure, and (ii) group-wise splitting that enforces complete separation of soil sources between training and testing sets.
Taking atrazine degradation as an example, under random splitting, models exhibit apparently strong predictive performance, characterized by near-zero mean bias and inflated coefficients of determination (R² against the 1:1 line = 0.835; RMSE = 0.037; MAE = 0.019). In contrast, group-wise splitting reveals a pronounced degradation in performance, with the coefficient of determination against the 1:1 line dropping to R² = 0.099, accompanied by substantially increased errors (RMSE = 0.093; MAE = 0.053) and systematic overestimation of μ, reflected by a positive bias of 0.013 (≈ 24%). A similar pattern emerges for atrazine adsorption. These results demonstrate that random data splitting can fundamentally overstate the predictive capability of ML models trained on literature-derived soil datasets when shared soil sources are present. Therefore, we argue that soil-based group-wise evaluation is essential for ensuring robust assessment of model generalizability in data-driven studies of soil biogeochemical processes.
How to cite: Chen, F. and Vanderborght, J.: Random data splitting of literature-derived data ignoring group structure leads to group-wise information leakage in machine-learning models: Evidence from atrazine adsorption and degradation in soils, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19636, https://doi.org/10.5194/egusphere-egu26-19636, 2026.
Obtaining representative soil samples is fundamental for robust soil research, agronomic decision-making, and evidence-based policy planning. Because soil sampling and laboratory analyses are resource-intensive, sampling campaigns typically rely on selecting a limited number of representative sites using methodologies that depend on the spatial scale of interest and the availability of prior information. A common approach is stratified sampling, in which sampling locations are allocated based on known drivers of soil variability. However, the choice of stratification variables and their effectiveness across scales remain open questions.
In this study, we compare three soil sampling strategies: simple random sampling, land-use-based stratification, and stratification based remote sensing products, across multiple field-scale and small-landscape-scale case studies. The performance of each strategy is evaluated in terms of its ability to capture spatial variability in key soil properties while minimizing sampling effort using design efficiency as the evaluation criterion. Our results provide insights into the relative efficiency and robustness of remote sensing-based stratification compared to more commonly applied approaches, and highlight the conditions under which each sampling strategy is most appropriate.
How to cite: Bernardini, L. G., Smit, E., Izquierdo-Verdiguier, E., Rosinger, C., Bodner, G., Wenzel, W., and Keiblinger, K.: Seeing variability before sampling: soil stratification from fields to landscapes, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-20856, https://doi.org/10.5194/egusphere-egu26-20856, 2026.
Global Navigation Satellite System-Reflectometry (GNSS-R), as a favorable technology to provide large-scale soil moisture estimates, contributes to studies in climatology, hydrology, and agriculture. The Tianmu-1 Meteorological Mission (TM-1), currently runs 23 satellites in orbit (including one experimental satellite) with multi-GNSS compatibility, achieve shorter revisit periods and higher data acquisition frequencies compared with single-satellite missions. The hourly TM-1 surface soil moisture (SSM) products, offer affluent information for global soil moisture monitoring. This study provides the first comprehensive characterization and performance evaluation of TM-1 SSM products based on in-situ measurements and products of Soil Moisture Active Passive (SMAP), European Space Agency Climate Change Initiative (ESA CCI), and Global Land Data Assimilation System (GLDAS). The TM-1 SSM demonstrates expected spatiotemporal patterns at both regional and global scales. The in-situ validation results reveal its landcover-dependent accuracy, with superior performance over bare soils (unbiased Root Mean Square Error, ubRMSE of about 0.02 m³/m³) compared to vegetated regions (ubRMSE of around 0.07 m³/m³). Furthermore, Extended Triple Collocation (ETC) assessments using (1) TM-1, active, and ground observations and (2) TM-1, model, and ground observations triplets are conducted. The ETC-derived results present that TM-1 SSM achieve global correlation coefficient of 0.75 and random error standard deviation of 0.035 m³/m³. Overall, this study demonstrates the reliable accuracy of TM-1 SSM product, and provides valuable insights for its refinement and potential applications.
How to cite: Wang, Q.: In-situ and triple collocation-based evaluations of Tianmu-1 global soil moisture products, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21463, https://doi.org/10.5194/egusphere-egu26-21463, 2026.
