Spatial Prediction of Saturated Soil Hydraulic Conductivity in Headwater Basins Using Machine Learning: A Transferable Methodological Framework

Spatial Prediction of Saturated Soil Hydraulic Conductivity in Headwater Basins Using Machine Learning: A Transferable Methodological Framework | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 31 December 2025 V1 Latest version Share on Spatial Prediction of Saturated Soil Hydraulic Conductivity in Headwater Basins Using Machine Learning: A Transferable Methodological Framework Authors : Leandro Campos Pinto 0000-0003-2071-7028 [email protected] and Carlos Mello Authors Info & Affiliations https://doi.org/10.22541/au.176717717.70038547/v1 211 views 120 downloads Contents Abstract 2.3 Terrain Attributes 3.3 Validation of the Ksat Map Based on Hydrological Indicators Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract This study presents a machine learning–based methodological framework for spatially predicting saturated soil hydraulic conductivity (Ksat) in headwater basins, using the Upper Grande River Basin (UGRB), Brazil, as a representative case study. Four Machine Learning models—Support Vector Machine (SVM), Generalized Linear Model (GLMnet), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) were assessed under three data-processing scenarios, including robust outlier filtering and low-Ksat thresholding. Terrain attributes, land use, and geomorphological indices were used as environmental covariates. Among the tested algorithms, the SVM model achieved the best predictive performance, presenting the lowest Root Mean Squared Error (RMSE = 0.602) and Mean Absolute Error (MAE = 0.523) in the outlier-filtered scenario. Although the explained variance was limited across all models, reflecting the intrinsic spatial variability of Ksat, variable importance analysis highlighted geomorphological classifications and curvature-related terrain metrics as the dominant predictors. The resulting Ksat map revealed coherent spatial patterns, with higher conductivity values associated with upland areas and native vegetation cover. Functional validation using independent hydrological indicators showed a strong positive correlation (r ≈ 0.84) between the predicted mean Ksat and the baseflow-to-total runoff ratio across five sub-basins, confirming the hydrological consistency of the spatial predictions. The proposed framework is transferable and provides a robust basis for integrating soil hydraulic properties into distributed hydrological modeling and land-use planning in headwater catchments. Spatial Prediction of Saturated Soil Hydraulic Conductivity in Headwater Basins Using Machine Learning: A Transferable Methodological Framework Leandro Campos Pinto; Carlos Rogério de Mello Corresponding Author: [email protected] Federal University of Lavras, Water Resources Department Abstract This study presents a machine learning–based methodological framework for spatially predicting saturated soil hydraulic conductivity (Ksat) in headwater basins, using the Upper Grande River Basin (UGRB), Brazil, as a representative case study. Four Machine Learning models—Support Vector Machine (SVM), Generalized Linear Model (GLMnet), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) were assessed under three data-processing scenarios, including robust outlier filtering and low-Ksat thresholding. Terrain attributes, land use, and geomorphological indices were used as environmental covariates. Among the tested algorithms, the SVM model achieved the best predictive performance, presenting the lowest Root Mean Squared Error (RMSE = 0.602) and Mean Absolute Error (MAE = 0.523) in the outlier-filtered scenario. Although the explained variance was limited across all models, reflecting the intrinsic spatial variability of Ksat, variable importance analysis highlighted geomorphological classifications and curvature-related terrain metrics as the dominant predictors. The resulting Ksat map revealed coherent spatial patterns, with higher conductivity values associated with upland areas and native vegetation cover. Functional validation using independent hydrological indicators showed a strong positive correlation (r ≈ 0.84) between the predicted mean Ksat and the baseflow-to-total runoff ratio across five sub-basins, confirming the hydrological consistency of the spatial predictions. The proposed framework is transferable and provides a robust basis for integrating soil hydraulic properties into distributed hydrological modeling and land-use planning in headwater catchments. Keywords : Saturated hydraulic conductivity; Machine learning; Digital soil mapping; Terrain attributes; Geomorphology; Headwater basins; Hydrological validation Highlights Machine learning models effectively predicted Ksat in a headwater catchment Terrain attributes and geomorphons were key predictors of Ksat spatial patterns Ensemble models outperformed linear approaches across data scenarios Data filtering improved model robustness and reduced prediction uncertainty The framework is transferable to other data-scarce headwater basins 1. INTRODUCTION Headwater basins play a critical role in hydrological systems due to their capacity to regulate water availability and sustain downstream ecosystems and human demands (Golden et al., 2025; Pinto et al., 2017). Runoff generation in these environments is highly sensitive to land-use change and soil management practices, reinforcing the global challenge of protecting headwater regions and improving the understanding of their hydrological functioning (Golden et al., 2025; Mello et al., 2025). In many tropical regions, deforestation and inadequate soil management have disrupted infiltration processes, groundwater recharge, and baseflow dynamics, leading to increased hydrological vulnerability (Germer et al., 2010; Viola et al., 2014; Wei et al., 2022). In southeastern Brazil, these challenges are particularly evident in the Upper Grande River Basin (UGRB), a strategic headwater region that supports the Brazilian hydropower system and supplies water to numerous municipalities (Mello et al., 2025). At the basin scale, soil hydraulic behavior is a key determinant of hydrological response units, especially in landscapes dominated by highly weathered soils with high infiltration potential, such as Oxisols (Silva et al., 2018). Within this context, saturated soil hydraulic conductivity (Ksat) is a fundamental parameter governing infiltration, groundwater recharge, and flow partitioning in headwater catchments. Saturated hydraulic conductivity is strongly influenced by soil texture, macroporosity, aggregation, and mineralogical composition, as well as by land use and management-induced structural changes (Vieira et al., 2025; Pinto et al., 2019a). In headwater environments characterized by complex terrain, geomorphological controls further modulate soil development and water redistribution, exerting a strong influence on spatial Ksat patterns (Pinto et al., 2015; 2016). However, direct field measurement of Ksat is costly, time-consuming, and challenged by pronounced spatial variability, limiting its applicability at watershed scales (Moustafa, 2000; Zimmermann and Elsenbeer, 2008; Hervé-Fernández et al., 2023). Digital Soil Mapping (DSM) approaches, combined with machine learning algorithms, have emerged as effective tools for predicting soil hydraulic properties using spatial covariates derived from terrain analysis, land cover, remote sensing data, and proximal soil sensing (Ma et al., 2019; Silva et al., 2021; Chen et al., 2022; Ghavami et al., 2023). Nevertheless, Ksat remains one of the most difficult soil properties to model due to its strong dependence on microscale structural features that are poorly captured by macro-scale predictors (Gupta et al., 2021). Consequently, robust modeling strategies and independent validation approaches are,required to ensure physically meaningful predictions. In this study, we propose and evaluate a transferable machine learning–based pipeline for spatial prediction of saturated soil hydraulic conductivity in headwater basins. Four algorithms—RF, SVM with radial kernel, GLMnet, and XGBoost—were compared under different data-processing scenarios to assess model robustness. Beyond statistical validation, the predicted Ksat maps were functionally evaluated using independent hydrological indicators derived from multiple sub-basins. The Upper Grande River Basin was selected as a case study due to its hydrological relevance, physiographic complexity, and data availability, while the proposed framework is designed to be applicable to other headwater regions. 2. Material and mETHODS 2.1 Study Area and Case Study Description The Upper Grande River Basin (UGRB) (Figure 1) is located in southeastern Brazil and represents a strategic headwater region of the Grande River system. The basin is characterized by complex terrain, with elevations ranging from approximately 800 to 2,400 m, and a predominance of deeply weathered tropical soils, mainly Oxisols and Inceptisols (Araújo et al., 2008). The regional climate is classified as Cwb (Köppen), with a marked wet season from October to March and a dry season from April to September. These physiographic and climatic conditions make the basin particularly suitable for evaluating the interaction between terrain controls, soil hydraulic properties, and hydrological response. The UGRB was adopted as a representative case study due to the availability of detailed soil hydraulic measurements, high-resolution digital elevation data, land-use information, and long-term hydrological records. While the analysis focuses on this basin, the methodological framework developed herein is designed to be transferable to other headwater catchments with similar data availability. Image (image1.png) is missing or otherwise invalid. Figure 1. Location of the study area at Upper Grande River Basin (UGRB), its digital elevation model (DEM), land use, sampling points for saturated soil hydraulic conductivity (Ksat), and watersheds for models validation. 2.2 Saturated Hydraulic Conductivity Data Saturated soil hydraulic conductivity (Ksat) data were obtained from field measurements conducted at multiple sampling sites across the basin, covering different soil types, land uses, and topographic positions. Measurements were performed using standardized field methods, ensuring consistency across sampling campaigns. Given the inherently skewed distribution and high spatial variability of Ksat, the dataset was carefully examined for extreme values prior to modeling. To assess model robustness, three data-processing scenarios were considered: (i) the complete dataset without filtering; (ii) an outlier-filtered dataset, in which extreme values were removed using robust statistical criteria; and (iii) a thresholded dataset including only values of Ksat ≤ 2 m day⁻¹, focusing on the dominant hydraulic behavior of the basin. These scenarios allowed the evaluation of model sensitivity to data distribution and extreme observations. 2.3 Terrain Attributes Terrain Attributes (TA) are widely recognized as effective predictors of soil properties, acting as auxiliary spatial variables (Mulder et al., 2011; Vaysse and Lagacherie, 2015; Menezes et al., 2016; Silva et al., 2016; Pinto et al., 2017; Bonfatti et al., 2018; Jena et al., 2022). Analyzing the relationship between TA and Ksat provides a better understanding of its spatial pattern. This approach is crucial because the spatial variability of Ksat at the field scale, linked to TA, establishes a connection between the variability and the water movement in the soil profile, reinforcing the importance of integrating TA into the spatial analysis of Ksat (Menezes et al., 2016; Pinto et al., 2019b; Jena et al., 2022). Based on a contour map on a scale of 1:50,000, a Digital Elevation Model (DEM) of the Upper Rio Grande Basin (UGRB-DEM) was constructed with a spatial resolution of 30 meters. It is a product of the USGS ASTER sensor (available at https://gdex.cr.usgs.gov/gdex/). To ensure the hydrological consistency of the DEM, its depressions (sinks) were filled, and used to calculate the terrain attributes (TA) (Gabrecht & Martz, 1999; Wu et al., 2008; Costabile et al., 2022). The Landuse map utilized in this study was sourced from the MapBiomas Project (Brazil). Specifically, we employed the high-resolution dataset from the 10-meter collection for the year 2023 (based on Sentinel-2 imagery). For the modeling, the following environmental covariates derived from the DEM were used: Topographic Position Index (tpi), Relief Texture (texture), Terrain Roughness Index (terr_rug_i), and Surface Area (surf_area). Flow Direction (fdr), Topographic Wetness Index (twi) (Beven; Kirkby, 1979); Slope, Plan Curvature (planc), Profile Curvature (profc) (Zevenbergen; Thorne, 1987); Aspect (aspect) (Gallant &Wilson, 2000); Cross-sectional Curvature (cros_sec_c), Maximum Curvature (max_curv), Minimum Curvature (min_curv) (Guisan et al., 1999). In addition, multiresolution plan indices such as the Valley Bottom Plan Index (mrvbf) and the Ridge Top Plan Index (mrrtf) (Gallant & Dowling, 2003) were included, along with various curvature metrics such as Longitudinal Curvature (long_curv), Convexity (convexity), General Curvature (gener_curv), and automated topography classifications (Iwa_pike) (Iwahashi and Pike, 2007) and Geomorphons (Jasiewicz & Stepinski, 2013), and hydrological attributes such as Drainage Basin Slope (catch_slop). The use of landform classification approaches, such as Iwahashi and Pike (2007) and Geomorphons, has proven to be particularly effective for representing hydropedological processes in mountainous headwater basins, especially when combined with soil hydraulic properties, such as Ksat (Pinto et al., 2016). 2.4 Definition of Modeling Scenarios To assess the stability and robustness of the model with respect to extreme values points of the target variable, three modeling scenarios were established (Table 1). Table 1. Overview of the three modeling scenarios developed to assess model stability and robustness. A Complete Data Use of the entire available dataset, serving as a reference. B Robust Outlier Filter Influential observations and multivariate outliers were removed using the Minimum Covariance Determinant Estimator (MCD) method (Rousseeuw & Leroy, 1987; Hubert et al., 2018). A cutoff threshold was established based on the chi-square (χ2) distribution at 97.5%. C Low Ksat value filter Sample selection on the original scale considering Ksat ≤ 2 m day-1 focusing on the dominant hydraulic behavior of the basin. 2.5 Training and Evaluation of the Models 2.5.1 Data Preprocessing and Splitting Each scenario dataset was partitioned into Training (80%) and Test (20%) subsets using the “createDataPartition” function from the caret package (Kuhn, 2015). The training then underwent a rigorous pre-processing to optimize data suitability for subsequent modeling. Categorical variables were transformed into binary dummy variables using the “dummyVars” function (or equivalent coding approach) in R. Numerical predictors were centered and scaled to achieve a mean of zero and a variance of one. Further address non-normal distributions in numerical features, the Yeo-Johnson transformation was applied (Yeo and Johnson, 2000). Lastly, all predictors were screened for near-zero variance to prevent the inclusion of predictors that offer minimal informational content to the models. 2.5.2. Machine Learning Models Four distinct Machine Learning (ML) algorithms were selected and compared across each modeling scenario to ensure a comprehensive evaluation of predictive performance: the Random Forest (RF) ensemble model (based on decision trees) (Breiman, 2001); the Support Vector Machine with a Radial kernel (SVM Radial), utilizing the principle of structural risk minimization (Vapnik, 2013); the Generalized Linear Model with Regularization (GLMnet), which incorporates both Lasso and Ridge regularization techniques (Friedman et al., 2010); and the high-performance eXtreme Gradient Boosting (XGBoost) algorithm (Chen and Guestrin, 2016). Model training employed the “trainControl” function from the caret package, configured for Repeated Cross-Validation (10-fold, 3 repeats) with a randomized hyperparameter search. Computational efficiency was achieved through parallel processing (“doParallel”) to minimize overall processing time. 2.5.3. Performance Evaluation The predictive capacity of each trained model was assessed on the independent testing set using a suite of statistical metrics. The primary metric for model optimization and selection was the Root Mean Squared Error (RMSE), calculated on the logarithmic scale, as it penalizes larger errors. Model fitting was further evaluated by the Coefficient of Determination (R 2 ) and the Mean Absolute Error (MAE), both calculated on the logarithmic scale. To provide a readily interpretable metric for the end-user, the Mean Absolute Percentage Error (MAPE) was computed on the original Ksat scale (m day -1 ). Consistent with established methodology, the model exhibiting the lowest overall RMSE was selected as the final, optimal predictive model (Elbisy, 2025). 2.6 Variables Importance Variable importance was calculated for all trained models using the “varImp” function from the R’s caret package. For categorical variables treated as dummy variables, the total importance of the related binary predictors was summed to present the unified contribution of the original variable. 2.7 Spatialization (Mapping) Spatial prediction of log(Ksat) was executed across the full stack of covariate rasters using the Terra package (Hijmans et al., 2022). The prediction procedure was optimized for computational efficiency and memory management by processing the raster stack in blocks or chunks, allowing handling large datasets without overwhelming RAM. Crucially, the covariate values within each block were subjected to the same pre-processing steps defined during model training (dummy coding, centering, scaling, and Yeo-Johnson transformation), ensuring consistency in variable levels and numerical scaling. The final optimized model was then applied to predict log(Ksat) values for every pixel. Post-processing involved the exportation of the resulting log(Ksat) raster (TIFF format) and a subsequent transformation back to the original Ksat scale (m day -1 ) by applying the exponential function (10^(log(Ksat)) – 0.001). All raster covariates were geometrically harmonized to a common grid using a reference raster, ensuring identical spatial extent, resolution, and coordinate reference system before spatial prediction. 2.8 Validation of the Ksat Map Based on Hydrological Indicators The validation of the machine learning models’ performance in mapping Ksat within the Upper Grande River Basin (UGRB) was conducted using hydrological indicators from five watersheds (see Figure 1). Among the indicators, the baseflow-to-runoff ratio (BF/TR) was considered as it is expected that watersheds exhibiting a higher baseflow contribution will demonstrate a higher average Ksat (Campling et al., 2002; Pinto et al., 2017; Soylu & Bras, 2022; Sarah et al., 2024). The use of baseflow-related indicators as proxies for groundwater recharge and subsurface storage has been widely applied in headwater basins, particularly in southeastern Brazil (Pinto et al., 2017; Ribeiro et al., 2024). 2.8.1 Determination of Base Flow Baseflow was determined using monthly time series of mean streamflow and precipitation, organized according to the hydrological year, defined as the period from October to September, following standard practice for southeastern Brazil. Monthly streamflow data were aggregated by hydrological year to compute total annual runoff (D total). The mean hydrological year was identified as the year whose total runoff showed the smallest deviation from the long-term mean of the series, ensuring representation of average hydrological conditions. Baseflow separation was performed using the Barnes method (Barnes, 1939; Durães and Mello, 2013; Hingray et al., 2014), which is based on identifying the inflection point of the hydrograph recession. This method assumes an exponential decay of baseflow, as described by the Maillet equation (Tallaksen, 1995): \(Q_{t}=Q_{0}\ exp(-\alpha.t)\) (1) Where \(Q_{0}\) is the initial baseflow rate; Qt is the baseflow rate at time t (monthly); α (month -1 ) is the recession coefficient. Baseflow dynamics were analyzed using the monthly hydrograph of the mean hydrological year, with emphasis on recession periods that reflect groundwater contributions. This approach ensures consistency among streamflow, precipitation, and runoff analyses and provides a hydrologically coherent estimation of baseflow for the study basin. Recent studies in headwater basins of Minas Gerais have demonstrated that long-term variations in baseflow and groundwater storage are strongly influenced by geomorphological controls and soil hydraulic properties, reinforcing the relevance of baseflow-based validation approaches (Ribeiro et al., 2024). Results and discussion Machine Learning Model Performance The performance evaluation of the four Machine Learning models (Random Forest - RF, Support Vector Machine - SVM, Generalized Linear Model with Regularization - GLMnet, Linear Model, and eXtreme Gradient Boosting - XGBoost) in predicting saturated soil hydraulic conductivity (log 10 Ksat) (Table 2) demonstrates the complexity inherent in mapping this soil hydrological property, which shows a high spatial variability, mainly in a basin scale, such as the Upper Grande River Basin (UGRB). Table 2. Predictive performance of the Random Forest (RF), Radial Kernel Support Vector Machine (SVM), Regularized Generalized Linear Model (GLMnet), and eXtreme Gradient Boosting (XGBoost) algorithms across distinct data-processing scenarios. A RF 0.888 -0.11 0.679 A SVM 0.811 0.07 0.667 A XGBoost 1.043 -0.54 0.814 B RF 0.612 0.31 0.505 B SVM 0.602 0.33 0.523 B XGBoost 0.780 -0.13 0.647 C RF 1.005 -0.12 0.681 C SVM 1.203 -0.60 0.894 C XGBoost 0.950 0.00 0.693 The complexity inherent in modeling saturated hydraulic conductivity (Ksat) is evident in the relatively low explanatory power observed for the coefficient of determination (R²) across the models tested. The highest R² log value achieved was 0.338, as observed for the Support Vector Machine (SVM) in Scenario B. This result indicates that approximately 67% of the variance in Ksat remained unexplained by the spatial covariates utilized, which included terrain attributes as well land use maps. Such limited explanatory capacity underscores the intrinsic difficulty of accurately predicting Ksat using these types of spatial information alone. This finding aligns with recent research emphasizing that Ksat is among the most spatially variable and challenging soil properties to model. The primary reason for this difficulty is Ksat’s strong dependence on microscale structural features, including macroporosity, the arrangement of soil aggregates, and degradation processes induced by land use changes. For example, Gupta et al. (2021) demonstrated that even when using globally trained machine learning models—incorporating over 6,000 georeferenced observations and a comprehensive set of terrain, climate, and vegetation covariates—the models could only account for a limited portion of Ksat variability. This highlights the persistent challenge of representing soil structure and preferential flow pathways at regional scales. Similarly, Costa et al. (2024) found that variations in soil structure and porosity have a significant impact on water movement and hydraulic behavior within subtropical watersheds. Their findings reinforce the necessity of integrating structural indicators when modeling soil hydraulic properties. Empirical studies conducted in South America (Seguel et al., 2020; Marín-Pimentel et al., 2023) and the Andes (Crovo et al., 2021; Hervé-Fernández et al., 2023) further support these observations. They reported pronounced spatial variability in saturated hydraulic conductivity under field conditions, which was closely linked to factors such as land use, soil compaction, and vegetation cover. Given the predictive focus of machine learning algorithms, root mean square error (RMSE_log) and mean absolute error (MAE_log) were selected as primary evaluation metrics. RMSE’s quadratic penalty on large errors makes it suitable for identifying optimal models, while MAE provides a robust measure of average prediction error (Willmott & Matsuura, 2005). Consequently, the SVM and RF models demonstrated the highest effectiveness, achieving the lowest RMSE values in scenario B. The SVM model also presented the lowest MAE (0.505), confirming its superior average predictive accuracy and robustness to outliers. The better performance of these models is likely attributable to their capacity to accommodate nonlinear interactions and complex feature spaces, while controlling overfitting through ensemble-based learning and random feature selection in Random Forest or kernel-based structural optimization in SVM Radial—making them less sensitive to the inherent noise and high variability of the Ksat data.The SVM model exhibited markedly lower performance (RMSE = 1.203 for Scenario C; Table 2), which was an unexpected result for a method widely recognized for its ability to capture nonlinear relationships in complex feature spaces. This outcome suggests that, under the low-Ksat filtering applied in Scenario C (Ksat ≤ 2 m day⁻¹), the signal-to-noise ratio of the remaining data was insufficient to allow the SVM to define stable and generalizable decision boundaries. As a consequence, the model performance approached that of a mean-based predictor, indicating limited effective learning (Pham & Won, 2022). Similar behavior has been reported in SVM applications to highly heterogeneous hydrological and soil datasets, where reduced data variability, sparse support vectors, and sensitivity to kernel parametrization constrain the model’s capacity to exploit nonlinear patterns (Deka, 2014; Lange & Sippel, 2020; He et al., 2019). These findings reinforce that, although advanced nonlinear methods often outperform simpler approaches, their predictive performance remains strongly dependent on data structure, the representativeness of extreme observations, and the balance between variance reduction and boundary definition within the feature space.The removal of outliers (Scenario B) resulted in a marked improvement in model performance across most evaluation metrics, as evidenced by the decline in RMSE for the majority of algorithms (Table 2). For instance, the SVM model’s RMSE decreased from 0.811 to 0.602, indicating that influential and extreme Ksat observations were introducing noise rather than informative structure, and that their removal allowed the model to learn more stable and generalizable nonlinear patterns.While extreme Ksat values may theoretically help delineate the limits of the solution space and reflect preferential flow dynamics (Seguel et al., 2020; Crovo et al., 2021), the results of this study indicate that such observations acted primarily as influential noise rather than informative signals. Their exclusion in Scenario B led to improved predictive performance, suggesting that the retained data better represent the dominant soil hydraulic behavior captured by the available covariates. This finding is consistent with other machine learning-based studies in soil science, where strict outlier filtering enhanced model robustness and predictive reliability (Maldaner et al., 2021). In Scenario C, where the dataset was restricted to low Ksat values, model performance deteriorated across all evaluated metrics (Table 2). Contrary to expectations, the reduction in the range of the response variable did not lead to improvements in absolute error measures. In particular, the SVM model exhibited a substantial increase in error, with RMSE_log rising to 1.203 and MAE_log reaching 0.894, alongside a strongly negative R²_log (−0.60). These results indicate that limiting the dataset to low Ksat values reduced the effective variability necessary for defining stable nonlinear relationships, thereby impairing model learning. Similar degradation, although less pronounced, was observed for Random Forest and XGBoost, suggesting that the exclusion of higher Ksat values constrained the representativeness of the training data and weakened the models’ predictive capability.Ksat remains a particularly difficult parameter to predict with precision, primarily due to its high spatial variability and its multiscale dependence on physical factors, such as macroporosity and soil structure (Deb & Shukla, 2012). Consequently, the relatively modest R 2 log values observed in this study are consistent with findings within the field of Digital Soil Mapping (DSM), where predictive performance is often limited, mainly in studies focusing on large watersheds or the complex characteristics of tropical soils. Furthermore, global and regional-scale Ksat prediction efforts using advanced machine learning techniques, including Random Forest algorithm, consistently have produced modest R 2 values (Gupta et al., 2021; Veloso et al., 2022). Consequently, the relatively modest R² log values observed in this study are consistent with findings within the broader field of Digital Soil Mapping (DSM), where predictive performance for various soil properties at large scales can be limited (Chen et al., 2022), especially in studies focusing on large watersheds or the complex characteristics of tropical soils. These findings underscore the inherent complexity and persistent difficulty in accurately modeling this soil hydraulic property. The analysis of the variable importance in the SVM model (the best performing model with the lowest RMSE in Scenario B) was crucial (Figure 2). It identified the most relevant spatial covariates that the model used to capture the spatial dependence structure of Ksat in the Upper Grande River Basin (UGRB). Image (image2.png) is missing or otherwise invalid. Figure 2. Importance of SVM model variables in scenario B for predicting Ksat. Figure 2 presents the ranking of covariates according to their relative importance in Ksat prediction, highlighting convexity, cross-sectional curvature (cross_sec_c), and geomorphological classes derived from Geomorphons as the most influential predictors. This hierarchy confirms the dominant role of geomorphological control and terrain-derived attributes in regulating Ksat within the Upper Grande River Basin. The prominence of curvature- and shape-related metrics indicates that local topographic configuration and surface form exert a strong influence on soil hydraulic behavior, likely through their effects on water redistribution, soil development, and structural organization. In addition, covariates related to plan curvature and landform classification based on the Iwahashi and Pike (2007) method also contributed to model performance, indicating that variations in flow convergence and divergence, as well as terrain segmentation into morphologically distinct units, capture micro- and meso-scale differences in effective soil depth, surface stoniness, and subsurface continuity, which collectively influence soil permeability. Spatial Pattern of Saturated Soil Hydraulic Conductivity (Ksat) The Ksat map generated by the best-performing model (SVM) is in Figure 3. It reveals a spatial pattern that demonstrates hydrological coherence and strong dependence on the most important landscape covariates, such as Terrain Attributes (TA). We can see the highest Ksat values matched with the upper watersheds, which has demonstrated high water yield capacity (Mello et al., 2025). Image (image3.png) is missing or otherwise invalid. Figure 3. Ksat map for the Upper Grande River Basin, southeast Brazil. A visual assessment of the predicted map reveals a distinct spatial pattern of Ksat across the UGRB. The highest Ksat values, indicative of elevated permeability, are predominantly situated in upland regions such as hilltops and interfluves (Figure 3). This topographical correspondence aligns with the hydrological aspects in headwater basins: elevated areas experience a significant sub-surface flow, resulting in reduced water residence time (Soulsby et al., 2006; Pinto et al., 2019; McMillan, 2020; Brutsaert, 2023). This drainage mechanism fosters the development of well-drained soils, which often exhibit coarser textures or highly porous structures not susceptible to compaction from prolonged saturation (Brutsaert, 2023). In the other words, areas associated with higher predicted Ksat values are consistent with environments exhibiting lower bulk density and higher macroporosity, which are typical attributes of soils under native vegetation or less intensive management, as reported for Oxisols in southeastern Brazil (Pinto et al., 2019a). Sub-surface limits the accumulation of fine particles like clay, thereby, preventing macropores blockage and promoting the maintenance of high Ksat values in the surface layer (Blanco-Canqui et al., 2002; Lin et al., 2008, 2010; Sobieraj et al., 2002). Furthermore, elevated Ksat values in these watersheds are strongly linked to the presence of remnant native vegetation, including Atlantic Forest, a relationship likely driven by enhanced macroporosity resulting from root systems and organic matter inputs (Pinto et al., 2015; 2017; Chandler et al., 2018; Oliveira et al., 2017; 2018). The accumulation of organic matter and the prevention of soil compaction, both influenced by agricultural and livestock activities, ensure that the Ksat of the surface layer is maximized (Menezes et al., 2016; Pinto et al., 2017; Oliveira et al., 2018). Price et al. (2010) examined soil physical properties under forest, natural pasture, and managed pasture in the southern Blue Ridge Mountains of southwestern North Carolina, USA. They found that converting forests to other landuses, such as pasture and managed pasture, decreased soil infiltration capacity and led to increased overland flow, significantly altering the water dynamics in the watershed. Also, Higher Ksat values in forested headwater areas reinforce the role of subsurface flow pathways in buffering hydrological responses and maintaining stream water quality, a process previously documented in Atlantic Forest environments of southeastern Brazil (Pinto et al., 2013). Owuor et al. (2018), while studying the Sondu basin in western Kenya, Africa, observed lower hydraulic conductivity in pasture soils compared to natural forest soils. They attributed this to soil compaction induced by animal trampling. Furthermore, they linked changes in landscape hydrology, river discharge patterns, and groundwater recharge rates to landuse changes in the watershed. These findings underscore the pivotal role native forest maintenance plays in regulating water movement within the soil profile of the headwater basins. Similar studies have highlighted that the conservation of adequate soil cover, preferably native vegetation in headwaters, will promote greater hydrological sustainability (Bonell et al., 2010; Pinto et al., 2017; Chandler et al., 2018; Oliveira et al., 2017; 2018; Leimer et al., 2021; Mozaffari et al., 2021; Blanchy et al., 2023), and increase resilience to impacts from climate change. In contrast, areas exhibiting low Ksat values are primarily located in valley bottoms and foothill areas (Figure 2), which function as regions of flow convergence and potential water accumulation. The transport and deposition of fine material from the slopes, along with the formation of clay horizons and prolonged saturation, naturally reduce Ksat (Brutsaert, 2023). Consequently, the predicted spatial pattern has significant implications for the hydrological modeling of the basins, thereby validating the utility of the generated map for macro-scale analyses. Furthermore, the pronounced visual correlation between predicted Ksat, altitude, and landuse (Figure 1) indicates that the SVM model accurately captured the influence of landscape features on soil permeability. These results demonstrate that the prediction of Ksat in the Upper Grande River region was both precise and functionally relevant, supporting its integration into distributed hydrological models, where the identification of areas with varying permeability is prioritized over absolute values at each sampling point. 3.3 Validation of the Ksat Map Based on Hydrological Indicators This section details the functional validation of the Ksat map generated by the SVM model, using long-term hydrological indicators from five representative watersheds of the UGRB. The validation process incorporates the relationships between the predicted average Ksat and the watershed-scale indices, including baseflow (BF), surface runoff (OF), and flow recession behavior (α), and is supplemented with data on predominant landuse (Table 3; Figure 4). Table 3. Summary of hydrological indicators, predicted Ksat (Ksat_pred), and predominant land use (P_landuse) per watershed. Aiuruoca 329 757 0.031 1496 0.67 Forest (64.2)* Andrelândia 353 496 0.033 1677 0.54 Rupestrian field (40.1) Carvalhos 339 542 0.039 1369 0.62 Forest (53.7) Fazenda Paraíba 255 307 0.018 1521 0.52 Pasture (42.1) Madre de Deus 389 327 0.008 1526 0.50 Pasture (29.2) *The predominant land use percentage is superscripted in parentheses. Image (image4.png) is missing or otherwise invalid. Figure 4. Functional validation of the Ksat map: relationships between predicted average Ksat and watershed hydrological indicators. The analysis of the relationship between Ksat and the Baseflow/Total Runoff (BF/TR) ratio (Figure 4A) revealed a pronounced positive linear trend (R² = 0.84; p = 0.0283), suggesting that watersheds exhibiting higher predicted Ksat values contribute proportionally greater baseflow to surface runoff. This finding underscores the physical consistency of the SVM model, as enhanced soil permeability increases infiltration and groundwater recharge, thereby sustaining streamflow during dry periods. This strong correlation indicates that the spatialized Ksat map is hydrologically coherent, reinforcing the concept that soil hydraulic properties and geomorphology exert a buffering effect on groundwater storage and baseflow dynamics, as also reported for headwater basins in Minas Gerais by Ribeiro et al. (2024). The Aiuruoca watershed, predominantly covered by native forest (64.2%), exhibited the highest base flow (757 mm) and BF/TR ratio (69.7%), in agreement with its elevated Ksat (0.67 m day⁻¹). Similarly, the Carvalhos watershed — also extensively forested (53.7%) — displayed comparably high Ksat values (0.62 m day⁻¹) and BF/TR = 61.5%, reflecting the influence of preserved vegetation and accumulated organic matter on soil structure and infiltration capacity (Pinto et al., 2017; Oliveira et al., 2017; 2018). In contrast, the Fazenda Paraíba and Madre de Deus watersheds, where pasture and agricultural lands are predominant (42.1% and 29.2%, respectively), exhibited the lowest Ksat values (0.52 and 0.50 m day⁻¹) and the lowest BF/TR ratios (54.7% and 45.7%). These findings reflect limited infiltration and groundwater recharge, which are characteristic of regions experiencing soil compaction and decreased macroporosity as a result of pasture and agricultural fields management (Centeri, 2022; Singh, 2025). Similarly, Andrelândia, with a dominance of rocky fields (40.1%), showed intermediate hydrological responses (Ksat = 0.54 m day⁻¹; BF/TR = 58.4%), suggesting moderate infiltration capacity in shallow, coarse-textured soils, typical of these environments (Liu et al., 2025). The inverse relationship observed between Ksat and Overland flow/Total Runoff (OF/TR) (Figure 4C; R² = 0.84; p = 0.0283) further supports this interpretation. Increased Ksat is associated with reduced OF, indicating that landscapes characterized by native vegetation and higher organic matter inputs enhance infiltration and reduce surface runoff. On the other hand, the predominance of pastures leads to greater surface runoff due to soil compaction and decreased permeability. The relationship between Ksat and the recession coefficient (α) (Figure 4B) was positive, but not significant (R² = 0.49; p = 0.187). Although not statistically significant, it remains physically consistent: higher Ksat values are linked to increased α values, reflecting a faster hydrological response after rainfall occurrence. Forested watersheds, such as Carvalhos (α = 0.039) and Aiuruoca (α = 0.031), showed more marked recession behavior, indicative of permeable soils and efficient groundwater-streamflow connectivity. In contrast, Madre de Deus (α = 0.008) and Fazenda Paraíba (α = 0.018) presented lower α and Ksat values, suggesting slower groundwater responses and greater surface runoff. The mean annual precipitation across the watersheds (1369–1677 mm year⁻¹) did not fully explain the observed hydrological differences. This suggests that factors like soil permeability and vegetation cover have a greater influence on how surface runoff is partitioned than rainfall volume alone. For instance, both Paraíba and Madre de Deus watersheds received comparable total annual rainfall (~1500 mm), yet their lower BF and Ksat values were mainly due to degraded soils and limited vegetation cover. Overall, the hydrological indicators in Table 3 and the relationships depicted in Figure 4 confirm the effectiveness of the Ksat map derived from the SVM model. The consistent patterns linking Ksat, BF/TR, OF/TR, and α across land use types indicate that the predicted Ksat distribution accurately represents large-scale trends in soil permeability gradients and their hydrological effects. Forest landscapes, with higher inputs of organic matter and well-structured soils, promote more infiltration and contribute to baseflow, whereas pastures and rocky regions exhibit reduced permeability and greater surface runoff. Therefore, the Ksat map not only aligns statistically with the observed hydrological patterns but also reflects the eco-hydrological functioning. conclusIONS This study developed and evaluated a transferable machine learning–based framework for the spatial prediction of saturated soil hydraulic conductivity (Ksat) in headwater basins, using the Upper Grande River Basin as a representative case study. The results demonstrate that the combination of robust data-processing strategies, nonlinear learning algorithms, and spatially explicit prediction workflows is essential to address the pronounced variability and skewed distribution typically associated with Ksat. Among the evaluated algorithms, the Support Vector Machine model exhibited the most consistent performance across the tested scenarios, particularly when trained on the outlier-filtered dataset. Although the explained variance remained moderate, the reduction in prediction errors and the spatial coherence of the predicted Ksat maps indicate that the proposed approach is capable of capturing meaningful patterns of soil hydraulic behavior at the basin scale. The analysis of predictor importance highlighted the dominant influence of geomorphological context and curvature-related terrain attributes on Ksat spatial variability. These findings reinforce the role of landform organization in regulating soil development and water redistribution processes in headwater landscapes, supporting a physically consistent interpretation of the modeled results. Independent functional validation using hydrological indicators provided strong evidence of the physical realism of the predicted Ksat maps. The significant positive relationship between mean predicted Ksat and the baseflow-to-total runoff ratio across multiple sub-basins confirms that areas with higher modeled conductivity contribute more effectively to groundwater recharge and sustained streamflow. Overall, the proposed framework offers a robust and transferable solution for integrating soil hydraulic properties into distributed hydrological modeling and land-use planning. By explicitly linking digital soil mapping with functional hydrological validation, this study advances the application of machine learning in soil–hydrology research and provides a practical methodology for headwater regions where direct measurements of Ksat are limited. Acknowledgements We gratefully acknowledge the financial support of the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) 175060/2023-1, and Water Resources Department at Federal University of Lavras (UFLA). ReferencES Barnes, B.S., 1939. The structure of discharge recession curves. Eos, Trans. Am. Geophys. Union 20 (4), 721–725. Beven, K.J., Kirkby, M.J., 1979. A physically based, variable contributing area model of basin hydrology. Hydrol. Sci. Bull. 24 (1), 43–69. Blanco-Canqui, H., Gantzer, C.J., Anderson, S.H., Alberts, E.E., Ghidey, F., 2002. Saturated hydraulic conductivity and its impact on simulated runoff for claypan soils. Soil Sci. Soc. Am. J. 66 (5), 1478–1485. Blanchy, G., Albrecht, L., Bragato, G., Garré, S., Jarvis, N., Koestel, J., 2023. Impacts of soil management and climate on saturated and near-saturated hydraulic conductivity: analyses of the Open Tension-disk Infiltrometer Meta-database (OTIM). Hydrol. Earth Syst. Sci. 27 (14), 2703–2724. Bonfatti, B.R., Hartemink, A.E., Vanwalleghem, T., Minasny, B., Giasson, E., 2018. A mechanistic model to predict soil thickness in a valley area of Rio Grande do Sul, Brazil. Geoderma 309, 17–31. Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32. Brutsaert, W., 2023. Hydrology. Cambridge University Press, Cambridge. Campling, P., Gobin, A., Beven, K., Feyen, J., 2002. Rainfall-runoff modelling of a humid tropical catchment: the TOPMODEL approach. Hydrol. Processes 16 (1), 231–253. https://doi.org/10.1002/hyp.341 Centeri, C., 2022. Effects of grazing on water erosion, compaction and infiltration on grasslands. Hydrology 9 (2), 34. Chandler, K.R., Stevens, C.J., Binley, A., Keith, A.M., 2018. Influence of tree species and forest land use on soil hydraulic conductivity and implications for surface runoff generation. Geoderma 310, 120–127. Chen, S., Arrouays, D., Mulder, V.L., Poggio, L., Minasny, B., Roudier, P., Walter, C., 2022. Digital mapping of GlobalSoilMap soil properties at a broad scale: A review. Geoderma 409, 115567. Chen, T., Guestrin, C., 2016. XGBoost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, New York, pp. 785–794. Costabile, P., Costanzo, C., Gandolfi, C., Gangi, F., Masseroni, D., 2022. Effects of DEM depression filling on river drainage patterns and surface runoff generated by 2D rain-on-grid scenarios. Water 14 (7), 997. Costa, A.G., Soares, M.F., Pinto, M.A.B., Fuentes-Guevara, M.D., Nunes, R.F., Kaiser, D.R., …, Timm, L.C., 2024. Exploring the Soil Structure Analysis of a Subtropical Watershed by Introducing a Blanket Physico-Hydraulic Health Index. Land Degrad. Dev. 35 (18), 5713–5728. Crovo, O., Aburto, F., da Costa-Reidel, C., Montecino, F., Rodríguez, R., 2021. Effects of livestock grazing on soil health and recovery of a degraded Andean Araucaria forest. Land Degrad. Dev. 32 (17), 4907–4919. Deb, S.K., Shukla, M.K., 2012. Variability of hydraulic conductivity due to multiple factors. Am. J. Environ. Sci. 8 (5), 489–502. Deka, P.C., 2014. Support vector machine applications in the field of hydrology: a review. Appl. Soft Comput. 19, 372–386. Durães, M.F., Mello, C.R., 2013. Groundwater recharge behavior based on surface runoff hydrographs in two basins of the Minas Gerais State. Ambi-Agua 8 (1), 57–66. Elbisy, M.S., 2025. Predictive Modeling of Saturated Hydraulic Conductivity using Machine Learning Techniques. Eng. Technol. Appl. Sci. Res. 15 (2), 21348–21355. Friedman, J., Hastie, T., Tibshirani, R., 2010. Regularization Paths for Generalized Linear Models via Coordinate Descent. J. Stat. Softw. 33 (1), 1–22. Gabrecht, J., Martz, L.W., 1999. Digital elevation model issues in water resources modeling. In: Proceedings of the 19th ESRI International User Conference, San Diego, California, pp. 254–257. Gallant, J.C., Dowling, T.I., 2003. A multiresolution index of valley bottom flatness for mapping depositional areas. Water Resour. Res. 39 (12), 1347–1359. Gallant, J.C., Wilson, J.P., 2000. Primary topographic attributes. In: Wilson, J.P., Gallant, J.C. (Eds.), Terrain Analysis: Principles and Applications. John Wiley & Sons, New York, pp. 51–85. Germer, S., Neil, C., Krusche, A.V., Elsenbeer, H., 2010. Influence of land-use change on near-surface hydrological processes: Undisturbed forest to pasture. J. Hydrol. 380 (3–4), 473–480. Ghavami, M.S., Ayoubi, S., Mosaddeghi, M.R., Naimi, S., 2023. Digital mapping of soil physical and mechanical properties using machine learning at the watershed scale. J. Mt. Sci. 20 (10), 2975–2992. Golden, H.E., Christensen, J.R., McMillan, H.K., Kelleher, C.A., Lane, C.R., Husic, A., …, Cheng, F., 2025. Advancing the science of headwater streamflow for global water protection. Nat. Water 3 (1), 16–26. Guisan, A., Weiss, S.B., Weiss, A.D., 1999. GLM versus CCA spatial modeling of plant species distribution. Plant Ecol. 143 (1), 107–122. Gupta, S., Lehmann, P., Bonetti, S., Papritz, A., Or, D., 2021. Global prediction of soil saturated hydraulic conductivity using random forest in a covariate‐based geoTransfer function (CoGTF) framework. J. Adv. Model. Earth Syst. 13 (4), e2020MS002242. He, J., Mattis, S.A., Butler, T.D., Dawson, C.N., 2019. Data-driven uncertainty quantification for predictive flow and transport modeling using support vector machines. Comput. Geosci. 23 (4), 631–645. Hervé-Fernández, P., Muñoz-Arriagada, R., Glucevic-Almonacid, C., Bahamonde-Vidal, L., Radic-Schilling, S., 2023. Influence of rangeland land cover on infiltration rates, field-saturated hydraulic conductivity, and soil water repellency in southern Patagonia. Rangel. Ecol. Manag. 90, 92–100. Hijmans, R.J., 2022. terra: Spatial Data Analysis. R package version 1.5-21. https://cran.r-project.org/package=terra Hingray, B., Picouet, C., Musy, A., 2014. Hydrology: a Science for Engineers. CRC Press, Taylor & Francis Group, Boca Raton, FL. Hubert, M., Debruyne, M., Rousseeuw, P.J., 2018. Minimum covariance determinant and extensions. Wiley Interdiscip. Rev. Comput. Stat. 10 (3), e1421. Iwahashi, J., Pike, R.J., 2007. Automated classifications of topography from DEMs by an unsupervised nested-means algorithm and a three-part geometric signature. Geomorphology 86 (3–4), 409–440. Jasiewicz, J., Stepinski, T., 2013. Geomorphons: a pattern recognition approach to classification and mapping of landforms. Geomorphology 182, 147–156. Jena, R.K., Bandyopadhyay, S., Pradhan, U.K., Moharana, P.C., Kumar, N., Sharma, G.K., …, Hossain, A., 2022. Geospatial modelling for delineation of crop management zones using local terrain attributes and soil properties. Remote Sens. 14 (9), 2101. Kuhn, M., 2015. caret: classification and regression training. Astrophys. Source Code Libr., ascl-1505. Lange, H., Sippel, S., 2020. Machine learning applications in hydrology. In: Forest-Water Interactions. Springer International Publishing, Cham, pp. 233–257. Leimer, S., Berner, D., Birkhofer, K., Boeddinghaus, R.S., Fischer, M., Kandeler, E., Wilcke, W., 2021. Land-use intensity and biodiversity effects on infiltration capacity and hydraulic conductivity of grassland soils in southern Germany. Ecohydrology 14 (6), e2301. Lin, H., 2010. Earth’s Critical Zone and hydropedology: concepts, characteristics, and advances. Hydrol. Earth Syst. Sci. 14 (1), 25–45. Lin, H., Brooks, E., McDaniel, P., Boll, J., 2008. Hydropedology and surface/subsurface runoff processes. In: Lin, H. (Ed.), Hydropedology: Synergistic Integration of Soil Science and Hydrology. Elsevier, Amsterdam, pp. 317–335. Liu, F., Wang, P., Zhou, Z., Liu, J.E., Wang, N., 2025. Effects of root traits of different diameter classes on soil infiltration in naturally restored grasslands on the northern Loess Plateau. Plant Soil. https://doi.org/10.1007/s11104-025-06860-4 (in press). Ma, Y., Minasny, B., Malone, B.P., McBratney, A.B., 2019. Pedology and digital soil mapping (DSM). Eur. J. Soil Sci. 70 (2), 216–235. Maldaner, L.F., Molin, J.P., Spekken, M., 2021. Methodology to filter out outliers in high spatial density data to improve maps reliability. Sci. Agric. 79, e20200178. Marín-Pimentel, G.E., Rueda-Saa, G., Menjivar-Flores, J.C., 2023. Evaluation of physicochemical properties in agricultural soils on the flat and piedmont areas of Valle del Cauca, Colombia with emphasis on degradation. Environ. Earth Sci. 82 (7), 157. McMillan, H., 2020. Linking hydrologic signatures to hydrologic processes: A review. Hydrol. Processes 34 (6), 1393–1409. Mello, C.R., Guzman, J.A., Vieira, N.P., Viola, M.R., Beskow, S., Guo, L., …, Rodrigues, A.F., 2025. Mitigating severe hydrological droughts in the Brazilian tropical high-land region: A novel land use strategy under climate change. Int. Soil Water Conserv. Res. (in press). Menezes, M.D., Silva, S.H.G., Mello, C.R., Owens, P.R., Curi, N., 2016. Spatial prediction of soil properties in two contrasting physiographic regions in Brazil. Sci. Agric. 73 (3), 274–285. Moustafa, M.M., 2000. A geostatistical approach to optimize the determination of saturated hydraulic conductivity for large-scale subsurface drainage design in Egypt. Agric. Water Manag. 42 (3), 291–312. Mozaffari, H., Moosavi, A.A., Sepaskhah, A.R., 2021. Land use-dependent variation of near-saturated and saturated hydraulic properties in calcareous soils. Environ. Earth Sci. 80 (23), 769. Mulder, V.L., de Bruin, S., Schaepman, M.E., Mayr, T.R., 2011. The use of remote sensing in soil and terrain mapping – a review. Geoderma 162 (1–2), 1–19. Oliveira, V., Mello, C., Viola, M., Srinivasan, R., 2017. Assessment of climate change impacts on streamflow and hydropower potential in the headwater region of the Grande river basin, Southeastern Brazil. Int. J. Climatol. 37 (14), 5005–5023. Oliveira, V.A.D., Mello, C.R.D., Viola, M.R., Srinivasan, R., 2018. Land-use change impacts on the hydrology of the upper Grande river basin, Brazil. Cerne 24 (4), 334–343. Owuor, S.O., Butterbach-Bahl, K., Guzha, A.C., Jacobs, S., Merbold, L., Rufino, M.C., Pelster, D., Díaz-Pinés, E., Breuer, L., 2018. Conversion of natural forest results in a significant degradation of soil hydraulic properties in the highlands of Kenya. Soil Tillage Res. 176, 36–44. Pham, K., Won, J., 2022. Enhancing the tree-boosting-based pedotransfer function for saturated hydraulic conductivity using data preprocessing and predictor importance using game theory. Geoderma 420, 115864. Pinto, L.C., Mello, C.R.D., Ferreira, D.F., vila, L.F., 2013. Water quality index in two land use situations in the Mantiqueira Range. Ciênc. Agrotec. 37, 338–342. Pinto, L.C., Mello, C.R., Norton, L.D., Owens, P.R., Curi, N., 2015. Role of Inceptisols in the hydrology of mountainous catchments in Southeastern Brazil. J. Hydrol. Eng. 21 (8), 05015007. Pinto, L.C., Mello, C.R., Norton, L.D., Owens, P.R., Curi, N., 2016. Spatial prediction of soil-water transmissivity based on fuzzy logic in a Brazilian headwater watershed. Catena 143, 26–34. Pinto, L.C., Mello, C.R., Norton, L.D., Silva, S.H.G., Taveira, L.R.S., Curi, N., 2017. Land-use effect on hydropedology in a mountainous region of Southeastern Brazil. Ciênc. Agrotec. 41 (4), 413–427. Pinto, L.C., Mello, C.R., Norton, L.D., Poggere, G.C., Owens, P.R., Curi, N., 2018. A hydropedological approach to a mountainous Clayey Humic Dystrudept in the Mantiqueira Range, southeastern Brazil. Sci. Agric. 75 (1), 60–69. Pinto, L.C., Chagas, W.F.T., Amaral, F.H.C., 2019a. Physical attributes of a Dystroferric Red Latosol (Oxisol) under different management systems. Rev. Agrogeoambiental 11 (1), 111–119. Pinto, L.C., Mello, C.R.D., Norton, L.D., Curi, N., 2019b. Land-use influence on the soil hydrology: An approach in upper Grande River basin, Southeast Brazil. Ciênc. Agrotec. 43, e015619. Price, K., Jackson, C.R., Parker, A.J., 2010. Variation of surficial soil hydraulic properties across land uses in the southern Blue Ridge Mountains, North Carolina, USA. J. Hydrol. 383 (3–4), 256–268. Ribeiro, C.L., de Mello, C.R., Guzman, J.A., 2024. Groundwater storage trend in headwater basins: shreds of evidence from the last decades in Minas Gerais state, Brazil. Environ. Earth Sci. 83 (10), 327. Rousseeuw, P.J., Leroy, A.M., 1987. Robust Regression and Outlier Detection. John Wiley & Sons, New York. Sarah, S., Shah, W., Somers, L.D., Deshpande, R.D., Ahmed, S., 2024. Saturated hydraulic conductivity (Ksat) and topographic controls on baseflow contribution in high-altitude aquifers with complex geology. J. Hydrol. 641, 131763. Seguel, O., Díaz, D., Acevedo, E., Silva, P., Homer, I., Seitz, S., 2020. Hydraulic conductivity in a soil cultivated with wheat-rapeseed rotation under two tillage systems. J. Soil Sci. Plant Nutr. 20 (4), 2304–2315. Silva, S.H., Weindorf, D.C., Faria, W.M., Pinto, L.C., Menezes, M.D., Guilherme, L.R., Curi, N., 2021. Proximal sensor-enhanced soil mapping in complex soil-landscape areas of Brazil. Pedosphere 31 (4), 615–626. Silva, S.H.G., de Menezes, M.D., Owens, P.R., Curi, N., 2016. Retrieving pedologist’s mental model from existing soil map and comparing data mining tools for refining a larger area map under similar environmental conditions in Southeastern Brazil. Geoderma 267, 65–77. Silva, V.P.R., Silva, M.T., Singh, V.P., Souza, E.P., Braga, C.C., Holanda, R.M., Almeida, R.S.R., Sousa, F.A.S., Braga, A.C.R., 2018. Simulation of stream flow and hydrological response to land-cover changes in a tropical river basin. Catena 162, 166–176. Singh, N., 2025. Perenniality Impacts on Soil Physical and Hydraulic Properties and Ecosystem Services: A Review. Sustainability 17 (24), 10988. Sobieraj, J.A., Elsenbeer, H., Coelho, R.M., Newton, B., 2002. Spatial variability of soil hydraulic conductivity along a tropical rainforest catena. Geoderma 108 (1–2), 79–90. Soulsby, C., Tetzlaff, D., Rodgers, P., Dunn, S., Waldron, S., 2006. Runoff processes, stream water residence times and controlling landscape characteristics in a mesoscale catchment: An initial evaluation. J. Hydrol. 325 (1–4), 197–221. Soylu, M.E., Bras, R.L., 2022. Global Shallow Groundwater Patterns From Soil Moisture Satellite Retrievals. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 15, 89–98. Tallaksen, L.M., 1995. A review of baseflow recession analysis. J. Hydrol. 165 (1–4), 349–370. Vapnik, V., 2013. The Nature of Statistical Learning Theory. Springer Science & Business Media. Vaysse, K., Lagacherie, P., 2015. Evaluating Digital Soil Mapping approaches for mapping GlobalSoilMap soil properties from legacy data in Languedoc-Roussillon (France). Geoderma Reg. 4, 20–30. Veloso, M.F., Rodrigues, L.N., Fernandes Filho, E.I., 2022. Evaluation of machine learning algorithms in the prediction of hydraulic conductivity and soil moisture at the Brazilian Savannah. Geoderma Reg. 30, e00569. Vieira, C.B., Silva, G.H.M.C., Almeida, B.G.D., Pessoa, L.G.M., Freire, F.J., de Souza Junior, V.S., …, Freire, M.B.G.D.S., 2025. Saturated Hydraulic Conductivity of Nine Soils According to Water Quality, Soil Texture, and Clay Mineralogy. Agronomy 15 (4), 864. Viola, M.R., Mello, C.R., Beskow, S., Norton, L.D., 2014. Impacts of land-use changes on the hydrology of the Grande River basin headwaters, Southeastern Brazil. Water Resour. Manag. 28 (13), 4537–4550. Wei, C., Dong, X., Ma, Y., Zhao, W., Yu, D., Tayyab, M., Bo, H., 2022. Impacts of land use types, soil properties, and topography on baseflow recharge and prediction in an agricultural watershed. Land 12 (1), 109. Willmott, C.J., Matsuura, K., 2005. Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance. Clim. Res. 30 (1), 79–82. Wu, S., Li, J., Huang, G.H., 2008. A study on DEM-derived primary topographic attributes for hydrologic applications: Sensitivity to elevation data resolution. Appl. Geogr. 28 (3), 210–223. Yeo, I.-K., Johnson, R.A., 2000. A new family of power transformations to improve normality or symmetry. Biometrika 87 (4), 954–959. Zimmermann, B., Elsenbeer, H., 2008. Spatial and temporal variability of soil saturated hydraulic conductivity in gradients of disturbance. J. Hydrol. 361 (1–2), 78–95. Image (Figure 1.png) is missing or otherwise invalid. Image (Figure 2.png) is missing or otherwise invalid. Image (Figure 4.png) is missing or otherwise invalid. Image (Figure_3.png) is missing or otherwise invalid. Information & Authors Information Version history V1 Version 1 31 December 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords digital soil mapping headwater basins hydrological validation machine learning saturated hydraulic conductivity terrain attributes Authors Affiliations Leandro Campos Pinto 0000-0003-2071-7028 [email protected] Federal University of Lavras View all articles by this author Carlos Mello Federal University of Lavras View all articles by this author Metrics & Citations Metrics Article Usage 211 views 120 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Leandro Campos Pinto, Carlos Mello. Spatial Prediction of Saturated Soil Hydraulic Conductivity in Headwater Basins Using Machine Learning: A Transferable Methodological Framework. Authorea . 31 December 2025. DOI: https://doi.org/10.22541/au.176717717.70038547/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text {"doi":"10.22541/au.176717717.70038547/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fee7dcadf59df94',t:'MTc3OTMxMzEzOA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();