Tropical forests are significant carbon sinks and their soils’ carbon storage potential is immense. However, little is known about the soil organic carbon (SOC) stocks of tropical mountain areas whose complex soil-landscape and difficult accessibility pose a challenge to spatial analysis. The choice of methodology for spatial prediction is of high importance to improve the expected poor model results in case of low predictor-response correlations. Four aspects were considered to improve model performance in predicting SOC stocks of the organic layer of a tropical mountain forest landscape: Different spatial predictor settings, predictor selection strategies, various machine learning algorithms and model tuning. Five machine learning algorithms: random forests, artificial neural networks, multivariate adaptive regression splines, boosted regression trees and support vector machines were trained and tuned to predict SOC stocks from predictors derived from a digital elevation model and satellite image. Topographical predictors were calculated with a GIS search radius of 45 to 615 m. Finally, three predictor selection strategies were applied to the total set of 236 predictors. All machine learning algorithms—including the model tuning and predictor selection—were compared via five repetitions of a tenfold cross-validation. The boosted regression tree algorithm resulted in the overall best model. SOC stocks ranged between 0.2 to 17.7 kg m^-2, displaying a huge variability with diffuse insolation and curvatures of different scale guiding the spatial pattern. Predictor selection and model tuning improved the models’ predictive performance in all five machine learning algorithms. The rather low number of selected predictors favours forward compared to backward selection procedures. Choosing predictors due to their indiviual performance was vanquished by the two procedures which accounted for predictor interaction.

As the largest terrestrial carbon pool, soil organic carbon (SOC) has the potential to influence and mitigate climate change; thus, SOC monitoring is of high importance in the frameworks of various international treaties. Therefore, high-resolution SOC maps are required. Machine learning (ML) offers new opportunities to develop these maps due to its ability to data mine large datasets. The aim of this study was to apply three algorithms commonly used in digital soil mapping – random forest (RF), boosted regression trees (BRT), and support vector machine for regression (SVR) – on the first German agricultural soil inventory to model the agricultural topsoil (0–30 cm) SOC content and develop a two-model approach to address the high variability in SOC in German agricultural soils. Model performance is often limited by the size and quality of the soil dataset available for calibration and validation. Therefore, the impact of enlarging the training dataset was tested by including data from the European Land Use/Cover Area frame Survey for agricultural sites in Germany. Nested cross-validation was implemented for model evaluation and parameter tuning. Grid search and the differential evolution algorithm were also applied to ensure that each algorithm was appropriately tuned . The SOC content of the German agricultural soil inventory was highly variable, ranging from 4 to 480 g kg⁻¹. However, only 4 % of all soils contained more than 87 g kg⁻¹ SOC and were considered organic or degraded organic soils. The results showed that SVR produced the best performance, with a root-mean-square error (RMSE) of 32 g kg⁻¹ when the algorithms were trained on the full dataset. However, the average RMSE of all algorithms decreased by 34 % when mineral and organic soils were modelled separately, with the best result from SVR presenting an RMSE of 21 g kg⁻¹. The model performance was enhanced by up to 1 % for mineral soils and by up to 2 % for organic soils. Despite the ability of machine learning algorithms, in general, and SVR, in particular, to model SOC on a national scale, the study showed that the most important aspect for improving the model performance was to separate the modelling of mineral and organic soils.

Soil organic carbon (SOC) plays a major role concerning chemical, physical, and biological soil properties and functions. To get a better understanding of how soil management affects the SOC content, the precise monitoring of SOC on long-term field experiments (LTFEs) is needed. Visible and near-infrared (Vis–NIR) reflectance spectrometry provides an inexpensive and fast opportunity to complement conventional SOC analysis and has often been used to predict SOC. For this study, 100 soil samples were collected at an LTFE in central Germany by two different sampling designs. SOC values ranged between 1.5 % and 2.9 %. Regression models were built using partial least square regression (PLSR). In order to build robust models, a nested repeated 5-fold group cross-validation (CV) approach was used, which comprised model tuning and evaluation. Various aspects that influence the obtained error measure were analysed and discussed. Four pre-processing methods were compared in order to extract information regarding SOC from the spectra. Finally, the best model performance which did not consider error propagation corresponded to a mean RMSE_MV of 0.12 % SOC (R²=0.86). This model performance was impaired by ΔRMSE_MV=0.04 % SOC while considering input data uncertainties (ΔR²=0.09), and by ΔRMSE_MV=0.12 % SOC (ΔR²=0.17) considering an inappropriate pre-processing. The effect of the sampling design amounted to a ΔRMSE_MV of 0.02 % SOC (ΔR²=0.05). Overall, we emphasize the necessity of transparent and precise documentation of the measurement protocol, the model building, and validation procedure in order to assess model performance in a comprehensive way and allow for a comparison between publications. The consideration of uncertainty propagation is essential when applying Vis–NIR spectrometry for soil monitoring.

Societal demands on soil functionality in agricultural soil-landscapes are confronted with yield losses and environmental impact. Soil functional information at national scale is required to address these challenges. On behalf of the well-known theory that soils and their site-specific characteristics are the product of the interaction of the soil-forming factors, pedometricians seek to model the soil-landscape relationship using machine learning. Following the rationale that similarity in soils is reflected by similarity in landscape characteristics, we defined soil functional types (SFTs) which were projected into space by machine learning. Each SFT is described by a multivariate soil parameter distribution along its depth profile. SFTs were derived by employing multivariate similarity analysis on the dataset of the Agricultural Soil Inventory. Soil profiles were compared on behalf of differing sets of soil properties considering the top 100 and 200 cm, respectively. Various depth weighting coefficients were tested to attribute topsoil properties higher importance. Support vector machine (SVM) models were then trained employing optimization with a distributed multiple-population hybrid Genetic algorithm for parameter tuning. Model training, tuning, and evaluation were implemented in a nested k-fold cross-validation approach to avoid overfitting. With regards to the SFTs, organic soils were differentiated from mineral soils of various particle size distributions being partly influenced by waterlogging and groundwater. Further SFTs reflect soils with a depth limitation within the top 100 cm and high stone content. Altogether, with SVM predictive model accuracies between 0.7 and 0.9, the agricultural soil-landscape of Germany was represented with eight SFTs. Soil functionality with regards to the soil’s capacity to store plant-available water and soil organic carbon is well characterized. Four additional soil functions are described to a certain extent. An extension of the approach to fully cover soil functions such as nutrient cycling, agricultural biomass production, filtering of contaminants, and soil as a habitat for soil biota is possible with the inclusion of additional soil properties. Altogether, the developed data product represents the 3D multivariate soil parameter space. Its agglomerated simplicity into a limited number of spatially allocated process units provides the basis to run agricultural process models at national scale (Germany).

Machine-learning algorithms are good at computing non-linear problems and fitting complex composite functions, which makes them an adequate tool for addressing multiple environmental research questions. One important application is the development of pedotransfer functions (PTFs). This study aims to develop water retention PTFs for two remote tropical mountain regions with rather different soil landscapes: (1) those dominated by peat soils and soils under volcanic influence with high organic matter contents and (2) those dominated by tropical mineral soils. Two tuning procedures were compared to fit boosted regression tree models: (1) tuning with grid search, which is the standard approach in pedometrics; and (2) tuning with differential evolution optimization. A nested cross-validation approach was applied to generate robust models. The area-specific PTFs developed outperform other more general PTFs. Furthermore, the first PTF for typical soils of Páramo landscapes (Ecuador), i.e., organic soils under volcanic influence, is presented. Overall, the results confirmed the differential evolution algorithm's high potential for tuning machine-learning models. While models based on tuning with grid search roughly predicted the response variables' mean for both areas, models applying the differential evolution algorithm for parameter tuning explained up to 25 times more of the response variables' variance.