Search

The Soil Land Inference Model (SoLIM) was primarily proposed by Zhu et al. (Zhu A X, Band L, Vertessy R, Dutton B. 1997. Derivation of soil properties using a soil land inference model (SoLIM). Soil Sci Soc Am J. 61: 523–533.) and was based on the Third Law of Geography. Based on the assumption that the soil property value at a location of interest will be more similar to that of a given soil sample when the environmental condition at the location of interest is more similar to that at the location from which the sample was taken, SoLIM estimates the soil property value of the location of interest using the soil property values of known samples weighted by the similarity between those samples and the location of interest in terms of an attribute domain of environmental conditions. However, the current SoLIM method ignores information about the spatial distances between the location of interest and those of the sample. In this study, we proposed a new method of soil property mapping, SoLIM-IDW, which incorporates spatial distance information into the SoLIM method by means of inverse distance weighting (IDW). The proposed method is based on the assumption that the soil property value at a location of interest will be more similar to that of a known sample both when the environmental conditions are more similar and when the distance between the location of interest and the sample location is shorter. Our evaluation experiments on A-horizon soil organic matter mapping in two study areas with independent evaluation samples showed that the proposed SoLIM-IDW method can obtain lower prediction errors than the original SoLIM method, multiple linear regression, geographically weighted regression, and regression-kriging with the same modeling points. Future work mainly includes the determination of optimal power parameter values and the appropriate setting of the parameter under different application contexts.

Portable X-ray fluorescence (pXRF) spectrometry and magnetic susceptibility (MS) via magnetometer have been increasingly used with terrain variables for digital soil mapping. However, this methodology is still emerging in many countries with tropical soils. The objective of this study was to use proximal soil sensor data associated with terrain variables at varying spatial resolutions to predict soil classes using the Random Forest (RF) algorithm. The study was conducted on a 316-ha area featuring highly variable soil classes and complex soil-landscape relationships in Minas Gerais State, Brazil. The overall accuracy and Kappa index were evaluated using soils that were classified at 118 sites, with 90 being used for modeling and 28 for validation. Digital elevation models (DEMs) were created at 5-, 10-, 20-, and 30-m resolutions using contour lines from two sources. The resulting DEMs were processed to generate 12 terrain variables. Total Fe, Ti, and SiO2 contents were obtained using pXRF, with MS determined via a magnetometer. Soil class prediction was performed using the RF algorithm. The quality of the soil maps improved when using only the five most important covariates and combining proximal sensor data with terrain variables at different spatial resolutions. The finest spatial resolution did not always provide the most accurate maps. The high soil complexity in the area prevented highly accurate predictions. The most important variables influencing the soil mapping were MS, Fe, and Ti. Proximal sensor data associated with terrain information were successfully used to map Brazilian soils at variable spatial resolutions.

The overview of publications on soil cartography in “Pochvovedenie”/“Eurasian Soil Science” journal for the period 1899–2020 demonstrates a high diversity of themes and certain trends in the number and dynamics of papers. Their total number (365), calculated per 5-year-long intervals, was distributed rather evenly among these 121 years: approximately 10–15 papers in each interval, although three maximums are rather clear. The first one fell on the post-war interval and was followed by 1965–1970 and 2010–2015 maximums. Discussion of large-scale maps dominated the early publications, many of them tackled soil surveys and applied problems; in the mid-century papers, soil maps of various regions of the country were described since it was time of extensive terrain investigations; numerous were also papers concerning methodology of soil mapping. New approaches and technique were actively discussed in the papers at the turn of centuries, such as remote sensing or digital soil mapping. Along with map compilation issues, there are publications on applying information provided by soil maps for both traditional and novel purposes: schemes of zoning in the former case and development of prognostic maps or assessment of pedodiversity in the latter case. The majority of papers on zoning, concern soil-geographical (later soil-ecological) schemes, whereas the derived types of zoning, for example, ameliorative or erosional, are discussed in few papers. The performed overview may be regarded as summing up the results of traditional soil mapping development with emphasizing its most valuable achievements, as well as indicating the initial signs of new trends.

Precision agriculture provides detailed information on the spatial variability of soil properties, including nutrient content, allowing for local-specific decision making. Recently, proximal sensors have been used to accurately predict soil properties, contributing to reduce costs of conventional wet-chemistry analyses for soil characterization. However, further investigations on this approach in tropical soils are needed. This work aimed to use portable X-ray fluorescence (pXRF) spectrometry data for prediction of exchangeable Ca2+ and available K+ and P contents in soils of a highly heterogeneous tropical area and evaluating its practical applications. 90 samples from soil A horizon were collected in a regular grid design, and analyzed through pXRF and for nutrient contents. Such data were split into modeling (63 samples) and validation (27 samples) datasets. Linear regression (LR), polynomial regression (PR), power regression (PwR) and stepwise multiple linear regression (SMLR) were tested for predictions. The models were used to spatially represent nutrient contents across the area and to compare the practical effects of varying regression models. PXRF elemental data provided reliable predictions of exchangeable Ca2+ and available P via SMLR and PwR, respectively, reaching root mean square errors (RMSE) of 5.66 cmolc dm−3 for Ca2+ and 9.13 mg dm−3 for P. Available K+ predictions were not successful. Different models yielded contrasting maps showing the classes of soil fertility across the area, drawing attention to the importance of testing multiple prediction models and using the best one for precision agriculture. Fusion of data from different proximal sensors may enhance available K+ predictions.

The soil organic carbon (SOC) pool is larger than the biotic and atmospheric carbon pools, and understanding cultivated land SOC is critical for evaluating soil fertility and local agricultural man...

Accurate soil information is critically important for forest management planning and operations but is challenging to map. Digital soil mapping (DSM) improves upon the limitations of conventional soil mapping by explicitly linking a variety of environmental data layers to spatial soil point data sets to continuously predict soil variability across a landscape. Thus far, much DSM research has focussed on the development of ultrafine-resolution soil maps within agricultural systems; however, increasing availability of light detection and ranging (LiDAR) data presents new opportunities to apply DSM to support forest resource applications at multiple scales. This project describes a DSM workflow using LiDAR-derived elevation data and machine learning models (MLMs) to predict key forest soil attributes. A case study in the Hearst Forest in northeastern Ontario, Canada, is used to illustrate the workflow. We applied multiple MLMs to the Hearst Forest to predict soil moisture regime and textural class. Both qualitative and quantitative assessment pointed to the random forest MLM producing the best maps (63% accuracy for moisture regime and 66% accuracy for textural class). Where error occurred, soils were typically misclassified to neighbouring classes. This standardized, flexible workflow is a valuable tool for practitioners that want to undertake DSM as part of forest resource management and planning.

We performed a systematic mapping of validation methods used in digital soil mapping (DSM), in order to gain an overview of current practices and make recommendations for future publications on DSM studies. A systematic search and screening procedure, largely following the RepOrting standards for Systematic Evidence Syntheses (ROSES) protocol, was carried out. It yielded a database of 188 peer-reviewed DSM studies from the past two decades, all written in English and all presenting a raster map of a continuous soil property. Review of the full-texts showed that most publications (97%) included some type of map validation, while just over one-third (35%) estimated map uncertainty. Most commonly, a combination of multiple (existing) soil sampe datasets was used and the resulting maps were validated by single data-splitting or cross-validation. It was common for essential information to be lacking in method descriptions. This is unfortunate, as lack of information on sampling design (missing in 25% of 188 studies) and sample support (missing in 45% of 188 studies) makes it difficult to interpret what derived validation metrics represent, compromising their usefulness. Therefore, we present a list of method details that should be provided in DSM studies. We also provide a detailed summary of the 28 validation metrics used in published DSM studies, how to interpret the values obtained and whether the metrics can be compared between datasets or soil attributes.

Soil degradation resulting from water erosion poses a serious threat to food and environmental security, therefore the research of soil erosion features and soil erosion mapping do not lose their relevance. The paper presents the results of large-scale digital mapping of the erosion degree of the arable soils in the Prokhorovsky district of the Belgorod region (85 thousand hectares), based on two approaches: (1) linking the factors of erosion-accumulative processes and the erosion degree of soil directly (factor -property model), and (2) due to imitation erosion model WaTEM/SEDEM (factor - process - property model). The inclusion of the process component into the digital soil mapping algorithm allows taking into account not only the spatial but also the temporal soil erosion features. It was revealed that the agricultural development of the Prokhorovsky district was primarily carried out on lands that are weakly prone to erosion, with the rate of erosion almost two times lower than on younger arable lands. As a result, the soil erosion maps, based on the factor - process - property model, with and without taking into account the duration of agricultural use, largely correspond to each other. Dominant soil categories (the map pixel corresponds to one soil taxa - noneroded and slightly eroded, medium, highly eroded), mapping by factor -property and factor - process - property models, have a high degree of correspondence to each other (prediction identity for 90% of pixels), while the soil combinations (the map pixel has information on the proportion of soils with different erosion degrees of soil) more significant (identity for less than 60% of pixels). The areas of zonal, erosion-zonal, and weakly eroded soil combinations differ 1.5-2 times, in the direction of a greater degree of soil erosion on the factor - process - property map.

Selecting a proper set of covariates is one of the most important factors that influence the accuracy of digital soil mapping (DSM). The statistical or machine learning methods for selecting DSM covariates are not available for those situations with limited samples. To solve the problem, this paper proposed a case-based method which could formalize the covariate selection knowledge contained in practical DSM applications. The proposed method trained Random Forest (RF) classifiers with DSM cases extracted from the practical DSM applications and then used the trained classifiers to determine whether each one potential covariate should be used in a new DSM application. In this study, we took topographic covariates as examples of covariates and extracted 191 DSM cases from 56 peer-reviewed journal articles to evaluate the performance of the proposed case-based method by Leave-One-Out cross validation. Compared with a novices’ commonly-used way of selecting DSM covariates, the proposed case-based method improved more than 30% accuracy according to three quantitative evaluation indices (i.e., recall, precision, and F1-score). The proposed method could be also applied to selecting the proper set of covariates for other similar geographical modeling domains, such as landslide susceptibility mapping, and species distribution modeling.

Digital soil mapping (DSM) is an effective mapping technique that supports the increased need for quantitative soil data. In DSM, soil properties are correlated with environmental characteristics using statistical models such as regression. However, many of these relationships are explicitly described in mechanistic simulation models. Therefore, the mechanistic relationships can, in theory, replace the statistical relationships in DSM. This study aims to develop a mechanistic model to predict soil organic matter (SOM) stocks in Natura2000 areas of the Cantabria region (Spain). The mechanistic model is established in four steps: (a) identify major processes that influence SOM stocks, (b) review existing models describing the major processes and the respective environmental data that they require, (c) establish a database with the required input data, and (d) calibrate the model with field observations. The SOM stocks map resulting from the mechanistic model had a mean error (ME) of -2 t SOM ha−1 and a root mean square error (RMSE) of 66t SOM ha-1. The Lin's concordance correlation coefficient was 0.47 and the amount of variance explained (AVE) was 0.21. The results of the mechanistic model were compared to the results of a statistical model. It turned out that the correlation coefficient between the two SOM stock maps was 0.8. This study illustrated that mechanistic soil models can be used for DSM, which brings new opportunities. Mechanistic models for DSM should be considered for mapping soil characteristics that are difficult to predict by statistical models, and for extrapolation purposes. This research was financially supported by the Environmental Hydraulics Institute ‘IH Cantabria of Universidad de Cantabria’ and the CGIAR Research Programme on Climate Change, Agriculture and Food Security (CCAFS). The CCAFS project is carried out with support from CGIAR Fund Donors and through bilateral funding agreements. Besides the financial support, we would like to thank Sara Alcalde Aparicio for collaboration in the collection and analyses of soil samples.

Decision tree algorithms such as Random Forest have become a widely adapted method for mapping soil properties in geographic space. However, implementing explicit geographic relationships into these methods has proven problematic. Using x- and y-coordinates as covariates gives orthogonal artefacts in the maps, and alternative methods using distances as covariates can be inflexible and difficult to interpret. We propose instead the use of coordinates along several axes tilted at oblique angles to provide an easily interpretable method for obtaining a realistic prediction surface. We test the method for mapping topsoil organic matter contents in an agricultural field in Denmark. The results show that the method provides accuracies on par with the most reliable alternative methods, namely kriging and the use of buffer distances to the training points. Furthermore, the proposed method is highly flexible, scalable and easily interpretable. This makes it a promising tool for mapping soil properties with complex spatial variation. We believe that the method will be highly useful for mapping soil properties in larger areas, and testing it for this purpose is a logical next step.

In digital soil mapping, machine learning (ML) techniques are being used to infer a relationship between a soil property and the covariates. The information derived from this process is often translated into pedological knowledge. This mechanism is referred to as knowledge discovery. This study shows that knowledge discovery based on ML must be treated with caution. We show how pseudo-covariates can be used to accurately predict soil organic carbon in a hypothetical case study. We demonstrate that ML methods can find relevant patterns even when the covariates are meaningless and not related to soil-forming factors and processes. We argue that pattern recognition for prediction should not be equated with knowledge discovery. Knowledge discovery requires more than the recognition of patterns and successful prediction. It requires the pre-selection and preprocessing of pedologically relevant environmental covariates and the posterior interpretation and evaluation of the recognized patterns. We argue that important ML covariates could serve the purpose of providing elements to postulate hypotheses about soil processes that, once validated through experiments, could result in new pedological knowledge. Highlights: We discuss the rationale of knowledge discovery based on the most important machine learning covariates We use pseudo-covariates to predict topsoil organic carbon with random forest Soil organic carbon was accurately predicted in a hypothetical case study Pattern recognition by random forest should not be equated to knowledge discovery.

Process-based models of soil organic carbon (C) define soil organic matter by conceptual C pools. In the Roth C model these pools are represented by the particulate, humus and resistant organic C fractions (POC, HOC, and ROC). Here, we used three different sets of estimates of the C fractions to initialise Roth C and assessed their effect on the model predictions of the 0–0.3 m C stocks in a 30-year simulation. We estimated the stocks of POC, HOC and ROC with (1) pedotransfer functions (PTF), (2) proximal sensing, and (3) digital soil mapping with data from sensing. We conducted experiments in three C estimation areas of a cattle grazing farm in Australia. We found that in all cases, the model predicted increasing total organic C (TOC) stocks by up to 15.2 Mg C ha−1. Using the PTF estimates, the uncertainty in the model predictions ranged from 4.9 to 7.1 Mg C ha−1. Using proximal sensing and digital soil mapping, the uncertainty decreased to 2.3–3.6 Mg C ha−1 and 3.2–5.4 Mg C ha−1, respectively. Our results show that model initialisation with proximal sensing and digital soil maps well represent the spatial variation of TOC and the C pools, so that simulation of changes in C stocks were more reliable. Using the PTF estimates resulted in biased predictions that underestimated the TOC stocks by 2.0–4.5 Mg C ha−1 over the simulation period. Informing a soil C model with measurements is key to building confidence in model-predictions of TOC stocks, providing useful feedback to management practices.

The methods of digital mapping are promising for creating soil maps on difficultly accessible territories. This study was aimed at searching of optimal approaches for digital mapping of the soil cover in poorly studied western part of the Bol’shezemel’skaya tundra on different scales. Medium-scale (1 : 200 000) and small-scale (1 : 1 M) soil maps served as the source of initial information about soils of this region; actual information of the state of the territory was obtained from remote sensing data (Landsat 8 scenes, Aug. 14, 2013) and digital elevation model ASTER GDEM v.2. After extraction of information and the choice of predictors, the analysis of digital soil cover models obtained with the use of different algorithms – Random Forest (RF), Multinomial Logistic Regression (MLR) and Linear Discriminant Analysis (LDA) – was performed. The coefficient of agreement between the newly developed digital models and the initial paper-based soil maps (kappa) was calculated. This test demonstrated that the RF algorithm ensures the best results, so the final digital maps were obtained using it. Averaged kappa values for the compared small- and medium-scale models were as follows: RF – 0.39 and 0.36; MLR – 0.31 and 0.31; and LDA – 0.28 and 0.18, respectively. After the preliminary correction of the initial medium-scale map, the kappa values somewhat increased (RF – 0.39, MLR – 0.35, LDA – 0.30). At the stage of evaluation of digital soil maps obtained with the use of RF algorithm, these maps and the initial soil maps were compared with independent point-size terrain data. The degree of agreement between these data and the new digital soil maps proved to be no less than that for the initial maps. For the initial and digital small-scale maps, it reached 24 and 26 %, respectively; for the initial and digital medium-scale maps, 54 and 43 %, respectively. After the preliminary correction of the initial medium-scale map, the degree of agreement between the digital model and terrain data improved considerably and reached 61 %. This method of digital soil mapping on the basis of analogous data seems to be optimal.

The simultaneous effect of the environmental factors as determinants of soil invertebrate community structures is little understood. This study was carried out at two depths (0–10, 10–20 cm) in the forest ecosystem located in Golestan province, north of Iran. The abundance, diversity, and richness of twelve taxonomic groups of soil invertebrates were modeled using three groups of environmental attributes consisted of topographic attributes and vegetation indices, soil properties and tree diversity indices as predictors. Genetic algorithm was applied to identify and select the effective environmental attributes; the non-linear machine learning, random forest (RF), was employed to predict abundance, diversity, and richness of invertebrates and to determine the most important variables controlling the horizontal and vertical distribution of invertebrates in the forest ecosystem. Results showed that RF was an accurate model for predicting abundance (RMSE = 0.15), diversity (RMSE = 0.18) and richness (overall accuracy = 0.51, kappa = 0.34) of soil invertebrates at the first depth. Similarly, at the second depth, RF showed a good prediction for abundance (RMSE = 0.20), diversity (RMSE = 0.18) and richness (overall accuracy = 0.23, kappa = 0.12) of soil invertebrates. Horizontal distribution of invertebrates' abundance was affected by soil microbial respiration (Resp), land surface temperature (LST), Normalized Difference Vegetation Index (NDVI), Shannon index of trees and soil moisture, while vertical distribution of invertebrates' abundance was more related to soil organic carbon (SOC) and soil moisture. Variable importance analysis confirmed that horizontal and vertical types of distribution of Shannon index of soil invertebrates were affected by the same factors, including land temperature, NDVI, soil respiration and moisture. Also, soil moisture, soil respiration, NDVI and richness of trees had significant effects on horizontal distribution of soil invertebrates' richness, but silt, Normalized Ratio Vegetation Index (NRVI), soil respiration and moisture had positive effects on vertical distribution of richness. In conclusion, that digital soil mapping (DSM) model could be apparently considered as a powerful and economical method in biodiversity modeling. Other supplementary data such as invertebrates feeding web are also suggested as the input dataset for further investigations.

Classification of fields into management zones based on variability of soil fertility parameters is under use in precision agriculture. The study was conducted in west Wollega zone of Ethiopia covering nearly 40 km2 of agricultural land, with the aim to explain variability of soils in the field, classify soils into mapping units and produce a map of these soils at a scale of 1:10000 using geostatistics. In this paper, soil mapping units (SMUs) were interchangeably used with management zones. Ten SMUs were identified in the study site. The SMU mean pH value varied between 5.3 in the SMU6 and 6.4 in the SMU7. Variation in soil texture, pH, exchangeable acidity, exchangeable potassium (K), calcium (Ca), magnesium (Mg), sodium (Na), cation exchange capacity (CEC), organic carbon (OC), total nitrogen (TN), and plant available phosphorus (AvP), were observed within and among SMUs. Variability of soil parameters ranged from 1.06% (Na) in the SMU5 to 172.94% (AvP) in the SMU8. The OC and AvP content of all SMUs were found in the very low and low critical ranges. Intensive grid sampling had helped to produce digital soil maps and exposed soil heterogeneity across the landscape. We suggest that classification of fields into management zones was a remedy to low crop production occurred from uniform applications of nutrients and chemical amendments to significantly varying soils.

This article reviews the current research and applications of various digital soil mapping (DSM) techniques used to map Soil Organic Carbon (SOC) concentration and stocks following a systematic mapping approach from 2013 until present (18 February 2019). It is intended that this review of relevant literature will assist prospective researchers by identifying knowledge clusters and gaps in relation to the digital mapping of SOC. Of 120 studies, most were clustered in some specific countries such as China, Australia and the USA. The highest number publications were in 2016 and 2017. Regarding the predictive models, there was a progression from Linear Models towards Machine Learning (ML) techniques, and hybrid models in Regression Kriging (RK) framework performed better than individual models. Multiple Linear Regression (MLR) was the most frequently used method for predicting SOC, although it was outperformed by other ML techniques in most studies. Random Forest (RF) was found to perform better than MLR and other ML techniques in most comparative studies. Other common and competitive techniques were Cubist, Neural Network (NN), Boosted Regression Tree (BRT), Support Vector Machine (SVM) and Geographically Weighted Regression (GWR). Due to the inconsistency in various comparative studies, it would be advisable to calibrate the competitive algorithms using specific experimental datasets. This review also reveals the environmental covariates that have been identified as the most important by RF technique in recent years in regard to digital mapping of SOC, which may assist in selecting optimum sets of environmental covariates for mapping SOC. Covariates representing organism/organic activities were among the most frequent among top five covariates, followed by the variables representing climate and topography. Climate was reported to be influential in determining the variation in SOC level at regional scales, followed by parent materials, topography and land use. However, for mapping at a resolution that represents smaller areas such as a farm- or plot-scale, land use and vegetation indices were stated to be more influential in predicting SOC. Furthermore, unlike a previous review work, all recent studies in this review incorporated validation and 41% of them estimated spatially explicit prediction of uncertainty. Only 9.16% studies performed external validation, whereas most studies used data-splitting and cross-validation techniques which may not be the best options for datasets obtained through non-probability sampling.

Data Mining techniques play an important role in the prediction of soil spatial distribution in systematic soil surveying, though existing methodologies still lack standardization and a full understanding of their capabilities. The aim of this work was to evaluate the performance of preprocessing procedures and supervised classification approaches for predicting map units from 1:100,000-scale conventional semi-detailed soil surveys. Sheets of the Brazilian National Cartographic System on the 1:50,000 scale, ?Dois Córregos? (?Brotas? 1:100,000-scale sheet), ?São Pedro? and ?Laras? (?Piracicaba? 1:100,000-scale sheet) were used for developing models. Soil map information and predictive environmental covariates for the dataset were obtained from the semi-detailed soil survey of the state of São Paulo, from the Brazilian Institute of Geography and Statistics (IBGE) 1:50,000-scale topographic sheets and from the 1:750,000-scale geological map of the state of São Paulo. The target variable was a soil map unit of four types: local ?soil unit? name and soil class at three hierarchical levels of the Brazilian System of Soil Classification (SiBCS). Different data preprocessing treatments and four algorithms all having different approaches were also tested. Results showed that composite soil map units were not adequate for the machine learning process. Class balance did not contribute to improving the performance of classifiers. Accuracy values of 78 % and a Kappa index of 0.67 were obtained after preprocessing procedures with Random Forest, the algorithm that performed best. Information from conventional map units of semi-detailed (4th order) 1:100,000 soil survey generated models with values for accuracy, precision, sensitivity, specificity and Kappa indexes that support their use in programs for systematic soil surveying. Made available in DSpace on 2020-01-11T00:41:14Z (GMT). No. of bitstreams: 1 APPreprocessingprocedures.pdf: 277257 bytes, checksum: d01e5c88c5102c5df419373a97c8aa99 (MD5) Previous issue date: 2019

Over the last decades extensive work has been done on sampling optimization. Many of the related papers focused on the optimization of sampling for only one soil property. However, there is a necessity to prepare a sampling strategy which is optimized for multivariate digital soil mapping (DSM) purposes. The aim of our work was to elaborate a sampling optimization methodology for multivariate DSM considering the demands on economic efficiency. We presented and tested it through a real-time survey at Tokaj Wine Region, Hungary. The soil properties of interest were pH, soil organic matter (SOM), and calcium carbonate (CaCO3) content. The end-users defined the minimal requested precision for the DSM products (in terms of the average range of the 90% prediction interval), and priority areas on which more detailed survey was requested. We planned a two-phase soil survey based on regression kriging (RK). The results from the first-phase sampling were used to parameterize the second-phase sampling in which spatial simulated annealing (SSA) was applied. The spatially averaged range of the 90% prediction interval was the pre-survey quality measure which can be readily derived from the RK variance. The workflow can be summarized as follows: (1) rank the soil properties considering their spatial variabilities, and precision requests, (2) optimize the sampling design by SSA for the dominant soil property, (3) optimize the sampling by the invers application of SSA for the next soil property using the optimized design from the previous step, and (4) repeat the previous step until all soil property are being selected. In our case, SOM was the dominant property. According to the plot of the sample size vs. quality measure, the optimized design with 500 samples will ensure the minimal requested precision for the SOM map (i.e. 0.5%). In the next step, the optimal removal of those sampling points was targeted which have less information content. In the cases of pH and CaCO3, 100 and 175 could be removed from the 500 samples, and the remaining 400 and 325 samples will ensure the requested precision for the pH (i.e. 1.2) and CaCO3 (i.e. 5%) maps. We computed the relative sampling density on priority and non-priority areas for each sampling designs which showed that densities on priority areas were at least 1.5 times higher than on non-priority areas. We could conclude that the methodology is able to optimize the sampling design for multivariate DSM purposes considering numerous sampling constraints such as the predefined precision, priority areas, and economic efficiency.

Red spruce (Picea rubens) ecosystems in the high elevations of the central Appalachians of the eastern United States are the focus of ongoing restoration efforts due to the valuable ecosystem services these forests provide. Recent research has shown that spodic materials still present in the soil represent the pedomemory of the historic extent of red spruce forests in the region. A dataset containing 221 points with varying spodic intensities and 29 environmental variables collected from the Monongahela National Forest in West Virginia, USA, was used to evaluate the utility of a species distribution model, Maximum Entropy (MaxEnt), for predicting the presence of spodic properties. MaxEnt was selected for evaluation because, as a presence-only model, it inherently omits absence locations and thereby reduces the risk of including false absences (i.e., herein, locations that have undergone some level of depodzolization) unlike other models previously used to predict pedomemory. Model outputs that employed three different spodic intensity class inputs—very weak to strong expression, weak to strong expression, and strong expression—resulted in similar spodic probability predictions, though there was less area mapped as transitional probabilities in the strong expression model than the two models that included weaker spodic intensity input data. Permutation importance indicated that no single or small subset of environmental variables controlled the three model outputs, perhaps because the environmental covariates may have been too coarse or not strongly enough associated with podzolization processes to be very important. When the output from the MaxEnt model using the full range of spodic intensities (very weak to strong) was compared to an output produced using a presence-absence model (random forests), there was approximately 62% agreement (where both models predicted presence or both predicted absence) for the cells in the top 40% of the predicted probabilities.

Digital soil mapping (DSM) often relies on existing soil samples obtained from various sources. However, the spatial distribution of such soil samples can be biased, for example, towards areas of better accessibility. Such biased coverage over the geographic space (i.e., spatial bias) often leads to biased coverage of the soil samples over the environmental covariate space. As a result, spatial bias degrades the correlation or statistical relationship between samples and covariates in the study area and impedes DSM accuracy. This paper presents a representativeness heuristic for mitigating spatial bias in existing soil samples for improving DSM accuracy. The key idea of the heuristic was to define and quantify sample representativeness as the goodness-of-coverage of the soil samples over the environmental covariate space. Spatial bias was then mitigated by weighting the samples towards maximizing their representativeness. Determination of the sample weights was conceived as an optimization problem and accordingly the optimal weights were determined using a genetic algorithm. To evaluate the effectiveness of the representativeness heuristic, a case study of mapping soil organic matter (SOM) content using existing soil samples was conducted in Heshan study area, northeastern China. Results showed that weighting soil samples using the optimal weights determined from the representativeness heuristic improved SOM content mapping accuracy. Moreover, a positive relationship between sample representativeness and mapping accuracy was observed, suggesting sample representativeness is an effective indicator of mapping accuracy. Additionally, the determined optimal weights were informative of individual sample importance and thus can be used as guidance to filter existing soil samples to improve DSM accuracy.

In low relief region such as plains, applied digital soil mapping has a controvertible issue, therefore, this study was aimed to digital mapping of soil classes at family levels by appropriate Geomorphometric variables along with fuzzy logic with area of 16,600 hectares in Qazvin Plain. Based on the geomorphologic map, the plain and pen plain are dominant landscape units. In this regards, 61 soil profiles were dogged. According to the expert’s opinion, covariates including diffuse insolation, standardized height, catchment area, valley depth and multiresolution valley bottom flatness (MrVBF) had the most important in order to generating soil map. Also, 19 fuzzy soil class maps were generated through using sample-based in ArcSIE software. Validation were carried out using achieved overall accuracy (OA) and Kappa index through error matrix. Subsequently, both ignorance and exaggerating uncertainty of hardened soil map were also done. The results showed that 19 soil families class were found. Accordingly, OA and the Kappa index were 54% and 46% respectively. The uncertainty of ignorance and exaggeration were obtained from 0 to 0.64 and 0 to 1, respectively. Moreover, the results indicated that exaggerated uncertainty was the highest in the northern and the lowest in the southern regions. Generally, applied geomorphometric parameters had the specific importance in the low relief areas for mapping of soils that have not been assessed properly so far.

International audience

Aim of study: To use artificial neural networks (ANN) to predict the values and spatial distribution of soil chemical attributes from apparent soil electrical conductivity (ECa) and soil clay contents.Area of study: The study was carried out in an area of 1.2-ha cultivated with cocoa, located in the state of Bahia, Brazil.Material and methods: Data collections were performed on a sampling grid containing 120 points. Soil samples were collected to determine the attributes: clay, silt, sand, P, K+, Ca2+, Mg2+, S, pH, H+Al, SB, CTC, V, OM and P-rem. ECa was measured using the electrical resistivity method in three different periods related to soil sampling: 60 days before (60ECa), 30 days before (30ECa) and when collecting soil samples (0ECa). For the prediction of chemical and physical-chemical attributes of the soil, models based on ANN were used. As input variables, the ECa and the clay contents were used. The quality of ANN predictions was determined using different statistical indicators. Thematic maps were constructed for the attributes determined in the laboratory and those predicted by the ANNs and the values were grouped using the fuzzy k-means algorithm. The agreement between classes was performed using the kappa coefficient.Main results: Only P and K+ attributes correlated with all ANN input variables. ECa and clay contents in the soil proved to be good variables for predicting soil attributes.Research highlights: The best results in the prediction process of the P and K+ attributes were obtained with the combination of ECa and the clay content.

In Colombia, the rise of agricultural and pastureland expansion continues to exert increasing pressure on the structure and ecological processes of savannahs in the Eastern Plains. However, the effect of land use change on soil properties is often unknown due to poor access to remote areas. Effective management and conservation of soils requires the development spatial approaches that measure and predict dynamic soil properties such as soil organic carbon (SOC). This study estimates the SOC stock in the Eastern Plains of Colombia, with validation and uncertainty analyses, using legacy data of 653 soil samples. A random forest model of nine environmental covariate layers was used to develop predictions of SOC content. Model validation was determined using the Taylor series method, and root-mean-squared error (RMSE) and mean error (ME) were calculated to assess model performance. We found that the model explained 50.28% of the variation within digital SOC content map. Raster layers of SOC content, bulk density, and coarse rock fragment within the Eastern Plains were used to calculate SOC stock within the region. With uncertainty, SOC stock in the topsoil of the Eastern Plains was 1.2 G t ha−1. We found that SOC content contributed nearly all the uncertainty in the SOC stock predictions, although better determinations of SOC stock can be obtained with the use of a more geomorphological diverse dataset. The digital soil maps developed in this study provide predictions of extant SOC content and stock in the topsoil of the Eastern Plains, important soil information that may provide insight into the development of research, regulatory, and legislative initiatives to conserve and manage this evolving ecosystem.

Knowledge, data, and understanding of soils is key for advancing agriculture and society. There is currently a critical need for sustainable soil management tools for enhanced food security on Native American Tribal Lands. Tribal Reservations have basic soil information and limited access to conservation programs provided to other U.S producers. The objective of this study was to create first ever high-resolution digital soil property maps of Quapaw Tribal Lands with limited data for sustainable soil resource management. We used a digital soil mapping (DSM) approach based on fuzzy logic to model the spatial distribution of 24 soil properties at 0–15 and 15–30 cm depths. A digital elevation model with 3 m resolution was used to derive terrain variables and a total of 28 samples were collected at 0–30 cm over the 22,880-ha reservation. Additionally, soil property maps were derived from Gridded Soil Survey Geographic Database (gSSURGO) for comparison. When comparing properties modeled by DSM to those derived from gSSURGO, DSM resulted in lower root mean squared error (RMSE) for percent clay and sand at 0–15 cm, and cation exchange capacity, percent clay, and pH at 15–30 cm. Conversely, gSSURGO-derived maps resulted in lower RMSE for cation exchange capacity, pH, and percent silt at the 0–15 cm depth, and percent sand and silt at the 15–30 cm depth. Although, some of the soil properties derived from gSSURGO had lower RMSE, spatial soil property patterns modeled by DSM were in better agreement with the topographic complexity and expected soil-landscape relationships. The proposed DSM approach developed property maps at high-resolution, which sets the baseline for production of new spatial soil information for Quapaw Tribal soils. It is expected that these maps and future versions will be useful for soil, crop, and land-use decisions at the farm and Tribal-level for increased agricultural productivity and economic development. Overall, this study provides a framework for developing DSM on Tribal Lands for improving the accuracy and detail of soil property maps (relative to off the shelf products such as SSURGO) that better represents soil-forming environments and the spatial variability of soil properties on Tribal Lands.

In order to improve the signal-to-noise ratio of the hyperspectral sensors and exploit the potential of satellite hyperspectral data for predicting soil properties, we took MingShui County as the study area, which the study area is approximately 1481 km2, and we selected Gaofen-5 (GF-5) satellite hyperspectral image of the study area to explore an applicable and accurate denoising method that can effectively improve the prediction accuracy of soil organic matter (SOM) content. First, fractional-order derivative (FOD) processing is performed on the original reflectance (OR) to evaluate the optimal FOD. Second, singular value decomposition (SVD), Fourier transform (FT) and discrete wavelet transform (DWT) are used to denoise the OR and optimal FOD reflectance. Third, the spectral indexes of the reflectance under different denoising methods are extracted by optimal band combination algorithm, and the input variables of different denoising methods are selected by the recursive feature elimination (RFE) algorithm. Finally, the SOM content is predicted by a random forest prediction model. The results reveal that 0.6-order reflectance describes more useful details in satellite hyperspectral data. Five spectral indexes extracted from the reflectance under different denoising methods have a strong correlation with the SOM content, which is helpful for realizing high-accuracy SOM predictions. All three denoising methods can reduce the noise in hyperspectral data, and the accuracies of the different denoising methods are ranked DWT > FT > SVD, where 0.6-order-DWT has the highest accuracy (R2 = 0.84, RMSE = 3.36 g kg−1, and RPIQ = 1.71). This paper is relatively novel, in that GF-5 satellite hyperspectral data based on different denoising methods are used to predict SOM, and the results provide a highly robust and novel method for mapping the spatial distribution of SOM content at the regional scale.

International audience; Spatial information of soil carbon storage at national and global level is essential for soil quality and environmental management. Improved knowledge on the amount and spatial distribution of the carbon stock in soils is crucial in estimating changes in the terrestrial carbon dynamics and management options for carbon-storing. A study was conducted to map the soil organic carbon stock (SOC) over 56,763 km2 area of Western Ghats of south India using a digital soil mapping approach. Landsat data, terrain attributes, and bioclimatic variables were used as covariates. Equal-area quadratic splines were fitted to soil profile datasets to estimate soil organic carbon stock at six standard soil depths (0–5, 5–15, 15–30, 30–60, 60–100 and 100–200 cm) and Quantile Regression Forest (QRF) algorithm was used to predict the SOC stocks. Prediction of SOC stock was better for surface layer (R2 = 31–43%) and the performance was decreasing with depth (R2 = 7–21%). The modal performance was also compared with SoilGrids products. Although the spatial patterns were similar, the present predicted SOC maps outperformed SoilGrids products in terms of both R2 and RMSE. The predicted total soil organic stock in the Western Ghats ranged from 7.1 kg m−2 to 30.9 kg m−2 and the total estimated SOC was 917 Tg. The present high resolution SOC maps help to assess and monitor the soil health and preparation of proper land use planning.

International audience; Soil available water capacity (SAWC) is a key soil indicator that plays a major role in many ecosystem services, such as food production, irrigation management, soil drought, flood control, and climate and gas regulation. Digital soil mapping (DSM) can be used to obtain needed SAWC maps. However, SAWC differs from the usual soil properties considered in DSM in that it involves several soil properties determined at several soil layers. Therefore, a specific approach is required to obtain SAWC maps and the associated uncertainty predictions.The objective of this study was to build a SAWC mapping approach that could predict SAWC values at three maximum rooting depths (60, 100 and 200 cm) and their associated prediction uncertainties.The approach was tested in the Languedoc-Roussillon region (southern France). Elementary available water capacities of each layers (in cm.cm−1) and soil layer thicknesses were first mapped separately at 0–30, 30–60, 60–100 and 100–200 cm and then aggregated to estimate the SAWCs at the three mentioned maximum rooting depths. SAWC uncertainty was estimated with an error propagation model that used a first-order Taylor analysis. This analysis considered the mapping errors of each involved property, which were estimated by the quantile regression forest algorithm. We tested different error propagation models that differently considered the correlations between these mapping errors: no correlation considered, correlations between soil layer thicknesses and elementary water capacities per soil layer only, correlations between soil layers only, or all correlations considered.The performances of both SAWC predictions and their uncertainties were assessed with a 10-fold cross validation that was iterated 20 times. The SAWC predictions showed poor accuracies (percentages of explained variance ranged from 0.12 to 0.13). The uncertainties of SAWC predictions were best estimated when the correlations between the soil layer errors were considered in the error propagation model whereas the uncertainties of SAWC predictions were severely underestimated when these correlations were neglected.In spite of the poor performance in predicting SAWC at the punctual level due to the low density of soil observations (1/19 km2), the SAWC approach appeared promising since it produced maps that agreed with the available pedological knowledge and precisely estimated the uncertainties.

Topographic features of territory have a significant impact on the spatial distribution of soil properties. This research is focused on digital soil mapping (DSM) of main agrochemical soil properties—values of soil organic carbon (SOC), nitrogen, potassium, calcium, magnesium, sodium, phosphorus, pH, and thickness of the humus-accumulative (AB) horizon of arable lands in the Trans-Ural steppe zone (Republic of Bashkortostan, Russia). The methods of multiple linear regression (MLR) and support vector machine (SVM) were used for the prediction of soil nutrients spatial distribution and variation. We used 17 topographic indices calculated using the SRTM (Shuttle Radar Topography Mission) digital elevation model. Results showed that SVM is the best method in predicting the spatial variation of all soil agrochemical properties with comparison to MLR. According to the coefficient of determination R2, the best predictive models were obtained for content of nitrogen (R2 = 0.74), SOC (R2 = 0.66), and potassium (R2 = 0.62). In our study, elevation, slope, and MMRTF (multiresolution ridge top flatness) index are the most important variables. The developed methodology can be used to study the spatial distribution of soil nutrients and large-scale mapping in similar landscapes.

The soil cover of the forest-steppe and steppe zones of the East European Plain is characterized by diverse soil combinations revealed during large-scale and detailed soil mapping against the background of a traditional zonal sequence of dominant automorphic soils alternating from the north to the south and clearly displayed on small-scale soil maps. The composition, configuration and functioning of particular soil cover patterns are determined by the soil forming factors acting within a given area. The elementary soil areas (detailed scale) and elementary soil cover patterns maps (large scale) of the Central Russian, Kalach, and Volga Uplands are created by both traditional and digital soil mapping methods. Low-contrasting soil combinations with the background Haplic Chernozems (Loamic or Clayic, Pachic) alternating with zooturbated Haplic Chernozems (Loamic or Clayic, Pachic) on convex elements of the microtopography and Luvic Chernozems (Loamic or Clayic, Pachic) on concave elements of the microtopography prevails under conditions of thick clay loamy parent materials and free drainage. Under conditions of shallow embedding by low-permeable clayey sediments, the soil cover includes Chernozems or Chernic Phaeozems with stagnic features in some part of the soil profile or even Mollic Stagnosols. The presence of shrink-swell clays of different ages leads to the formation of Bathyvertic Chernozems, Vertic Chernozems, Vertic Chernic Phaeozems and/or Pellic Vertisols. The presence of soluble salts in the parent material leads to the development of solonetzic soil complexes consisting of Protosodic or Sodic Chernozems and different types of Solonetzes.

This study aimed to evaluate the performance of three spatial association models used in digital soil mapping and the effects of additional point sampling in a steep-slope watershed (1,200 ha). A soil survey was carried out and 74 soil profiles were analyzed. The tested models were: Multinomial logistic regression (MLR), C5 decision tree (C5-DT) and Random forest (RF). In order to reduce the effects of an imbalanced dataset on the accuracy of the tested models, additional sampling retrieved by photointerpretation was necessary. Accuracy assessment was based on aggregated data from a proportional 5-fold cross-validation procedure. Extrapolation assessment was based on the multivariate environmental similarity surface (MESS). The RF model including additional sampling (RF*) showed the best performance among the tested models (overall accuracy = 49%, kappa index = 0.33). The RF* allowed to link soil mapping units (SMU) and, in the case of less-common soil classes in the watershed, to set specific conditions of occurrence on the space of terrain-attributes. MESS analysis showed reliable outputs for 82.5% of the watershed. SMU distribution across the watershed was: Typic Rhodudult (56%), Typic Hapludult* (13%), Typic Dystrudept (10%), Typic Endoaquent + Fluventic Dystrudept (10%), Typic Hapludult (9.5%) and Rhodic Hapludox + Typic Hapludox (2%).

Fine scale maps of soil properties enable efficient land management and inform earth system models. Recent efforts to create soil property maps from field observations tend to use similar tree-based machine learning interpolation approaches, but often deal with depth of predictions, validation, and uncertainty differently. One of the main differences in approaches is whether to model individual depths of interest separately as ‘2D’ models, or to create models that incorporate depth as a predictor variable creating a ‘3D’ model that can make predictions for all depths. It is unclear how choice of 2D or 3D approach influences model accuracy and uncertainty due to lack of direct comparison and inconsistent presentation of results in past studies. This study compares 2D and 3D methods for mapping soil electrical conductivity (salinity), pH, sum of fine and very fine sands, and organic carbon at 30 m resolution for the upper 432,000 km2 of the Colorado River Watershed of the United States of America. A new, simple, model-agnostic relative prediction interval (RPI) approach to report uncertainty is presented that scales prediction interval width to the 95% interquantile width of the original training sample distribution. The RPI approach enables direct comparison of uncertainty between properties and depths and is easily interpretable by end users. Results indicate that 3D mapping of soil properties with strong variation with depth can result in substantial areas with much higher uncertainty that coincide with unrealistic predictions relative to 2D models, even though 3D models had slightly better global cross-validation scores. Maps and global model summaries of RPI proved helpful in identifying these issues with 3D models. These results suggest that the use of RPI or similar approaches to evaluate models can identify accuracy problems not evident in global validation diagnostics.

In digital soil mapping (DSM) there is an increasing interest in incorporating multiscale variations of soil and environmental covariates to improve mapping accuracy. Empirical mode decomposition (EMD) is a frequently used method for multiscale analysis but has been rarely applied to DSM. This study examines if DSM with EMD covariates results in a more accurate map of a test area than DSM with original covariates. First, EMD was used to decompose terrain attributes of an area. These EMD components were then used to map soil organic carbon (SOC) concentration, pH, clay content and cation exchange capacity (CEC) by two DSM methods: multiple linear regression (MLR) and regression kriging (RK). The same two DSM methods were used with original covariates; in addition, ordinary kriging (OK) maps were produced without covariates. The resulting maps were compared by various accuracy measures. The use of EMD components for DSM improved correlations between soil properties and terrain attributes, and improved the goodness‐of‐fit of MLR models significantly and prediction accuracies by 2.5–18% of root mean square error (RMSE) for SOC, clay and CEC. By contrast, the use of EMD components for pH increased the goodness‐of‐fit of the MLR model slightly and decreased the prediction accuracy by 30% of RMSE. This study concluded that the use of EMD components for DSM improved mapping accuracies if the goodness‐of‐fit of the MLR models was significantly improved by EMD components. It is possible that sampling based on EMD components may provide a sample set that would improve mapping accuracy further. HIGHLIGHTS: Empirical mode decomposition (EMD) components of terrain attributes were applied to soil mapping. EMD components improved correlations between soil properties and terrain attributes. EMD components improved the goodness‐of‐fit of multiple linear regression (MLR). Significant improved goodness‐of‐fit of MLR resulted in improved mapping accuracies.

Mapping the extent and locations of peatland at landscape scale has implications for carbon inventories, conservation and ecosystem services assessments. The main aim of this paper was to model and map the extent of northern peat soils while taking into account its uncertainty, and in particular exploring: 1. the use of radar Sentinel 1 as alternative to optical sensors to reduce problems due to clouds while taking advantage of the seasonality changes and 2. the use of deep learning for peat classification and as application of Digital Soil Mapping. The data sets defining presence or absence of peat in the soil were obtained from different sources and different sampling schemes, densities and distributions. Scotland was used as test case, because of its cloudy weather and fragmented distribution of different types of peat. An extension of the scorpan-kriging approach was used. The trend was estimated with different approaches: Generalized Additive Models, RandomForest and deep learning (convolutional neural networks). Each approach produced the probability of belonging to a class and the predicted class for each pixel. The results were assessed using out-of-sample measures. In this study 108 combinations of data sets and models (including trend approaches, sets of covariates and modelling of the spatial structure) were assessed. Overall, spatially explicit models performed better. The choice of the statistical method can have a significant impact on the predictive performances, while the sets of environmental covariates had a lower impact. Sentinel-1 with morphological features proved to be a good alternative to optical data for peat mapping. It is important to have balanced data sets representing the distribution of the data, because merging heterogeneous sources of data from different populations does not necessarily improve predictions. The use of deep learning and convolutional neural network provided initial promising results. There were large differences in the modelling approaches. These differences and uncertainties need to be taken into account for further modelling such as earth surface modelling or carbon accounting.

A better understanding of the spatial pattern of soil organic matter (SOM) is important for scientific soil management. As multisource secondary data become increasingly cheap and readily available, numerous methods have been established to incorporate secondary information; however, these methods exhibit limitations under certain conditions due to their relatively strict requirements on secondary data. In this study, we tried to integrate sampled soil data and secondary data more effectively within the framework of Bayesian maximum entropy (BME). Specifically, multiple linear regression (MLR) and geographically weighted regression (GWR) were run 100 times based on environmental covariates such as terrain indices, vegetation indices and categorical variables obtained from soil maps. Then, the 95% confidence interval was derived from the multiple prediction values at each of the soft data points. For comparison, some conventional techniques, including ordinary kriging (OK), regression kriging (RK) and geographically weighted regression kriging (GWRK), were also applied. The results showed that BME exhibited a prediction accuracy comparable to that of OK and maintained the prediction uncertainty at a low level, while other studied methods (MLR, GWR, RK and GWRK) were all significantly inferior to BME and OK. The proposed methodology in this study represents a promising scenario for the digital soil mapping, especially when the relationships between the target soil attributes and various secondary information are not strong or residuals of trend models show insignificant spatial autocorrelation.

Soil mappers have traditionally relied on tacit knowledge and qualitative assessment of soil-landscape relationships to obtain the physiographic context necessary to predict soil distribution and spatial patterns. This assessment implicitly utilizes a nested hierarchical approach based on differences in the phenomenon scale of soil forming factors where climate, landscape, parent material, and topography are examined in sequence to create a model of soil-landscape relationships. Our objective was to predict parent material distribution using expert knowledge paired with quantitative digital terrain attributes as part of a nested approach to digital soil mapping. The study took place at an 890-hectare research farm in Logan County, Arkansas, which is part of the Arkansas River Valley. Two major groups of parent material are identifiable in the Arkansas River Valley: residual sandstone and shale on erosional uplands, and silty/clayey pedisediment in depositional areas. A 5-m digital elevation model was used to derive thirteen terrain attributes for the study site. Three of the terrain attributes, namely topographic position index, multi-resolution valley bottom flatness, and vertical distance to channel network, were utilized as part of a rule-based approach to model parent material distribution based on preliminary reconnaissance and expert knowledge of the area. The model was validated by sampling 20 locations using a conditioned Latin hypercube sampling (cLHS) design to evaluate the prediction accuracy. Seventy-five percent of cLHS samples were accurately predicted to be residuum or pedisediment. The resulting map also had 90% agreement with the National Cooperative Soil Survey map; however, the digital map was able to provide more spatially explicit information especially on inclusions. Incorporating parent material distribution as part of a nested hierarchical approach to digital soil mapping aids in constraining and predicting soil properties, enables a more straightforward examination of physiographic context and can ultimately lead to more accurate digital soil maps for land management.

Spatial information on soil particle size distribution and soil organic carbon (SOC) are important for land-use management, environmental models and policy-making. Digital soil mapping (DSM) techniques can quantitatively predict these soil properties using minimal resources. However, DSM has not been adequately evaluated at the farm-scale. The aim of this study was to optimise the DSM framework to produce farm-scale soil maps for 366 ha in the Sandspruit catchment, Western Cape, South Africa. Four feature selection techniques and eight predictive models were evaluated on their ability to predict particle size distribution and SOC. A boosted linear feature selection produced the highest accuracy for all but one soil property. The top-performing predictive models were robust linear models for gravel (ridge regression, RMSE 9.01%, R 2 0.75), sand (support vector machine, RMSE 4.69%, R 2 0.67), clay (quantile regression, RMSE 2.38%, R 2 0.52) and SOC (ridge regression, RMSE 0.19%, R 2 0.41). Random forest was the best predictive model for silt content (RMSE 4.12%, R 2 0.53). This approach appears to be robust for farm-scale soil mapping where the number of observations is often small but high-resolution soil data are required.

Soil organic carbon (SOC) is an important indicator to evaluate agricultural soil quality. Precise mapping SOC can help to facilitate soil and environmental management decisions. This study applied...

It is impossible to make pedological maps without understanding subsurface attributes. Several strategies can be used for soil mapping, from a tacit knowledge to mathematical modeling. However, there are still gaps in knowledge regarding how to optimize subsurface mapping. This work aimed to quantify subsurface soil attributes using satellite spectral reflectance and geographically weighted regression (GWR) techniques. The study was carried out in Sao Paulo, Brazil, in an area spanning 47,882 ha. Multitemporal satellite images (Landsat-5) were initially processed in order to retrieve spectral reflectance from the bare soil surface. Based on a toposequence method, 328 points were then distributed across the area (at depths between 0 and 20 cm and 80 and 100 cm) and analyzed for their soil chemical and physical attributes (including the reflectance spectra (400 to 2500 nm)) in the laboratory. We achieved 67.72% of bare soil for the whole study area, with the remaining 32.28% of the unmapped surface being filled by kriging interpolation. All 328 samples were modeled using surface (Landsat-5 TM spectral reflectance) and subsurface (acquired in the laboratory) data, reaching up to 0.72 R2adj. The correlation between the spectra of both depths was significant and the soil attributes prediction reached an R2adj of validation above 0.6 for clay, hue, value, and chroma at 0–20 and 80–100 cm depths. The satellite soil surface reflectance allowed the estimation of soil subsurface attributes. These results demonstrate that diagnostic soil attributes can be quantified based on spectral pedotransfer (SPEDO) functions to assist digital soil mapping and soil monitoring. Despite our efforts to determine soil subsurface properties using digital soil mapping approach, this task still need considerable refinement. Thus, research must continue to aggregate outcomes from other techniques.

A digital mapping approach was applied to soils in Scotland, producing maps at 100‐m resolution and different levels of classification. This used neural networks to predict fuzzy soil class weightings based upon site descriptors from existing soil survey data. The intention of this work was to produce a set of soil maps for Scotland, which provide greater spatial resolution mapping than currently available, and provide fuzzy data to be used in mapping of ecosystem services using a novel approach explained in a second paper. When selecting the class with the highest weighting for test data points, Producer's Accuracy of 70.7% was achieved in mapping the broadest classification (five classes), while Producer's Accuracy levels of 59.9% and 43.7% were achieved for 11 and 30 classes, respectively. Evaluation of confusion matrices generated from the neural network model tests showed that, particularly for the classification systems with 11 and 30 classes, misclassification errors were more common between similar soil classes than between classes that were very different from one another. This implies that straightforward estimation of classification accuracy that measures “correct” and “incorrect” gives misleading results, and that the fuzzy classification maps are more useful than the crisp classification accuracy values. These fuzzy classification maps are therefore potentially useful for evaluating ecosystem services relating to soil type. We conclude that the mapping approach used here provides a classification accuracy comparable to other approaches and that the outputs can be used for mapping soil properties, processes, functions and ecosystem services.

Particle-size distribution (PSD) plays an important role in influencing a number of soil physical, chemical, and biological properties. Currently, digital soil mapping (DSM) methods based on empirical observations have been widely used in mapping PSD. However, DSM methods rarely consider soil genetic processes. This paper investigated the use of a mechanistic soil evolution model, State Space Soil Production and Assessment Model (SSSPAM), to simulate the spatiotemporal evolution of PSD in the Hunter Valley, NSW, Australia. SSSPAM simulates the spatial and temporal variation of PSD within a landscape based on erosion due to overland flow, deposition, and physical weathering within the soil profile. We conducted a simulation over the 144 km2 area using a 30 m digital elevation model (DEM) as an input. The model simulated soil evolution over 70,000 years to ensure that the PSD had reached steady-state. To validate and analyze the influence of different process parameters on particle size dynamics, we carried out a parametric study in a field within the study area and found a strong relationship between runoff excess generation, exponential weathering rate, and soil particle-size distribution. As expected, higher discharge rates produced coarser particles and larger weathering rates produced finer PSD. We further explored the feasibility of combining the mechanistic SSSPAM and empirical DSM approaches by comparing simulation results with observed sand content. We found limitations of the SSSPAM model to predict sand fraction accurately in the study area due to incomplete process coverage. The output of SSSPAM can be improved by integrating it with DSM techniques. Overall, SSSPAM can explore how particle size will change through time and identify areas with risks of erosion and deposition. Such a model can be used to inform large-scale management to ensure our soil is secured in the future.

The Hydrology of Soil Types (HOST) classification scheme is based on national soil maps in the UK and has been developed by using expert knowledge to link soil hydrological indicators with conceptual models of surface and subsurface flow pathways through the soil profile. HOST has been optimised by regressing HOST class proportions against the Base Flow Index (BFI) and has been used to estimate catchment hydrological response and as input to land and environmental risk mapping. In this study, we performed spatial disaggregation of HOST class map unit polygons at a 1:250,000 scale in Scotland using the Random Forests (RF) classifier and selected environmental covariates to downscale HOST class mapping to 100 m grid resolution. Training datasets were developed using an area-weighted sampling scheme and by extracting HOST class information simultaneously by both single-class and multiple-class (complex) polygons by using the estimated within-polygon HOST class proportions as weights. The most-probable HOST-RF class maps along with surfaces of HOST class probability of occurrence were generated using 100 realisations from respective RF model runs each using a new training dataset. The performance of the disaggregated HOST-RF class maps for predicting catchment hydrological response was assessed by comparing BFI calculated using RF-predicted HOST class proportions with BFI calculated using flow data from gauges in 90 selected catchments. Results show that in this study using a larger number of training samples increases the risk of overfitting the RF model whereas lower sampling densities of 1 point per km2 were found to be sufficient and capable of promoting effective disaggregation of both single-class and complex polygons. Hydrological response predictions for the study catchments made with the disaggregated HOST-RF classes improved when class proportions were weighted using the generated probability surfaces and were very similar with the ones made with the original HOST class polygons, with small prediction improvements observed in the more low-lying and more slowly-responsive catchments when the HOST-RF classes were used. Overall, validation results indicate that the RF classifier was successful in predicting the spatial distribution of HOST classes in the study area, with the exception of HOST classes found within complex polygons areas dominated by soils with organic horizons of variable depth due to the difficulty of the algorithm to identify clear distinctions based on available covariate values. This study demonstrates the potential of applying digital soil mapping techniques to map soil hydrological functions by using an integrative hydropedological approach which links soils and hydrology at a landscape scale.

Detailed knowledge of agricultural soil properties is a key element for high-quality food production. However, high-resolution soil data covering a large agricultural region are generally unavailable. This study explores a demand-driven cooperative framework for soil data sourcing that connects individual farmers to several stakeholders by means of a centralised database containing more than 16,000 records of soil information collected within the framework of an integrated production program for intensively managed permanent crops in the Adige/Etsch and Venosta/Vinschgau valleys in South Tyrol, Italy. Data for soil pH, soil organic matter (SOM), and soil texture were used to produce digital soil maps with a RMSE of 0.21, 1.25% and a cross-validation of 43%, respectively. Spatialisation was conducted using either regression-kriging or multinomial logistic regression. Collaboration among farmers, public administrators, and researchers provided a successful cooperative framework for digital soil mapping. The maps highlight the complex interplay of the postglacial evolution of these valleys due to the presence of a cluster of large alluvial fans and the anthropogenic influences of intense farming on pH, SOM, and soil texture. This study regarded a subset of the available soil properties, which can be dealt with using the geostatistical approaches presented herein. Thus, a long-term soil monitoring program and the combination of all available variables will allow digital assessment of the spatial patterns of nutrient availability, ecological risk assessments, change detection studies, and an overall long-term plan for soil security at larger spatial scales.

The main trends in the development of soil mapping methods are discussed, and the major problems are identified. By the present time, the transition from the paper-based soil maps to digital soil-geographical databases has already been completed. The digital mapping of soils and their properties is now accepted as the main method at all the levels of generalization. The approaches of digital soil mapping, as well as of the traditional one, are based on the ideas of V.V. Dokuchaev about the dependence of soils on soil-forming factors. However, in digital soil mapping, new achievements of mathematical statistics and mathematical modeling are being widely applied. This provides for a greater objectivity and reproducibility of the digital soil maps in comparison with the traditional soil maps. At the same time, all unsolved problems of soil cartography related to the lack of field observation data, scale, soil taxonomy, spatial microheterogeneity, and mapping of individual soil properties are preserved. Partially, these problems can be solved by using remote sensing data. When the soil geographical information is used to assess the quality of soil resources, the interpretation of remote sensing data for mapping purposes seems to be more preferable in comparison with the methods of digital soil mapping.

Terrain attributes are commonly used as predictors of soil organic carbon (SOC) in digital soil mapping. However, there are no fixed rules in the selection of suitable grid size and models with different attribute combinations. Past studies have used a few empirical, as well as pedological guidelines to determine scale dependency of terrain attributes on SOC prediction. The aim of this paper was to evaluate the scale dependency of terrain attributes using varying grid sizes and select the most important attributes and optimum grid size for SOC prediction. A 7500 km 2 area located in Denmark was selected; a total of 2,514,820 prediction models were generated in Cubist data-mining tool in which 8570 SOC observations and 22 terrain attributes at 71 different grid sizes ranging from 12.8 m to 2304 m were used as inputs. Terrain attributes were derived from the Light Detection And Ranging (LiDAR) based digital elevation model (DEM) (1.6 m × 1.6 m grid size) and was subsequently resampled to different resolutions by simple mean aggregation. Relative importance and usage of each terrain attribute in each prediction model were computed and only the top 5 attributes were reported for different attribute combinations. The results showed that the relative contribution of terrain attributes to predict SOC distribution varied by grid sizes, and by grid size and attribute combinations. Overall, Relative Slope Position (rsp), Channel Altitude (chnl_alti), and Standard Height (standh) were the three most important terrain attributes in the five-attribute-model at all grid resolutions and the remaining two attributes being Normalized Height (normalh) and Valley Depth (vall_depth) at resolutions finer than 30 m, and elevation and Channel Base (chnl_base) at resolutions coarser than 30 m. The models at 88 m and 92.8 m grid size (nearest to the 90 m SRTM data resolutions) and 30.4 m (nearest to the 30 m TM satellite image resolution) were validated. We observed that the model performance was dependent on grid size, and by attribute combinations. For example, for a 4-attribute model that used rsp, chnl_alti, elevation and vall_depth, the best performance was for 30.4 m compared to 88 m and 92.8 m grid sizes. We found that for modeling SOC distribution, the three terrain attributes rsp, chnl_alti, and standh were found to be the most important at all resolutions and should be considered as important variables in future SOC modeling studies in young moraine landscapes.

Our understanding of the spatial distribution of soil carbon (C) pools across diverse land uses, soils, and climatic gradients at regional scale is still limited. Research in digital soil mapping and modeling that investigates the interplay between (i) soil C pools and environmental factors (“deterministic trend model”) and (ii) stochastic, spatially dependent variations in soil C fractions (“stochastic model”) is just emerging. This evoked our motivation to investigate soil C pools in the State of Florida covering about 150,000 km2. Our specific objectives were to (i) compare different soil C pool models that quantify stochastic and/or deterministic components, (ii) assess the prediction performance of soil C models, and (iii) identify environmental factors that impart most control on labile and recalcitrant pools and soil total C (TC). We used soil data (0–20 cm) from a research collected at 1014 georeferenced sites including measured bulk density, recalcitrant carbon (RC), labile (hot-water extractable) carbon (HC) and TC. A comprehensive set of 327 geospatial soil-environmental variables was acquired. The Boruta method was employed to identify “all-relevant” soil-environmental predictors. We employed eight methods - Classification and Regression Tree (CaRT), Bagged Regression Tree (BaRT), Boosted Regression Tree (BoRT), Random Forest (RF), Support Vector Machine (SVM), Partial Least Square Regression (PLSR), Regression Kriging (RK), and Ordinary Kriging (OK) – to predict soil C fractions and TC. Overall, 36, 20 and 25 predictors stood out as “all-relevant” to estimate TC, RC and HC, respectively. We predicted a mean of 5.29 ± 3.58 kg TC m−2 in the top 20 cm with the best model. The prediction performance assessed by the Ratio of Prediction Error to Inter-quartile Range for TC stocks was as follows: RF > SVM > BoRT > BaRT > PLSR > RK > CART > OK. The best models explained 71.6%, 71.7% and 30.5% of the total variation for TC, RC and HC, respectively. Biotic and hydro-pedological factors explained most of the variation in soil C pools and TC; lithologic and climatic factors showed some relationships to soil C pools and TC, whereas topographic factors faded from soil C models.

Previous studies on soil organic carbon content or stock mapping mostly use natural environmental covariates and do not consider the soil management practice factor. However, human activities have become an important influencing factor for soil organic carbon, especially for agricultural soils. Crop species/crop rotations and management practices significantly affect the amount and spatial variation of soil organic carbon in croplands, but have not been considered for mapping soil organic carbon. In this study, we used direct crop rotation information and variables generated using Fourier transform on HJ-1A/1B NDVI time series data to capture the periodic effect of crop rotation, and explored the effectiveness of incorporating such information in predicting topsoil organic carbon content in cropland. A case study applied such method in a largely agricultural area in Anhui province, China. Crop rotation information was obtained through field investigation. Various combinations of predictive environmental variables were experimented for mapping soil organic carbon. The results were validated using field samples. Results showed that the combination of natural environment variables with both crop rotation type and variables derived through Fourier transform yielded the highest accuracy. In addition, only using the Fourier decomposed variables and crop rotation information were able to achieve a similar accuracy with using only soil formative natural environmental variables. This indicates that crop rotation information has comparable predictive power of soil organic carbon as natural environment variables. This study demonstrates the effectiveness of including agricultural practice information in digital soil mapping in agricultural landscapes with differences in crop rotation.

The ability to accurately and precisely perform soil nutrient mapping over large areas is essential in the decision-making processes for precision agriculture. However, existing grid-based or non-grid-based digital soil mapping (DSM) can lead to the problem of mixed units of input information, which causes the mapping results to be unsuitable for direct use in guiding the implementation of precision agriculture. Instead, the goal of this study was to achieve DSM based on land parcels, which are the basic units of agricultural management and have practical geographical significance for precision mapping in agricultural areas. This study established a convolutional neural network-based automatic extraction model to extract land parcels from high resolution remote sensing images. Thirty environmental covariates were chosen and calibrated at land parcels to establish the relationships between soils and landscapes. Four prediction algorithms, namely, ordinary kriging, cokriging, random forest and artificial neural network, were combined with the land-parcel-based DSM framework to develop and evaluate their effectiveness in predicting four topsoil nutrient properties in an alluvial-diluvia plain agricultural region located in Ningxia province, China. The results of comparisons show that, overall, the land-parcel-based RF model achieved the best prediction accuracy; its relative improvement (RMSE%) values over the competing models were 1.27, 4.23, 3.19 and 9.01 for soil organic matter, soil total nitrogen, available phosphorus and available potassium, respectively. In addition, land-parcel-based mapping can improve algorithmic efficiency by approximately 4 times by effectively reducing the mapping units for complex agricultural areas compared with the grid-based mapping results when using the same algorithm, and it also achieves a better performance at the detail level. Overall, the land-parcel-based DSM approach achieved good results in plain agricultural areas, but the model still needs improvement for land-parcel-based DSM in mountainous and hilly agricultural areas, and a challenge remains in selecting the most appropriate environmental covariates.

Increasing demand for food production, global change, and growing population are the enormous challenges in recent decades. Accurate soil maps and adequate models are indispensable tools to assist managers, scientists, and decision-makers in addressing these challenges. Legacy soil polygon maps at national and regional scales are available widely, but lack detail, and therefore effective methods such as digital soil mapping (DSM) are needed to disaggregate these maps. The objective of this study was to disaggregate a legacy 1:1,000,000 soil map by three methods of disaggregation: a supervised classification method (DSMART algorithm) and two unsupervised classification methods including fuzzy c-means (FCM) and k-means (KM) clustering in Borujen region, Chaharmahal-Va-Bakhtiari Province, Central Iran for both great group and subgroup Taxonomic levels. Although field validation indicated that the accuracy of the disaggregated soil maps was lower than that of the conventional soil map at both levels of Soil Taxonomy, disaggregated approaches produced more detailed soil maps when compared with the first, second, and third most probable soil classes of the conventional soil map. The higher overall accuracy of the conventional soil map was due to soil association units which consist of more than one soil taxonomic class. FCM and DSMART methods produced more accurate and detailed disaggregated soil maps than KM clustering algorithm at the great group and subgroup levels, respectively. We concluded that the decision on what method to use depends on the map, the level of available information (map detail), available expert knowledge, and the availability of the soil unit composition percentages in the soil map legend.