Digital Soil Mapping Models Research Articles

Predictive soil mapping using random forest models: Applications in pH and soil organic matter assessment

Digital Soil Mapping (DSM) presents a highly scalable and efficient alternative to traditional soil analysis, which is typically limited by its labor-intensive processes, time constraints and low spatial resolution. By utilizing advanced computational techniques such as machine learning and remote sensing, DSM overcomes these limitations and improves the accuracy, efficiency and scalability of soil property assessments. This study, conducted across Tamil Nadu, India, applied DSM and Random Forest (RF) models to predict 2 key soil properties: pH and Soil Organic Matter (SOM). We employed Conditioned Latin Hypercube Sampling (cLHS) for optimized sampling point selection and utilized the Boruta algorithm to identify the most relevant covariates for accurate modeling. The RF models were fine-tuned using a comprehensive grid search, with the optimal configuration spanning from 500 to 2000 trees (ntree) and mtry from 1 to 11. The best-performing model was found with 2000 trees and mtry set to 1 yielding superior prediction for SOM and pH with Root Mean Square Error (RMSE) values of 0.71 and 0.60 respectively, showcasing a high level of predictive accuracy. Our findings emphasize the critical role that remote sensing indices play in predicting SOM, while pH was influenced by both terrain features and remote sensing data. In comparison to previous studies, this research offers novel improvements in both sampling optimization and model configuration, leading to enhanced predictive performance. These results hold significant potential for sustainable land-use planning, agricultural productivity and environmental management.

Contribution of Sentinel-2 spring seedbed spectra to the digital mapping of soil organic carbon concentration

Soil organic carbon (SOC) is central to the functioning of terrestrial ecosystems, has climate mitigation potential and provides several benefits for soil health. Understanding the spatial distribution of SOC can help formulate sustainable soil management practices. Digital soil mapping (DSM) uses advanced statistical and geostatistical methods to estimate soil properties across large areas. DSM integrates climate data, topographic features, geology, legacy soil maps, land management and remote sensing data. Bare soil spectra may reflect the presence of particular soil components, making satellite derived spectra suitable predictors of SOC. Bare soil spectra derived from Sentinel-2 were used to estimate SOC concentration (SOC%) and granulometric fractions in the plough layer (0–30 cm) of agricultural parcels in northern Belgium. Thereafter, the estimation performance of SOC% was compared for three DSM models: one with bare soil spectra, one with environmental covariates (topography, granulometry and vegetation), and a combined model with bare soil spectra and environmental covariates. The estimation performance of sand, silt and clay fractions using bare soil spectra from the spring seedbed (R2: 0.53–0.74; RPD: 1.49–2.05; RPIQ: 1.52–2.39) was higher than that of SOC% (R2: 0.16; RPD: 1.08; RPIQ: 1.32). The highest estimation performance of SOC% was obtained for a DSM model including all covariates (R2: 0.28; RPD: 1.18; RPIQ: 1.44), but the contribution of spring seedbed spectra to a model containing environmental covariates was small. The results provide valuable insights for refining soil property estimation using DSM with spectral and environmental covariates.

Digital soil mapping of soil burn severity

AbstractFire alters soil hydrologic properties leading to increased risk of catastrophic debris flows and post‐fire flooding. As a result, US federal agencies map soil burn severity (SBS) via direct soil observation and adjustment of rasters of burned area reflectance. We developed a unique application of digital soil mapping (DSM) to map SBS in the Creek Fire which burned 154,000 ha in the Sierra Nevada. We utilized 169 ground‐based observations of SBS in combination with raster proxies of soil forming factors, pre‐fire fuel conditions, and fire effects to vegetation to build a digital soil mapping model of soil burn severity (DSMSBS) using a random forest algorithm and compared the DSMSBS map to the established SBS map. The DSMSBS model had a cross‐validation accuracy of 48%. The established technique had 46% agreement between field observations and pixels. However, since the established technique is manual, it could not be compared to the DSMSBS model via cross‐validation. We produced SBS class uncertainty maps, which showed high prediction probabilities around field observations, and low probabilities away from field observations. SBS prediction probabilities could aid post‐fire assessment teams with sample prioritization. We report 107 km2 more area classified as high and moderate SBS compared to the established technique. We conclude that blending soil forming factors based mapping and vegetation burn severity mapping can improve SBS mapping. This represents a shift in SBS mapping away from validating remotely sensed reflectance imagery and toward a quantitative soil landscape model, which incorporates both fire and soils information to directly predict SBS.

Evaluating the quality of soil legacy data used as input of digital soil mapping models

AbstractMost of the Digital Soil Mapping (DSM) products now available across the globe have been developed from the deposits of punctual soil observations inherited from several decades of soil survey activity. By using these legacy data as inputs for calibrating our DSM models, we implicitly make the assumption that these legacy soil data are accurate and therefore do not affect significantly our DSM products. However, this assumption has never been tested. The objectives of this study were to evaluate the accuracies of soil property measurements retrieved from legacy soil profiles, to analyse the different sources of error that may affect these measurements and to examine their impacts on the soil property predictions delivered by DSM models. The study was focused on a control sampling within the coastal plain of Languedoc (Southern France) at 129 locations where legacy measured soil profiles were collected between 1955 and 1992. At each location, four topsoil (0–20 cm) samples were collected at increasing distances (0, 5, 25 and 100 m) to characterize the local variabilities of the soil properties. Six soil properties—Clay, Silt, Sand, Soil Organic Carbon, Calcium Carbonate contents and Cation Exchange Capacity—were determined for each sample using certified soil laboratory methods. The results revealed that legacy soil property values had large overall errors and large biases. Biases likely induced by differences in soil analysis protocols could be corrected by linear functions calibrated onto the reference data obtained from the control sampling. The contributions of the errors propagated from the manual geo‐referencing errors (mean = 31 m) represented on average 52% of the errors after analytical bias corrections. These errors exhibited large variations from one property to another due to differences in the short‐range spatial variations (0–100 m) of these soil properties. A DSM exercise conducted on our control sampling revealed that the errors of legacy soil data were propagated to the soil property predictions provided by the DSM models. However, this propagation could be largely mitigated by applying the above‐evoked corrections. This study highlights the need to better control the quality of the legacy soil data used in DSM and to account for this source of uncertainty in the DSM models.

Effect of training sample size, sampling design and prediction model on soil mapping with proximal sensing data for precision liming

Site-specific estimation of lime requirement requires high-resolution maps of soil organic carbon (SOC), clay and pH. These maps can be generated with digital soil mapping models fitted on covariates observed by proximal soil sensors. However, the quality of the derived maps depends on the applied methodology. We assessed the effects of (i) training sample size (5–100); (ii) sampling design (simple random sampling (SRS), conditioned Latin hypercube sampling (cLHS) and k-means sampling (KM)); and (iii) prediction model (multiple linear regression (MLR) and random forest (RF)) on the prediction performance for the above mentioned three soil properties. The case study is based on conditional geostatistical simulations using 250 soil samples from a 51 ha field in Eastern Germany. Lin’s concordance correlation coefficient (CCC) and root-mean-square error (RMSE) were used to evaluate model performances. Results show that with increasing training sample sizes, relative improvements of RMSE and CCC decreased exponentially. We found the lowest median RMSE values with 100 training observations i.e., 1.73%, 0.21% and 0.3 for clay, SOC and pH, respectively. However, already with a sample size of 10, models of moderate quality (CCC > 0.65) were obtained for all three soil properties. cLHS and KM performed significantly better than SRS. MLR showed lower median RMSE values than RF for SOC and pH for smaller sample sizes, but RF outperformed MLR if at least 25–30 or 75–100 soil samples were used for SOC or pH, respectively. For clay, the median RMSE was lower with RF, regardless of sample size.

Evaluating the extrapolation potential of random forest digital soil mapping

Spatial soil information is essential for informed decision-making in a wide range of fields. Digital soil mapping (DSM) using machine learning algorithms has become a popular approach for generating soil maps. DSM capitalises on the relation between environmental variables (i.e., features) and a soil property of interest. It typically needs a training dataset that covers the feature space well. Mapping in areas where there are no training data is challenging, because extrapolation in geographic space often induces extrapolation in feature space and can seriously deteriorate prediction accuracy. The objective of this study was to analyse the extrapolation effects of random forest DSM models by predicting topsoil properties (OC, clay, and pH) in four African countries using soil data from the ISRIC Africa Soil Profiles database. The study was conducted in eight experiments whereby soil data from one or three countries were used to predict in the other countries. We calculated similarities between donor and recipient areas using four measures, including soil type similarity, homosoil, dissimilarity index by area of applicability (AOA), and quantile regression forest (QRF) prediction interval width. The aim was to determine the level of agreement between these four measures and identify the method that had the strongest agreement with common validation metrics. The results indicated a positive correlation between soil type similarity, homosoil and dissimilarity index by AOA. Surprisingly, we observed a negative correlation between dissimilarity index by AOA and QRF prediction interval width. Although the cross-validation results for the trained models were acceptable, the extrapolation results were unsatisfactory, highlighting the risk of extrapolation. Using soil data from three countries instead of one increased the similarities for all measures, but it had a limited effect on improving extrapolation. Also, none of the measures had a strong correlation with the validation metrics. This was particularly disappointing for AOA and QRF, which we had expected to be strong indicators of extrapolation prediction performance. Results showed that homosoil and soil type methods had the strongest correlation with validation metrics. The results for this case study revealed limitations of using AOA and QRF as measures of extrapolation effects, highlighting the importance of not relying on these methods blindly. Further research and more case studies are needed to address the effects of extrapolation of DSM models.

Estimating natural soil drainage classes in the Wisconsin till plain of the Midwestern U.S.A. based on lidar derived terrain indices: Evaluating prediction accuracy of multinomial logistic regression and machine learning algorithms

Accurate spatial prediction of natural soil drainage condition is not only important for agriculture and hydrological modeling but also for installing subsurface drainage and onsite waste disposal systems. For this research, 154 sites were selected based on a stratified random sampling method. For each site, drainage class was identified based on visual examination of soil cores. A digital elevation model developed from lidar data was used to derive seven terrain indices. Terrain indices were used to predict drainage class using four prediction models: multinomial logistic regression and three machine learning algorithms (random forest, C5.0, and artificial neural network). Based on 30% random hold-back validation data, all digital soil mapping (DSM) models provided similar results. The artificial neural network provided relatively higher overall accuracy of 70%, a kappa coefficient of 0.59, and a Brier score of 0.34. The rest of the DSM models as well as SSURGO provided an accuracy that ranged between 64 and 66%, a kappa coefficient that ranged between 0.52 and 0.54, and a Brier score that ranged between 0.41 and 0.46.

Incorporation of Fused Remote Sensing Imagery to Enhance Soil Organic Carbon Spatial Prediction in an Agricultural Area in Yellow River Basin, China

To overcome spatial, spectral and temporal constraints of different remote sensing products, data fusion is a good technique to improve the prediction capability of soil prediction models. However, few studies have analyzed the effects of image fusion on digital soil mapping (DSM) models. This research fused multispectral (MS) and panchromatic Landsat 8 (L8) bands, and MS Sentinel 2 (S2) and panchromatic L8 bands using the Brovey, Intensity–Hue–Saturation and Gram–Schmidt methods in an agricultural area in Yellow River Basin, China. To analyze the effects of image fusion on DSM models, various SOC prediction models derived from remote sensing image datasets were established by the random forest method. Soil salinity indices and spectral reflectance from all the remote sensing data had relatively strong negative correlations with SOC, and vegetation indices and water indices from all the remote sensing data had relatively strong positive correlations with SOC. Soil moisture and vegetation were the main controlling factors of the SOC spatial pattern in the study area. More spectral indices derived from pansharpened L8 and fused S2–L8 images by all three image fusion methods had stronger relationships with SOC compared with those from MS L8 and MS S2, respectively. All the SOC models established by pansharpened L8 and fused S2–L8 images had higher prediction accuracy than those established by MS L8 and MS S2, respectively. The fusion between S2 and L8 bands had stronger effects on enhancing the prediction accuracy of SOC models compared with the fusion between panchromatic and MS L8 bands. It is concluded that digital soil mapping and image fusion can be utilized to increase the prediction performance of SOC spatial prediction models.

Integrating additional spectroscopically inferred soil data improves the accuracy of digital soil mapping

Digital soil mapping has been increasingly advocated as an efficient approach to deliver fine-resolution and up-to-date soil information in evaluating soil ecosystem services. Considering the great spatial heterogeneity of soils, it is widely recognized that more representative soil observations are needed for better capturing the soil spatial variation and thus to increase the accuracy of digital soil maps. In reality, the budget for the field work and soil laboratory analysis is commonly limited due to its high cost and low efficiency. In the last two decades, being an alternative to wet chemistry, soil spectroscopy, such as visible-near infrared (Vis-NIR), mid-infrared (MIR) spectroscopy has been developed in measuring soil information in a rapid and cost-effective manner and thus enable to collect more soil information for digital soil mapping (DSM). However, spectroscopically inferred (SI) data are subject to higher uncertainties than reference laboratory analysis. Many DSM practices integrated SI data with soil observations into spatial modelling while few studies addressed the key question that whether these non-errorless soil data improve map accuracy in DSM. In this study, French Soil Monitoring Network (RMQS) and Land Use and Coverage Area frame Survey Soil (LUCAS Soil) datasets were used to evaluate the potential of SI data from Vis-NIR and MIR in digital mapping of soil properties (i.e. soil organic carbon, clay, and pH) at a national scale. Cubist and quantile regression forests were used for spectral predictive modelling and DSM modelling, respectively. For both RMQS and LUCAS Soil dataset, different scenarios regarding varying proportions of SI data and laboratory observations were tested for spectral predictive models and DSM models. Repeated (50 times) external validation suggested that adding additional SI data can improve the performance of DSM models regardless of soil properties (gain of R2 proportion at 3–19%) when the laboratory observations are limited (≤50%). Lower proportion of SI data used in DSM model and higher accuracy of spectral predictive models led to greater improvement of DSM. Our results also showed that a greater proportion of SI data lowered the prediction intervals which may result in an underestimation of prediction uncertainty. The determination of accuracy threshold on SI data for the use in DSM needs to be explored in future studies.

Extrapolation of Digital Soil Mapping Approaches for Soil Organic Carbon Stock Predictions in an Afromontane Environment

Soil scientists can aid in an essential part of ecological conservation and rehabilitation by quantifying soil properties, such as soil organic carbon (SOC), and is stock (SOCs) SOC is crucial for providing ecosystem services, and, through effective C-sequestration, the effects of climate change can be mitigated. In remote mountainous areas with complex terrain, such as the northern Maloti-Drakensberg in South Africa and Lesotho, direct quantification of stocks or even obtaining sufficient data to construct predictive Digital Soil Mapping (DSM) models is a tedious and expensive task. Extrapolation of DSM model and algorithms from a relatively accessible area to remote areas could overcome these challenges. The aim of this study was to determine if calibrated DSM models for one headwater catchment (Tugela) can be extrapolated without re-training to other catchments in the Maloti-Drakensberg region with acceptable accuracy. The selected models were extrapolated to four different headwater catchments, which included three near the Motete River (M1, M2, and M3) in Lesotho and one in the Vemvane catchment adjacent to the Tugela. Predictions were compared to measured stocks from the soil sampling sites (n = 98) in the various catchments. Results showed that based on the mean results from Universal Kriging (R2 = 0.66, NRMSE = 0.200, and ρc = 0.72), least absolute shrinkage and selection operator or LASSO (R2 = 0.67, NRMSE = 0.191, and ρc = 0.73) and Regression Kriging with cubist models (R2 = 0.61, NRMSE = 0.184, and ρc = 0.65) had the most satisfactory outcome, whereas the soil-land inference models (SoLIM) struggled to predict stocks accurately. Models in the Vemvane performed the worst of all, showing that that close proximity does not necessarily equal good similarity. The study concluded that a model calibrated in one catchment can be extrapolated. However, the catchment selected for calibration should be a good representation of the greater area, otherwise a model might over- or under-predict SOCs. Successfully extrapolating models to remote areas will allow scientists to make predictions to aid in rehabilitation and conservation efforts of vulnerable areas.

Transferability of Covariates to Predict Soil Organic Carbon in Cropland Soils

Precise knowledge about the soil organic carbon (SOC) content in cropland soils is one requirement to design and execute effective climate and food policies. In digital soil mapping (DSM), machine learning algorithms are used to predict soil properties from covariates derived from traditional soil mapping, digital elevation models, land use, and Earth observation (EO). However, such DSM models are trained for a specific dataset and region and have so far only allowed limited general statements to be made that would enable the models to be transferred to different regions. In this study, we test the transferability of SOC models for cropland soils using five different covariate groups: multispectral soil reflectance composites (satellite), soil legacy data (soil), digital elevation model derivatives (terrain), climate parameters (climate), and combined models (combined). The transferability was analyzed using data from two federal states in southern Germany: Bavaria and Baden-Wuerttemberg. First, baseline models were trained for each state with combined models performing best in both cases (R2 = 0.68/0.48). Next, the models were transferred and tested with soil samples from the other state whose data were not used during model calibration. Only satellite and combined models were transferable, but accuracy declined in both cases. In the final step, models were trained with samples from both states (mixed-data models) and applied to each state separately. This process significantly improved the accuracies of satellite, terrain, and combined models, while it showed no effect on climate models and decreased the models based on soil covariates. The experiment underlines the importance of EO for the transfer and extrapolation of DSM models.

Digital Soil Mapping of Cadmium: Identifying Arable Land for Producing Winter Wheat with Low Concentrations of Cadmium

Intake of cadmium (Cd) via vegetable food poses a possible health risk. Cereals are one of the major sources of Cd, and the Cd concentration in the soil has a great effect on the levels in the grain. The aim of the study was to produce decision support for identification of areas suitable for low-Cd winter wheat production in the form of a detailed digital soil map covering an important agricultural region in southern Sweden. A two-step approach was used: (1) we increased the number of soil Cd observations by combining two sets of soil samples, one with laboratory Cd analyses (304 samples) and one with predicted Cd from a portable x-ray fluorescent (PXRF) sensor (2097 samples); and (2) a digital soil mapping (DSM) model (gradient boosting regression) was calibrated on all 2401 soil samples to create a soil Cd concentration map using a number of covariates, of which airborne gamma ray data was identified as the most important. In the first step, cross-validation of the PXRF model obtained a model efficiency (E) of 0.82 and mean absolute error (MAE) of 0.08 mg kg−1. The DSM model had an E of 0.69 and MAE of 0.11 mg kg−1. The map of predicted soil Cd concentrations were compared against 307 winter wheat (Triticum aestivum L.) grain samples with laboratory-analyzed Cd concentrations. Areas in the map with low soil Cd concentrations had a high frequency of lower grain Cd concentrations. The map thus seemed to have potential for finding areas suitable for production of low-Cd winter wheat; e.g., for baby food.

Machine Learning Models for Prediction of Soil Properties in the Riparian Forests

Spatial variability of soil properties is a critical factor for the planning, management, and exploitation of soil resources. Thus, the use of different digital soil mapping models to provide accuracy plays a crucial role in providing soil physicochemical properties maps. Soil spatial variability in forest stands is not well-known in Iran. Meanwhile, riparian buffers are important for several services such as providing high water quality, nutrient recycling, and buffering agricultural production. Accordingly, in this research, 103 soil samples were taken using the Latin hypercubic method in the Maroon riparian forest of Behbahan and agricultural lands in the vicinity of the forest to evaluate the spatial variability of soil nitrogen, potassium, organic carbon, C:N ratio, pH, calcium carbonate, sand, silt, clay, and bulk density. Different machine learning models, including artificial neural networks, random forest, cubist regression tree, and k-nearest neighbor were used to compare the estimation of soil properties. Moreover, three main sources of spatial information including remote sensing images, digital elevation model, and climate parameters were used as ancillary data. Our results indicated that the random forest model has the best results in estimating soil pH, nitrogen, potassium, and bulk density. In contrast, the cubist regression tree indicated the best estimation for organic carbon, C:N ratio, phosphorous, and clay. Further, artificial neural networks showed the best estimation for calcium carbonate, sand, and silt contents. Our results revealed that geospatial information such as terrain parameters, climate parameters, and satellite images could be well used as ancillary data for the spatial mapping of soil physiochemical properties in riparian forests and agricultural lands. In conclusion, a specific machine learning model needs to be used for each soil property to provide highly accurate maps with less error.

Using homosoils for quantitative extrapolation of soil mapping models

AbstractSince the early 2000s, digital soil maps have been successfully used for various applications, including precision agriculture, environmental assessments and land use management. Globally, however, there are large disparities in the availability of soil data on which digital soil mapping (DSM) models can be fitted. Several studies attempted to transfer a DSM model fitted from an area with a well‐developed soil database to map the soil in areas with low sampling density. This usually is a challenging task because two areas have hardly ever the same soil‐forming factors in two different regions of the world. In this study, we aim to determine whether finding homosoils (i.e., locations sharing similar soil‐forming factors) can help transferring soil information by means of a DSM model extrapolation. We hypothesize that within areas in the world considered as homosoils, one can leverage on areas with high sampling density and fit a DSM model, which can then be extrapolated geographically to an area with little or no data. We collected publicly available soil data for clay, silt, sand, organic carbon (OC), pH and total nitrogen (N) within our study area in Mali, West Africa and its homosoils. We fitted a regression tree model between the soil properties and environmental covariates of the homosoils, and applied this model to our study area in Mali. Several calibration and validation strategies were explored. We also compared our approach with existing maps made at a global and a continental scale. We concluded that geographic model extrapolation within homosoils was possible, but that model accuracy dramatically improved when local data were included in the calibration dataset. The maps produced from models fitted with data from homosoils were more accurate than existing products for this study area, for three (silt, sand, pH) out of six soil properties. This study would be relevant to areas with very little or no soil data to carry critical soils and environmental risk assessments at a regional level.Highlights Soil mapping models were fitted with soil data within the homosoils of Mali. The fitted models were applied to our study area. Model accuracy dramatically improved when including local data. Homosoil maps were more accurate for 3 out of 6 soil properties compared to global and continental maps. New opportunity to map the regional soil pattern of areas with limited soil data coverage.

Semi-supervised learning for the spatial extrapolation of soil information

Digital soil mapping (DSM) can be used to predict soils at unvisited sites, but problems arise when predictions are needed in areas without any soil observations. In such situations, DSM can still extend the results from reference areas with soil data to target areas that are alike in terms of soil-forming factors and obey the same rules. Such DSM methods have low accuracy due to the complexity of spatial variation in soil, and the difficulty of matching soil-forming factors exactly between reference and target areas. A new approach for extrapolating soil information from reference to target areas is proposed in the current research. We evaluated the ability of a semi-supervised learning (SSLR→T) approach compared to a supervised learning (SLR→T) approach for extrapolating soil classes in two areas (reference and target areas) in central Iran. The SSLR→T used soil observations from the reference area and covariates from both areas. Then, the learned knowledge produced by SSLR→T was transferred to the target area to estimate soil classes. The findings revealed that SSLR→T resulted in higher overall accuracy (0.65) and kappa index (0.44) in the target area compared to the SLR→T (overall accuracy = 0.40 and kappa index = 0.18). Furthermore, the SSLR→T produced the lower values of the confusion index (mean = 0.66) compared to the SLR→T (mean = 0.80). This indicated that the SSLR→T could not only increase the accuracy but also decrease the uncertainty of the soil class predictions, compared to the spatial extrapolation predictions derived from the SLR→T. Generally, these findings indicated that leveraging covariate information from the target area during the training of DSM models in the reference area could successfully improve the generalization power of the models, indicating the effectiveness of SSLR→T for spatial extrapolation.

Digital soil mapping of copper in Sweden: Using the prediction and uncertainty as decision support in crop micronutrient management

Digital soil mapping (DSM) of topsoil copper (Cu) concentrations and prediction intervals covering 90% of agricultural land in Sweden was performed, in order to identify areas at risk of Cu deficiency. A total of 12,527 soil samples were used to calibrate the DSM model, using airborne gamma radiation data, climate data, topographical data and soil texture class data. Among the samples included, 11,093 had no laboratory-analysed Cu concentrations, so their Cu concentrations were predicted using portable X-ray fluorescence (PXRF) measurements. Cross-validation of the PXRF model resulted in Nash-Sutcliffe model efficiency coefficient (E) of 0.66 and mean absolute error (MAE) of 3.3 mg kg−1. Cross-validation of the DSM model showed somewhat lower performance (E = 0.57, MAE = 4.1 mg kg−1). Based on the lower bound of the prediction interval (5th percentile), 48% of agricultural soils in Sweden are most likely not at risk of Cu deficiency (>7 mg kg−1). The Cu map was also validated against concentrations in soil samples from five fields (25–47 ha in size; four samples per ha). The field means were predicted with a MAE of 1.0 mg kg−1 and within-field variation was reproduced with a field-wise squared Pearson correlation coefficient (r2) of 0–0.36. The classification metric ‘recall’ showed that the map of soil Cu concentrations might not predict all possible areas at risk of being Cu deficient, as observational data indicates that about 22% of soils in the mapped area should have Cu concentrations below the risk limit. However, the metric ‘precision’ showed that when the soil map predicted a concentration at or below 7 mg kg−1, it was generally correct. Increasing the limit resulted in the recall and precision increasing rapidly. The remaining 52% of agricultural soils at risk of being below the Cu concentration limit can be targeted by laboratory analysis or monitoring.

Effects of multi‐temporal environmental variables on SOC spatial prediction models in coastal wetlands of a Chinese delta

AbstractMapping soil organic carbon (SOC) distributions in coastal wetlands plays an important role in assessing ecosystem services and investigating the global carbon cycle. Little research has explored the effects of environmental variables with seasonal variations on digital soil mapping (DSM). Our research utilized machine learning methods and established multiple prediction models of SOC based on multi‐temporal data from dry and wet seasons, and mono‐temporal data from April. The results showed that the relationships between SOC and environmental variables in different months varied significantly in coastal wetlands of the Yellow River Delta (YRD). In general, the environmental variables in the wet season showed stronger correlations and higher importance scores with SOC compared with those in the dry season. In addition, SOC prediction models in wet season and April had stronger prediction performance compared with those in the dry season. As a result, data fusion of multi‐temporal data did not necessarily contribute to the model performance enhancement. Relative homogenous soil‐landscape attributes and spectral characteristics in dry season could not accurately explain the strong spatial variation of SOC in this area, and it might be the major reason that caused the stronger model performance of soil prediction models in wet season than those in dry season. Therefore, the accurate spatial prediction of soil properties requires the characterization of the seasonal dynamics of soil‐landscape relationships. In general, the findings of this research demonstrated that seasonal variation of environmental variables should be considered in the establishment of a DSM model in coastal wetland.

Improving the Prediction of Soil Organic Matter in Arable Land Using Human Activity Factors

Detailed spatial distribution of soil organic matter (SOM) in arable land is essential for agricultural management and decision making. Based on digital soil mapping (DSM) theory, much attention has been focused on the selection of environmental covariates. However, the importance of human activity factors in SOM prediction has not received enough attention, especially in arable soil. Moreover, due to the insufficient amount of soil sampling data used to train and validate the DSM model, the prediction results may be questionable, and some even contradictory. This paper explores the effectiveness of the human footprint, amount of fertilizer application, agronomic management level, crop planting type, and irrigation guarantee degree in SOM mapping of arable land in Heilongjiang Province. The results show that the model only including environmental covariates accounts for 41% of the variation in SOM distribution. The model combining the five human activity factors increases the SOM spatial prediction by 39% in terms of R2 (coefficient of determination), 12% in terms of RMSE (root mean square error), 15% in terms of MAE (mean absolute error), and 11% in terms of LCCC (Lin’s concordance correlation coefficient), showing better prediction accuracy and performance. This indicates that human activity factors play a crucial role in determining SOM distribution in arable land. In the SOM prediction, soil moisture is the most important environmental covariate, and the amount of fertilizer application with a relative importance of 11.36% (ranking 3rd) is the most important human activity factor, higher than the annual average precipitation and elevation. From a spatial point of view, the Sanjiang Plain is a difficult area for prediction.

Machine Learning Techniques for Estimating Hydraulic Properties of the Topsoil across the Zambezi River Basin

It is critical to produce more crop per drop in an environment where water availability is decreasing and competition for water is increasing. In order to build such agricultural production systems, well parameterized crop growth models are essential. While in most crop growth modeling research, focus is on gathering model inputs such as climate data, less emphasis is paid to collecting the critical soil hydraulic properties (SHPs) data needed to operate crop growth models. Collection of SHPs data for the Zambezi River Basin (ZRB) is extremely labor-intensive and expensive, thus alternate technologies such as digital soil mapping (DSM) must be explored. We evaluated five types of DSM models to establish the best spatially explicit estimates of the soil water content at pF0.0 (saturation), pF2.0 (field capacity), and pF4.2 (wilting point), and of the saturated hydraulic conductivity (Ksat) across the ZRB by using estimates of locally calibrated pedotransfer functions of 1481 locations for training and testing the DSM models, as well as a reference dataset of measurements from 174 locations for validating the DSM models. We produced coverages of environmental covariates from various source datasets, including climate variables, soil and land use maps, parent materials and lithologic units, derivatives of a digital elevation model (DEM), and Landsat imagery with a spatial resolution of 90 m. The five types of models included multiple linear regression and four machine learning techniques: artificial neural network, gradient boosted regression trees, random forest, and support vector machine. Where the residuals of the initial DSM models were spatially autocorrelated, the models were extended/complemented with residual kriging (RK). Spatial autocorrelation in the model residuals was observed for all five models of each of the three water contents, but not for Ksat. On average for the water content, the R2 ranged from 0.40 to 0.80 in training and test datasets before adding kriged model residuals and ranged from 0.80 to 0.95 after adding model residuals. Overall, the best prediction method consisted of random forest as the deterministic model, complemented with RK, whereby soil texture followed by climate and topographic elevation variables were the most important covariates. The resulting maps are a ready-to-use resource for hydrologists and crop modelers to aliment and calibrate their hydrological and crop growth models.

Impacts of Soil Properties, Topography, and Environmental Features on Soil Water Holding Capacities (SWHCs) and Their Interrelationships

Soil water holding capacities (SWHCs) are among the most important factors for understanding the water cycle in forested catchments because they control available plant water that supports evapotranspiration. The direct determination of SWHCs, however, is time consuming and expensive, so many pedotransfer functions (PTFs) and digital soil mapping (DSM) models have been developed for predicting SWHCs. Thus, it is important to select the correct soil properties, topographies, and environmental features when developing a prediction model, as well as to understand the interrelationships among variables. In this study, we collected soil samples at 971 forest sites and developed PTF and DSM models for predicting three kinds of SWHCs: saturated water content (θS) and water content at pF1.8 and pF2.7 (θ1.8 and θ2.7). Important explanatory variables for SWHC prediction were selected from two variable importance analyses. Correlation matrix and sensitivity analysis based on the developed models showed that, as the matric suction changed, the soil physical and chemical properties that influence the SWHCs changed, i.e., soil structure rather than soil particle distribution at θS, coarse soil particles at θ1.8, and finer soil particle at θ2.7. In addition, organic matter had a considerable influence on all SWHCs. Among the topographic features, elevation was the most influential, and it was closely related to the geological variability of bedrock and soil properties. Aspect was highly related to vegetation, confirming that it was an important variable for DSM modeling. Information about important variables and their interrelationship can be used to strengthen PTFs and DSM models for future research.