Digital Soil Mapping Techniques Research Articles

Prediction of soil organic carbon using machine learning techniques and geospatial data for sustainable agriculture

Sustainable agriculture revolves around soil organic carbon (SOC), which is essential for numerous soil functions and ecological attributes. Farmers are interested in conserving and adding additional soil organic carbon to certain fields in order to improve soil health and productivity. The relationship between soil and environment that has been discovered and standardized throughout time has enhanced the progress of digital soil-mapping techniques; therefore, a variety of machine learning techniques are used to predict soil properties. Studies are thriving at how effectively each machine learning method maps and predicts SOC, especially at high spatial resolutions. To predict SOC of soil at a 30 m resolution, four machine learning models—Random Forest, Support Vector Machine, Adaptive Boosting, and k-Nearest Neighbour were used. For model evaluation, two error metrics, namely R2 and RMSE have been used. The findings demonstrated that the calibration and validation sets’ descriptive statistics sufficiently resembled the entire set of data. The range of the calculated SOC content was 0.06 to 1.76 %. According to the findings of the study, Random Forest showed good results for both cases, i.e. evaluation using cross validation and without cross validation. Using cross validation, RF confirmed highest R2 as 0.5278 and lowest RMSE as 0.1683 for calibration dataset while without cross validation it showed R2 as 0.8612 and lowest RMSE as 0.0912 for calibration dataset. The generated soil maps will help farmers adopt precise knowledge for decisions that will increase farm productivity and provide food security through the sustainable use of nutrients and the agricultural environment.

High-resolution digital mapping of soil erodibility in China

Soil erodibility (K) is the intrinsic susceptibility of a soil to water erosion. Currently, its detailed and accurate spatial distribution information especially over large areas is urgently required for national and regional soil erosion assessment and soil conservation decision making. This study combined pedotransfer function with digital soil mapping techniques to develop a high-resolution map of soil erodibility across China. The First, based on a recent national soil survey, we adopted the erosion-productivity impact calculator (EPIC) to calculate soil erodibility values at 4710 soil sampling points. Then, with the caclulated values of points, we used five techniques including polygon linking (PL), ordinary kriging (OK), Cubist, extreme gradient boosting (XGBoost), and random forests (RF) to generate spatial distribution of soil erodibility. The three latter machine learning techniques modeled the quantitative relationships between soil erodibility and a set of environmental covariates. The results showed that machine learning methods exhibited much more spatial details than the PL and OK did. Among the five techniques the RF achieved the highest accuracy with R2 of 0.49 and RMSE of 0.0077 t ha h ha−1 MJ−1 mm−1 based on 10-fold cross-validation. Spatial uncertainty analysis of the RF predictions showed that high uncerntainty occurred in northwestern China and low uncertainty in the center and southeast. We found that topographical and climatic variables are major environmental factors indirectly controlling spatial variation of soil erodibility while the soil particle composition and SOC contents directly influence the variation.

Digital Mapping of Soil pH and Electrical Conductivity Using Geostatistics and Machine Learning

This study investigates the spatial variability of soil pH and electrical conductivity (EC) in Suryapet district of Southern Telangana Zone through various digital soil mapping approaches. The 202 surface (0-15cm) soil samples were collected and analysed for pH and EC. The analysed data was further divided into calibration set and validation set in the ratio of 75:25. The geostatistical techniques like Ordinary Kriging, Inverse Distance Weighting (IDW) and Regression Kriging and data mining technique like random forest technique were used to predict the spatial distribution of pH and EC (dSm-1) over the study area. The accuracy of these methods was assessed using validation data set by calculating RMSE, ME and R2 values. The results showed that among all the approaches, random forest (RF) technique performed better with lower RMSE, ME and higher R2 values for spatial prediction of soil pH (RMSE=0.014, ME=0.28 and R2=0.81) and EC (RMSE=0.134, ME=0.022 and R2=0.73). The RF predicted maps show that the pH of soils varied from neutral (6.5-7.5) to slightly alkaline (7.5-8.5) and the soils of Suryapet district were considered as non-saline (EC: 0-2 dSm-1). The findings of the current study shows that among digital soil mapping techniques, random forest model can be an effective tool for assessing spatial variability of soil pH and EC for further studies.

The Montes del Plata Soil Spatial Inference System

Soil is one of the prior components of a forest site, hence determining its final productivity. Effective soil management requires an understanding of soil distribution patterns within the landscape. The ability to accurately predict and map soil properties over Montes del Plata (MDP) landholdings can improve silvicultural management decisions that can result in higher and sustainable forest productivity. Today's decisions are based mainly on interpreting conventional soil maps (discrete semi-detailed) produced from local empirical experts' knowledge. Creation of conventional soil maps is a high time-consumable non-parametrizable process which result does not represent soil heterogeneous diversity, both in location and depth. In this work, we applied Digital Soil Mapping (DSM) techniques to produce continuous soil maps showing different soil-predicted property values with a concomitant measured uncertainty grade over 1,200 hectares. We were able to produce continuous soil properties maps with an average precision accuracy of 60% by developing a DSM protocol in three representative farms of MDP landholdings, in areas to be forested with Eucalyptus on its second cycles. A set of 53 environmental covariates was developed and harmonized at 30 m2 of spatial resolution. A field sampling strategy was defined based on the Latin Hypercube sampling method, and 180 soil profiles were collected. In laboratory conditions, we perform textural and chemical analyses on 345 soil samples from A and B horizons. Vertical harmonization was applied, and properties were estimated at three standardized depths (0-30, 30-60, and 60-100). A Random Forests Model (RMF) is fitted to predict texture, organic matter, organic carbon, pH_H2O, pH_KCl, N, P, K, Ca, Mg, Al, Na, cation exchange capacity, base saturation, and profile soil depth. Our results demonstrate that DSM techniques can be successfully applied to representative MDP's landholding soils. We anticipate our preliminary work as the starting point for using DSM techniques that result in acceptable soil continuous maps in less time. However, the developed protocol performance must be tested in similar soil conditions, and an approach must be undertaken to look for a lesser sampling density. These results, in addition to focusing on only those soil parameters linked to operative silvicultural decisions (viz. P, K, Ca, and soil depth), may lead to lower final associated costs.

Spatial assessment of the soil organic carbon content under different types of land use in the Ohrid valley

Spatial assessment of key soil properties is a basic prerequisite for the evidence-based decision making and sustainable use and management of soil. The aim of this work was to estimate the spatial distribution of SOC under different types of land use, by the means of Digital Soil Mapping techniques. A site-specific soil data collection for the Ohrid valley was integrated with continuous and discrete datasets of environmental covariates, serving as predictors. The selected test area outlines the variability of factors influencing the SOC content and spatial distribution. Soil sampling locations were randomly distributed within a predefined mesh with a 1-sq.km spatial resolution and further stratified to outline different types of land use within each mash square. Soil samples were collected from 93 locations at three depths, each 20 cm apart, covering the total area of 10 thousand ha of arable land, forestland, and land under natural vegetation. A set of additional environmental dataset was collected, namely the soil map, land use map, geology map, digital terrain model and its derivatives, satellite images, climate data, as well as a set of indices NDVI, SAVI, BI etc., developed from the remote sensing datasets. Multiple linear regression was used for evaluating the regression pattern between the environmental predictors and the target variable. To estimate spatial variability, several regression tree methods were used. The results obtained using this approach have given a better spatial overview of the most vulnerable areas regarding SOC depletion. Out of 21 locations examined, the content of soil organic carbon in the top layer (0-20 cm.) of forestland was on average 6.81%, while at 22 locations examined under grassland, the average content was 4.07%. The arable land, which is under continuous human impact, had the lowest content of SOC of 2.5% under field crops and 2.61% under perennials.

Digitally mapping soil carbon of the uThukela headwater catchment in the Maloti-Drakensberg, a remote Afromontane mountain region

ABSTRACT Although regional soil mapping is commended, site-specific studies are required for mapping and quantifying soil organic carbon (SOC) stocks at a landscape scale for management and rehabilitation purposes. Site-specific studies are especially important in remote mountainous areas where soil data are largely absent. The aim of this study was to quantify and map the SOCs of an alpine region in the Maloti-Drakensberg. The samples collected in-field along with digital soil mapping techniques were used to map the SOCs. The models were SoLIM’s rule based (RB) and sample based (SB), random forest (RF), least absolute shrinkage and selection operator (LASSO), regression kriging with cubist (RK-CB) and universal kriging. From the results, the mean SOC for the validation dataset of the study area was 12.44 kg organic carbon (OC) m−2. The best model was RK-CB (mean = 12.43 kg OC m−2), with R 2 and root mean square error (RMSE) of 0.61 and 4.01 kg OC m−2, respectively. The underperforming model was SoLIM-RB (mean = 13.27 kg OC m−2), with R 2 = 0.27 and RMSE = 5.75 kg OC m−2. The RK-CB model from this study significantly outperformed a region-scaled model, proving that site-specific studies in small catchments should be preferred to, especially if there are no soil data available for that area.

Predictive geochemical mapping using machine learning in western Kenya

Digital soil mapping techniques represent a cost-effective method for obtaining detailed information regarding the spatial distribution of chemical elements in soils. Machine learning (ML) algorithms using random forest (RF) models have been developed for classification, pattern recognition and regression tasks, they are capable of modelling non-linear relationships using a range of datasets, identifying hierarchical relationships, and determining the importance of predictor variables. In this study, we describe a framework for spatial prediction based on RF modelling where inverse distance weighted (IDW) predictors are used in conjunction with ancillary environmental covariates. The model was applied to predict the total concentration (mg kg−1) and assess the prediction uncertainty of 56 elements, soil pH and organic matter content using 466 soil samples in western Kenya; the results of iodine (I), selenium (Se), zinc (Zn) and soil pH are highlighted in this work. These elements were selected due to contrasting biogeochemical cycles and widespread dietary deficiencies in sub-Saharan Africa, whilst soil pH is an important parameter controlling soil chemical reactions. Algorithm performance was evaluated determining the relative importance of each predictor variable and the model's response using partial dependence profiles. The accuracy and precision of each RF model were assessed by evaluating out-of-bag predicted values. The models R2 values range from 0.31 to 0.64 whilst CCC values range from 0.51 to 0.77. The IDW predictor variables had the greatest impact on assessing the distribution of soil properties in the study area, however, the inclusion of ancillary environmental data improved model performance for all soil properties. The results presented in this paper highlight the benefits of ML algorithms which can incorporate multiple layers of data for spatial prediction, uncertainty assessment and attributing variable importance. Additional research is now required to ensure health practitioners and the agri-community utilise the geochemical maps presented here for assessing the relationship between environmental geochemistry, endemic diseases and preventable micronutrient deficiency.

CLSoilMaps: A national soil gridded database of physical and hydraulic soil properties for Chile

Spatially explicit soil information is crucial for comprehending and managing many of Earth´s processes related to carbon, water, and other biogeochemical cycles. We introduced a gridded database of soil physical properties and hydraulic parameters at 100 meters spatial resolution. It covers the continental area of Chile and binational basins shared with Argentina for six standardized depths following the specifications of the GlobalSoilMap project. We generated soil maps based on digital soil mapping techniques based on more than 4000 observations, including unpublished data from remote areas. These maps were used as input for the pedotransfer function Rosetta V3 to obtain predictions of soil hydraulic properties, such as field capacity, permanent wilting point, total available water capacity, and other parameters of the water retention curve. The trained models outperformed several other DSM studies applied at the national and regional scale for soil physical properties (nRMSE ranging from 6.93% to 15.7%) and delivered acceptable predictions (nRMSE ranging from 10.4% to 15.6%) for soil hydraulic properties, making them suitable for countless environmental studies.

Spatial modelling of topsoil properties in Romania using geostatistical methods and machine learning.

Various research topics from the field of soil science or agriculture require digital maps of soil properties as input data. Such maps can be achieved by digital soil mapping (DSM) techniques which have developed consistently during the last decades. Our research focuses on the application of geostatistical methods (including ordinary kriging, regression-kriging and geographically weighted regression) and machine learning algorithms to produce high resolution digital maps of topsoil properties in Romania. Six continuous predictors were considered in our study (digital elevation model, topographic wetness index, normalized difference vegetation index, slope, latitude and longitude). A tolerance test was performed to ensure that all predictors can be used for the purpose of digital soil mapping. The input soil data was extracted from the LUCAS database and includes 7 chemical properties (pH, electrical conductivity, calcium carbonate, organic carbon, N, P, K) and the particle-size fractions (sand, silt, clay). The spatial autocorrelation is higher for pH, organic carbon and calcium carbonate, as indicated by the partial sill / nugget ratio of semivariograms, meaning that these properties are more predictable than the others by kriging interpolation. The optimal DSM method was selected by independent sample validation, using resampled statistics from 100 samples randomly extracted from the validation dataset. Also, an additional independent sample of soil profiles, comprising legacy soil data, and the 200k Romania soil map were used for a supplementary validation. The results show that machine learning and regression-kriging are the optimal methods in most cases. Among the machine learning tested algorithms, the best performance is associated with Support Vector Machines and Random Forests methods. The geographically weighted regression is also among the optimum methods for pH and calcium carbonates spatial prediction. Good predictions were achieved for pH (R2 of 0.417-0.469, depending on the method), organic carbon (R2 of 0.302-0.443), calcium carbonates (R2 of 0.300-0.330) and moderate predictions for electric conductivity, total nitrogen, silt and sand (R2 of 0.155-0.331), while the lowest prediction characterizes the phosphorous content (R2 of 0.015-0.044). LUCAS proved to be a reliable and useful soil database and the achieved spatial distributions of soil properties can be further used for national and regional soil studies.

Mercury Prediction in Urban Soils by Remote Sensing and Relief Data Using Machine Learning Techniques

This article aims to explore the use of machine learning (ML) methods for mapping the distribution of mercury (Hg) content in topsoil, using the city of Ufa (Russia) and adjacent areas as an example. For this purpose, a soil dataset of 250 points sampled from a 0–20 cm depth on different land uses, including residential, industrial and undisturbed (forests and parks), was used. Random Forest (RF), Extreme Gradient Boosting (XGboost), Cubist and k-Nearest Neighbor (kNN) ML techniques were employed to model and map the Hg concentrations. We used remote sensing data (RSD) and topographic attributes as explanatory variables. ML models were calibrated and validated using the leave-one-out cross-validation approach. The Hg content varied from 0.005 to 0.58 mg/kg and was characterized by very high variability. According to the MAE and RMSE metrics, the RF method resulted in the most accurate spatial prediction for the Hg content (0.029 and 0.065 mg/kg, respectively), while the XGBoost approach showed the lowest prediction efficiency (0.032 and 0.073 mg/kg, respectively). The results showed that the slope map, spectral index MSI and Sentinel-2A band B11 were the key variables in explaining the variability of Hg content. We found that higher uncertainty values of soil Hg were found in croplands, urban residential and industrial areas, which supports the view that spatial modelling of HM in urban landscapes is challenging. The present study provides insights into the potential of digital soil mapping techniques in combination with RSD and terrain variables for identifying areas at risk of Hg contamination in urban areas, which can inform land-use planning and management strategies to protect human health and the environment.

Modeling the spatial variation of calcium carbonate equivalent to depth using machine learning techniques

Inorganic carbon is the largest source of carbon in terrestrial surface, particularly in arid and semiarid regions, including the Chahardowli Plain in western Iran. Inorganic carbon plays an equal or greater role than organic soil carbon in these areas, although less attention has been paid in quantifying their variability. The objective of this study was to model and map calcium carbonate equivalent (CCE) presenting inorganic carbon in soil using machine learning and digital soil mapping techniques. Chahardowli Plain in foothills of the Zagros Mountains in the southeast of Kurdistan Province in Iran was taken as a case study area. CCE was measured at 0-5, 5-15, 15-30, 30-60, and 60-100cm depths following GloalSoilMap.net project specifications. A total of 145 samples were collected from 30 soil profiles using the conditional Latin hypercube (cLHS) method of sampling. Relationships between CCE and environmental predictors were modeled using random forest (RF) and decision tree (DT) models. In general, the RF model performed slightly superior than the DT model. The mean value of CCE increased with soil depth, from 3.5% (0-5cm) to 63.8% (30-60cm). Remote sensing (RS) variables and terrestrial variables were equally important. The importance of RS variables was higher at the surface than terrestrial variables, and vice versa. The most significant variables were Channel Network Base Level (CNBL) variable and Difference Vegetation Index (DVI) with the same variable importance value (21.1%). In areas affected by river activities, the use of the CNBL and vertical distance to channel networks (VDCN) as variables in digital soil mapping (DSM) could increase the accuracy of soil property prediction maps. The VDCN played a principal role in soil distribution in the study area by affecting the rate of discharge and, thus, erosion and sedimentation. A high percentage of carbonate in parts of the region could exacerbate nutrient deficiencies for most crops and provide information for sustainably managing agricultural activity.

Digital mapping of the soil available water capacity: tool for the resilience of agricultural systems to climate change

Soil available water capacity (AWC) is a key function for human survival and well-being. However, its direct measurement is laborious and spatial interpretation is complex. Digital soil mapping (DSM) techniques emerge as an alternative to spatial modeling of soil properties. DSM techniques commonly apply machine learning (ML) models, with a high level of complexity. In this context, we aimed to perform a digital mapping of soil AWC and interpret the results of the Random Forest (RF) algorithm and, in a case study, to show that digital AWC maps can support agricultural planning in response to the local effects of climate change. To do so, we divided this research into two approaches: In the first approach, we showed a DSM using 1857 sample points in a southeastern region of Brazil with laboratory-determined soil attributes, together with a pedotransfer function (PTF), remote sensing and DSM techniques. In the second approach, the constructed AWC digital soil map and weather station data were used to calculate climatological soil water balances for the periods between 1917–1946 and 1991–2020. The result showed the selection of covariates using Shapley values as a criterion contributed to the parsimony of the model, obtaining goodness-of-fit metrics of R2 0.72, RMSE 16.72 mm m−1, CCC 0.83, and Bias of 0.53 over the validation set. The highest contributing covariates for soil AWC prediction were the Landsat multitemporal images with bare soil pixels, mean diurnal, and annual temperature range. Under the current climate conditions, soil available water content (AW) increased during the dry period (April to August). May had the highest increase in AW (∼17 mm m−1) and decrease in September (∼14 mm m−1). The used methodology provides support for AWC modeling at 30 m resolution, as well as insight into the adaptation of crop growth periods to the effects of climate change.

Assessment of macronutrients status using digital soil mapping techniques: a case study in Maru'ak area in Lorestan Province, Iran.

The present study was conducted to compare generalized linear model (GLM), random forest (RF), and Cubist to produce available phosphorus (AP) and potassium (AK) maps and to identify the covariates that control mineral distribution in Lorestan Province, Iran. To this end, the locations for collecting 173 soil samples were determined through the conditioned Latin hypercube sampling (cLHS) method, at four different land-uses (orchards, paddy fields, agricultural, and abandoned fields). The performance of the models was assessed by coefficient of determination (R2), root-mean-square error (RMSE), and mean absolute error (MAE) indices. The results showed that the RF model fitted better than GLM and Cubist models and could explain 40 and 57% of AP and AK distribution, respectively. The R2, RMSE, and MAE of the RF model were 0.4, 2.81, and 2.43 for predicting AP and equal to 0.57, 143.77, and 116.61 for predicting AK, respectively. The most important predictors selected by the RF model were valley depth and soil-adjusted vegetation index (SAVI) for AP and AK, respectively. The maps showed higher AP and AK content in apricot orchards compared to other land-uses. No difference was observed between AP and AK content on paddy fields, agricultural, and abandoned areas. The higher AP and AK contents were related to orchard management practices, such as failure to dispose of plant residuals and fertilizer consumption. It can be concluded that the orchards (by increasing soil quality) was the best land-use in line with sustainable management for the study area. However, generalizing the results needs more detailed research.

Mapping Brazilian soil mineralogy using proximal and remote sensing data

Minerals control many soil functions and play a crucial role in addressing global existential issues. Measuring the abundance of soil minerals is a laborious, costly, and time-consuming task; however, soil spectroscopy can be a useful tool to overcome this issue. This work aimed to map the abundance of major mineralogical components of soils in Brazil from surface to 1 m deep and at a spatial resolution of 30 m. Spectral data of the Brazilian Soil Spectral Library with Vis-NIR-SWIR was used to estimate the abundance of haematite, goethite, kaolinite, and gibbsite. These minerals were spatialized using digital soil mapping techniques. We also developed a novel framework to obtain bare soil reflectance for areas without natural or anthropic soil exposure (continuous image) and used it as covariate. Soil minerals and their abundances were successfully estimated by Vis-NIR-SWIR reflectance. Haematite predictions presented the most accurate results with Random Forest models, followed by gibbsite, kaolinite, and goethite. The spatial validation with reference mineralogical data found R2 of 0.64 (haematite), 0.40 (goethite), 0.20 (kaolinite/Kt), 0.29 (gibbsite/Gbs), and 0.40 (Kt/Kt + Gbs). The resulting maps of soil minerals were in accordance with the geology, pedology, climate, and relief of Brazil and revealed the spatial distribution of mineral abundances at a finer resolution than existing geological and pedological maps, reaching a farm level detail.

Comparison of Machine Learning-Based Prediction of Qualitative and Quantitative Digital Soil-Mapping Approaches for Eastern Districts of Tamil Nadu, India

The soil–environmental relationship identified and standardised over the years has expedited the growth of digital soil-mapping techniques; hence, various machine learning algorithms are involved in predicting soil attributes. Therefore, comparing the different machine learning algorithms is essential to provide insights into the performance of the different algorithms in predicting soil information for Indian landscapes. In this study, we compared a suite of six machine learning algorithms to predict quantitative (Cubist, decision tree, k-NN, multiple linear regression, random forest, support vector regression) and qualitative (C5.0, k-NN, multinomial logistic regression, naïve Bayes, random forest, support vector machine) soil information separately at a regional level. The soil information, including the quantitative (pH, OC, and CEC) and qualitative (order, suborder, and great group) attributes, were extracted from the legacy soil maps using stratified random sampling procedures. A total of 4479 soil observations sampled were non-spatially partitioned and intersected with 39 environmental covariate parameters. The predicted maps depicted the complex soil–environmental relationships for the study area at a 30 m spatial resolution. The comparison was facilitated based on the evaluation metrics derived from the test datasets and visual interpretations of the predicted maps. Permutation feature importance analysis was utilised as the model-agnostic interpretation tool to determine the contribution of the covariate parameters to the model’s calibration. The R2 values for the pH, OC, and CEC ranged from 0.19 to 0.38; 0.04 to 0.13; and 0.14 to 0.40, whereas the RMSE values ranged from 0.75 to 0.86; 0.25 to 0.26; and 8.84 to 10.49, respectively. Irrespective of the algorithms, the overall accuracy percentages for the soil order, suborder, and great group class ranged from 31 to 67; 26 to 65; and 27 to 65, respectively. The tree-based ensemble random forest and rule-based tree models’ (Cubist and C5.0) algorithms efficiently predicted the soil properties spatially. However, the efficiency of the other models can be substantially increased by advocating additional parameterisation measures. The range and scale of the quantitative soil attributes, in addition to the sampling frequency and design, greatly influenced the model’s output. The comprehensive comparison of the algorithms can be utilised to support model selection and mapping at a varied scale. The derived digital soil maps will help farmers and policy makers to adopt precision information for making decisions at the farm level leading to productivity enhancements through the optimal use of nutrients and the sustainability of the agricultural ecosystem, ensuring food security.

South Africa needs a hydrological soil map: a case study from the upper uMngeni catchment

Accurate hydrological modelling to evaluate the impacts of climate and land use change on water resources is pivotal to sustainable management. Soil information is an important input in hydrological models but is often not available at adequate scale with appropriate attributes for direct parameterisation of the models. In this study, conducted in three quaternary catchments in the midlands of KwaZulu-Natal, three different soil information sets were used to configure SWAT+, a revised version of the Soil and Water Assessment Tool (SWAT). The datasets were: (i) the Land Type database (currently the only soil dataset covering the whole of South Africa), (ii) disaggregation of the Land Type database using digital soil mapping techniques (called DSMART), and (iii) a dataset where DSMART were complemented by field observations and interpretations of the hydropedological behaviour of the soils (DSMART+). Simulated streamflow was compared with measured streamflow at three weirs with long-term measurements, and the impact of the soil datasets on water balance simulations was evaluated. In general, the simulations were acceptable when compared to other studies, but could be improved through calibration and including small reservoirs in the model. The DSMART+ dataset yielded more accurate simulations of streamflow in all three catchments with Nash-Sutcliffe efficiencies increasing by between 9% and 67% when compared to the Land Type dataset. The value of the improved soil maps is, however, highlighted through the enhanced spatial detail of streamflow generation mechanisms and water balance components. The internal catchment processes are represented more accurately, and we argue that South Africa needs a detailed hydrological soil map for effective water resource management.

Spatial prediction of soil properties through hybridized random forest model and combination of reflectance spectroscopy and environmental covariates

Spatial information on land and soil resources are critical towards addressing land degradation for ensuring sustainable soil and crop management. To address these needs, digital soil mapping techniques have emerged as an efficient and low-cost solution. Although digital soil mapping has typically leveraged geospatial environmental variables (e.g. remote sensing), the application and integration of spectroscopic data with those environmental variables remain limited. Hence, this study combines visible and near-infrared (Vis-NIR) spectroscopy, remote sensing, and topographic data and applies random forests, hybridized with particle swarm optimization algorithm (RF + PSO), to predict the spatial variability of soil clay content, electrical conductivity (EC), and calcium carbonate equivalent (CCE) for 370 km2 of agricultural land in western Iran. Using a conditioned Latin hypercube approach, 220 soil samples at the 0–20 cm depth increment were acquired throughout the study area. Three sets of environmental covariates were tested: Scenario A (Vis-NIR spectroscopy data), Scenario B (environmental data), and Scenario C (Vis-NIR spectroscopy + environmental data). According to the 10-fold cross-validation procedure with 100 replications, the RF + PSO model showed an acceptable level accuracy for all scenarios, although the accuracy of the RF + PSO model using the Scenario C data was higher than all other scenarios: the Lin’s Concordance Correlation Coefficient values were 0.77, 0.83, and 0.74 for the clay contents, EC, and CCE, respectively. The results demonstrated that the combination of Vis-NIR spectroscopic data and commonly available environmental covariates provided the best input data for the hybridized model and enhanced its performance.

Digital mapping of soil pH and carbonates at the European scale using environmental variables and machine learning

Soil pH and carbonates (CaCO3) are important indicators of soil chemistry and fertility, and the prediction of their spatial distribution is critical for the agronomic and environmental management. Digital soil mapping (DSM) techniques are widely accepted for the geospatial analysis of the soil properties. They are rapid and cost-efficient approaches that can provide quantitative prediction. However, the digital mapping of soil pH and CaCO3 are not well studied, especially at a continental scale. In this research, we mapped the soil pH and CaCO3 at the European scale using multisource environmental variables and machine learning approaches. Moderate Resolution Imaging Spectroradiometer (MODIS) products, terrain attributes, and climatic variables were considered. Meanwhile, nine machine learning algorithms, namely, three linear and six nonlinear models, were used for the spatial prediction of soil pH and CaCO3. The land use and cover area frame statistical survey (LUCAS) 2015 topsoil dataset provided by the European Soil Data Centre was utilised. The performances of different models were compared and analysed in terms of coefficient of determination (R2), root mean square error (RMSE), and ratio of performance to deviation (RPD). Specifically, nonlinear machine learning models outperformed the linear ones, and extremely randomized trees (ERT) gave the most satisfactory result for soil pH (R2 = 0.70, RMSE = 0.75, and RPD = 1.84) and CaCO3 (R2 = 0.53, RMSE = 93.49 g/kg, and RPD = 1.46). The results revealed that MODIS products and climatic variables were important in predicting soil pH and CaCO3. Moreover, spatial distribution of soil pH and CaCO3 in Europe were mapped at 250 m resolution, and the areas with high CaCO3 content always showed high soil pH value.

Modeling of total and active organic carbon dynamics in agricultural soil using digital soil mapping: a case study from Central Nova Scotia

Monitoring the changes in soil organic carbon (SOC) pools is critical for sustainable soil and agricultural management. This case study models total and active organic carbon dynamics (2015/2016 to 2019/2020) using digital soil mapping (DSM) techniques. Model predictors include topographic variables generated from light detection and ranging data; soil and vegetation indices derived from Landsat satellite images; and soil and crop inventory information from Agriculture and Agri-Food Canada to predict total organic carbon (TOC) and permanganate oxidizable carbon (POXc) at the 0–15 cm depth increment for a 37 km2 study area in Truro, Nova Scotia. Quantile Regression Forest and stochastic Gradient Boosting Model were utilized for prediction. Although both models performed equally well for predicting TOC and POXc, the accuracy of TOC predictions (e.g., concordance correlation coefficient (CCC) = 0.67) was better than POXc predictions (e.g., CCC = 0.53). The Landsat variables and crop inventory were dominant predictors, while topographic variables across the relatively homogeneous terrain had relatively little influence. During the study period, changes in POXc were predicted across 98% of the study area, with a mean absolute loss of 5.77 (±11.48) mg/kg/year, and in TOC on 27% of the area, with a mean absolute loss of 0.15 (±0.09) g/kg/year. While the annual crop fields observed the highest loss of TOC and POXc, the decline in pasture–grassland–forage fields was relatively low. The study reinforced the effectiveness of DSM for modeling multiple SOC pools at the farm to landscape scales.

Effects of sample size and covariate resolution on field-scale predictive digital mapping of soil carbon

Soil carbon is a key soil property that regulates numerous soil processes and affects soil chemical, physical and biological properties. Given its importance, prediction of soil carbon using digital soil mapping (DSM) techniques has seen increased interest at scales ranging from global to local. The accuracy of predictive models is reliant on sample size, covariate resolution, and the algorithms being used to predict the target variable. Despite this, study of these components within a DSM workflow has received considerably less attention in the scientific literature. In this study, we examined the effects of sample size and covariate resolution on the performance of the Cubist and random forest (RF) algorithms, and compared these to ordinary kriging, for predicting total soil carbon in the 0–15 cm depth interval of a 26-hectare agricultural field in southwestern Ontario, Canada, originally sampled using a 20-m grid (689 samples). Using the conditioned Latin hypercube sampling algorithm, we created replicated sampling plans of increasing size (10, 25, and then 50 to 600 in steps of 50) across six different resolutions of environmental covariates. Overall, the kriging model marginally outperformed both the Cubist and RF models. All three models showed quick increases in model performance indicators (Lin’s concordance correlation coefficient, CCC; root mean square error, RMSE) at the smallest sample sizes with diminishing returns thereafter. Using the unit invariant knee technique, we determined the optimal sample size based on CCC and RMSE to be between 100 and 150 samples. We also observed no trend in optimal sample size across the six covariate resolutions for CCC or RMSE. The optimal sample sizes were confirmed by a review of the predicted carbon maps and the lack of change in the spatial distribution of soil carbon after reaching the optimal sample size. Effects of covariate resolution might not be apparent in this study given the field-scale nature of the work and narrow range of resolutions; however, we recommend this be evaluated for larger study areas to gain more insight about the interaction of sample size and covariate resolution.