Conditioned Latin Hypercube Sampling 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.

Soil sampling design matters - Enhancing the efficiency of digital soil mapping at the field scale

Optimisation of sampling design (methods chosen to select the samples) and sample size (number of samples) remains a key challenge in digital soil mapping, especially in the area of precision farming with the expected economic benefits from the introduction of new technologies. As the existing information is available in the form of relevant environmental covariates, its combination with non-parametric machine learning techniques requires careful planning from the initial field sampling to the final production of digital soil maps. The aim of this study is to compare widely used covariate-wise sampling designs combined with variable sample sizes for supervised prediction of common soil drivers of agricultural productivity (pH, soil organic carbon, soil macronutrients) in a real case study of a field (35 ha) with heterogeneous soil properties. From a total of 200 samples, we evaluated different sample sets where 10, 30 and 60 field samples were selected by conditioned Latin Hypercube Sampling (cLHS) and Feature Space Coverage Sampling (FSCS) to calibrate random forest (RF) models. The evaluation was performed on independently in-situ sampled test points. In addition to these datasets, we also compared the investigated methods with Simple Random Sampling (SRS) in a numerical benchmark experiment with increasing sample size, comparing the global accuracies of the predicted maps on the test points, but using interpolated maps as the artificial true population for each soil characteristic. The results of the study in both the field experiment and the numerical experiment showed slightly better results for the FSCS method, especially when the number of samples was small. At smaller training sample sizes, the risk of insufficiently accurate prediction models was slightly lower for FSCS and the difference decreased as the sample size increased. Nevertheless, sample size proved to be the most important factor in the accuracy of RF models, regardless of the sampling technique. The results suggest that a sample size between 18 and 30 training samples (0.6 to 1 sample ha−1) seems plausible for covariate-wise predictions using RF at field scale in our case study. The relative importance of each auxiliary variable for each RF calibration was also assessed for the field experiment. The results showed that the additional introduction of spatial proxies overshadowed the importance of other covariates, but only significantly improved the model calibration at larger sample sizes. The calibrated models without spatial proxies showed the strongest effect of remotely sensed surface characteristics.

A conditioned Latin hypercube sampling design methodology for ground-truthing transient EM resistivity models

Translating 3D resistivity models of the subsurface to relevant hydrogeological parameters such as sediment texture requires applying a rock physics transform that is ideally calibrated using co-located 1D sediment texture logs. It is difficult to select optimal logging sites in the model that span the range of observed resistivity values using expert opinion alone. Instead, we demonstrate a methodology for selecting logging sites using conditioned Latin hypercube sampling (cLHS), a maximally stratified sampling design methodology. cLHS employs simulated annealing to select sampling sites that minimize an objective function by reproducing the statistics of the resistivity model. We implement numerous preprocessing steps to isolate desirable regions within the model for logging, and we use principal component analysis to reduce the dimensionality of the model to aid the convergence of the simulated annealing procedure while preserving the majority of the model variance. Additionally, we impose an additional objective function condition to penalize the selection of sites in the resistivity model that exhibit a high degree of lateral variance. This condition is used to increase the likelihood that the sediment texture logs acquired at the selected sites correlate well with the resistivity model. We tested our approach on a resistivity model obtained using a towed transient EM system in an almond orchard in California’s Central Valley, comparing our selected sets of logging sites against the five cone penetrometer testing logs originally acquired at the site, as well as randomly selected logging sites. The sample sets selected with cLHS were consistently able to reproduce the statistics of the resistivity model and did so more effectively than the true sampling locations or the random sampling locations.

Sample Size Optimization for Digital Soil Mapping: An Empirical Example

In the evolving field of digital soil mapping (DSM), the determination of sample size remains a pivotal challenge, particularly for large-scale regional projects. We introduced the Jensen-Shannon Divergence (DJS), a novel tool recently applied to DSM, to determine optimal sample sizes for a 2790 km2 area in Ontario, Canada. Utilizing 1791 observations, we generated maps for cation exchange capacity (CEC), clay content, pH, and soil organic carbon (SOC). We then assessed sample sets ranging from 50 to 4000 through conditioned Latin hypercube sampling (cLHS), feature space coverage sampling (FSCS), and simple random sampling (SRS) to calibrate random forest models, analyzing performance via concordance correlation coefficient and root mean square error. Findings reveal DJS as a robust estimator for optimal sample sizes—865 for cLHS, 874 for FSCS, and 869 for SRS, with property-specific optimal sizes indicating the potential for enhanced DSM accuracy. This methodology facilitates a strategic approach to sample size determination, significantly improving the precision of large-scale soil mapping. Conclusively, our research validates the utility of DJS in DSM, offering a scalable solution. This advancement holds considerable promise for improving soil management and sustainability practices, underpinning the critical role of precise soil data in agricultural productivity and environmental conservation.

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.

A Proposed Methodology for Determining the Economically Optimal Number of Sample Points for Carbon Stock Estimation in the Canadian Prairies

Soil organic carbon (SOC) sequestration assessment requires accurate and effective tools for measuring baseline SOC stocks. An emerging technique for estimating baseline SOC stocks is predictive soil mapping (PSM). A key challenge for PSM is determining sampling density requirements, specifically, determining the economically optimal number of samples for predictive soil mapping for SOC stocks. In an attempt to answer this question, data were used from 3861 soil organic carbon samples collected as part of routine agronomic soil testing from a 4702 ha farming operation in Saskatchewan, Canada. A predictive soil map was built using all the soil data to calculate the total carbon stock for the entire study area. The dataset was then subset using conditioned Latin hypercube sampling (cLHS), both conventional and stratified by slope position, to determine the total carbon stocks with the following sampling densities (points per ha): 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, and 0.8. A nonlinear error function was then fit to the data, and the optimal number of samples was determined based on the number of samples that minimized soil data costs and the value of the soil carbon stock prediction error. The stratified cLHS required fewer samples to achieve the same level of accuracy compared to conventional cLHS, and the optimal number of samples was more sensitive to carbon price than sampling costs. Overall, the optimal sampling density ranged from 0.025 to 0.075 samples per hectare.

On the efficacy of conditioned and progressive Latin hypercube sampling in supervised machine learning

In this paper, Latin Hypercube Sampling (LHS) method is compared as per its effectiveness in supervised machine learning procedures. Employing LHS saves computer's processing time and in conjunction with Latin hypercube design properties and space filling ability, is considered as one of the most advanced mechanisms in terms of sampling. Although more data usually deliver better results, when using LHS techniques, same quality outputs can be produced with less data and, as a result, storage cost and training time are reduced. Conditioned Latin Hypercube Sampling (cLHS) is one of those techniques, successfully performing in supervised machine learning tasks. Unfortunately, the minimum sufficient training dataset size cannot be known in advance. In this case, progressive sampling is recommended since it begins with a small sample and progressively increases its size until model accuracy no longer improves. Combining Latin hypercube sampling and the idea of sequentially incrementing sampling, we test Progressive Latin Hypercube Sampling (PLHS) while monitoring the performance of the sampling-based training as the sample size grows. PLHS and cLHS algorithms are applied in datasets with discrete variables securing that each sample is provided with the Latin hypercube design properties and preserves the principal ability of LHS for space filling, as illustrated in respective sample projecting diagrams. The performance of the above LHS methods in supervised machine learning is evaluated by the degree of training of the model, which is certified through the accuracy of the produced confusion matrices in test files. The results from the use of the above Latin Hypercube Sampling techniques compared against benchmark sampling method empirically prove that machine learning training process becomes less costfull, while remaining reliable.

Unravelling spatial drivers of topsoil total carbon variability in tropical paddy soils of Sri Lanka

This study aimed to map and identify the spatial drivers of total carbon (TC) concentration in topsoil (0–15 cm) across paddy-growing regions in tropical climates using Sri Lanka as a case study. For model calibration, a total of 888 sampling locations were sampled using the conditioned Latin Hypercube sampling approach with a sample density of one sample per 11 km2. Additionally, 99 sampling sites were selected using a design-based (probabilistic) stratified random strategy for independent evaluation of the developed models. Two distinct spatial random forest (RF) models were developed using a variety of environmental covariates: Model 1: using all environmental covariates without variable selection; Model 2: only incorporated covariates selected based on the forward selection process. Evaluation of model quality using fully independent validation sites revealed that both Model 1 and Model 2 performed similarly. Based on the spatial estimates of Model 1 across the paddy-growing regions of Sri Lanka, the predicted TC concentration varied from 0.89% to 13.15%. The highest predicted TC concentration range was in the Wet zone (2.06% to 13.15%), followed by the Intermediate zone (1.18% to 7.23%), and the lowest was reported in the Dry zone (0.86% to 4.30%). In the spatial estimates of Model 2, the predicted values varied between 0.86% and 13.29% and were similar to Model 1. The highest predicted TC concentration range was in the Wet zone (2.09% to 13.29%), followed by the Intermediate zone (1.08% to 6.99%), and the lowest was reported in the Dry zone (0.86% to 4.30%) following the similar pattern to Model 1. In fact, this clearly showed the importance of mean annual rainfall on the dynamics of TC in tropical rice production systems. Furthermore, the variable importance plot of the RF models revealed that out of all considered environmental covariates, the mean annual rainfall was identified as being the most important variable in the developed spatial prediction function. Moreover, we deployed an area of applicability (AOA) calculation to quantify and identify regions where prediction is less reliable and quantified the prediction uncertainty using a bootstrapping approach. Additionally, we assessed the influence of increasing the number of calibration sites on model prediction quality and reliability using user defined sequence of calibration sites. Independent evaluations of each model indicated that model performance quality indices were improved up to n = 400 and thereafter stagnated. For AOA results, an improvement in model reliability is observed for Wet and Intermediate zones when models are developed using 400 calibration sites. Derived estimates of TC can be used for regional-scale planning to enhance the soil carbon and provide a baseline for designing a future land-based carbon accounting system for Sri Lanka.

Optimal sampling using Conditioned Latin Hypercube for digital soil mapping: An approach using Bhattacharyya distance

Soil properties are important because they determine the soil’s suitability for different types of plant growth, ecosystems and biota functioning. Soil properties influence nutrient cycling, carbon sequestration and soil management. Digital Soil Mapping (DSM) is a procedure to map soil properties. Soil sampling for DSM is a foundational step in building prediction accuracy and essential for incorporating variability in terms of environmental covariates (ancillary variables). Conditioned Latin Hypercube (CLH) sampling is a method for generating a sample of points from a multivariate distribution that has been conditioned on one or more covariates. It is an extension of Latin Hypercube sampling, which is a popular technique for generating samples from a multivariate distribution in a way that ensures that each dimension is sampled uniformly. CLH sampling carries the benefit of selecting sampling locations covering the feature space and forming a Hypercube of the original sample. However, determining the optimum sample size is crucial in soil survey exercises constrained by budget and time limits. For this purpose, a study was carried out on Scotland's Finzean Estate (44.8 km2) location. A dataset of 21 independent features (16 continuous and five categorical) and 17,932 sampling locations was created using Digital Elevation Model (DEM) derivatives, soil classes map, land cover map, the peat depth map and parent material map to further generate sub-samples and compare the generated sub-samples with the original population. Two hundred CLH sampling datasets were extracted from the original population (17932 data points) with different sizes (5, 10, 15, 20, …, 100) and each size was given 10 repetitions e.g. (5_1, 5_2, …, 5_10). The sample datasets were analysed by comparing the mean, standard deviation, boxplot and estimates of the probability density function (pdf) for all the 16 continuous independent features. All the mentioned comparisons suggest that the impact of increasing sample size on the distribution of covariates can be observed up to a certain point, beyond which further increases in sample size may not yield noticeable differences. Bhattacharyya distance, a statistical measurement that quantifies the similarity between two probability distributions, was calculated between every quantitative and qualitative element of respective sampling size and for the original population. In contrast, as the CLH sample dataset size increased, the Bhattacharyya distance value decreased and became constant.. The optimum number of samples based on the study was determined for the spatial extent of the Finzean Estate in Scotland and a range of 25–50 CLH samples was suggested based on the study. This work, therefore, achieved both reductions in sampling location numbers compared to classical approaches and identification of the precise location of these sample location to achieve optimal DSM.

Predictive performance of machine learning model with varying sampling designs, sample sizes, and spatial extents

Using machine learning and earth observation data to capture real-world variability in spatial predictive mapping depends on sample size, design, and spatial extent. Nonetheless, there is still ambiguity in answering some basic questions: a) How many samples are necessary for fitting the model? b) Which sampling techniques are suitable for modeling? c) Do results vary with changes in spatial extents? These questions are crucial for spatial modeling projects and require proper investigation. In the present study, we evaluated two sampling designs with different sample sizes, considering three nested spatial extents. Specifically, we adopted the conditioned Latin Hypercube Sampling and Simple Random Sampling designs. Based on this, a Random Forest model was used to predict Above-Ground forest Biomass at local, regional, and national spatial extents, comparing different sample sizes (n = 25, 50, 100, 200, 300, and 500). We defined one national extent, five regional extents within the national extent, and a local extent inside each regional extent. Each sampling design and size combination was tested 100 iterations. The results showed that there was no significant difference between the different sampling designs. The accuracy metrics showed marginal differences for 25 and 50 sample sizes, which were then reduced to minimal and provided similar results. However, a deeper analysis of all 100 repetitions exposed a noteworthy pattern: cLHS outperformed the SRS in terms of RMSE and variability. Regarding the sampling size, the R2 values increased with increasing sample size. Nevertheless, beyond a minimum of 300 to 500 samples, the improvement in accuracy became insignificant, emphasizing the diminishing returns with excessively large sample sizes. Moreover, increasing the size of the spatial extent reduced the accuracy of the model, possibly due to the effect of environmental factors or landscape nature. Therefore, this study demonstrates the potential impact of sample size, sampling design, and spatial extents on model accuracy and emphasizes the importance of reducing the sample size to reduce the model's complexity.

A Multi-Variable Sentinel-2 Random Forest Machine Learning Model Approach to Predicting Perennial Ryegrass Biomass in Commercial Dairy Farms in Southeast Australia

One of the most valuable and nutritionally essential agricultural commodities worldwide is milk. The European Union and New Zealand are the second- and third-largest exporting regions of milk products and rely heavily on pasture-based production systems. They are comparable to the Australian systems investigated in this study. With projections of herd decline, increased milk yield must be obtained from a combination of animal genetics and feed efficiencies. Accurate pasture biomass estimation across all seasons will improve feed efficiency and increase the productivity of dairy farms; however, the existing time-consuming and manual methods of pasture measurement limit improvements to utilisation. In this study, Sentinel-2 (S2) band and spectral index (SI) information were coupled with the broad season and management-derived datasets using a Random Forest (RF) machine learning (ML) framework to develop a perennial ryegrass (PRG) biomass prediction model accurate to +/−500 kg DM/ha, and that could predict pasture yield above 3000 kg DM/ha. Measurements of PRG biomass were taken from 11 working dairy farms across southeastern Australia over 2019–2021. Of the 68 possible variables investigated, multiple simulations identified 12 S2 bands and 9 SI, management and season as the most important variables, where Short-Wave Infrared (SWIR) bands were the most influential in predicting pasture biomass above 4000 kg DM/ha. Conditional Latin Hypercube Sampling (cLHS) was used to split the dataset into 80% and 20% for model calibration and internal validation in addition to an entirely independent validation dataset. The combined internal model validation showed R2 = 0.90, LCCC = 0.72, RMSE = 439.49 kg DM/ha, NRMSE = 15.08, and the combined independent validation had R2 = 0.88, LCCC = 0.68, RMSE = 457.05 kg DM/ha, NRMSE = 19.83. The key findings of this study indicated that the data obtained from the S2 bands and SI were appropriate for making accurate estimations of PRG biomass. Furthermore, including SWIR bands significantly improved the model. Finally, by utilising an RF ML model, a single ‘global’ model can automate PRG biomass prediction with high accuracy across extensive regions of all seasons and types of farm management.

Divergence metrics for determining optimal training sample size in digital soil mapping

Digital soil mapping (DSM) typically requires three common ingredients: georeferenced samples, environmental covariates, and a model. Of the three, sample design, or the selection of sample size and locations, has received considerably less attention. This is not surprising given that most studies are primarily limited by budget, the result being a focus on stratification of sampling locations in covariate (feature) space with less emphasis placed on the sample size. At the very least, determining the optimal sample size, regardless of whether it is achievable within a given budget, provides critical information about the loss of information from not collecting enough samples for a given study area. In this study, we evaluated the use of the Kullback-Leibler divergence (DKL), the Jensen-Shannon divergence (DJS), the Jenson-Shannon distance (DistJS), and the normalized variance in determining an optimal sample size for predicting total soil carbon at the field scale. The divergence metrics were computed for replicated (n = 10) sample plans using the conditioned Latin hypercube sampling algorithm across increasing samples sizes of 10, 25, and 50 to 400 in steps of 50 to determine an optimal sample size; the sensitivity of the divergence metrics to increasing the number of covariates and the number of bins for their computations were evaluated. The random forest algorithm was used to train predictive models using the same replicated sample sizes to determine the required sample size to optimize model performance based on root mean square error and Lin’s concordance correlation coefficient. The divergence metrics were insensitive to the number of covariates, but very sensitive to the number of bins specified for their calculation. On average, optimal sample size increased linearly (two additional samples per additional bin) regardless of the number of covariates used. The optimal sample sizes were 124, 133 and 220 for the DKL, DJS and DistJS divergence metrics, respectively, while the variance technique proved to be unreliable. Based on the model performance metrics from model validation, the optimal sample size ranged from 146 to 154 samples. The DistJS overestimated the optimal sample size considerably, while the DKL and DJS were quite similar to the optimal sample size determined from model validation. Future work should evaluate the use of divergence metrics for determining optimal sample size for multiple soil properties or classes, using various machine learning models, across different project scales, and with other sampling algorithms.

APPLICATION OF SEMI-VARIOGRAM ANALYSIS IN MEASURING SPATIAL VARIABILITY AND DISTRIBUTION OF SELECTED SOIL PROPERTIES IN NORTHEAST AKWA IBOM STATE, NIGERIA

Application of semi-variogram analysis in measuring spatial variability and distribution of selected soil properties in Northeast Akwa Ibom State, was carried out.The aim was to assess spatial variability and distribution of selected soil properties in the study area for effective site-specific soil management and precision agriculture using semi-variogram analysis. A digital elevation model (DEM) of the study area was acquired from United State Geological Survey (USGS) at 30m resolution. Slope gradient map that is capable of capturing the short-scale spatial variability of soil properties in the study area was generated from the DEM to guide field sampling. Modified conditioned Latin hypercube sampling technique was used in selecting observation points. Soil samples were collected from each observation point at 0-30 cm and 30-60 cm depths using soil auger. A total of 152 soil samples were collected for laboratory analysis. Analysed data of depth interval of 0-30cm and 30-60 cm were integrated to form depth interval of 0-60 cm. The data were subjected to normality test to ascertain the normal distribution of the data. Selected soil properties were subjected to semi-variogram analysis. The study revealed that slope gradient was able to captured short scale spatial variation in some soil properties understudy. Soil texture of the flat/nearly flat was sand in both surface and subsurface soil and sand in the surface soil and loamy sand in the subsurface soil in gently sloping and sloping. Soil pH was slightly acid in in flat/nearly flat and gently sloping and strongly acid in the sloping area in both surface and subsurface soil. Organic carbon was very high in the flat/nearly flat and gently sloping and high in sloping topography in both surface and subsurface soil. Total N was low in the sloping area and moderately low in gently sloping and nearly flat /flat. Base saturation was very high in the sloping topography and high in the gently sloping and nearly flat /flat. The results of semi-variogram analysis showed that all the selected soil properties exhibited spatial dependence within some distances. The range was 136.2 m for sand, 76.4m for silt, 1.6 m for clay, 1.7m for soil pH, 9.4m for organic carbon, 7.1m for total N, 9.2m for available P and 7.8 m for exchangeable K in the study area. Beyond this range, there is no longer relationship between sample points and sample values are not related to one another. The strength of the spatial dependence of sand, silt, soil pH, organic carbon, total N and available P was moderate; exchangeable K was strong while clay was weak. The semi- variance (sill) was 57.4 for sand, 23.8 for silt, 7.15 for clay, 0.21 for soil pH, 1.18 for organic carbon, 0.002 for total N, 85.1 for available P, and 0.03 for exchangeable K. The nugget variance or nugget effect was 25.9 for sand, 10.2 for silt, 5.9 for clay, 0.06 for soil pH, 0.60 for organic carbon, 0.001 for total N, 33.4 for available P and 0.003 for exchangeable K. The best fitted models were Exponential for sand and silt; Gaussian for available P and Spherical for clay, pH, organic carbon, total N and exchangeable K.

Soil organic carbon stock prediction using multi-spatial resolutions of environmental variables: How well does the prediction match local references?

We evaluated how the spatial resolution of environmental variables (n = 47) altered their ability to predict soil organic carbon (SOC) stocks (0–30 cm depth) using training data from Gridded Soil Survey Geographic-gSSURGO and SoilGrids databases. Training and validation subsamples (1,629) were selected using a conditioned Latin hypercube sampling (cLHS) design based on environmental variables in Vermont, U.S. The predictive relationships between environmental variables and SOC stock (t C ha−1) were developed using machine learning algorithms. The algorithms were trained (70 %) and evaluated (30 %) using a random subset of database subsamples, respectively, with an additional evaluation step using local, independent SOC reference data (n = 272). The Random Forest (RF) algorithm outperformed other algorithms at all spatial resolutions in estimating SOC stocks. As spatial resolution increased, model performance with the gSSURGO database increased (R2 = 0.33–0.62 and RMSE = 42.42–34.92), while no such trend was observed for the SoilGrids database. The best SOC stock model prediction using the SoilGrids database was achieved with a 10 m resolution (R2 = 0.54 and RMSE = 4.67). Evaluation of modeled results using the external, or independent, reference data showed a significant decrease compared to the internal validation in prediction accuracy (R2 = 0.11–0.14 for gSSURGO and, R2 = −0.19 for SoilGrids). The gSSURGO database showed that soil maps (including suborders, drainage classes, temperature, and moisture) and geology/landform maps had a greater influence than other environmental variables at all spatial resolution scales. In contrast, climatic- and DEM-related variables were more significant for the SoilGrids database. Our study suggested that the origin of the SOC stock database and the sampling scheme largely affects the importance of environmental variables assigned in the machine learning algorithm. Our results confirmed that the variable and data sources, model type, and combination of environmental variables significantly influenced prediction accuracy. In conclusion, DSM products should be re-evaluated with local references when used for spatial extents that are different from those for which they were initially designed.

Farm Household Typology Based on Soil Quality and Influenced by Socio-Economic Characteristics and Fertility Management Practices in Eastern Kenya

The smallholder farming systems in Sub-Saharan Africa (SSA) are highly diverse and heterogeneous in terms of biophysical and socio-economic characteristics. This study was conducted in upper Eastern Kenya (UEK) to categorize farm households and determine the influence of socio-economic characteristics (SeC) and soil fertility management practices (SFMP) on soil fertility across farms. Conditioned Latin hypercube sampling (cLHS) was performed to determine 69 soil sampling sites within Meru and Tharaka Nithi counties. From each household (whose field soil sample was obtained), data relating to resource endowment and soil fertility management were collected through a household questionnaire survey. Standard laboratory procedures were used to analyse soil samples. Data reduction was performed using categorical principal component analysis (CATPCA) (for SeC and SFMP) and standard principal component analysis (PCA) (for soil properties). Two-step cluster analysis identified three distinct farm categories or farm types (FT), namely, low fertility farms (FT1), moderately fertile farms (FT2), and fertile farms (FT3). The correlation of clusters against soil properties was significant across pH, soil organic carbon (SOC), cation exchange capacity (CEC), available P, plant available K, and exchangeable bases. FT1 had low SOC, pH, CEC and available P (soil characteristics), low usage of fertilizer and manure (soil fertility management), and smaller household size, lower income, and smaller farm size (socio-economic). FT2 had lower SOC (compared to FT3) and available P. In terms of soil fertility management, FT2 had higher cases of fallowing and composting with moderate fertilizer usage. Households in this category had moderate income, family size, and land size (socio-economic). FT3 had relatively high SOC, pH, CEC, and mineral nutrients. This farm type was characterized by high fertilizer use (soil fertility management) as well as larger household size, higher income, and larger farm size (socio-economic). The results indicate the importance of nutrient management in enhancing soil quality. Delineation and characterization of farms based on the various parameters including resource endowment reveal imbalanced farm resource flows, suggesting a need for locally tailored interventions suited for location-specific conditions to facilitate improved targeting of soil fertility-enhancing technologies and sustainable crop production regimes. While fertilizer is one of the most critical inputs for enhancing agricultural production, it is a major contributor to nitrous oxide emissions from agriculture and can have negative environmental effects on soil biota and water sources. Farmers’ knowledge on the use of fertilizer is thus necessary in developing strategies (such as integrated approach) to promote its efficient use and minimize its detrimental impacts.

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.

Optimizing Sampling Strategies for Near-Surface Soil Carbon Inventory: One Size Doesn’t Fit All

Soils comprise the largest pool of terrestrial carbon yet have lost significant stocks due to human activity. Changes to land management in cropland and grazing systems present opportunities to sequester carbon in soils at large scales. Uncertainty in the magnitude of this potential impact is largely driven by the difficulties and costs associated with measuring near-surface (0–30 cm) soil carbon concentrations; a key component of soil carbon stock assessments. Many techniques exist to optimize sampling, yet few studies have compared these techniques at varying sample intensities. In this study, we performed ex-ante, high-intensity sampling for soil carbon concentrations at four farms in the eastern United States. We used post hoc Monte-Carlo bootstrapping to investigate the most efficient sampling approaches for soil carbon inventory: K-means stratification, Conditioned Latin Hypercube Sampling (cLHS), simple random, and regular grid. No two study sites displayed similar patterns across all sampling techniques, although cLHS and grid emerged as the most efficient sampling schemes across all sites and strata sizes. The number of strata chosen when using K-means stratification can have a significant impact on sample efficiency, and we caution future inventories from using small strata n, while avoiding even allocation of sample between strata. Our findings reinforce the need for adaptive sampling methodologies where initial site inventory can inform primary, robust inventory with site-specific sampling techniques.

Sampling and classifier modification to DSMART for disaggregating soil polygon maps

DSMART has been widely used to disaggregate multi-component soil polygon map units into raster maps. The algorithm randomly selects (simple random sampling approach) an equal number of synthetic sample points from all map units and assigns soil classes proportionate to the map unit composition using the C5.0 algorithm. Conditional Latin Hypercube sampling (cLHS) is a maximally stratified random sampling that guarantees full coverage of multivariate distribution. cLHS has shown success in digital soil mapping as it optimizes sample point selection based on environmental covariates. Information-driven stratification and sampling may help improve the assignment of soil classes within map units. The objective of this study was to compare cLHS against simple random sampling (SRS) approach using three classifier algorithms in DSMART to disaggregate the soil great group and soil series maps of a sub-watershed in Southern Ontario, Canada, as a case study example. SRS and cLHS were applied in two methods, sampling “by polygon” and “by area”, and three classifiers selected were C5.0, random forest (RF), and k-nearest neighbor (K-NN) (a total of twelve combinations − 4 sampling approaches and 3 classifier algorithms). The original R-code of the DSMART package was modified to incorporate cLHS sampling and change classifier algorithms. For the soil great group predictions, RF performed better in combination with “by polygon” SRS and “by polygon” cLHS approaches (Kappa scores 0.57). For the soil series predictions, C5.0 and “by area” cLHS approach (Kappa score 0.50) performed well compared to other combinations (Kappa score 0.25–0.44). Overall, cLHS performed marginally better with C5.0 and RF while, SRS performed better with K-NN. Disaggregated maps using cLHS generated lower prediction uncertainty values.

REGRESSION MODEL OF TOPOGRAPHY WITH THE DISTRIBUTION OF SELECTED SOIL PROPERTIES IN NORTHEAST AKWA IBOM STATE, NIGERIA

Regression model of topography with the distribution of selected soil properties in Northeast Akwa Ibom State, Nigeria was carried out. The aim was to establish a regression model of topography with the distribution of selected soil properties in Northeast Akwa Ibom State for soil properties predictions and management. Topographic map (or elevation map) of the study area was generated from digital elevation model (DEM) at 30m resolution acquired from United State Geological Survey (USGS). It was classified into three elevation classes of lower elevation (0-50 masl), middle elevation (50-100 masl) and higher elevation (100-150 masl) to guide field sampling. With the aid of Global Positioning System (GPS), the classes obtained from topographic map were cross-checked (ground-truthing) in the field. Modified Conditioned Latin Hypercube Sampling Method was used in selecting observation points. Each observation point was purposively selected to fall within the classes of topographic map to give a good coverage of both feature space (classes of topographic map) and geographical space (study area). A total of 120 soil samples were collected at a depth of 0-30cm and 30-60 cm using soil auger. The samples were taken to the laboratory for analysis. The results revealed variation in soil properties among the three topographic classes under study. Sand fraction was significantly higher (p > 0.05) in lower elevation than other elevation classes while silt and clay fractions were significantly higher (p > 0.05) in the middle and higher elevations than lower elevation in the study area. Soil pH was significantly higher (p > 0.05) (slightly acid) in the lower elevation than middle and higher elevations (moderately acid). Electrical conductivity and base saturation were significantly higher (p > 0.05) in the lower elevation than middle and higher elevations. Organic carbon, total N, ECEC, exchangeable Mg and exchangeable K were significantly higher (p > 0.05) in the middle elevation than that of lower and higher elevations. Available P of higher elevation was significantly higher (p > 0.05) than that of lower and middle elevations. The study also showed that topography accounted for 3 % variation of sand and silt fractions; 5 % variation of clay fraction and 4 % variation of soil pH. Topography also explained 19 % variation of available P, 8.4 % variation of exchangeable K and 5.5 % variation of organic carbon and total N in the study area. The remaining fractions of variation may be attributed to other factors of soil formation such as parent material, climate, organism and time.

Assessment of soil organic carbon stocks in Alberta using 2-scale sampling and 3D predictive soil mapping

A three-dimensional predictive soil mapping approach for predicting soil organic carbon (SOC) stocks (t/ha) at high spatial resolution (30 m) for Alberta for 2020–2021 is described. A remote sensing data stack was first prepared covering Alberta’s agricultural lands. A total of 404 sampling locations were distributed across Alberta using 2-scale sampling: (1) 22 pilot farms representing main climatic zones and (2) conditioned Latin hypercube sampling at each farm. Soil samples were taken at four standard depths (0–15, 15–30, 30–60, 60–100 cm) using soil probes and analyzed for SOC. Predictive models for SOC content and bulk density were built separately and then used to predict at 0, 15, 30, 60, and 100 cm and calculate aggregated SOC stocks per pixel. The SOC content and bulk density models had R squares of 0.61 and 0.68, respectively. Based on these mapping results, grassland soils were consistently associated with higher SOC stocks across all soil types as compared to croplands. The average SOC stock increase for grassland soils compared to cropland soils was 2.1 Mg per hectare, ranging from 2.17 to 6.09 Mg per hectare depending on soil type. Results also showed that >15 % of total SOC stocks were located in subsoil, which was higher than expected.