Prediction Of Soil Properties Research Articles

New hybrid evolutionary models for spatial prediction of soil properties in Kurdistan

AbstractWith intense human activity causing constant environmental change, there is a greater need than ever to have accurate and frequently updated soil information. Traditional soil maps are large‐scale polygon maps of soil type and implicitly assume no variation in soil properties within polygons. DSM approaches can increase the accuracy of modelling. Therefore, the current study was conducted to evaluate the applicability of a classic adaptive neuro‐fuzzy inference system (ANFIS) and three hybridized ANFIS models (i.e. particle swarm optimization—PSO, the genetic model—GA and the artificial bee colony model—ABC) for predicting eight soil properties in western Iran. The ANFIS+ABC and ANFIS+PSO approaches showed the best performance in predicting soil properties such as soil organic carbon (SOC) and total nitrogen (N). The ABC+ANFIS‐hybrid model increased R2 values by 0.32, 0.24, 0.23 and 0.37 and decreased RMSEs by 11%, 32.7 mg kg−1, 0.08 mmhos cm−1 and 0.3% when predicting SOC, K, EC and SP, respectively, compared with the ANFIS model alone. Also, the use of the PSO+ANFIS‐hybrid model led to an increase in R2 values of 0.44, 0.29, 0.34 and 0.36 and a decrease in RMSEs of 0.02, 0.29 ppm, 0.34% and 0.36% for predicting N, P, pH and TNV, respectively, compared with the ANFIS model alone. In general, the ANFIS+ABC and ANFIS+PSO improved the prediction accuracies for SOC and N by 32% and 24%, respectively, compared with the classic ANFIS model. As a whole, our findings demonstrated the ability of hybrid models for improving the accuracy of the ANFIS approach for predicting soil features in western Iran, which supports our main hypothesis.

Exploration of machine learning methods for prediction and assessment of soil properties for agricultural soil management: a quantitative evaluation

Soil is a heterogeneous and complex natural resource that is the factual basis of almost all agriculture production activities. The soil’s inherent nutrients or physiochemical properties help the researchers better understand the soil ecosystem dynamics and play a crucial role in guiding farmland decision-makers in their routine decisions. Therefore, the accurate forecasting of soil leads to improved and better soil health management (SHM). The recent advances in sensing and computational technologies have led to the expanding accessibility of farmland data either obtained distantly or proximally. The increasing availability of massive data and unreservedly accessible open-source algorithms have prompted a quickened use of machine learning (ML) procedures to investigate soil conditions. Therefore, to understand the usage of ML techniques in exploring soil properties and related applications, this paper concentrates on reviewing and analyzing ML techniques precisely to predict and assess soil properties for improved decisions on agricultural SHM. The article also explores various other vital factors like algorithms, implementation tools, and performance metrics employed in numerous soil assessment application domains and different challenges and future research directions for SHM using ML techniques. The detailed assessment concludes that the response for ML in the prediction and evaluation of soil properties for SHM is very promising for the sustainable growth of agriculture.

Machine learning-based prediction and performance study of transparent soil properties

An indispensable process of geotechnical modeling with transparent soils involves analyzing images and soil property simulations. This study proposes an objective framework for quantitative analysis of the influential mechanism of three key factors, namely, different aggregate proportions (DAP), solvent ratio (SR), and solute solution ratio (SSR) on transparent soils' transparency and shear strength. 125 groups of transparent soil samples considering these three factors were prepared to investigate their impact on transparency and shear strength through Elastic Net regression. Spearman correlation analysis was performed for transparency and shear strength. Furthermore, by comparing the performance of XGBoost, GBDT, Random Forest, and SVR after hyperparameter tuning in predicting transparency and shear strength, XGBoost proved to be the optimal machine learning model with the lowest MSE of 0.0048 and 0.0306 and was innovatively adopted to analyze how various factors affect the transparency and shear strength, thus enhancing the interpretability of machine learning. A ranking system, according to the importance scores of XGBoost, shows that SSR was the most important factor affecting both shear strength and transparency of transparent soils, with importance scores being 0.45 and 0.57, respectively. Our study may shed light on the preparation and performance study of transparent soils.

Soil properties: Their prediction and feature extraction from the LUCAS spectral library using deep convolutional neural networks

Soil, as a non-renewable resource, should be monitored continuously to prevent its degradation and promote sustainable agriculture. Soil spectroscopy in the visible-near infrared range is a fast and cost-effective analytical technique to predict soil properties. Although traditional machine learning methods are widely used for modeling soil spectral data, large spectral datasets may require better analytical methods for big data. Here, we explored the modeling potential of deep convolutional neural networks (DCNNs) for soil properties based on a large soil spectral library. The European topsoil dataset provided by the Land Use/Cover Area frame Survey (LUCAS) was used for DCNN modeling with the original absorbance spectra. Two single-task 16-layer DCNN models (LucasResNet-16 and LucasVGGNet-16) were used to make regression predictions of seven soil properties and classification predictions of soil texture. The effects of data pre-processing on single-task and multi-task DCNN modeling were assessed. The SHapley Additive exPlanations method was used to interpret the output of a DCNN model (LucasResNet-16). The DCNN models produced accurate predictions for most soil properties, and were superior to a single-task shallow convolutional neural network and traditional machine learning methods. Spectral transformation was effective for predicting some soil properties, while spectral downsampling led to a reduction in the modeling accuracy. The performance of a multi-task DCNN model built on the basis of LucasResNet-16 was improved compared with the performance of the single-task model. Soil organic carbon content, nitrogen content, cation exchange capacity, pH, and calcium carbonate content were well predicted, with the root mean squared error of 19.130 g∙kg−1, 0.971 g∙kg−1, 6.614 cmol(+)∙kg−1, 0.326, and 24.526 g∙kg−1, respectively. The overall classification accuracy of soil texture was 0.749 (four groups) and 0.566 (12 levels). The position of feature wavelengths differed among the soil properties, for which multiple characteristic peaks were common. This study fully demonstrates the modeling potential of deep learning with soil ultraspectral data, which could enhance precision agriculture.

Evaluating the synergy of three soil spectrometers for improving the prediction and mapping of soil properties in a high anthropic management area: A case of study from Southeast Brazil

Precision agriculture aspires to manage the soil spatial variability within a field. However, agricultural practices can decrease the spatial autocorrelation of soil properties at a short distance, which requires an increase in sampling density to improve the reliability of interpolated surfaces. Data fusion of several soil spectrometers is a promising alternative that can increase knowledge about soil physicochemical variability. Therefore, this research aimed to evaluate whether the synergy of portable X-ray fluorescence, Vis-NIR, and Mid-Infrared spectrometers can successfully increase sampling density and improve the quality of the soil attribute maps, in a high fertilization management study area. We conducted a case study on a sugarcane field in Southeast Brazil. Our spectral data fusion approach did not increase the prediction performance for soil properties, achieving the same accuracy (α ≤ 5%) of the single spectral ranges (Vis-NIR or MIR or pXRF) individually. Thus, Vis-NIR was chosen to evaluate if spectroscopy could improve soil properties prediction and mapping. We used three dataset scenarios to build soil property maps: two were based on soil wet-chemical laboratory analysis using 1 sample per ha, and 1 sample per 0.25 ha and a third used a spectroscopy model with 0.25 samples per ha calibrated using soil wet-chemical laboratory analysis at the rate of 1 sample per ha. In a field with strong anthropic influences, due to liming and soil fertilization, Vis-NIR spectroscopy cannot predict well the amount of phosphorus and pH at a sampling density of 1 sample per 0.25 ha. Increased soil sampling enhances soil mapping accuracy only for chemical properties (e.g., potassium) that are correlated with soil texture. Nevertheless, soil spectra can classify areas of similar soil texture and help farmers with the site-specific management of crops.

Relationships and Spatial Structure of Soil Physical Properties in a Ten-Year Corn-Cotton Rotation Field

This study quantified the relationships between soil, textural, and hydraulic properties at the field-scale for a conventional tilled Memphis silt loam that had undergone a 10-year corn and cotton rotation and described their spatial variability. Composite soil samples collected from the plow layer at 272 nodes on 15 x 15 m grids were analyzed for texture and bulk density. These values were used as pedotransfer functions to predict unsaturated (Ko) and saturated hydraulic (Ks) conductivities as well as the van Genuchten curve shape parameters α and n. Regression analyses quantified relationships between the measured and model predicted soil properties. While correlations between textural and model predicted soil properties including bulk density were significant (p<0.05), those between sand and clay, clay and n, clay and α were not. Sand and silt appeared to be better predictors of soil hydraulic conductivity and the van Genuchten curve shape parameters for the soil investigated. Spatial dependence was strong for sand, silt, bulk density, Ko, α and n, and moderate for clay and Ks.

Machine Learning Strategy for Soil Nutrients Prediction Using Spectroscopic Method

The research presented in this paper is based on the hypothesis that the machine learning approach improves the accuracy of soil properties prediction. The correlations obtained in this research are important for understanding the overall strategy for soil properties prediction using optical spectroscopy sensors. Several research results have been stated and investigated. A comparison is made between six commonly used techniques: Random Forest, Decision Tree, Naïve Bayes, Support Vector Machine, Least-Square Support Vector Machine and Artificial Neural Network, showing that the best prediction accuracy cannot always be achieved by the most common and complicated method. The influence of the chosen category for nutrient characterization was investigated, indicating better prediction when a multi-component strategy was used. In contrast, the prediction of single-component soil properties was less accurate. In addition, the influence of category levels was not as significant as expected when choosing between 3-level, 5-level or 13-level nutrient characterization for some nutrients, which can be used for a more precise nutrient characterization strategy. A comparative analysis was performed between soil from a local farm with similar texture and soils collected from different locations in Slovenia, which gave a better prediction for a local farm. Finally, the influence of principal component analysis was validated using 5, 10, 20 and 50 first principal components, indicating the better performance of machine learning when using the 50 principal components.

Evaluation of variable rate nitrogen fertilization scenarios in cereal crops from economic, environmental and technical perspective

Optimizing variable rate nitrogen fertilization (VRNF) is required to minimize inputs, maximize outputs, and reduce environmental footprints. However, this is not possible in commercial farms, since it would require to intervene in regular agricultural practices with potential economic and environmental consequences and would take years of costly experiments. This study aims at the evaluation of different VRNF scenarios based on simulations that include the crop nitrogen need, soil variability, nitrogen limits set by legislations and the technical specification of variable-rate machinery. A simulation software was designed and developed to allow automatic prediction and mapping of soil and crop properties, and calculate recommendations for VRNF. The software was used to compare traditional uniform rates nitrogen (N) fertilization (URNF) against three VRNF treatments [e.g., map-based (MB), sensor-based (SB) and map-sensor-based (MSB)] in four commercial fields. Under each of the three treatments two different application schemes were evaluated, e.g., applying more N fertilizer to the more fertile zones (Kings scheme - KS) and more N fertilizer to the least fertile zone (Robin Hood scheme - RHS). Simulations were made after imposing N legislation limits and without imposing them. Finally, VRNF application was evaluated for a full boom sprayer, a section control sprayer and a nozzle control sprayer. Results indicated that the VRNF did not exceed the traditional URNF treatment, but only if the N limit by legislation is imposed. The RHS consumed 16.4–118.1 % less N fertilizer than the KS and 33.3–56.2 % than URNF treatment. However, the KS without N limit always exceeded the applied N fertilizer compared to the URNF treatment, generating high risk of N leaching. When imposing the N limit, both KS and RHS consumed less N than the traditional URNF treatment, except for the MSB under the KS, which used the same N as the URNF treatment. Spatial variability can be observed in the MB treatment, temporal variability in the SB treatment and both variabilities in the MSB treatment. Regarding the spatial variability, no significant differences between section and nozzle control options could be observed, whereas variability was minimized when using the full boom control. We concluded that the optimal fertilization scheme is VRNF based on a combination of RHS and section control, which is expected to result in reducing N fertilizer application cost, minimising risk of N leaching, and ensuring N applied is under the set legislation limits. Future work should include the crop response in the analysis to provide a comprehensive evaluation of the optimal VRNF treatment.

Soil Organic Matter Prediction Model with Satellite Hyperspectral Image Based on Optimized Denoising Method

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

Using remote sensors to predict soil properties: Radiometry and peat depth in Dartmoor, UK

Remote sensors provide high resolution data over large spatial extents that can potentially be used to map soil properties such as the concentration of organic carbon or its moisture content. The sensors rarely measure the property of interest directly but instead measure a related property. There is a need to make ground measurements of the property of interest to calibrate a model or relationship between the soil property and the sensor data.We develop a framework for optimizing the locations and number of ground measurements of a soil property for surveys incorporating sensor data. The data are used to estimate a linear mixed model of the property where the fixed effects are a flexible spline-based function of the sensor measurements.The framework is used to map peat depth across a portion of Dartmoor National Park using radiometric potassium data measurements from an airborne survey. The most accurate maps result from using a geostatistical predictor to combine the relationship with the sensor data and the spatial correlation amongst the peat depth measurements. The optimal sampling designs suggest that ground measurements should be focussed where peat depths are largest and most uncertain. When measurements are made at 25 optimally selected sites, predictions that do not utilise the sensor data have 20% larger root mean square errors than those that do. For 200 ground measurements this benefit is 14%. The maps produced using the sensor data and 25 ground measurements have smaller root mean square errors than those based only upon 200 ground measurements.

Pedotransfer Functions For Estimation Of Soil Moisture Constants From Penetration Resistance Measurements And Some Soil Properties

Studies to prediction of soil moisture constants and other soil properties rather than direct measurements were never dwindle importance. Models were derived from other soil properties obtained easily. Therefore, in this study focused on the predictability of some moisture constants, whose determination was often difficult and time-consuming, from penetration resistance measurements. In the improvement of alternative models for the estimation of moisture constants; in addition to penetration resistance, textural fractions (sand, clay and silt), bulk density, CaCO3 % and organic matter contents were included. The models were created according to soil groups with different textures (sandy, loamy, clay) for moisture constants at 0.1, 0.33, 0.5 and 15 bar. In the models for estimation of 0.1, 0.33, 0.5 and 15 bar moisture content, the highest differences in R2 values (0.61, 0.60, 0.64 and 0.59) between the actual and the predicted data was obtained for loamy soils. For this group, the root means square error (RMSE) ranged between 1.32 and 1.90 %, and in addition, the mean error (ME) was determined to be in a range from 1.53 to 2.05 %. For the estimation of moisture content at different soil moisture tensions using organic matter, bulk density, clay and penetration resistance properties, the coefficient of determination ranged from 71 to 77 %. Therefore, it is concluded that the alternative models, developed using penetration resistance or by the addition of some other soil properties, could be used safely in the loamy texture soils.

Do ARMA Models Provide Better Gap Filling in Time Series of Soil Temperature and Soil Moisture? The Case of Arable Land in the Kulunda Steppe, Russia

The adoption of climate-smart agriculture requires the comprehensive development of environmental monitoring tools, including online observation of climate and soil settings. They are often designed to measure soil properties automatically at different depths at hour or minute intervals. It is essential to have a complete dataset to use statistical models for the prediction of soil properties and to make short-term decisions regarding soil tillage operations and irrigation during a vegetation period. This is also important in applied hydrological studies. Nevertheless, the time series of soil hydrological measurements often have data gaps for different reasons. The study focused on solving a problem of gap-filling in hourly time series of soil temperature and moisture, measured at the 30 cm depth using a weighted gravitation lysimeter station while meteorological data were recorded simultaneously by a weather station. The equipment was installed in the Kulunda Steppe in the Altai Krai, Russia. Considering that climate conditions affect soil temperature and moisture content directly, we did a comparative analysis of the gap-filling performance using the three imputation methods—linear interpolation, multiple linear regression, and extended ARMA (p,q) models with exogenous climatic variables. The results showed that, according to the minimum of the mean absolute error, ARMA (p,q) models with optimally selected order parameters, and an adaptive window, had some advantages compared to other single-imputation methods. The ARMA (p,q) model produced a good quality of gap-filling in time series with the mean absolute error of 0.19 °C and 0.08 Vol. % for soil temperature and moisture content, respectively. The findings supplemented the methodology of hydrological data processing and the development of digital tools for the online monitoring of climate and soil properties in agriculture.

Predicting soil moisture content over partially vegetation covered surfaces from hyperspectral data with deep learning

AbstractPrevious studies for retrieving soil moisture content (SMC) from visible and near‐infrared hyperspectral data over vegetation‐covered surfaces using spectral unmixing, non‐negative matrix factorization, and albedo/vegetation coverage in trapezoid spaces have required mass spectral preprocessing and offered only limited improvements in prediction accuracy. Recently, deep learning has triggered some improvements in soil properties prediction from hyperspectral data because of its automatic feature extraction and high accuracy. In this study, hyperspectral data in a simulation experiment with different vegetation coverages, SMCs, and soil types were acquired. Deep learning models, one‐dimensional convolutional neural network (1D‐CNN), and long short‐term memory network (LSTM) are proposed to predict SMC. The results showed that two deep learning models achieved excellent predictions (residual prediction deviation [RPD] > 2.5) using the unpreprocessed mixed spectra and partial least squares regression (PLSR) had a good prediction (RPD = 1.88). The 1D‐CNN (R2p = .91) and LSTM (R2p = .90) significantly outperformed PLSR (R2p = .72), which demonstrated that deep learning could improve SMC prediction over partially vegetation‐covered surfaces. However, when only using bare soil spectra, the prediction accuracy was commensurate, whether through the 1D‐CNN, LSTM, or PLSR models; additionally, 1D‐CNN and LSTM had better performance on all mixed spectra than bare soil spectra. These results indicated that deep learning had no advantage on smaller datasets. We also found that SMC prediction with deep learning was affected by vegetation coverage and soil type but was still very good. The 1D‐CNN and LSTM are effective models for predicting SMC with large hyperspectral datasets acquired from complex soil surface conditions.

Using bias correction and ensemble modelling for predictive mapping and related uncertainty: A case study in digital soil mapping

In this study, we explored the potential benefits of using bias correction and ensemble modelling for the prediction of soil properties and assessment of related uncertainty. The proposed approach combines resampling techniques applied to soil observations, covariates and hyperparameters to generate a set of simulated values at the same location. The ensemble predictions resulting from the resampling are then used to generate deterministic predictions for the final mapping product along with the related uncertainty. We also introduced bias correction into the modelling framework in order to overcome conditional bias that is commonly encountered in digital soil mapping products. We compared the accuracy of our predictions resulting from bias correction and ensemble modelling with previously published global soil mapping products. Our results demonstrated that bias correction improves the linearity and the ratio of the variance between simulated and observed values and reduces conditional bias by a factor of 25 to 50% for different soil properties. The performance of the deterministic predictions obtained from ensemble modelling is better than most of its individual component models, and is always located in the first quantile of the performance of all members. The analysis of uncertainty suffers from underdispersion, which means that local uncertainty tends to be underestimated by our approach 40 to 60% of the time. A comparison with the performance achieved by global soil mapping products in our area, indicates that global mapping products achieved low performance (R2: −0.48–0.13) and suffered from an important conditional bias (alpha: 0.23–0.59, where alpha is the ratio of variance between predicted and observed values), leading to unrealistic predictions at the local scale. Ultimately, the combination of bias correction and ensemble modelling appears to be both useful and relevant for digital soil mapping and helps to address three common problems: equifinality, assessment of uncertainty and, correction of conditional bias in simulated values. The procedure described in this study is relatively easy to implement and is not computationally intensive. In operational use, the combination of bias correction and ensemble modelling should increase the quality of the information produced for environmental management and modelling, while additionally providing uncertainty maps.

WITHDRAWN: An efficient method for determining DEM parameters of a loose cohesive soil modelled using Hysteretic Spring and Linear Cohesion contact models

Abstract Appropriate determination of discrete element modelling (DEM) parameters is essential for accurate predictions of soil properties and disturbance behaviours in agriculture. In this paper, DEM parameters of a loose cohesive soil modelled using Hysteretic Spring and Linear Cohesion contact models were determined by a combination of Plackett-Burman (PB) test, Steepest Ascent test and Central Composite test. The accuracies of DEM models developed under different soil moisture contents (ranging from 0.27 to 22 %) were evaluated using slumping angle of repose (SAOR) test and funnelling angle of repose (FAOR) test. The results showed that heap angles under different soil moisture contents decreased quadratically with time with determination coefficients ranging from 0.750 2 to 0.969. For the given soil, friction or rolling friction coefficient of soil-soil changed with a cubic function as moisture content increased. Similar coefficients of friction (from 0.23 to 0.25) and rolling friction (from 0.038 to 0.049) of soil-soil were found for soils with moisture contents from 12 to 22 %. Based on the analysis of variance (ANOVA) outputs, soil properties tested were significantly affected by soil moisture content at p

Evaluating validation strategies on the performance of soil property prediction from regional to continental spectral data

Visible-near infrared (vis–NIR) spectroscopy has been widely used to characterize soil information from field to global scales. Before applying a calibrated spectral predictive model to acquire soil information, either independent validation or k-fold cross validation is used to evaluate model performance. However, there is no consensus on which validation strategy is more suitable and robust when evaluating model performance for the studies in different scales. The objective of this study is to evaluate and compare the model performance of two validation strategies coupling different calibration sizes (a ratio of calibration to validation of 2:1, 4:1 and 9:1) and calibration sampling strategies (random sampling (RS), rank, Kennard-Stone (KS), rank-Kennard-Stone (RKS) and conditioned Latin hypercube sampling (cLHS)) across scales. A total of 17,272 vis–NIR spectra of mineral soils from LUCAS data (continental scale) and their soil organic carbon (SOC) and clay contents were used in this study, and the dataset was further split into national (2761 samples in France) and five regional datasets (110 to 248 samples from five French administrative regions). To eliminate the effect of changing validation set on the model performance, a consistent test set (20% of total samples at each scale) was split to evaluate all the combinations involved in two validation strategies. The Lin’s concordance correlation coefficient (CCC) of the cubist model were stable for both SOC and clay for different calibration sizes, calibration sampling and validation strategies for a large calibration size (>1400) at the national and continental scales. A larger calibration size can potentially improve model performance for a small dataset (<300) at the regional scale, and a wider calibration range would result in better model performance. No silver bullet was found among the different calibration sampling strategies at the regional scale. For five French regions (small data set), we found a high variation (95th percentile minus the 5th percentile) in the CCC among the models built from 50 repeated RS (0.10–0.44 for SOC, 0.16–0.52 for clay) and cLHS (0.08–0.40 for SOC, 0.12–0.36 for clay). This finding indicates that a one-time RS or cLHS for selecting the calibration set has high uncertainty in model evaluation for a small dataset and therefore should be used with caution. Therefore, we suggest the following: (1) for a large data set (thousands), either one-time random sampling for independent validation or k-fold cross validation would be appropriate; (2) for a small data set (dozens to hundreds), k-fold cross validation and/or repeated random sampling for independent validation would be more robust for spectral predictive model evaluation.

Enhancing the accuracy of machine learning models using the super learner technique in digital soil mapping

Digital soil mapping approaches predict soil properties based on the relationships between soil observations and related environmental covariates using techniques such as machine learning (ML) models. In this research, a wide range of ML models (12 base learners) were tested in predicting and mapping soil properties. Furthermore, a super learner approach was used to improve model accuracy by combining the predictions of the base learners. A major challenge of using super learner and complex models is that the exact contribution of individual covariates in the overall prediction is not always known. To address this issue, permutation feature importance (PFI) analysis was applied as a model-agnostic interpretation tool. The weights assigned to each ML base learner obtained from super learner, and feature importance values obtained from each ML base learner were used to quantify the contribution of individual covariates on the final prediction. The super learner and PFI techniques were tested by predicting a variety of soil physical and chemical properties of the Urmia Lake playa in Iran. As expected, the results indicated that the super learner had substantially higher accuracies for predicting soil properties in comparison to the individual base learners. For instance, the super learner showed an improved performance in comparison to linear regression by decreasing the root mean square error by an average of 46%. The PFI analysis revealed the important contribution of geomorphic and groundwater data in predicting soil properties. Overall, the proposed approach may be used for improving accuracy of ML models in digital soil mapping.

Using a One-Dimensional Convolutional Neural Network on Visible and Near-Infrared Spectroscopy to Improve Soil Phosphorus Prediction in Madagascar

As a proximal soil sensing technique, laboratory visible and near-infrared (Vis-NIR) spectroscopy is a promising tool for the quantitative estimation of soil properties. However, there remain challenges for predicting soil phosphorus (P) content and availability, which requires a reliable model applicable for different land-use systems to upscale. Recently, a one-dimensional convolutional neural network (1D-CNN) corresponding to the spectral information of soil was developed to considerably improve the accuracy of soil property predictions. The present study investigated the predictive ability of a 1D-CNN model to estimate soil available P (oxalate-extractable P; Pox) content in soils by comparing it with partial least squares (PLS) and random forest (RF) regressions using soil samples (n = 318) collected from natural (forest and non-forest) and cultivated (upland and flooded rice fields) systems in Madagascar. Overall, the 1D-CNN model showed the best predictive accuracy (R2 = 0.878) with a highly accurate prediction ability (ratio of performance to the interquartile range = 2.492). Compared to the PLS model, the RF and 1D-CNN models indicated 4.37% and 23.77% relative improvement in root mean squared error values, respectively. Based on a sensitivity analysis, the important wavebands for predicting soil Pox were associated with iron (Fe) oxide, organic matter (OM), and water absorption, which were previously known wavelength regions for estimating P in soil. These results suggest that 1D-CNN corresponding spectral signatures can be expected to significantly improve the predictive ability for estimating soil available P (Pox) from Vis-NIR spectral data. Rapid and accurate estimation of available P content in soils using our results can be expected to contribute to effective fertilizer management in agriculture and the sustainable management of ecosystems. However, the 1D-CNN model will require a large dataset to extend its applicability to other regions of Madagascar. Thus, further updates should be tested in future studies using larger datasets from a wide range of ecosystems in the tropics.

Prediction of various soil properties for a national spatial dataset of Scottish soils based on four different chemometric approaches: A comparison of near infrared and mid-infrared spectroscopy

Infrared spectroscopic techniques, in combination with chemometric approaches, have been widely used to estimate different physical and chemical properties in soil samples. This study aims to assess the performance of diffuse reflectance spectroscopy in the near-infrared (NIR) region and Attenuated Total Reflection Fourier Transform infrared spectroscopy (ATR-FTIR) in the mid-infrared (MIR) region to predict nine different soil properties: total carbon, total nitrogen, bulk density, clay, sand, silt, pH (in H2O), exchangeable Mg and exchangeable K using chemometric approaches. The predictive performance of four different regression methods i.e., Partial least square (PLS), support vector regression (SVR), Cubist and convolutional neural network (CNN) in combination with different pre-processing approaches were investigated. For CNN, the FTIR/NIR spectra were converted to spectrograms and fed into the CNN to examine the prediction accuracy of two-dimensional convolutional neural network (CNN-2D) and to compare its performance with one-dimensional convolutional neural networks (CNN-1D, spectral data as input). To achieve these objectives, we used a spectral library of 650 samples from National Soils Inventory of Scotland (NSIS) dataset collected in a 20 km grid throughout Scotland between 2007 and 2009. The FTIR and NIR data were both split into calibration (nc = 520) and validation (nv = 130) sets. Our results show that the regression models with FTIR data have better predictive performance than those created using NIR data for all the studied soil properties (improvement in root mean square error of prediction (RMSEP) of 3–61%), except for pH. Comparing the different chemometric approaches, for both the NIR and FTIR, the results indicated that the CNN-1D models performs better than PLS, SVR and CNN-2D for all the studied soil components in terms of RMSEP. We found that the CNN-1D models created using the NIR spectral dataset were superior (lower RMSEP values) to those developed using the Cubist approach for total carbon, total nitrogen, clay, silt, pH and exchangeable K, whereas the Cubist models for sand and exchangeable Mg performed slightly better than CNN-1D. For the FTIR data, Cubist models for total carbon and silt performed best (with lowest RMSEP and highest Residual Prediction Deviation (RPD) values), while for the rest of the components, CNN-1D outperformed the PLS, SVR, Cubist and CNN-2D. We also calculated the impact of the NIR and FTIR variables used by the Cubist and CNN-1D models in predicting different soil properties. We used the Shapley Additive Explanation (SHAP) values, a game theoretic approach, to interpret the output of the CNN-1D models.

Improving the spatial prediction accuracy of soil alkaline hydrolyzable nitrogen using GWPCA‐GWRK

AbstractPrincipal component analysis‐multiple linear regression (PCA‐MLR) is usually used to weaken the multi‐collinearity effects among auxiliary variables in a regression prediction. However, both PCA and MLR in this model are only built on variable space rather than geographical space. When used in the spatial prediction of soil properties, PCA‐MLR usually cannot effectively capture the spatially non‐stationary structures among auxiliary variables and spatially non‐stationary relationships between the target variable and principal component scores. Moreover, PCA‐MLR may ignore the potentially valuable regression residual. To address these limitations, this study first proposed geographically weighted principal component analysis‐geographically weighted regression kriging (GWPCA‐GWRK) for the spatial prediction of soil alkaline hydrolyzable nitrogen (AN) in Shayang County, China. Then, the spatial prediction accuracy of GWPCA‐GWRK was compared with those of the following five models: ordinary kriging (OK), co‐kriging (CoK), PCA‐MLR, PCA‐graphically weighted regression (PCA‐GWR), and GWPCA‐GWR. Results showed that (i) eight variables were determined as auxiliary data by a geodetector; (ii) the spatially non‐stationary relationships among the eight auxiliary variables were revealed by the results of the local correlation analysis, Monte Carlo test, and GWPCA; (iii) GWPCA‐GWRK provided the lowest prediction error (RMSE = 18.80 mg kg−1, MAE = 12.79 mg kg−1) and highest Lin's concordance correlation coefficient (LCCC; 0.75); (iv) relative improvement accuracies over the traditionally‐used OK were 19.74% for GWPCA‐GWRK, 16.42% for GWPCA‐GWR, 8.09% for PCA‐GWR, −3.67% for PCA‐MLR, and 4.70% for CoK. It is concluded that the proposed GWPCA‐GWRK model is an effective spatial predictor, which can adequately extract the main information of the multiple auxiliary variables in a large‐scale area.