Proximal sensor-enhanced soil mapping in complex soil-landscape areas of Brazil

Sérgio H.G.SILVA,David C.WEINDORF,Wilson M.FARIA,Leandro C.PINTO,Michele D.MENEZES,Luiz R.G.GUILHERME and Nilton CURIDepartment of Soil Science,Federal University of Lavras,Lavras 3700-000(Brazil)

2Department of Earth and Atmospheric Sciences,Central Michigan University,Mount Pleasant 48858(USA)

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

Key Words:magnetic susceptibility,magnetometer,portable X-ray fluorescence spectrometer,soil class,soil spatial analysis,spatial resolution,terrain variables

INTRODUCTION

Soils vary with climate,organisms,parent material,time,and landscape positions reflected by topography(relief)(Jenny,1941).Such variations can be represented on soil maps,which show the spatial distribution of soil classes and properties of a region.Soil maps provide support for decisionmaking about sustainable soil management at specific places across an area(Papadopoulou-Vryniotiet al.,2014;Pelegrinoet al.,2019).

Many studies regarding digital soil mapping which achieved successful predictions of soil classes and properties rely on correlation between the target soil variables to be predicted with terrain models(predictor variables)derived from digital elevation models(DEMs)(Mooreet al.,1993;Papadopoulou-Vryniotiet al.,2013;Adhikariet al.,2014;Silvaet al.,2016a;Henglet al.,2017).Results of prediction accuracy are also influenced by the spatial resolution of the DEMs and the terrain models derived from them(Sørensen and Seibert,2007;Cavazziet al.,2013).However,the best spatial resolution(smallest pixel size)does not always deliver optimal results(Hengl,2006;Mulderetal.,2011;Cavazziet al.,2013).Moreover,in areas with complex soil-landscape relationships,to create soil maps that display a variety of soil classes and properties based on terrain may not produce reliable results.This is especially apparent where soil variability is better explained by soil parent material(Curiet al.,1990).

Soil parent material information is not readily available on large-scale maps for many tropical countries,which hinders the prediction of soil classes(Manciniet al.,2019).Furthermore,many tropical soils are very deep,sometimes reaching tens of meters to the parent rock(Oliveiraet al.,2017).Thus,it is of critical importance to develop alternative methods that can infer parent material variability of an area for the creation of more detailed soil maps(Silvaet al.,2016b).Importantly,different soil parent materials can easily be detectedviachanges in fundamental elemental composition within soils(Stockmannet al.,2016).

In this sense,proximal sensors can be used in field and laboratory settings to provide soil scientists with rapid information on soil properties(Hartemink and Minasny,2014;Hortaet al.,2015;Dudaet al.,2017).Promising tools in this regard are the portable X-ray fluorescence(pXRF)spectrometer,which is capable of detecting and quantifying diverse elemental contents in seconds(Weindorfet al.,2014;Ribeiroet al.,2017),and the magnetometer,which determines the magnetic susceptibility(MS)of materials through the ratio of induced magnetization of an object to the intensity of the magnetizing field(Mullins,1977).While the potential of MS for diverse soil-related studies has been well documented(Schmidtet al.,2005;Haneschet al.,2007;Barrioset al.,2012;Camargo et al.,2014;Siqueiraet al.,2015),only recently has the number of studies involving pXRF sharply increased(Zhuet al.,2011;Weindorfet al.,2012;Sharmaet al.,2015;Dudaet al.,2017;Silvaet al.,2018;Teixeiraet al.,2018;Chakrabortyet al.,2019).

The pXRF is capable of detecting and quantifying multiple elements in seconds,ex situorin situ,at a low cost and without the generation of chemical residues(Weindorfet al.,2014;Hortaet al.,2015;Ribeiroet al.,2017).A pXRF spectrometer measures the secondary fluorescence energy emitted by each element when being hit by low energy X-rays exhibiting a characteristic fluorescence.Portable X-ray fluorescence sensors provide precise analytical results(Wang,2013;Westet al.,2013;Stockmannet al.,2016)and have been successfully used in recent pedological and environmental studies(Radu and Diamond,2009;Weindorfet al.,2012,2015;Aldabaaet al.,2015;Sharmaet al.,2015;Silvaet al.,2016b,2018;Dudaet al.,2017;Wanget al.,2017;McGladderyet al.,2018).

Several methods used to predict soil classes and properties include decision trees(Giassonet al.,2011;Pelegrinoet al.,2016;),support vector machine(Brungardet al.,2015;Taghizadeh-Mehrjardiet al.,2015),fuzzy logic(Shiet al.,2009;de Menezeset al.,2014;Pintoet al.,2017),and the Random Forest(RF)algorithm.The RF algorithm has been increasingly used for predictions(Radet al.,2014;Silvaet al.,2017)as it allows using both numerical and categorical variables,assessing the importance of each variable in the generation of the final model,and modeling non-linear relationships(Breiman,2001;Liaw and Wiener,2002).However,good results are not guaranteed,mainly due to the complexity of some areas(ten Catenet al.,2011),such as the campus of the Federal University of Lavras(Brazil),where this current study was conducted.This area contains variable geology and complex soil-landscape relationships(Curiet al.,2017).

The possibility of improving soil mapping in areas with complex soil-landscape relationships by combining proximal sensors,terrain models,and the RF algorithm for detailed soil mapping within tropical conditions is the justification for the present study.The objective of this study was to investigate the feasibility of using the RF algorithm to improve the prediction of soil classes at subgroup taxonomic levels using proximal soil sensor data(pXRF and magnetometer)associated with terrain variables at varying spatial resolutions and sources of terrain information.Furthermore,we sought to compare the effects of four DEM spatial resolutions and their related terrain maps on such predictions.We hypothesized that proximal sensor data associated with terrain information will improve the predictions relative to those based only on terrain information since proximal sensors can identify and quantify soil physicochemical properties that reflect soil parent material.Hence,soil mapping accuracy will be increased by including related parent material information(Fe,Ti and Si contents and MS data)into the modeling process.

MATERIALS AND METHODS

General occurrence and features

The study was conducted on the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil.The area has complex soil-landscape relationships(Fig.1).The regional climate is Cwb(subtropical highland),as per the Köppen classification,with mean annual rainfall and temperature of 1 529 mm and 19.5°C,respectively(Curiet al.,2017).Native vegetation features cerrado(Brazilian savannah)and tropical forest.Contemporary land use includes agricultural crops,such as coffee and corn,pastures,eucalyptus and pine plantations.In addition,there are research areas with agricultural crops in addition to pear,peach,and olive trees(Curiet al.,2017).The geology of the area includes gabbro,leucocratic and mesocratic gneisses from the base of slopes to the summits,and alluvial sediments on floodplains(Curiet al.,1990,2017).

Soil sampling and laboratory analyses

A total of 118 sampling sites were selected across the study area(Fig.1)and georeferenced using a Garmin handheld GPS model GPSMAP 78s(Olathe,USA),with an accuracy ofca.4 m.Soils were morphologically described and samples from the A and B horizons(one sample per horizon per location)were collected for laboratory characterization.Samples from 90(76%)of the 118 sites were collected in a regular grid design(200 m×200 m)and used for modeling,while those from the 28 remaining sites(24%)were used for validation of the prediction maps.Some of the 90 sampling points had to be relocated due to inaccessibility of the predefined sampling location.In these cases,soil samples were collected at a spot as close as possible to the original site.

Fig.1 Location of the study area of the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil,and sampling sites for modeling and validation.

The collected samples were air-dried,passed through a 2-mm sieve,and subjected to laboratory analyses required for soil classification(Curiet al.,2017).The main soil orders in the area include Oxisols(Ferralsols),Ultisols(Acrisols),Inceptisols(Cambisols),Entisols(Regosols),and,at smaller proportion,Histosols(Histosols)according to Keys to Soil Taxonomy(Soil Survey Staff,2014)and,in parenthesis,IUSS Working Group WRB(2015).The results of these analyses were used to taxonomically classify the soils at each site to the subgroup levels according to Keys to Soil Taxonomy(Soil Survey Staff,2014).

Soil sensor analyses

Air-dried and sieved samples were scanned with a Bruker®model S1 Titan LE pXRF(Billerica,USA)operated at 50 keV with a 100 μA X-ray tube.Scanning was conducted in Trace mode for 60 s in triplicate per sample according to the method described by Weindorf and Chakraborty(2016).These analyses allowed for the determination of Fe,Ti,and SiO2contents,which were subsequently used in model predictions.Instrument performance was confirmedviascanning of certified reference materials by the National Institute of Standards and Technology(NIST)(2710a and 2711a).Soil material certified by the pXRF manufacturer was also analyzed.All pXRF results were compared with the certified Fe,Ti,and SiO2contents to confirm optimal instrument performance.The recovery values(100×pXRF result/certified result)for Fe,Ti,and SiO2in NIST 2710a,NIST 2711a,and the manufacturer check sample were 88%,106%,and 98%,N/A,165%,116%,and 94%,103%,and 115%,respectively.The closer to 100%the recovery value,the more correct the pXRF quantification of Fe,Ti,and SiO2.

Magnetic susceptibility per unit of mass(χBF)was determined using a Bartington MS2B magnetometer(Bartington Instruments Ltd.,England)on the air-dried samples at low frequency(0.47 kHz).Calculations were made through the expressionχBF=10κ/m,whereκ(volumetric magnetic susceptibility)is dimensionless andmis the mass of the sample.To allow pXRF and MS data to be used in soil class predictions across the entire study area,these discreet observations were spatially interpolated through inverse distance weighting(IDW)interpolation using ESRI ArcGIS 10.3(Redlands,USA).This interpolated information was used in the models along with the other covariates(terrain data)for soil class prediction at each pixel throughout the area.To aid comprehension of the whole methodological process performed herein,a flowchart is presented(Fig.2).

Fig.2 Flowchart showing the methodological process adopted in this work.DEMs=digital elevation models;LCMS=local 1-m contour map source;IBMS=Institute of Geography and Statistics(IBGE)20-m contour map source;pXRF=portable X-ray fluorescence;IDW=inverse distance weighting.

Terrain variables

To obtain terrain variables commonly used for predicting and mapping of soil classes(Adhikariet al.,2014;de Menezeset al.,2014;Vågenet al.,2016),DEMs were generated from contour lines in ArcGIS 10.3 software,using the Topo to Raster function.To learn the effect of resolution on soil class predictions,DEMs were created with 5-,10-,20-,and 30-m resolutions.Additionally,for the study area,contour lines with 1 m of vertical distance were available from the local city hall(local 1-m contour map source(LCMS)).However,for most of the Brazilian territory,contour lines with 20 m of vertical distance are available from the Institute of Geography and Statistics(IBGE).As contour lines with larger vertical distance generate less detailed DEMs and derivative maps,less accurate soil class predictions would be expected using IBGE contour lines than using LCMS.To evaluate the influence of both contour lines on the generation of DEMs and derivative terrain maps at various spatial resolutions and in turn the predicted soil maps,DEMs with the aforementioned spatial resolutions from both IBGE 20-m contour map source(IBMS)and LCMS were created,comprising a total of eight DEMs plus their derivative maps.Thus,for each spatial resolution,soil maps were generated to:i)evaluate the differences provoked by the different spatial resolutions(5,10,20,and 30 m)and data sources of the contour lines(IBMS,with 20 m of vertical distance,and LCMS,with 1 m of vertical distance)on the accuracy of the predicted soil maps;and mainly ii)test whether proximal sensor variables in addition to terrain variables could enhance the performance of models to predict soil classes(mainly those models that used IBMS data).If confirmed,proximal sensor variables would contribute to improving soil mapping when used together with IBMS data widely and freely available in Brazil.

For each DEM,12 terrain variables were generated.Eleven of these variables were obtained using System for Automated Geoscientific Analysis(SAGA)GIS software(Conradet al.,2015)(Table I).The twelfth terrain variable was Geomorphons,which identifies the ten most frequent and common landform elements from a typical terrestrial landscape(flat,summit,ridge,shoulder,spur,slope,hollow,footslope,valley,and depression)(Jasiewicz and Stepinski,2013),allowing for correlations between each of these elements with the soil classes and properties typical of each landform(Silvaet al.,2016c).Geomorphons were derived from the DEMs using Geographic Resources Analysis Support System(GRASS)GIS 7.0 software(GRASS Development Team,2017).In addition to the 12 terrain variables,Fe,Ti,and SiO2data from pXRF and MS data in the diagnostic B horizon of each soil formed the 16 covariates employed to predict soil classes.

Soil class prediction and map accuracy

At each of the 118 sampling sites,the values of the 12 terrain variables were extracted and added to the pXRF and MS results.Next,the modeling and validation data sets were defined.The former was composed of data from 90 randomly chosen sampling sites(76%of the total),except for the inclusion of the three soil classes that were only found at one site(Mollic Endoaquent,Typic Udifolist,and Typic Udorthent).These three soil classes were inserted into the modeling data set to guarantee that the model would be able to predict all soil classes in the area.The validation data set was composed of data from the remaining 28 sampling sites(24%of the total),which were used to assess the accuracy of the predicted maps.

The RF package(Liaw and Wiener,2015)in R software(R Development Core Team,2009)was next applied to predict soil classes.Soil maps were generated with two types of data sets to evaluate the effects of covariate data from proximal sensors on soil class predictions:a)terrain variables(total of 12 variables)only and b)data from proximal sensors plus terrain variables(total of 16 variables).Thereafter,the 16 maps generated by RF(two sources of contour lines to generate DEMs×two types of data sets×four spatial resolutions)were validated by comparing the soil classes of the 28 places(validation data set)with their correspondingmodel-predicted soil classes using the overall accuracy(OA)and Kappa index.

TABLE ICharacterization and descriptive statistics of the covariates used to model and predict soil classes

The OA was calculated as the sum of the main diagonal components of the confusion matrix(correctly predicted sampling places)divided by the total number of sampling sites(118)(Alberget al.,2004):

wherejjrepresents the elements in the main diagonal that indicates the number of correctly predicted samples per soil class,Cis the number of soil classes,xjjis the number of correctly predicted sampling places,andNis the total number of sampling sites,which is 118.

The Kappa index was calculated taking into account the number of soil classes within the study area,the number of correctly classified places,and the total number of sampling sites(Congalton,1991):

wherePo is the proportion of correctly classified sampling sites andPe is the probability of random agreement.Kappa index ranges from-1 to 1,indicating increasing accuracy as it approaches 1(Landis and Koch,1977).

For a final evaluation,the most important covariates for each model were obtained from RF.For that,we ran the RF classification 30 times per model and recorded the importance rankings of the top five important variables for each of the 16 prediction maps,following the suggestion of Millard and Richardson(2015).Then,using only the five most important covariates per model,we created and validated 16 other maps,as previously described,to assess the accuracy changes by using only the most important covariates in the models.

RESULTS AND DISCUSSION

Soil class characterization by proximal sensors

The soils found in the study area were classified as Typic Dystrudept(Sideralic Dystric Cambisol(Loamic)),Mollic Endoaquent(Dystric Umbric Tidalic Gleysol(Clayic,Uterquic)),Typic Endoaquent(Dystric Tidalic Gleysol(Clayic,Inclinic,Uterquic)),Xanthic Hapludox(Xanthic Ferralsol(Dystric,Clayic)),Typic Hapludox(Rhodic Ferralsol(Dystric,Clayic)),Anionic Acrudox(Ferritic Gibbsic Rhodic Geric Ferralsol(Dystric,Clayic)),Rhodic Kandiudult(Rhodic Acrisol(Clayic,Cutanic)),Typic Udifolist(Tidalic Dystric Histosol(Claynovic)),Inceptic Hapludult(Xanthic Acrisol(Clayic,Differentic)),Typic Rhodudult(Acric Ferralic Dystric Nitisol(Densic)),Typic Hapludult(Chromic Acrisol Clayic,Differentic)),and Typic Udorthent(Dystric Regosol(Loamic Ochric)).The largest MS values were associated with Anionic Acrudox and Rhodic Kandiudults(Tables II and SI,see Supplementary Material for Table SI),as these soils developed from gabbro and contain higher amounts of magnetic minerals such as maghemite and magnetite(Curi and Franzmeier,1987).The smallest MS values were found for lowland soils(Mollic Endoaquents,Typic Endoaquents,and Typic Udifolists)due to the low presence of maghemite and magnetite under reducing conditions.Intermediate values were found for the other soils,all of which were derived from gneiss.

The B horizons of soils derived from gabbro(Anionic Acrudox and Rhodic Kandiudults)also had the highest Fe and Ti values and the lowest SiO2value(Tables II and S1).In the A horizons,these soils,plus the Typic Udifolists,had the highest Fe and Ti values and the lowest SiO2content.The Typic Udifolists featured an A horizon formed by colluvium from higher elevation areas,which are dominated by soils derived from gabbro.In addition,mean Fe contents in Typic Udifolists were lower than those found for Anionic Acrudox and Rhodic Kandiudults probably because under such low elevation conditions part of the Fe-bearing minerals might have been dismantled due to Fe reduction(Schaetzl and Anderson,2005;Resendeet al.,2014).These results indicate the potential of MS and pXRF to aid in separating soil classes in an easy way since they reflect soil parent material information and pedogenic processes.

Soil map validation

The maps generated solely with terrain variables of 5-,10-,20-,and 30-m spatial resolution showed OA values of 32%,14%,29%,and 25%,respectively,and Kappa index values of 0.15,-0.02,0.16,and 0.08,respectively,for the DEM IBMS source(Table III).However,when terrain variables were associated with proximal sensor data,these maps had OA values of 46%,25%,39%,and 43%,respectively,and Kappa index values of 0.32,0.11,0.28,and 0.30,respectively.These results reveal an average increase of 50%in OA and 154%in Kappa index when using proximal sensor data in association with terrain variables for the DEM IBMS source.This same trend was verified when LCMS was used for developing DEMs and their derivative maps,in which terrain variables associated with proximal sensor data presented superior accuracy.Thus,the data of pXRF and MS can be considered as important variables to aid in digital soil mapping.Figure 3 shows the maps with optimal accuracy.

The quality of soil maps increased for both contour line sources when using proximal sensor data associated with terrain variables at all DEM resolutions(Table III).The best OA and Kappa index values were obtained at resolution of 5 m.However,the second best prediction maps for the IBMSsourced contour lines had a resolution of 30 m,while for LCMS it had a resolution of 10 m,with an accuracy very similar to that of the 30-m resolution map.This indicates an absence of a trend that is directly related to spatial resolution,i.e.,prediction maps do not necessarily become more accurate with increasing spatial resolution(Silvaet al.,2016c).Similar results have been reported by Cavazziet al.(2013),who found that the optimal spatial resolution for mapping soils varied according to the topographic features of the areas mapped.Additionally,the difficulty in defining the proper DEM spatial resolution was related to finding the scale at which soil forming processes occur so that they could be well represented in a specific spatial resolution.Samuel-Rosaet al.(2015)evaluated the effects of increasing spatial resolution on mapping soil properties and found that the best results for clay content were not achieved using the best resolution DEM.For other mapped properties,increases in mapping accuracy were modest and did not justify the use of better resolution DEMs.

TABLE IIDescriptive statistics of magnetic susceptibility(MS)values and Fe,Ti,and SiO2 contents measured with a portable X-ray fluorescence(pXRF)spectrometer for the subsurface(B)horizon of the studied soil classes on the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil

TABLE IIIOverall accuracy(OA)and Kappa index values of the soil class prediction maps for the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil,created using terrain data of the Institute of Geography and Statistics(IBGE)20-m contour map source(IBMS)or local 1-m contour map source(LCMS)and based on terrain variables(T)only or on terrain variables associated with proximal soil sensors(T+S)at different spatial resolutions of digital elevation models

Fig.3 Optimal soil class prediction maps for the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil based on terrain data generated using local 1-m contour map source(LCMS)(a–d)or the Institute of Geography and Statistics(IBGE)20-m contour map source(IBMS)(e–h)at different spatial resolutions of 5(a and e),10(b and f),20(c and g),and 30(d and h)m of digital elevation models.1=Typic Dystrudept;2=Mollic Endoaquent;3=Typic Endoaquent;4=Xanthic Hapludox;5=Anionic Acrudox;6=Typic Hapludox;7=Rhodic Kandiudult;8=Typic Udifolist;9=Inceptic Hapludult;10=Typic Hapludult;11=Typic Rhodudult;12=Typic Udorthent.

The best results were always obtained with the association of terrain and proximal sensor data instead of using terrain data solely.However,even with this association of variables,the general accuracy of the maps was moderate.This is likely due to the high soil variability in this area(Curiet al.,2017),where soil-landscape relationships cannot be readily established to represent the soils.For example,Anionic Acrudox generally occurs in gently sloping areas,but in the studied area it was found on slopes with gradients from 3%to 40%.Contrary to many countries in temperate regions,most of Brazil contains soils millions of years old(Kämpf and Curi,2012).Such aging has promoted alterations on the landscape without the same rate of alterations in soils at the order level.There are reports of Oxisols found on steep slopes in Minas Gerais,Brazil,that can be explained as being formed from the weathering of parent materials such as tuffite during dry conditions on nearly flat topography,which eventually steepened(Ker,1998).Additionally,there are reports of Inceptisols that occur on gently sloping areas developed from the erosion of ancient soils(Skorupaet al.,2017).In an attempt to improve the results,modeling using only the most important covariates(terrain variables and proximal sensor data)indicated by RF was performed.It was based on the findings of Millard and Richardson(2015),who found that predictions for imagery classification were improved when using only the most important variables instead of using all the predictor variables for generating prediction modelsviaRF.

Covariate importance

For all DEM resolutions,MS,Fe content,and Ti content,in this order,were the three most important covariates for all soil class prediction maps(Table IV).Since MS and soil chemical composition are related to parent material,in areas with highly variable geology,the use of proximal sensors for identifying and quantifying these properties can improve soil mapping precision considering that these soil properties can be a proxy of soil parent material information when scanning is performed in the laboratory on air-dried and ground samples(Silvaet al.,2016b).The other variables that were among the five most important included(4thand 5thpositions):multiresolution index of valley bottom flatness(four times),topographic wetness index,profile curvature,aspect,elevation,slope(two times each),vertical distance to channel network,and SAGA wetness index(one time each).However,the following variables were not among the five most important for any modeling:SiO2,multiresolution index of valley bottom flatness,profile curvature,terrain surface classification,and geomorphons.

TABLE IVFive most important covariates(I,II,III,IV,and V,from most to least)for each soil class prediction model according to the Random Forest algorithm for the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil,developed using terrain data of the Institute of Geography and Statistics(IBGE)20-m contour map source(IBMS)or local 1-m contour map source(LCMS)at spatial resolutions of 5,10,20,and 30 m

There are some areas where the variety of soil classes and properties are better explained by soil parent material than topography(Skorupaet al.,2016).In this sense,among all the covariates,MS has a greater influence on the predicted maps(Fig.4)and the accuracy of models when using proximal sensor data(Table IV),regardless of the spatial resolution(5,10,20,or 30 m)for both DEM sources.Using the RF algorithm,Akpaet al.(2014)found a large geologic influence on soil particle-size fractions in Nigeria,demonstrating the influence of lithology on soil properties.Siqueiraet al.(2015)found that MS and soil-landscape relationships were effective in delineating soil map units at a detailed scale,identifying areas with different patterns of pedogenetic variability.

Fig.4 Soil class prediction maps for the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil based on terrain data generated using local 1-m contour map source(LCMS)(a–d)or the same terrain data associated with proximal soil sensors(e–h)at different spatial resolutions of 5(a and e),10(b and f),20(c and g),and 30(d and h)m of digital elevation models.1=Typic Dystrudept;2=Mollic Endoaquent;3=Typic Endoaquent;4=Xanthic Hapludox;5=Anionic Acrudox;6=Typic Hapludox;7=Rhodic Kandiudult;8=Typic Udifolist;9=Inceptic Hapludult;10=Typic Hapludult;11=Typic Rhodudult;12=Typic Udorthent.

Refined predictions

Following Millard and Richardson(2015),the five most important covariates per model,i.e.,for different DEM sources and resolutions(Table V),were used to generate and validate 16 new soil maps based on terrain variables(T)only and terrain variables associated with proximal sensor data(T+S).The soil map accuracies tended to improve when using only the five most important covariates for prediction(Table V).Essentially,such a technique is a practical means for avoiding data overfitting.However,for the resolution of 5 m(IBMS),there was a 4%reduction in accuracy when using the T+S association.Figure 5 shows the best maps generated using only the five most important covariates.

In general,the accuracy indices increased when using only the most important covariates(Table V)instead of using several covariates(Table III)during the RF analyses for both DEM sources(IBMS and LCMS).The same trend was observed for the covariates encompassing terrain variables and proximal sensor data(T+S),mainly for the LCMS data source(Table III).The T+S association improved soil class prediction as shown in the results of OA and Kappa index,especially for the LCMS data source.

The LCMS data source resulted in the best indices(OA and Kappa index),and the maps with resolutions of 30,20,and 5 m had higher values of accuracy when using the T+S association(Table V).The best result for the IBMS data source was achieved with the 30-m resolution map,likely because generating greater spatial resolution DEMs(smaller pixel sizes)from those less detailed contour lines may increase the errors of the DEMs and their derivative maps and,hence,decrease the performance of the prediction models.Penížeket al.(2016)obtained optimal results with resolutions of 2–10 m when evaluating the effect of DEM spatial resolution(1–30 m)in association with different algorithms to delineate colluvial soil areas.Specifically,the finest resolution contained localized features that impaired the models,whereas coarser resolution neglected some important terrain features.In both cases,quality was reduced.

Since proximal sensors in digital soil mapping are relatively novel in soil science,these results confirm that using

TABLE VOverall accuracy(OAs)and Kappa index values of the soil class prediction maps generated using only the five most important covariates for the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil based on the terrain data obtained from the Institute of Geography and Statistics(IBGE)20-m contour map source(IBMS)and local 1-m contour map source(LCMS),created based on terrain variables(T)only or on terrain variables associated with proximal soil sensors(T+S)at different spatial resolutions of digital elevation models

Fig.5 Optimal soil class prediction maps for the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil based on terrain data associated with proximal soil sensors obtained from local 1-m contour map source(LCMS)(a–d)or the Institute of Geography and Statistics(IBGE)20-m contour map source(IBMS)(e–h)at different spatial resolutions of 5(a and e),10(b and f),20(c and g),and 30(d and h)m of digital elevation models.1=Typic Dystrudept;2=Mollic Endoaquent;3=Typic Endoaquent;4=Xanthic Hapludox;5=Anionic Acrudox;6=Typic Hapludox;7=Rhodic Kandiudult;8=Typic Udifolist;9=Inceptic Hapludult;10=Typic Hapludult;11=Typic Rhodudult;12=Typic Udorthent.

pXRF and MS can be very promising for digital soil mapping,as postulated by Mahmoodet al.(2012)and Silvaet al.(2016b).Some previous studies in Brazil presented adequate mapping accuracy,mostly without using proximal sensor data as predictor variables(Souzaet al.,2016;Menezeset al.,2018),but others argued that accurate mapping in some areas would not be possible without proximal sensor data(Silvaet al.,2016b).The approach presented here can contribute to future accurate and detailed digital soil mapping in Brazil,mainly because pXRF and MS are able to provide detailed information related to parent material,which is not easily obtained in Brazil.Other proximal sensors may also contribute to this purpose,such as the visible-near-infrared(Vis-NIR)diffuse reflectance spectroscopy,which has proven useful in predicting soil class and properties in Brazil(Terraet al.,2015;Rizzoet al.,2016;Ramirez-Lopezet al.,2019)and other parts of the world(Bagheri Bodaghabadiet al.,2015;Wanget al.,2015;Dudaet al.,2017).

Although the general accuracy in the present study was somewhat low due to the complexity of the area,the use of a dense sampling scheme(one sample per 3.5 ha),a robust algorithm,high resolution terrain models,and proximal sensors,enabled inferences of soil parent material changes within an area that lacks detailed lithological information.This may help to understand soil class distribution patterns across the landscape,improving class prediction and,ultimately soil mapping.Although the use of proximal sensor data required the interpolation of their values from the places samples were collected in order to make them continuous to the whole study area(a procedure that may lead to accumulated errors),incorporating such data into the models improved the overall results compared to using terrain data solely.Therefore,according to the presented results,pXRF and MS data have high potential to be used in association with terrain models for predicting soil classes and to improve the scale of existing maps by using sensor data as covariates for prediction,despite the need for spatially interpolating the collected sensor data.

CONCLUSIONS

Based on the RF algorithm,MS and pXRF-determined Fe and Ti contents were the three most important covariates for soil class prediction for the 316-ha campus of the Federal University of Lavras,Lavras County,Minas Gerais State,Brazil.The quality of the predicted soil maps increased when integrating proximal sensor data with terrain variables for all spatial resolutions evaluated.The most accurate maps for both LCMS and IBMS data sources were at 5 m of spatial resolution,followed by maps using 30-m resolution.The IBMS data source was generally considered optimal.Soil map accuracy improved when using only the five most important covariates in predictions instead of all variables in the RF algorithm.However,optimal maps were not always reliant upon the finest spatial resolution.Summarily,the approaches presented here offer a simple alternative for increasing map accuracy in support of digital soil mapping in tropical countries using proximal sensor data as predictor variables concomitantly with terrain information.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge the assistance from the BL Allen Endowment in Pedology at Texas Tech University,USA and the Brazilian funding agencies National Council for Scientific and Technological Development(CNPq)(Nos.301930/2019-8 and 306389/2019-7),the Coordination for the Improvement of Higher Education Personnel(CAPES),Brazil(No.590-2014),and Research Support Foundation of the State of Minas Gerais(FAPEMIG),Brazil(No.PPM 00305-17)for the financial support provided.

SUPPLEMENTARY MATERIAL

Supplementary material for this article can be found in the online version.