Digital soil mapping at local scale using a multi-depth Vis ... - FEAGRI

Geoderma 274 (2016) 18–27

Contents lists available at ScienceDirect

Geoderma journal homepage: www.elsevier.com/locate/geoderma

Digital soil mapping at local scale using a multi-depth Vis–NIR spectral library and terrain attributes Rodnei Rizzo a, José A.M. Demattê b,⁎, Igo F. Lepsch c, Bruna C. Gallo b, Caio T. Fongaro b a b c

Environmental Analysis and Geoprocessing Laboratory, Center for Nuclear Energy in Agriculture, University of São Paulo, Av. Centenário, 303, PO BOX 96, 13416-000 Piracicaba, SP, Brazil College of Agriculture “Luiz de Queiroz”, University of São Paulo, ESALQ/USP, Department of Soil Science, Av. Padua Dias 11, PO Box 9, CEP 13418-900 Piracicaba, SP, Brazil Visiting Soil Scientist, ESALQ/USP, Department of Soil Science. Av. Padua Dias 11, PO Box 9, CEP 13418-900 Piracicaba, SP, Brazil

a r t i c l e

i n f o

Article history: Received 11 June 2015 Received in revised form 10 February 2016 Accepted 20 March 2016 Available online xxxx Keywords: Regional soil spectral library Regression tree Fuzzy c-means Munsell color Brazilian Soil Classification System

a b s t r a c t Conventional soil mapping is costly and time consuming. Therefore, the development of quick, cheap, but accurate methods is required. Several studies highlight the importance of developing regional soil spectral libraries for digital soil mapping, but few studies report on the use of these libraries to aid digital mapping of soil types. This study aims to produce a digital soil map using as training set Visible and Near Infra-Red (Vis–NIR) spectra from local soil samples, a regional spectral library and terrain attributes. The soils were sampled in 162 locations on a 270-ha farm in the municipality of Piracicaba, São Paulo, Brazil. Spectra from topsoil and subsoil were measured in laboratory (400–2500 nm) and arranged as multi-depth spectra. Information was summarized by principal component analysis. Regression tree models were calibrated to predict principal components (PC) scores based on terrain attributes. After calibration, the models were applied to the entire study site, resulting in PC score maps. Fuzzy c-means and PC maps were used to define the soil mapping units (MU). Based on fuzzy centroids, representative samples (RS) were defined to the MU. Munsell soil color and soil order were predicted from soil spectra and used to characterize the MU. The regression tree model had a good fit for PC1, with an r2 of 0.92, and a satisfactory r2 for PC3, PC4, and PC5, respectively 0.58, 0.66 and 0.53. The fuzzy clustering defined seven MU. The R2 for Munsell color predictions were 0.94 (hue), 0.96 (value) and 0.73 (chroma). Soil order had good agreement in validation, with kappa coefficient of 0.41. The methodology indicates the potential of Vis–NIR spectra to improve soil mapping campaigns and consequently provides a product similar to a conventional soil map. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Choosing species and crop varieties adapted to the various kinds of soils delineated in maps could provide higher crop yields. However, soil maps with adequate scale for crop management are scarce (Mendonça-santos and Dos Santos, 2006). Ben-Dor et al. (2008) reported on high cost associated with soil surveying and mapping. One alternative to reduce costs could be the adoption of Digital Soil Mapping (DSM). A number of studies describe DSM techniques to create maps of soil attributes or even soil types (identified as taxonomic classes). In most cases, these maps are derived from a calibration set (punctual information related to chemical and physical properties or soil classification) and environmental covariates, such as terrain attributes and satellite images (Adhikari et al., 2014; Lagacherie et al., 2012; Vasques et al., 2015).

⁎ Corresponding author. E-mail addresses: [email protected] (R. Rizzo), [email protected] (J.A.M. Demattê), [email protected] (I.F. Lepsch), [email protected] (B.C. Gallo), [email protected] (C.T. Fongaro).

http://dx.doi.org/10.1016/j.geoderma.2016.03.019 0016-7061/© 2016 Elsevier B.V. All rights reserved.

Visible and near-infrared (Vis–NIR) spectroscopy can be a useful indicator of soil variability (Demattê and Terra, 2014). The ability to obtain a large number of information at lower costs or short time allows increasing the number of observations and consequently improving digital soil mapping (Viscarra Rossel et al., 2009). Recently, the joint effort of researchers from several countries has resulted in the establishment of a global soil spectral library (Viscarra Rossel et al., 2016). These databases have a great potential to improve accuracy of digital soil maps, providing information about the most relevant soil attributes, enabling spatio-temporal monitoring of soils in many regions worldwide. Given that soil spectra carry information about many soil attributes (Soriano-Disla et al., 2014), studies have suggested that the spectra could also be used to measure similarities between soil types and consequently provide soil classification (Vasques et al., 2014; Viscarra Rossel and Webster, 2011). Bellinaso et al. (2010) used a regional soil spectral library to describe and classify soil profiles according to the Brazilian Soil Classification System (SiBCS) (Embrapa, 2013). Ben-Dor et al. (2008) developed the 3S-HeD, a device able to improve reflectance data measurement on the field. The authors attached a field spectrometer to this device and performed a quantitative profile description based on

R. Rizzo et al. / Geoderma 274 (2016) 18–27

Vis–NIR spectra. Vasques et al. (2015) elaborated a digital soil map of SiBCS suborder level based on satellite images, terrain attributes and interpolated average reflectance from soil Vis–NIR spectra. While many studies apply Vis–NIR spectra to improve digital mapping of soil attributes, the synergy between proximal Vis–NIR sensing and soil types has been little explored. Clearly, there is a need for strategies using Vis–NIR spectra on DSM of soil types (identified as soil classes). The aim of our study was to test a digital mapping technique that uses as training set (i) soil spectra from local samples, (ii) a regional spectral library and (iii) terrain attributes. Spectra of local samples are used in many steps of the mapping process to (a) define the mapping units (MU), (b) select representative samples (RS) from each MU and (c) classify the soil types according to SiBCS order level (Embrapa, 2013). 2. Material and methods 2.1. Study site The study site is a 270 ha farm located in the municipality of Piracicaba, São Paulo State, Brazil, between the coordinates 22°42′30″– 22°43′27″S and 47°33′32″–47°34′45″W (Fig. 1). Lithology is diabases from the Serra Geral Formation, argillaceous siltstone and argillites from the Tatuí Formation and argillites from the Irati Formation (Vidal-Torrado et al., 1999). The climate is “Cwa” subtropical with dry winters and rainy summers (Koppen classification). Annual rainfall ranges from 1250 to 1500 mm. Relief consists of two interconnected hills with dominantly convex slopes ranging from 2% (on the hilltops) to 12% (on the foothills). 2.2. Data acquisition A 30-meter resolution digital elevation model (DEM) of the study site was obtained from a topographical chart (1:10,000 scale) (Hutchinson, 1993). Later, the SAGA GIS (System for Automated

19

Geoscientific Analyses) was used to derive the following terrain attributes: Altitude Above Channel Network (AACN), Aspect (ASP), Catchment Area (CA), Channel Network Base Level (CNBL), Curvature (CUR) (Zevenbergen and Thorne, 1987), Hillshade (HIL), Topographic factor (LSF), Slope (SLOP) (Horn, 1981), Stream Power (SP) (Moore et al., 1993), Terrain Roughness (TR), Topographic Wetness Index (TWI) (Moore et al., 1993), Vector Terrain Roughness (VTR) (Hoffman and Krotkov, 1990) and Wetness Index (WI) (Moore et al., 1993). Soils were sampled with an auger along five toposequences, in a 30meter interval, at two depths (0–20 cm and 80–100 cm). We collected 324 samples that were dried at 50 °C and sieved through a 2-mm mesh. Fractions smaller than 2 mm were used for laboratorial analyses. The color of dry soil was measured with a Minolta colorimeter (CR– 300), adjusted to the Munsell color system (Campos et al., 2003). The soils were classified at the suborder level according to the Brazilian Soil Classification System (SiBCS) (Embrapa, 2013). The corresponding World Reference Base (IUSS Working Group WRB, 2014) and Soil Taxonomy (Soil Survey Staff, 2014) classes are shown in Table 1. The soil samples spectra were measured in the laboratory using a FieldSpec Pro spectrometer (Analytical Spectral Devices, Boulder, CO) considering a spectral range between 400 and 2500 nm. The system geometry corresponded to the perpendicular position of the sensor in relation to the sample at a distance of 27 cm. The light source was positioned at 61 cm from the sample and at an angle of 20° with the zenith. The absolute reference standard used was a white spectralon plate.

2.3. Spatial modeling of soil multi-depth spectra Spectra from the two sampled soil depths were joined in sequence to create a pseudo multi-depth soil spectrum (Vasques et al., 2014). The principal component analysis (PCA) was applied to summarize information in the spectra, resulting in 5 principal components (PC). The PCA was performed using the interactive NIPALS algorithm (Martens and Naes, 1989) implemented in Parles 3.01 (Viscarra Rossel, 2008).

Fig. 1. Location of the study site, the samples collected in the site and samples from the regional spectral library.

20

R. Rizzo et al. / Geoderma 274 (2016) 18–27

Table 1 Soil suborders according to the Brazilian Soil Classification System (SiBCS; EMBRAPA, 2013) and corresponding World Reference Base (WRB; IUSS Working Group WRB, 2014) and Soil Taxonomy (Soil Survey Staff, 2014) classes. Source: adapted from Vasques et al. (2014). Identif.

SiBCS classification

WRB classification

Soil taxonomy classification

CX LV LVA NV NX PA PV PVA R

Cambissolo Háplico Latossolo Vermelho Latossolo Vermelho-Amarelo Nitossolo Vermelho Nitossolo Háplico Argissolo Amarelo Argissolo Vermelho Argissolo Vermelho-Amarelo Neossolo

Cambisol Ferralsol Ferralsol Nitisol Nitisol Lixisol Lixisol Lixisol Leptosol, Regosol

Udepts Udox Udox Udalfs, Udults Udalfs, Udults Udalfs, Udults Udalfs, Udults Udalfs, Udults Lithic Udorthents,Lithic Udipsamments, Udorthents, Udipsamments

The next step consisted of calibrating a model to estimate the PC according to the predictor variables, i.e. terrain attributes. A model tree based on the M5 method (Quinlan, 1992) and implemented in R package Cubist (Kuhn et al., 2014) was used. The algorithm applies a recursive partitioning on the predictor variable space using a divide-andconquer strategy to build a tree like model (Viscarra Rossel and Chen, 2011). The partitions are generated according to a set of binary decision rules to reduce and minimize the intra-subset variation at each node (Henderson et al., 2005). This technique divides the variable space in smaller regions, where interactions are usually easier to describe and it creates a multivariate linear least-squares model for each of the partitions. During the calibration process, the effect of model parameters on prediction performance was tested based on a 10-fold cross validation. In this case, 130 (80%) observations from the dataset were used. During the process, up to 20 committee and 100 rules were tested to select the model with the best performance. The selected model was validated with an independent set, that is, the remaining 32 (20%) observations. The statistics considered in the predictions assessment were the Spearman's rank correlation coefficient (Cr), average error (AE), relative error (RE) and root mean-square error (RMSE). Cr measures the linear relationship between observed and predicted values, AE corresponds to the calculation of the mean error and indicates how close predictions are from the measured value, RE corresponds to: 1 XN jy −yi j i¼1 i RE ¼ N X 1 N jy −yj i¼1 i N and RMSE to: vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u N u1 X RMSE ¼ t ðy −yi Þ2 N i¼1 i where yi and yi are the observed and predicted values, y is the average of predicted values, N is the number of data. In general, useful models have an RE value smaller than 1, otherwise, there is a little predictive capacity. The cubist model also provides the set of rules and use frequency of each predictor in both conditions and linear models. This information indicates the importance of each terrain attribute to explain the variation of multi-depth spectra in the study site. After testing, the models with a reasonable fitting were generalized to the entire site using the terrain attributes. 2.4. Digitally mapping soil units Assuming the combined spectral information as a good indicator of soil variability, spectra multi-depth maps were used in a clustering

technique. A fuzzy c-means algorithm (Bezdek, 1981) was applied, dividing the landscape into different groups, i.e. defining the MU according to multi-depth spectra. The fuzzy was applied several times, varying the number of clusters from 5 to 10. Two indexes were evaluated to establish the best number of groups, the Xie and Beni index (XB) (Xie and Beni, 1991) and the partition coefficient (PTC) proposed by Bezdek (1981). The best partition corresponded to the lowest XB value and the highest PTC value. After calculating the validation indexes and defining the optimal number of clusters, the map representing the spatial distribution of fuzzy groups (i.e.: MU) was defined.

2.5. Characterizing soil mapping units After the MU were defined, the next step consisted on identify the soil order and Munsell color of each unit (Fig. 2). Soil color is an important characteristic and is used as a criterion in SiBCS suborder (2nd level) classification. First, a representative soil sample had to be defined to each MU. Considering that fuzzy algorithm creates centroids to every MU, soil samples most similar to the centroids were chosen as RS. A soil sample (superficial and sub-superficial) for each MU was selected based on the highest fuzzy membership values. The spectral library developed from Bellinaso et al. (2010) was the basis to identify the SiBCS soil order of the RS. This library was built using soil samples collected from field surveys in the Piracicaba region. A subset of this library was defined (SRSSL) and used in our study. In this subset, soil samples developed from the same parental materials of our study site were selected. A total of 191 sampled locations (including soil pits and auger borings) (Fig. 1) were used in the SRSSL. To predict soil order, the SRSSL information from both diagnostic horizons (superficial and subsuperficial) was combined (multidepth spectra). Later, soil order centroids were calculated averaging the characteristics values (principal component scores) of each soil class on the database. The classification was based on the Manhattan distance between the RSs and soil order centroids from the SRSSL, where the representative sample adopts the soil order of the nearest SRSSL centroid. The predictions of Munsell color required only the spectra from subsoils, here considered as the subsurface diagnostic horizon. Munsell color prediction combined methodologies described by Torrent and Barrón (1993) and ViscarraRossel et al. (2006). Briefly, RS spectra in the visible range were converted to the XYZ color system, integrating spectra and color-matching functions (x, y, z) of the Standard Observer D65 (CIE, 1931). The XYZ coordinates were converted to L*a*b*, coordinates a* and b* allowed the calculation of hue angles and chroma, while value was estimated based on L*. Using a LUT color conversion table, the hue angle was converted to Munsell notation (ViscarraRossel et al., 2006).

R. Rizzo et al. / Geoderma 274 (2016) 18–27

21

and subsoil. Soil color predictions were compared to colorimeter measurements. Soil orders were validated by comparing (in each 32 location sampled) the digital soil map polygons identifications with the soil classification performed by a pedologist. A confusion matrix and the kappa coefficient (Congalton et al., 1983) were calculated.

2.7. Comparison between conventional and digital soil maps A conventional soil map (CM) (1:10,000 scale) was designed and compared with the digital map. To allow a consistent comparison, the digital soil map classification was adjusted to the suborder level by following specifications of the 2nd categorical level of SiBCS (predicted soil order and Munsell color). Later, the agreement rate between maps was calculated. Note that the CM was not considered as a “ground truth”, but rather as an alternative soil map from the same site.

3. Results and discussion 3.1. Spatial variability of soil multi-depth spectra

Fig. 2. Flowchart of the methodology employed on digital soil mapping.

2.6. Validation To measure the fuzzy clustering ambiguity, the confusion index (CI) was calculated (Burrough and McDonnell, 1998) for each cell of the spatial grid. Burrough et al. (1997) defined the CI as the ratio between the second and the first highest membership values at an observation. If CI value is close to 0, the observation is strongly associated to the first dominant class and consequently there is little confusion. However, if CI is close to 1, the difference between the two membership values is small, implying more confusion. Higher confusion indexes might be associated to an intergrade between the two dominant classes. An independent validation set corresponding to 32 sampled locations were used to validate the technique. The soil color of each MU was predicted from the representative samples at 80–100 cm depth. On the other hand, the validation considered predictions from topsoil

The PCA analysis was performed on the multi-depth spectra and PC 1 to 5 were selected. These PC explained, respectively, 98.5%, 0.97%, 0.21%, 0.13% and 0.05% of data variability. The PC prediction had good to satisfactory results (Table 2). The r2 ranged from 0.29 to 0.92 in the independent validation. PC1 had the best r2 (0.92), while r2 of PC2 was 0.29. The relative error had better results for PC1 and PC3, PC4 presented RE values of 0.93 and in PC 2 and PC 5, RE values were higher than 1. Due to the unsatisfactory performance of PC2 estimates (RE = 1.43), this information was removed from the analysis. On the other hand, PC5 presented a reasonable r2 and error besides a coherent spatial distribution. Therefore, it was decided to keep it in the analysis despite the RE value 1.05. The PC loadings (Fig. 3) are important indicators of the influence of each wavelength on the PC scores variability, highlighting which soil properties are represented by the scores. The influence of soil attributes on the spectra is well discussed in several studies (Ben-Dor et al., 1999), Clark et al., 1990), Stenberg et al., 2010), thus, we provide only a brief description. The first PC loadings (Fig. 3a) indicate that at both soil depths, the spectral albedo explains most of PC variation. In general, the soil albedo is related to texture, organic carbon (OC) and mineralogy (magnetite and ilmenite), however, since the site has been cultivated with sugarcane in the last 20 years, low OC rates are observed. Consequently, only texture and mineralogy could be responsible for differences in the albedo. Variability on these attributes is conditioned by parental material, which, in turn, is closely related to elevation. Soils located at higher elevations were formed from diabases and showed higher contents of opaque sand minerals and Fe oxides. At intermediate to low elevations, argillite and siltstones predominates, forming soils with medium-to-high clay contents with low contents of iron oxides (Vidal-Torrado et al., 1999). This agrees with the PC1 map (Fig. 4), which has low PC1 values at the summit and higher scores on the foothill sloping locations.

Table 2 Tree model performance for prediction of 5 principal components, considering an independent validation set. The parameters considered were r2, RMSE, correlation coefficient, average error and relative error.

PC1 PC2 PC3 PC4 PC5

Comm.

r2

Cr

RMSE

AE

RE

10 10 20 20 10

0.92 0.29 0.58 0.66 0.53

0.96 0.02 0.72 0.53 0.43

1.25 0.82 0.12 0.26 0.20

1.53 1.16 0.16 0.41 0.20

0.27 1.43 0.75 1.05 0.92

22

R. Rizzo et al. / Geoderma 274 (2016) 18–27

Fig. 3. Loadings from (a) PC1, (b) PC3, (c) PC4 and (d) PC5, for sampled depths. A (0– 20 cm) and B (80–100 cm).

Fig. 4. Principal component maps of the multi-depth spectra and aerial photo of the field.

The most important terrain attributes for PC1 prediction (Fig. 5a) were elevation and altitude above channels, accounting for 88% and 62% to linear models. Other attributes with relevant contribution were slope (49%), aspect and hillshade (41%) (Fig. 5a). According to PC3 loadings (Fig. 3b), superficial and sub-superficial soil spectra contributed to variation in this PC, however, subsoils had higher loading values. Loadings from topsoil spectra are pronounced between 500 and 1000 nm, mainly at 680 nm, indicating influence of soils with high Fe oxides contents. Moreover, within range 1200–2400 nm, the observed small positive values are probably related to variations in clay contents. In subsoils, loadings from 500 to 1800 nm presented a similar pattern for surface loadings. Due to kaolinite and 2:1 minerals, features at 1900–2200 nm were more pronounced in the subsoil.

Comparison between PC3 map (Fig. 4) and an aerial-photo (bare soil) (Fig. 4) show that lower PC3 values occur where the red soils are, that is, at the summit, while higher PC3 values occur where the yellow soils are. Terrain attributes with higher contribution to the linear models were DEM (75%), AACN (60%), slope (40%), wetness index (33%) and channel network base level (31%) (Fig. 5b). Besides elevation, which is an indicative of geology, terrain attributes such as slope and wetness index are important to understand water dynamics in these soils and have strong influence on the weathering process, regulating soils mineralogy. PC4 has positive loadings from 400 to 1850 nm and negative values for wavelengths greater than 1900 nm. This pattern is observed both in topsoil and subsoil samples (Fig. 3c). Features at 530 and 880 nm

R. Rizzo et al. / Geoderma 274 (2016) 18–27

23

Terrain attributes that most contributed to the linear model were elevation (78%) and slope (49%) (Fig. 5c). The areas at the summits presented the highest scores while the lowest values were observed at the foothills (Fig. 4). In PC5 loadings (Fig. 3d), spectral features that most influenced topsoil scores were 530, 880, 2200 nm, besides the range from 1800 to 2400 nm. These wavelengths are related to hematite and kaolinite, while the range 1800–2400 nm also corresponds to variations in soil texture. Based on loadings from sub-superficial samples, three features influenced PC5, namely 530 nm, 700 nm (related to goethite) and 1900 nm (hygroscopic water in 2:1 clay minerals). Terrain attributes that most contributed to PC5 linear models where elevation (66%), altitude above channel network (59%), slope (53%) and terrain ruggedness (48%) (Fig. 5d). PC5 spatial distribution (Fig. 4) is mainly related to channel network and terrain ruggedness, where highest PC values are found at the stream bottom while low values occur at flat and high locations.

3.2. Mapping units and representative samples The MU were established based on the PC maps. The optimal number of units was based on the XB and PTC indexes (Fig. 6). According to the XB index, data clustering from 5 to 7 units tends to improve compactness and separation, however, dividing the area into 8 to 10 MU increases the index and consequently affects the results. For the PTC index, dividing the dataset from 5 to 7 units also provides better results, i.e. it creates units with lower fuzziness and reduces heterogeneity. From 8 to 10 MU, the PTC index reduces and presents a small variation, confirming that the best condition is reached at 7 units. The membership value maps (Fig. 7a) represent the spatial distribution and similarities between the MU. In general, the MU have a welldefined position in the landscape and low confusion. For example, MU 1 and 2 had high membership values at the summit and shoulder, respectively, and low values at the footslopes. MU 3 and 5 presented intermediary degree of similarity when compared to all other MU. Probably these MU are an intergrade between MU 1 and 7. MU 4, 6 and 7 occur in soils formed from argillites and siltstones and consequently have low similarity with MU 1 and 2, derived from diabase. Furthermore, MU 6 and 7 are located at a steeper foothill area, which explains not only the similarity between these two MU, but also the difference from summit soils. Based on fuzzy membership values, soil samples with the highest similarity to fuzzy centroids were selected as representative. The spatial distribution of these samples comprises soils derived from different geologic materials and diverse landscape features (Fig. 7b). In general, representative samples corresponded to areas with low confusion index, that is, samples from MU 1, 2, 4, 5 and 6 with CI values between

Fig. 5. Contribution of individual predictors (terrain attributes) to the model trees in both the conditionals and linear models used to predict (a) principal component 1, (b) principal component 3, (c) principal component 4 and (d) principal component 5.

confirm hematite influence, while features at 1900 and 2200 nm show influences of 2:1 clay minerals and kaolinite. Unlike PC3, the loadings ranging from 1000 nm to 1800 nm are not so relevant. However, the spectral range from 1900 nm to 2400 nm contributed to PC variability.

Fig. 6. Values of Xie & Beni index and Partition Coefficient, indicating the performance of fuzzy c-means when clustering the data from 5 to 10 partitions.

24

R. Rizzo et al. / Geoderma 274 (2016) 18–27

Fig. 7. Parameters of the fuzzy c-means analysis. (a) Membership value maps for the 7 mapping units; and (b) crisp soil units map, the confusion index map and representative samples selected for each Mu.

0.13 and 0.28. Samples from MU 3 and 7 had CI values of approximately 0.6, indicating that the 1st and 2nd highest membership values have a ratio lower than 0.4, therefore, in these sampling locations similarity between two of the MU is high. 3.3. MU identification and validation MU identification consisted in defining soil Munsell color and soil order. The soil color predictions were similar to the colorimeter measurements (Fig. 8) with R2 0.95 and 0.96 to hue number and value,

respectively, while Chroma had a determination coefficient of 0.73. The RMSE for Hue, Value and Chroma were 0.57, 0.19 and 0.29, respectively, indicating a good performance of the color prediction model, agreeing with Post et al. (1993). The soil order prediction had a good agreement with conventional classification with kappa coefficient 0.41 (Table 3). The method identified the three most prominent soil orders of the study site, Latossolos, Nitossolos and Cambissolos, which, altogether occupy more than 80% of the site. These orders were accurately identified in 70.6, 83.3 and 100% of the observations. In the validation set, there were also

Fig. 8. Comparison of the estimated hue number, value and chroma with the colorimeter measurements. Hue number correspondence: 10 = 10R; 12.5 = 2.5YR; 15 = 5YR; 17.5 = 7.5YR; 20 = 10YR; 22.5 = 2.5 Y.

R. Rizzo et al. / Geoderma 274 (2016) 18–27

25

Table 3 Confusion matrix and kappa coefficient of observed and predicted soil orders, calculated from the validation set. Soil classif.a

Predicted

Cambissolo Gleissolo Latossolo Nitossolo Argissolo Per cent correct

Observed Kappa Cambissolo

Gleissolo

Latossolo

Nitossolo

Argissolo

2 0 0 0 0 100.0

2 0 0 0 0 0.0

0 0 12 5 0 70.6

0 0 1 5 0 83.3

2 0 0 3 0 0.0

0.41

-

a

Soil orders and corresponding World Reference Base (WRB; IUSS Working Group WRB, 2014) and Soil Taxonomy (Soil Survey Staff, 2014) classes: Cambissolo–Cambisol, Udepts; Latossolo–Ferralsol, Udox; Nitossolo–Nitisol, Udalfs/Udults; Argissolo–Lixisol, Udalfs/Udults.

observations identified as Gleissolo and Argissolo, but the method was not able to accurately predict them. The SRSSL does not contain Gleissolos observations and consequently it was not possible to identify this soil type. On the other hand, the database has a considerable number of Argissolos observations, nevertheless, in the study site, this soil order was assigned as Nitossolo or Cambissolo. In the study site, there is little occurrence of Argissolos and seem to be intergrade between the Nitossolos and Cambissolos, which might explain the confusion. According to Vasques et al. (2014), the formation of both Argissolos and Nitossolos in this region is related to processes of clay accumulation in B horizon and consequently the soils display a good degree of similarity. 3.4. Comparison between traditional and digital soil maps Conventional and digital soil maps presented a good agreement with a similar spatial distribution of MU and soil orders (Fig. 9). In both maps, soils at the summit were classified as Latossolo Vermelho (LV), but the MU in the CM covered a larger area and corresponded to two different MU of the digital map. At the non-corresponding area, digital map classified the soils as Nitossolo Vermelho (NV) (Table 4). According to Cooper and Vidal-Torrado (2005), these soil classes might have high similarity and are distinguished basically by the structure of the diagnostic horizon. According to Bellinaso et al. (2010), distinction between LV and NV based on the spectrum is possible, however, it requires the feature evaluation at 2265 nm. LV usually contains a higher gibbsite content due to severe weathering processes, resulting in differences in the soil spectra. Another misclassification occurred between Gleissolos (GX) and Cambissolos (CX). During the field survey, an MU corresponding to soils with lithic contact (at approximately 2 m deep) was designed and located in slightly depressed areas. The MU presented a water

logging condition and consequently it is a hydromorphic soil. Although this soil had some similarities with the descriptions of the Cambissolos order, it was classified as Gleissolo Háplico (GX) due to criteria established by the SiBCS. The SRSSL showed no GX and consequently the method was not able to identify the MU as such. The method assigned the most similar class, that is, CX, which is coherent with soil survey observations. In the DSM, there is an MU defined as R (Neossolo), which was identified as CX in the conventional map (Fig. 9) (Table 4). The SRSSL from the municipality of Piracicaba presents observations of Neossolo Litólico mainly, which is a shallow soil with low weathering degree and without the B horizon. The CX from the study site are also soils with low weathering degree, however, different from R due to presence of a thin B horizon. Considering that the DSM used information from only two layers (0–20 cm and 80–100 cm), the sampling process may have missed the B horizon in these areas and affected the classification results. This work ratifies first observations of Demattê et al. (2001), where the use of aerial photographs (which evaluates relief) revealed greater number of poligons than a traditional survey. Indeed, Demattê et al. (2004), performed the pioneering practical method on using spectroscopy for soil mapping. These authors generated a soil map in a detailed scale using spectral sensing in various depths, associated with relief. These results are in accordance with our data and with a upgrade on the methodology and geotechnologies associated. 4. Conclusions The use of regional spectral libraries combined to local soil spectra allows pedologists to perform rapid predictions of soil attributes. Moreover, information from multiple depths enables the identification of

Fig. 9. (a) Soil units map, (b) digital soil map and (c) conventional soil map.

26

R. Rizzo et al. / Geoderma 274 (2016) 18–27

Table 4 Cross tabulation between conventional and digital soil map indicating the coincident and non-coincident areas, in percentage. Conventional soil map Soil classif.a

Digital soil map

PA

GX

NV

CX

LV

NX

PVA

FF

PA 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 GX 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 NV 0.0 3.2 63.7 3.7 39.4 37.9 9.7 93.9 CX 100.0 96.8 6.1 64.9 0.0 14.8 62.2 0.0 LV 0.0 0.0 4.9 0.1 59.8 0.0 0.0 0.0 NX 0.0 0.0 25.3 4.0 0.8 45.3 28.1 6.1 PVA 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 FF 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 R 0.0 0.0 0.0 27.3 0.0 2.0 0.0 0.0 Total 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0

a Soil suborders and corresponding World Reference Base (WRB; IUSS Working Group WRB, 2014) and Soil Taxonomy (Soil Survey Staff, 2014) classes: CX, Cambissolo– Cambisol, Udepts; LV, Latossolo Vermelho–Ferralsol, Udox; NV, Nitossolo Vermelho; NX, Nitossolo Háplico–Nitisol, Udalfs/Udults; PA, Argissolo Amarelo; PV, PVA, Argissolo Vermelho Amarelo–Lixisol, Udalfs/Udults; R, Neossolo–Leptosol, Lithic Udorthents/Lithic Udipsamments; Udorthents/Udipsamments.

soils according to a classification system. In this study, kappa coefficients of 0.41 indicated that the soil order classification had good agreement with the observed classes. Besides, Munsell color predictions for hue had excellent performance with R2 0.95 when compared to a colorimeter. On the other hand, SRSSL from the Piracicaba region still needs to be improved to include as many soil orders as possible and become more representative. The conventional soil mapping process is a complex activity where pedologist relies on tacit knowledge to delineate the MU and select sampling locations. Spatial explicit information, such as multi-depth spectra, could be used to help in decision making of soil scientists, in special for detailed scales. Validation of PC1 score map, for example, proves that such information is reliable. Considering that spectrometers are portable devices, further works should focus on applying this strategy on spectra measured in the field. Acknowledgments We thank the Department of Soil Science at the University of São Paulo, the Coordination for the Improvement of Higher Education Personnel (CAPES) at the Brazilian Ministry of Education and the State of São Paulo Research Foundation for financial support (proc. numbers 2009-54144-8 and 2014-22262-0). We also thank the Geotechnologies in Soil Science Group (GeoSS; http://esalqgeocis.wix.com/english) at the Soil Science Department at the University of São Paulo. References Adhikari, K., Minasny, B., Greve, M.B., Greve, M.H., 2014. Constructing a soil class map of Denmark based on the FAO legend using digital techniques. Geoderma 214–215, 101–113. Bellinaso, H., Dematte, J.A.M., Romeiro, S.A., 2010. Soil spectral library and its use in soil classification. Brazil J. Soil Sci. 34, 861–870. Ben-Dor, E., Irons, J.R., Epema, G.F., 1999. Soil reflectance. In: Rencez, A.N. (Ed.), Remote Sensing for the Earth Science: Manual of Remote Sensing. Jonh Wiley, New York, pp. 111–188. Ben-Dor, E., Heller, D., Chudnovsky, A., 2008. A novel method of classifying soil profiles in the field using optical means. Soil Sci. Soc. Am. J. 72, 1–13. Bezdek, J.C., 1981. Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum Press, New York. Burrough, P.A., McDonnell, R.A., 1998. Principles of Geographical Information Systems. Oxford University Press, New York. Burrough, P.A., Van Gaans, P.F.M., Hootsmans, R., 1997. Continuous classification in soil survey: spatial correlation, confusion and boundaries. Geoderma 77, 115–135. Campos, R.C., Demattê, J.A.M., Quartaroli, C.F., 2003. Determinação indireta do teor de hematita na fração argila de solos a partir de dados de colorimetria e radiometria. Pesq. Agrop. Brasileira 38, 521–528. Clark, R.N., Gallagher, A.J., Swayze, G.A., 1990. Material absorption band depth mapping of imaging spectrometer data using the complete band shape least-squares algorithm

simultaneously fit to multiple spectral features from multiple materials. Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) Workshop. JPL Publication, pp. 176–186. Commission Internationale de L'Eclairage, 1931. Proceedings of the eight session. Bureau Central de la CIE, Cambridge. Congalton, R.G., Oderwald, R.G., Mead, R.A., 1983. Assessing Landsat classification accuracy using discrete multivariate analysis statistical techniques. Photogramm. Eng. Rem. S. 49, 1671–1678. Cooper, M., Vidal-Torrado, P., 2005. Caracterização morfológica, micromorfológica e físico-hídrica de solos com horizonte B nítico [Morphological, micromorphological and hydro-physical characterization of soils with a nitic B horizon]. Braz. J. Soil Sci. 29, 581–595. Demattê, J.A.M., Demattê, J.L.I., Camargo, W.P., Fiorio, P.R., Nanni, M.R., 2001. Remote sensing in the recognition and mapping of tropical soils developed on topographic sequences. Mapp. Sci. Remote. Sens. 38, 79–102. Demattê, J.A.M., Campos, R.C., Alves, M.C., Fiorio, P.R., Nanni, M.R., 2004. Visible-NIR reflectance: a new approach on soil evaluation. Geoderma 121, 95–112. Demattê, J.A.M., Terra, F.S., 2014. Spectral pedology: a new perspective on evaluation of soils along pedogenetic alterations. Geoderma 217–218, 190–200. Henderson, B.L., Buib, E.N., Moranb, C.J., Simonb, D.A.P., 2005. Australia-wide predictions of soil properties using decision trees. Geoderma 124, 383–398. Hoffman, R., Krotkov, E., 1990. Terrain roughness measurement from elevation maps. Advances in Intelligent Robotics Systems Conference. International Society for Optics and Photonics, pp. 104–114. Horn, B.K.P., 1981. Hill shading and the reflectance map. Proceedings eIEEE Cambridge 69, pp. 14–47. Hutchinson, M.F., Lock, M.D., 1993. On thin plate splines and kriging. In: Tarter, M.E. (Ed.), Computing Science and Statistics 25. Interface Foundation of North America, Berkeley, pp. 55–62. Kuhn, M., Weston, S., Keefer, C., Coulter, N., 2014. Cubist: Rule- and Instance-Based Regression Modeling. (Rpackage version 0.0.18e. http://CRAN.R-project.org/ package=Cubist). Lagacherie, P., Bailly, J.S., Monestiez, P., Gomez, C., 2012. Using scattered hyperspectral imagery data to map the soil properties of a region. Eur. J. Soil Sci. 63, 110–119. Martens, H., Naes, T., 1989. Multivariate Calibration. John Wiley, New York. Mendonça-santos, M.L., Dos Santos, H.G., 2006. The state of the art of Brazilian soil mapping and prospects for digital soil mapping. In: Lagacherie, P., McBratney, A.B., Voltz, M. (Eds.), Developments in Soil Science. Elsevier, Amsterdam, pp. 39–54. Moore, I.D., Gessler, P.E., Nielsen, G.A., Peterson, G., 1993. Soil attribute prediction using terrain analysis. Soil Sci. Soc. Am. J. 57, 443–452. Post, D.F., Bryant, R.B., Batchily, A.K., Huete, A.R., Levine, S.J., Mays, M.D., Escadafal, R., 1993. Correlations between field and laboratory measurements of soil color. Soil Color 35–49. Quinlan, J.R., 1992. Learning with continuous classes. In: Adams, S. (Ed.), Proceedings of the 5th Australian Joint Conference on Artificial Intelligence. World Scientific, Singapore, pp. 343–348. EMBRAPA (Empresa Brasileira de Pesquisa Agropecuária), 2013. Sistema Brasileiro de Classificação de Solos. third ed. EMBRAPA, Brasília. Soil Survey Staff., 2014. Keys to Soil Taxonomy. 12th ed. USDA-Natural Resources Conservation Service, Washington, p. 362. Soriano-Disla, J.M., Janik, L.J., ViscarraRossel, R.A., MacDonald, L.M., McLaughlin, M.J., 2014. The performance of visible, near-, and mid-infrared reflectance spectroscopy for prediction of soil physical, chemical, and biological properties. Appl. Spectrosc. 49, 139–186. Stenberg, B., ViscarraRossel, R.A., Mouazen, A.M., Wetterlind, J., 2010. Visible and near infrared spectroscopy in soil science. Adv. Agron. 107, 163–215. Torrent, J., Barrón, V., 1993. Laboratory measurement of soil color: theory and practice. Soil color. Soil Sci. Soc. Agron. 31, 21–33. Vasques, G.M., Demattê, J.A.M., Viscarra Rossel, R.A., Ramírez-López, L., Terra, F.S., 2014. Soil classification using visible/near-infrared diffuse reflectance spectra from multiple depths. Geoderma 223–225, 73–78. Vasques, G.M.; Demattê, J.A.M., Viscarra-rossel, R.A; Ramirez-lopez, L. Terra, F.S.; Rizzo, R.; Souza Filho, B., 2015. Integrating geospatial and multi-depth laboratory spectral data for mapping soil classes in a geologically complex area in southeastern Brazil. Eur. J. Soil Sci. 66, 767–779. Vidal-Torrado, P., Lepsch, I.F., Castro, S.S., Cooper, M., 1999. Pedogênese em uma seqüência latossolo-podzólico na borda de um platô na depressão periférica paulista (Pedogenesis of an oxisol-ultisol-alfisol sequence on the border of a plateau at the paulista peripherical depression in Brazil). Braz. J. Soil Sci. 23, 909–921. Viscarra Rossel, R.A., 2008. ParLeS: software for chemometric analysis of spectroscopic data. Chemom. Intell. Lab. Syst. 90, 72–83. Viscarra Rossel, R.A., Chen, C., 2011. Digitally mapping the information content of visible– near infrared spectra of surficial Australian soils. Remote Sens. Environ. 115, 1443–1455. Viscarra Rossel, R.A., Webster, R., 2011. Discrimination of Australian soil horizons and classes from their visible–near infrared spectra. Eur. J. Soil Sci. 62, 637–647. Viscarra Rossel, R.A., Cattle, S.R., Ortega, A., Fouad, Y., 2009. In situ measurements of soil colour, mineral composition and clay content by vis–NIR spectroscopy. Geoderma 3–4, 253–266. ViscarraRossel, R.A., Minasny, B., Roudier, P., Mcbratney, A.B., 2006. Colour space models for soil science. Geoderma 133, 320–337. Viscarra Rossel, R.A., Behrens, T., Ben-Dor, E., Brown, D.J., Demattê, J.A.M., Shepherd, K.D., Shi, Z., Stenberg, B., Stevens, A., Adamchuk, V., Aïchi, H., Barthès, B.G., Bartholomeus, H.M., Bayer, A.D., Bernoux, M., Böttcher, K., Brodský, L., Du, C.W., Chappell, A., Fouad, Y., Genot, V., Gomez, C., Grunwald, S., Gubler, A., Guerrero, C., Hedley, C.B., Knadel, M., Morrás, H.J.M., Nocita, M., Ramirez-Lopez, L., Roudier, P., Rufasto Campos, E.M., Sanborn, P., Sellitto, V.M., Sudduth, K.A., Rawlins, B.G., Walter, C., Winowiecki, L.A.,

R. Rizzo et al. / Geoderma 274 (2016) 18–27 Hong, S.Y., Ji, W., 2016. A global spectral library to characterize the world's soil. EarthSci Rev. 155, 198–230. IUSS Working Group WRB, 2014. World Reference Base for Soil Resources 2014: International Soil Classification System for Naming Soils and Creating Legends for Soil Maps. World Soil Resources Report No 106. Rome, Food and Agriculture Organization.

27

Xie, X.L., Beni, G., 1991. A validity measure for fuzzy clustering. IEEE Trans. Pattern Anal. Mach. Intell. 13, 841–847. Zevenbergen, L.W., Thorne, C.R., 1987. Quantitative analysis of land surface topography. Earth Surf. Processes 12, 47–56.