Improving 3D Digital Soil Mapping Based on Spatialized Lab Soil Spectral Information

Abstract

Readily available environmental covariates in current digital soil mapping usually do not indicate the spatial differences between deep soil attributes. This, to a large extent, leads to a decrease in the accuracy of 3D soil mapping with depth, which seriously affects the quality of soil information generated. This study tested the hypothesis that spatialized laboratory soil spectral information can be used as environmental covariates to improve the accuracy of 3D soil attribute mapping and proposed a new type of environmental covariable. In the first step, with soil-forming environmental covariates and independent soil profiles, laboratory vis-NIR spectral data of soil samples resampled into six bands in Anhui province, China, were spatially interpolated to generate spatial distributions of soil spectral measurements at multiple depths. In the second step, we constructed three sets of covariates using the laboratory soil spectral distribution maps at multiple depths: conventional soil-forming variables (C), conventional soil-forming variables plus satellite remote sensing wavebands (C+SRS) and conventional soil-forming variables plus spatialized laboratory soil spectral information (C+LSS). In the third step, we used the three sets of environmental covariates to develop random forest models for predicting soil attributes (pH; CEC, cation exchange capacity; Silt; SOC, soil organic carbon; TP, total phosphorus) at multiple depths. We compared the 3D soil mapping accuracies between these three sets of covariates based on another dataset of 132 soil profiles (collected in the 1980s). The results show that the use of spatialized laboratory soil spectral information as additional environmental covariates has a 50% improvement in prediction accuracy compared with that of only conventional covariates, and a 30% improvement in prediction accuracy compared with that of the satellite remote sensing wavebands as additional covariates. This indicates that spatialized laboratory soil spectral information can improve the accuracy of 3D digital soil mapping.

1. Introduction

Environmental covariates are spatially continuous and have a synergistic relationship with soil spatial variability. They can be the variables characterizing factors such as the soil itself, climate, biology, topography, parent material, time and spatial location in the SCORPAN model [1].

Environmental covariates are key to digital soil mapping. Their ability to indicate soil spatial variability largely determines the upper limit of prediction accuracy and modeling complexity that can be achieved by soil mapping. The upper limit of soil prediction accuracy is higher when there are covariates with strong indication, which is high spatial synergy between soil attributes and environmental covariates. It is not necessary to have a complex model, and a simple model can achieve high accuracy. On the contrary, the upper limit is lower when there are covariates with weak indication, which is low spatial synergy between soil attributes and environmental covariates, and even complex models still achieve low accuracy.

Thus, in order to obtain the spatial distribution of prepared soil information, great emphasis has been put on the use of environmental covariates in digital soil mapping [1]. For example, Liu et al. [2,3] proposed a surface dynamic feedback approach to develop effective covariates to improve soil mapping in gentle topographic areas where easily accessible topographic and vegetation variables had weak synergistic relationships with soils. A review of 30 years of digital soil mapping showed that more than 70% of the works had used topography as covariates and more than 30% had used vegetation as covariates [4].

Improper selection of environmental covariates often does not adequately characterize soil-forming environmental conditions. Climate variables such as temperature and precipitation can be generated by spatial interpolation of weather station data. Topographic variables can be generated by digital terrain analysis of the DEM, and vegetation and land use variables are obtained by remote sensing observations. Many studies have shown that these existing covariates have a strong relationship with soil attributes at upper depths and a weak relationship with soil attributes at lower depths [5,6,7,8]. The soil parent materials have a strong relationship with the underlying soil. Chen et al. [4] argued that this was valid for soil organic carbon/soil organic matter, bulk density and particle size fractions (Clay, Silt, Sand), but not for pH and AWC. Meanwhile, it is difficult to obtain information on the spatial distribution of the parent materials through direct observation over a large area. Thus, the accuracy of 3D digital soil mapping of soil attributes usually decays with depth. This seriously affects the quality of soil information and adds great uncertainty to the simulations of hydrological, climate and soil pollution processes. Therefore, how to improve the accuracy of 3D digital soil mapping is an urgent research problem to be solved.

Developing new environmental covariates that can capture spatial differences in the subsoil is a direct solution. Heung et al. [9] developed a random forest model between parent material types and topographic variables for predictive mapping of the parent material types in a study area of British Columbia, Canada. Visible near-infrared (vis-NIR) remote sensing can only detect soil information at the depth of a few centimeters on the extreme surface in bare soil conditions. Geophysical remote sensing such as Gamma-ray spectroscopy can partially reflect the information of the parent material. Loiseau et al. [10] showed that airborne Gamma spectral remote sensing data were useful covariates for predictive mapping of surface soil texture in a plot in western France. However, the maximum detection depth of an airborne approach was 30 cm [11,12,13]. For sediments, Gamma spectra are mainly associated with the content of clay particles. For bedrock, Gamma spectral signals come from the minerals and are mainly applied to the detection of clay particle contents. Electromagnetic induction instruments can measure the electrical conductivity of soil particles. They are more sensitive to soil physicochemical attributes such as soil salinity, clay particle content, water content and nutrients. Møller et al. [14] used portable ground-based electromagnetic induction sensor measurements as covariates and performed high-resolution mapping of clay and organic matter content. However, for large-area applications, the consistency of measurement data is difficult to guarantee.

Laboratory soil spectra contain a wealth of information that can reflect a variety of soil attributes, including soil organic matter, total nitrogen, soil texture, conductivity [15], moisture content, iron oxide content, carbonate minerals and pollutant content [16,17]. Many regional and national soil spectral libraries have been established internationally [13,18] and global soil spectrum libraries [19] are under construction. They contain various spectral bands such as visible, near-infrared and mid-infrared. The spectral information was measured under controlled laboratory conditions for soil samples at different depths after air-drying and grinding treatment. They can reflect the physical and chemical information of underlying soil materials. However, laboratory soil spectra are mainly limited and discrete point data and cannot be used directly as environmental covariates to assist in the predictive mapping of soil attributes.

The purpose of this study is to propose a method for improving 3D digital soil mapping based on laboratory soil spectral information by testing the hypothesis that using spatialized laboratory soil spectral information can improve the accuracy of 3D soil mapping, that is, spatialized laboratory soil spectral information and its derived factors are developed and used as environmental covariates to assist in improving the prediction accuracy of 3D mapping.

2. Materials and Methods

2.1. Study Area

The study area is Anhui Province, China. Anhui Province lies between 114°54′–119°37′E and 29°41′–34°38′N, with a total area of 140,100 km². According to the Soil Series of Anhui Province [20], it is in a transition between a warm temperate zone and a subtropical zone. It has a humid climate and obvious seasonal changes because of the influence of the southeast subtropical monsoon. The average annual temperature is 14–17 °C and the average annual precipitation is 773–1670 mm. Precipitation gradually decreases from southeast to northwest. The study area has various types of landforms, including plains, terraces, hills and mountains. The northern part is the Huaibei plain area and the main land use is dryland. The central part is the hilly area of the Jianghuai Terrace and the plain area along the river has low elevation. Its main land uses are drylands and paddy fields. The southern part is a hilly mountain area with high elevation and undulating terrain. The main land uses are forests, grasslands and paddy fields.

The soil types in the study area include Calcaric Cambisols, Cutanic Luvisols, Eutric Cambisols, Eutric Regosols, Ferralic Cambisols, Ferric Acrisols, Ferric Lucisols, Haplic Fluvisols, Hydragric Anhtrosols and Lithic Leptosols. They are determined based on the World Reference Base for Soil Resources [21]. Among them, Hydragric Anhtrosols, Hydragric Anhtrosols and Calcaric Cambisols are the most widely distributed.

2.2. Soil Attributes Data

We used 132 typical soil profiles in this study (Figure 1). They originated from the Second National Soil Survey in the 1980s and represent the typical soil landscape in the study area. Soil attributes were measured for each horizon of the soil profiles. They are soil pH, cation exchange capacity, organic carbon content, soil texture and total phosphorus content.

We used equal-area quadratic spline functions [22,23] to simulate the depth distribution of the five soil attributes and computed the values of soil attributes at five fixed depths: 0–5, 5–15, 15–30, 30–60 and 60–100 cm.

Table 1. Statistical descriptions of the spline-fixed five soil attributes at different depths.

Soil Attributes	Depth (cm)	Mean (%)	SD (%)	CV	Skewness	Kurtosis
pH	0–5	6.385	1.282	0.201	0.437	−0.908
	5–15	6.484	1.221	0.188	0.404	−0.918
	15–30	6.757	1.150	0.170	0.045	−0.993
	30–60	6.923	1.136	0.164	−0.247	−0.924
	60–100	7.034	1.115	0.158	−0.413	−0.689
CEC	0–5	15.506	7.147	0.461	0.680	−0.304
	5–15	15.140	6.909	0.456	0.661	−0.381
	15–30	14.654	7.087	0.484	0.560	−0.524
	30–60	15.297	7.506	0.493	0.603	−0.706
	60–100	15.025	7.203	0.479	0.555	−0.431
Silt	0–5	34.122	12.252	0.359	0.284	−0.203
	5–15	34.115	11.326	0.332	0.164	−0.299
	15–30	33.860	11.032	0.326	0.034	−0.093
	30–60	33.494	12.209	0.365	0.403	1.441
	60–100	33.048	12.220	0.370	0.007	1.572
SOC	0–5	15.894	13.718	0.863	3.174	12.864
	5–15	14.407	11.473	0.796	2.821	11.265
	15–30	10.658	8.996	0.844	2.616	9.790
	30–60	7.528	10.499	1.395	5.666	43.163
	60–100	5.499	7.005	1.274	5.276	37.203
TP	0–5	0.557	0.319	0.573	2.020	5.743
	5–15	0.549	0.291	0.530	1.677	3.696
	15–30	0.529	0.303	0.573	2.588	13.015
	30–60	0.474	0.258	0.544	1.959	7.086
	60–100	0.449	0.264	0.589	2.240	9.941

lists the statistical description of the soil attributes at these depths. As the depth increases, soil pH and Silt show an increasing trend, while cation exchange capacity, soil organic carbon and total phosphorus all have different degrees of decrease. The standard deviations (SDs) of all soil attributes except soil pH are high and the data dispersion is significant. The dispersion of soil attributes increases with increasing depth; this leads to a decrease in predictability. Among them, soil organic carbon and total phosphorus have coefficients of variation (CV) greater than 0.5. The spatial variations between different soil profiles are obvious.

2.3. Laboratory Soil Spectral Data

An additional set of soil profile data were used to determine laboratory soil spectral information in this study; these data were derived from the project National Soil Series Survey and Soil Series of China [3]. The survey period was from 2010 to 2011, with a total of 114 soil profiles. Soil horizon samples were subjected to a continuous drying process, fully ground and passed through a 2 mm sieve. The vis-NIR reflectance data of each soil sample were measured using a Cary 5000 spectrophotometer (Agilent Technologies, Santa Clara, CA, USA) under controlled temperature and humidity conditions in the laboratory [24]. We recorded ten soil spectral repeats in each soil sample. The measured spectral band range was 350–2500 nm, with a resolution of 1 nm.

Denoising and dimensionality reduction were performed for laboratory soil spectral data. In the first step, large fluctuating noises usually occur at the beginning and end of the measurement phase due to the influence of the external environment. The noisy 350–399 nm and 2401–2500 nm bands were artificially removed and the soil spectral data at 400–2400 nm were retained. In the second step, a small number of information-rich spectral bands, which are sensitive to most soil attributes, were selected [24]. These are red band (R, 620–680 nm), green band (G, 520–600 nm), blue band (B, 430–460 nm), NIR short-wave band (SW-NIR, 840–890 nm), NIR mid-wave band (MW-NIR, 1550–1750 nm) and NIR long-wave band (LW-NIR, 2100–2300 nm). We averaged the spectral reflectance data within each band to obtain the mean reflectance data of the six bands. In the fourth step, we used equal-area quadratic spline functions to normalize the mean values of the six laboratory soil spectral characteristic bands and generated reflectance values at the five fixed depths of 0–5 cm, 5–15 cm, 15–30 cm, 30–60 cm and 60–100 cm [25].

2.4. Environmental Covariates

Table 2. Original environmental covariates and their related information collected in this study.

Category	Covariate	Description	Resolution
Parent material	pmcls	Type of parent materials	90 m
Terrain	elev	Elevation (m)	90 m
	anHis	Mountain shadow	90 m
	slp	Slope (%)	90 m
	asp	Aspect (°)	90 m
	curPln	Plan curvature	90 m
	curPrf	Profile curvature	90 m
	TWI	Topographic wetness index	90 m
	TPI	Topographic position index	90 m
	MrVBF	Multiscale Valley flatness index	90 m
	MrRTF	Multiscale ridge flatness index	90 m
Climate	MAT	Mean annual temperature (°C)	1000 m
	rangDiurnal	Mean diurnal range (°C)	1000 m
	rangAnnual	Mean annual range (°C)	1000 m
	isother	Isothermality (°C)	1000 m
	tempSeason	Air temperature seasonality (°C)	1000 m
	tempMaxWarm	Maximum air temperature of warmest month (°C)	1000 m
	tempMinCold	Minimum air temperature of coldest month (°C)	1000 m
	tempWettest	Air temperature of wettest season (°C)	1000 m
	tempDriest	Air temperature of driest season (°C)	1000 m
	tempMeanWarm	Mean air temperature of warmest month (°C)	1000 m
	tempMeanCold	Mean air temperature of coldest month (°C)	1000 m
	MAP	Mean annual precipitation (mm yr−1)	1000 m
	precSeason	Precipitation seasonality (mm yr−1)	1000 m
	precWettest	Precipitation of wettest month (mm mh−1)	1000 m
	precDriest	Precipitation of driest month (mm mh−1)	1000 m
	precWarm	Precipitation of warmest month (mm mh−1)	1000 m
	precCold	Precipitation of coldest month (mm mh−1)	1000 m
	solarRed	Mean annual solar radiation (J m−2 yr−1)	500 m
	windSpeed	Wind speed (m s−1)	1000 m
	vaporPress	Water vapor pressor (kpa)	1000 m
	evaTra	Terrestrial evaporation	500 m
	LST	Land surface temperature (°C)	500 m
Biological	NDVI	Mean NDVI	250 m
	EVI	Mean EVI	250 m
Satellite remote sensing	MODISb1	Reflectance of the red band (620–672 nm)	500 m
	MODISb2	Reflectance of near-infrared short wave (841–890 nm)	500 m
	MODISb3	Reflectance of the blue band (459–479 nm)	500 m
	MODISb4	Reflectance of the green band (545–565 nm)	500 m
	MODISb5	Reflectance of near-infrared medium wave (1230–1250 nm)	500 m
	MODISb6	Reflectance of near-infrared medium wave (1628–1652 nm)	500 m
	MODISb7	Reflectance of near-infrared long wave (2105–2155 nm)	500 m

lists the environmental covariates used in this study. They were classified into parent material variables, topographic variables, climatic variables, biological variables and satellite remote sensing waveband variables. In this study, we collected the environmental covariates that changed over time (such as climatic variables and biological variables) in two different periods. One period (2010s) of the environmental covariates was used to make spatial prediction of the laboratory soil spectral information. The other period (1980s) of environmental covariates was used to verify the improvement in spatialized laboratory soil spectral information on the accuracy of 3D mapping of soil attributes. For environmental covariates that did not change over time, we collected data in only one period.

Parent material variables were 1:21 million Soil Parent Materials map of China, obtained from Geographic Data Sharing Infrastructure, College of Urban and Environmental Science, Peking University (http://geodata.pku.edu.cn, accessed on 7 October 2021). Topographic variables were obtained based on 90 m spatial resolution SRTM DEM data (http://srtm.csi.cgiar.org/srtmdata, accessed on 7 October 2021), calculated using SAGA GIS software (http://www.saga-gis.org, accessed on 7 October 2021). Climate variables from the 1980s were 1 km spatial resolution climate data from the WorldClim database. They were all annual averages from 1970 to 2000 [26]. The climate variables from the 2010s are derived from the 1 km resolution climate and precipitation dataset of China drawn by Peng et al. [27]. Biological variables from the 2010s were NDVI and EVI from the MODIS data product (MOD13Q1) and biological variables from the 1980s were obtained by time downscaling using climate variables from the 1980s. Satellite remote sensing waveband variables from the 2010s were MODIS reflectance data from 7 bands (MOD09A1). Both biological variables and satellite remote sensing variables were obtained through the GEE platform and their mean values were calculated from the average data for the years 2000–2010. All environmental covariates were unified into the Albers projection coordinates system and resampled to 1 km resolution.

2.5. Methods

The method for testing the hypothesis consisted of three steps (Figure 2). In the first step, we spatialized the laboratory soil spectra to generate multiple-depth soil spectral-spatial distribution layers. In the second step, we constructed three sets of environmental covariates with the variables in Table 2 and the spatialized laboratory soil spectral variables. In the third step, we compared the three sets of covariates by assessing their 3D soil attribute mapping accuracies.

2.5.1. Spatialization of Laboratory Soil Spectra

We used a random forest model [28] to construct a quantitative relationship between laboratory soil spectral information and soil-forming environmental covariates (elev, anHis, slp, curPln, curPrf, TWI, TPI, MAT, tempSeason, tempWettest, windspeed, LST, NDVI, etc.). The random forest model is an integrated decision tree model that can reduce the overfitting of decision trees to a certain extent. It has the advantages of easy training and high accuracy and is currently a mainstream machine learning method [29,30,31]. This model randomly generates several decision trees for classification and regression problems. Each tree randomly draws data from training samples to include in the training process. The final prediction result is the average of the prediction results of all decision trees. Another advantage of the model is that the importance of input covariates can be judged by the out-of-bag mean square error values (%lncMSE) and the nodal bluntness values (lncNodePurity). Higher %lncMSE and lncNodePurity represent larger contributions of the covariate to the prediction results. We determined the optimal parameters of the model based on the out-of-bag prediction results and predicted spatial distributions of the laboratory soil spectral bands (R, G, B, SW-NIR, MW-NIR, LW-NIR) at 1 km spatial resolution at the 5 standard depths.

We selected four error metrics to calculate the accuracy of the spatialization of laboratory soil spectral information using the leave-one-out-cross validation (LOOCV) method [32]. These are mean error (ME), root mean square error (RMSE), coefficient of determination (R²) and Concordance Correlation Coefficient (CCC) [33], which were calculated as follows:

$$ME=\frac{1}{n}\sum _{i=1}^n\left(O_i-P_i\right)$$

(1)

$$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(O_i-P_i\right)}^2}$$

(2)

$$R^2=1-\frac{{\displaystyle {\sum }_{i=1}^n}{\left(O_i-P_i\right)}^2}{{\displaystyle {\sum }_{i=1}^n}{\left(O_i-\overline{O}\right)}^2}$$

(3)

$$CCC=\frac{2r{\sigma }_O{\sigma }_P}{{\sigma }_O^2+{\sigma }_P^2+{\left(\overline{O}-\overline{P}\right)}^2}$$

(4)

where $O_i$ and $P_i$ are the observed and predicted values of sample i; n is the total number of samples; $\overline{O}$ and $\overline{P}$ are the mean values of the sample observed values and predicted values; r is the correlation coefficient between the predicted and observed values of the samples; ${\sigma }_O$ and ${\sigma }_P$ are the standard deviations of the sample observation and predictions.

The construction of the random forest model is based on the open source R environment [34] and involves the packages “randomForest” [35], “caret” [36], “rgdal” [37], “raster” [38] and “ggplot2” [39].

2.5.2. Constructing Sets of Environmental Covariates

We constructed three sets of environmental covariates and compared their predictive performance for soil attributes. The first set is the conventional soil-forming variables (C), which includes terrain, climate and biological variables. The second set is the combination of conventional soil-forming variables and satellite remote sensing wavebands variables (C+SRS) with the purpose of checking the accuracy improvement in 3D soil attribute prediction when adding satellite remote sensing information in the covariates. The wavebands include short-wave NIR, blue and long-wave NIR. The third set is the combination of conventional soil-forming variables and laboratory soil spectral information variables (C+LSS) with the purpose of checking the accuracy improvement in 3D soil attribute prediction when adding laboratory soil spectral information to the covariates. We combined the 6 laboratory soil spectral bands to generate 12 soil spectral correlation factors, aiming to generate some laboratory soil spectral factor combinations by referring to the common combination of satellite remote sensing bands so as to expand the application range of soil spectral information.

Table 3. The expressions of soil spectral correlation factors.

Soil Spectral Correlation Factors	Computed Expression
T1	$R-G$
T2	$R+G+SWNIR$
T3	$\frac{SWNIR-R}{SWNIR+R}$
T4	$\frac{G-R}{G+R}$
T5	$\frac{SWNIR-G}{SWNIR+G}$
T6	$R\times G$
T7	$\sqrt{R^2+G^2}$
T8	$\frac{SWNIR}{LWNIR}$
T9	$\frac{SWNIR}{MWNIR}$
T10	$\frac{MWNIR}{G}$
T11	$\frac{MWNIR}{R}$
T12	$\frac{G}{R}$

lists the calculated expressions for the 12 soil spectral correlation factors.

2.5.3. Assessing Accuracy Improvements in 3D Soil Attribute Predictions

For each set of covariates and each of the soil attributes (pH, CEC, Silt, SOC, TP), we used a random forest model to construct their quantitative relationships at each depth. We trained all models with default parameters with the purpose of excluding the effect of different parameters on the model performance. We then used the LOOCV to assess the prediction accuracy of all models. The error metrics R² and RMSE were calculated using Equations (2) and (3). Then, we calculated the relative accuracy improvements in R² and RMSE of covariate sets C+SRS and C+LSS over covariate set C using the following equations [40].

$${RI}_{R^2}=\frac{R_{M_x}^2-R_{M_c}^2}{R_{M_c}^2}$$

(5)

$${RI}_{RMSE}=\frac{{RMSE}_{M_c}-{RMSE}_{M_x}}{{RMSE}_{M_c}}$$

(6)

where ${RI}_{R^2}$ and ${RI}_{RMSE}$ are the relative improvements, respectively, in R² and RMSE; $R_{M_x}^2$ and ${RMSE}_{M_x}$ are, respectively, the R² and RMSE for covariate set C+SRS or C+LSS; $R_{M_c}^2$ and ${RMSE}_{M_c}$ are, respectively, the R² and RMSE for covariate set C. Higher ${RI}_{R^2}$ or ${RI}_{RMSE}$ values mean a bigger accuracy improvement.

3. Results

3.1. Covariates of Laboratory Soil Spectra

Table 4. Accuracies of spatial predictions of six laboratory soil spectral bands at different depths.

Laboratory Soil Spectral Bands	Depth (cm)	ME	RMSE	R2	CCC
R	0–5	−0.129	2.982	0.209	0.436
	5–15	−0.098	2.890	0.239	0.499
	15–30	0.341	2.962	0.279	0.477
	30–60	−0.022	2.807	0.385	0.597
	60–100	−0.480	2.819	0.518	0.687
G	0–5	−0.261	2.748	0.142	0.398
	5–15	−0.338	2.702	0.176	0.401
	15–30	0.251	2.422	0.224	0.449
	30–60	0.334	2.564	0.286	0.587
	60–100	−0.203	2.434	0.426	0.584
B	0–5	−0.017	1.864	0.331	0.520
	5–15	−0.026	1.747	0.380	0.550
	15–30	0.064	1.794	0.349	0.498
	30–60	−0.038	1.680	0.332	0.510
	60–100	0.063	1.608	0.400	0.598
SW-NIR	0–5	−0.182	2.700	0.404	0.586
	5–15	−0.184	2.538	0.431	0.625
	15–30	−0.106	2.153	0.515	0.515
	30–60	0.100	2.423	0.515	0.682
	60–100	0.055	3.358	0.433	0.611
MW-NIR	0–5	−0.182	3.482	0.279	0.405
	5–15	−0.846	2.946	0.429	0.633
	15–30	−0.093	4.139	0.499	0.669
	30–60	−0.311	4.829	0.406	0.587
	60–100	−0.208	5.104	0.371	0.559
LW-NIR	0–5	−0.359	1.113	0.459	0.614
	5–15	−0.424	4.162	0.394	0.543
	15–30	−0.205	4.174	0.410	0.573
	30–60	−0.347	4.481	0.395	0.564
	60–100	−0.354	4.769	0.308	0.493

ME: mean error; RMSE: root mean square error; R²: coefficient of determination; CCC: concordance correlation coefficient.

lists the accuracies of spatial predictions of laboratory soil spectral bands R, G, B, SW-NIR, MW-NIR and LW-NIR. It shows that the ME values are close to 0 and the CCC values are mainly distributed between 0.4 and 0.7. This indicates a good agreement between the predicted and observed values. The R² values are mainly concentrated between 0.3 and 0.5, with an average of 0.36. This means the random forest algorithm is able to explain nearly 36% of the laboratory soil spectral information.

Figure 3 shows the spatialized laboratory soil spectral bands at five depth intervals. The laboratory soil spectral reflection values were relatively low in the northern plains and high in the southern hills. In the vertical dimension, the laboratory soil spectral reflection values showed an increasing trend with depth. In addition, the laboratory soil spectral reflectance showed an increasing trend with increasing wavelength. The lowest reflectance value was in the blue band. The highest reflectance value was in the mid-wave near-infrared band.

3.2. Correlations between Soil Attributes and Spatialized Laboratory Soil Spectral Information with Covariates

Figure 4 shows Pearson correlation matrices between soil attributes and environmental covariates at different depths. Overall, soil pH, CEC and SOC were strongly correlated with all covariates, while Silt and TP were weakly correlated with the covariates. The correlation between each soil attribute and covariate decreased with increasing depth.

Table 5. Significance tests of correlation coefficients between the soil attributes at different depths and satellite remote sensing waveband variables and laboratory soil spectral variables.

Soil Attributes	Depth (cm)	Satellite Remote Sensing	Lab Soil Spectral Information	p Value
pH	0–5	0.186 ± 0.114 b	0.364 ± 0.113 a	0.021
	5–15	0.117 ± 0.035 b	0.305 ± 0.139 a	0.034
	15–30	0.117 ± 0.031 a	0.285 ± 0.141 a	0.058
	30–60	0.158 ± 0.041 b	0.313 ± 0.111 a	0.030
	60–100	0.126 ± 0.007 b	0.199 ± 0.058 a	0.045
CEC	0–5	0.036 ± 0.021 b	0.165 ± 0.096 a	0.035
	5–15	0.053 ± 0.032 b	0.182 ± 0.097 a	0.038
	15–30	0.052 ± 0.017 b	0.145 ± 0.074 a	0.044
	30–60	0.079 ± 0.053 b	0.167 ± 0.057 a	0.022
	60–100	0.067 ± 0.071 b	0.124 ± 0.039 a	0.049
Silt	0–5	0.026 ± 0.028 a	0.049 ± 0.039 a	0.343
	5–15	0.025 ± 0.028 a	0.045 ± 0.034 a	0.352
	15–30	0.074 ± 0.047 b	0.177 ± 0.065 a	0.017
	30–60	0.050 ± 0.023 b	0.152 ± 0.061 a	0.012
	60–100	0.035 ± 0.041 b	0.157 ± 0.065 a	0.006
SOC	0–5	0.094 ± 0.019 b	0.249 ± 0.124 a	0.049
	5–15	0.122 ± 0.035 b	0.225 ± 0.079 a	0.042
	15–30	0.108 ± 0.025 b	0.203 ± 0.074 a	0.045
	30–60	0.112 ± 0.030 b	0.177 ± 0.045 a	0.017
	60–100	0.092 ± 0.055 b	0.163 ± 0.061 a	0.043
TP	0–5	0.022 ± 0.033 b	0.089 ± 0.035 a	0.006
	5–15	0.015 ± 0.020 b	0.149 ± 0.063 a	0.002
	15–30	0.019 ± 0.005 b	0.143 ± 0.090 a	0.030
	30–60	0.035 ± 0.025 b	0.107 ± 0.050 a	0.028
	60–100	0.127 ± 0.020 b	0.187 ± 0.044 a	0.029

Different lowercase letters indicate significant differences at p < 0.05 between the correlations of soil attributes with satellite remote sensing waveband variables and the correlations of laboratory soil spectral information variables.

lists the significance tests of correlation coefficients between the soil attributes at different depths and satellite remote sensing waveband variables or laboratory soil spectral variables. It can be seen that the correlation coefficients of the soil attributes with laboratory soil spectral variables were significantly higher than those with satellite remote sensing waveband variables at different depths (p < 0.05). This indicates that laboratory soil spectral information had a higher correlation with soil attributes than satellite remote-sensing wavebands.

3.3. Performance Improvement of 3D Soil Attribute Mapping

Figure 5 shows the improvement in 3D prediction accuracies of covariate set C+SRS and covariate set C+LSS over covariate set C. Covariate set C+LSS achieved 30~80% accuracy improvement (40% on average) compared with covariate set C; it achieved about 30% accuracy improvement compared with covariate set C+SRS. This indicates that the addition of laboratory soil spectral information can effectively improve the accuracy of 3D mapping of soil attributes relative to conventional soil-forming variables. Laboratory soil spectral information had a higher accuracy enhancement effect compared with satellite remote sensing information. Laboratory soil spectral information was obtained by direct spectral measurement of soil samples, which does not have the interference observed for other factors such as atmosphere, temperature and humidity. Thus, it is a more intuitive response to changes in soil attributes compared with conventional soil-forming and satellite remote-sensing variables.

The average R² improvements for the C+LSS covariate set compared with the C covariate set were, respectively, 31.14%, 57.26%, 46.08%, 40.78% and 32.90% for soil pH, CEC, Silt, SOC and TP, while those for the C+SRS covariate set were, respectively, 11.75%, 14.51%, 37.19%, 24.93%, and 11.98%. This indicates that using laboratory soil spectral information as covariates can produce higher accuracy compared with satellite remote sensing information. Satellite remote sensing sensors are susceptible to the influence of ground cover when capturing reflectance information in the bare soil state during ground observation. Soil CEC, Silt and SOC have direct correlations with soil spectral information, and thus their prediction accuracy with the C+LSS covariate set is better than other attributes.

Table 6. Average accuracy improvements in the five soil attributes at 0–60 cm and 60–100 cm depths when using satellite remote sensing variables and laboratory soil spectral variables.

Soil Attributes	Depth (cm)	Average of Accuracy Improvement
		M_C+SRS	M_C+LSS
pH	0–60	0.208	0.334
	60–100	0.106	0.303
CEC	0–60	0.189	0.505
	60–100	−0.038	0.867
Silt	0–60	0.523	0.633
	60–100	0.154	0.260
SOC	0–60	0.290	0.411
	60–100	0.223	0.479
TP	0–60	0.247	0.355
	60–100	−0.239	0.379

lists the means of the accuracy improvements in the soil attributes at 0–60 cm and 60–100 cm depths when using the C+SRS and the C+LSS covariate sets. It shows that the accuracy improvements tend to decrease with increasing depth when using the C+SRS covariate set. Accuracy improvements for CEC, SOC and TP tend to increase with the increase in depth when using the C+LSS covariate set. This is because satellite remote sensing can only capture soil information on the surface layer, while laboratory soil spectral information is the result of direct observation of soil samples and is capable of observing soil samples at any depth. Its prediction accuracy depends only on the correspondence between soil attributes and soil spectral information and it can effectively improve the situation such that the prediction accuracy of soil attributes decreases with depth.

Figure 6 shows the relative importance of the top 10 C+LSS covariates. blue represents conventional soil-forming variables and yellow represents laboratory soil spectral variables. It can be seen that the laboratory soil spectral information played an important role in the prediction of the spatial distribution of all the soil attributes. SW-NIR and LW-NIR were, respectively, the most important covariates for soil pH and CEC. G, T5 and T9 were the most important covariates for Silt at depths below 5 cm, which mainly respond to the effect of NIR short-wave band and green wave band on soil texture. LW-NIR and MW-NIR were the most important covariates for SOC. The soil spectral information in the near-infrared band has a good inversion capability for soil organic carbon. G, MW-NIR and T12 were the most important covariates for TP in soils below 5 cm depth.

Figure 7 shows the relative importance percentage of conventional soil-forming variables and laboratory soil spectral variables in the C+LSS covariate set. It can be seen that the laboratory soil spectral variables were more important than the conventional soil-forming variables at different depths for all five soil attributes. Figure 8 shows the scatter plot of the relative importance percentage of laboratory soil spectral information against the prediction accuracy improvement in the C+LSS covariate set. They had an overall positive correlation (p < 0.01). The higher the relative importance of laboratory soil spectral variables, the bigger the accuracy improvement for different soil attributes and depths.

4. Discussion

4.1. Predictability of Laboratory Soil Spectral Information

Table 4 and Figure 3 show that the prediction of laboratory soil spectral information in different bands can be realized using spatial speculation. The constructed model can predict the spectral information of soil in the laboratory well and has a certain accuracy (Table 4). At the same time, the spectral information of laboratory soil predicted by the random forest model has obvious spatial distribution characteristics.

The prediction results of soil spectral information in the laboratory have obvious horizontal and vertical distribution characteristics. They were affected by the geographical landscape, topographic position and soil attributes. The central alluvial areas have small soil particle sizes that influence the absorption characteristics of soil and result in a reduction in soil spectral reflectance [41,42]. Since soil organic carbon decreases with increasing depth, the soil in the lower depth usually has high spectral reflection [43,44].

At the same time, we found a relatively high correlation between laboratory soil spectral information and general soil attributes (Figure 4). The correlation coefficients of soil pH, CEC and SOC with laboratory soil spectral bands were relatively high. This is because soil organic matter content can influence soil spectral information in the near-infrared band [45]. The negative charge carried on the surface of soil organic matter can regulate soil acid buffering and thus affect soil pH [46,47]. The correlations between Silt and TP with each covariate at the 0–15 cm depth were weak. Due to the high agricultural activities in the study area, tillage and fertilization interfered with the nutrient composition of the topsoil. This led to a decrease in the correlation between the texture and nutrient content of the topsoil and the covariates; intermediate and deeper soil Silt and TP had high correlations with the vis-NIR bands of laboratory soil spectra and some spectral correlation factors.

4.2. Advantages of Laboratory Soil Spectral Information as Covariates

First, vis-NIR remote sensing can identify surface vegetation information and detect soil information only at a depth of about 3–5 cm on the extreme surface of the soil under bare soil conditions [48]. However, it is susceptible to interference from factors such as atmosphere and surface moisture. Although microwave remote sensing can penetrate the atmosphere and detect a certain depth of soil, it is also affected by the roughness of the ground surface, moisture and other factors. The laboratory soil spectral information is a direct hyperspectral detection of air-dried and ground soil samples under controlled laboratory conditions. The obtained spectral information is more representative of the physical and chemical attributes of the soil itself [15]. The soil samples were collected through soil surveys and the observation depth depends on the depth of the soil profile excavation. Many studies show high accuracy of soil spectral inversion of multiple soil attributes [49,50,51]. Using laboratory soil spectral information as covariates significantly increased the ability to capture soil attributes at lower depths and complemented the defects that remote sensing covariates primarily capture surface environmental information.

Second, the vis-NIR bands of laboratory soil spectral information are spectroscopically associated with many soil attributes. They can effectively differentiate soil types and predict soil attributes such as soil organic matter, soil texture, moisture content, iron oxide content, carbonate minerals and contaminant content [52,53,54]. The absorption bands at 400–800 nm, 1700–1900 nm and 2100–2400 nm can characterize the soil organic matter content. The absorption bands at 1300–1500 nm and 1800–2500 nm are sensitive to mineral content. The LW-NIR band at 1800–2100 nm can respond to information associated with soil moisture content. The binding band at 400–1000 nm reflects the distribution of iron oxides in the soil. The absorption band at 400–700 nm is an important basis for soil color [15,44]. Therefore, laboratory soil spectral information has great potential to provide effective environmental covariates for the predictive mapping of soil attributes.

4.3. Limitations of This Study

First, the laboratory soil spectral covariates are generated by the spatialization of soil spectral information of a number of soil sampling points and the predictive accuracy depends on a variety of factors such as machine learning models. The prediction error would be transferred to the soil attribute prediction results. If the prediction angle of soil spectral information in the laboratory is higher, the prediction result of soil attributes may be realized one step closer. This may be more dependent on the development of models and the indicative nature of environmental covariates used in predicting the spectral information of soils in the laboratory. The next study will quantitatively analyze this error transfer.

Second, the laboratory soil spectral information was reorganized by referring to the wavelength ranges of the vis-NIR bands of remote sensing satellite sensors [55]. The purpose of the reorganization was to compare laboratory soil spectral information with satellite remote sensing spectral information in 3D soil predictions, but the reorganization did not fully consider the association of spectral bands with specific soil attributes and converted the rich hyperspectral information into several simple spectral broadbands. This leads to the loss of a lot of spectral information and reduces the sensitivity of laboratory spectra to soil attributes, which may influence the performance of soil attribute predictions. In the next study, we will fully consider the characteristic bands of different soil attributes in the soil spectrum. Aiming at different soil attribute predictions, different feature band combinations were proposed to improve the 3D prediction accuracy of soil attributes.

5. Conclusions

The study demonstrated the effectiveness of an approach for improving the accuracy of 3D soil attribute mapping using spatialized laboratory soil spectral information as environmental covariates. Compared with the use of only conventional soil-forming covariates, the addition of laboratory soil spectral information can improve the prediction accuracy by about 50% and up to nearly 90% for the 3D spatial prediction of soil attributes. Compared with satellite remote sensing information, it can improve the prediction accuracy by about 30%. The approach in this study enhances the ability to quantitatively characterize the soil-forming environment and enriches the study of environmental covariates. It provides an effective way of improving the prediction accuracy of 3D digital soil mapping and has important implications for the applications of the existing soil spectral libraries.