Spatial Estimation of Soil Organic Matter and Total Nitrogen by Fusing Field Vis–NIR Spectroscopy and Multispectral Remote Sensing Data

Abstract

Accurate and timely acquisition of soil information is crucial for precision agriculture, food security, and environmental protection. Proximal visible near-infrared reflectance (vis–NIR) spectroscopy has been widely employed for rapid and accurate soil measurement, but its point measurement nature limits its direct applicability for large-scale soil surveys. On the other hand, remote sensing techniques can provide soil information at a larger scale, but their resolution is relatively coarse. While both techniques have been used independently for soil analyses, integrating vis–NIR spectroscopy with remote sensing remains a challenge and is underexplored, especially at the field scale. This study addresses this gap by combining field vis–NIR spectra with Gaofen-1 remote sensing data to spatially analyze soil organic matter and total nitrogen at the field scale. Unlike previous work, we first applied Gaofen-1 data and 10 derived spectral indices to estimate soil organic matter and total nitrogen using partial least squares regression and random forest, identifying the optimal combination of spectral indices. Then, we integrated the proximal vis–NIR spectra with this optimal spectral index combination for improved soil property estimation. This integration advanced existing methodologies by leveraging the high spatial resolution of Gaofen-1 data and the detailed spectral information from vis–NIR spectroscopy. The results showed the following: (1) the coefficient of variation across different crop growth stages of Gaofen-1 data was more crucial for modeling these two properties compared to bare soil Gaofen-1 data; (2) integrating proximal vis–NIR spectra with Gaofen-1 data improved model performance, yielding Lin’s concordance correlation coefficient (${\rho }_c$) values of 0.63 and 0.72 and ratios of performance to interquartile distance (RPIQ) of 1.99 and 1.59 for soil organic matter and total nitrogen, respectively; and (3) the combined use of vis–NIR spectra and Gaofen-1 data provided higher spatial estimation accuracy (R² of 0.68 and 0.57 for soil organic matter and total nitrogen) compared to ordinary kriging (R² of 0.63 and 0.31 for soil organic matter and total nitrogen). These results demonstrate that the synergistic use of remote sensing and proximal soil sensing is a practical approach for spatially estimating soil organic matter and total nitrogen at the field scale.

1. Introduction

Soil organic matter (SOM) and total nitrogen (TN) are critical soil fertility indicators. SOM improves soil structure, the availability of soil phosphorus and potassium, and soil water retention. Thus, it is of great significance to soil fertility, crop growth, and sustainable agricultural systems [1]. TN is the crucial nutrient element for crop growth and is necessary for maintaining a sufficient soil nitrogen supply to ensure the healthy growth of crops [2]. However, the excessive application of nitrogen fertilizer causes soil degradation (i.e., soil acidification) and a series of environmental problems, such as N leaching in groundwater and emissions of nitrous oxide [3]. Therefore, it is vital to acquire the spatial distribution of SOM and TN in a timely manner and as accurately as possible for precision agriculture and environmental protection.

Visible near-infrared reflectance (vis–NIR) spectroscopy, as a rapid, non-destructive, cost-effective proximal soil sensing technique, has been widely used in fast soil measurement [4,5,6,7,8]. Currently, vis–NIR has been successfully used to estimate soil physical, chemical, and biological properties, such as SOM, soil organic carbon (SOC), particle size fractions (i.e., clay, silt, and sand), iron and aluminum oxides, and TN [9,10,11,12]. SOM/SOC, clay, and TN are the most successfully estimated properties due to corresponding spectral absorption characteristics in the vis–NIR spectral region [13]. For example, the absorption characteristics of SOM in the near-infrared region are affected by stretching vibration and angular vibration of groups of C–H, N–H, O–H, C–O, C–N, N–O, C–C, and C=O [14]. The band characteristics near 1100 nm, 1600 nm, 700~1800 nm, 2000 nm, and 2200~2400 nm are mainly related to the SOC and TN contents [4,15,16]. Although vis–NIR can achieve rapid and accurate estimation of SOM and TN, its measurement is still based on point samples. Therefore, we cannot acquire spatial information on soil properties without spatial predictive models.

Remote sensing technology has the advantages of wide coverage, strong timeliness, and rapid information acquisition. Thus, it is widely used in large-area ground information acquisition, soil monitoring, and digital soil mapping [17,18]. Alabbas et al. (1972) conducted aerial experiments to study the relationship between soil organic matter and clay content with visible and near-infrared reflectance [19]. Since then, researchers have carried out a series of studies using remote sensing for spatial estimation of soil attributes, especially for measuring SOM/SOC and soil texture. For example, Chen et al. estimated SOC concentrations in a 115 ha field using remote sensing imagery, establishing a logarithmic linear relationship (R² = 0.93) between SOC and RGB band intensities [20]. Fox et al. developed a soil line Euclidean distance technique using red and near-infrared band intensities to map SOM, demonstrating its effectiveness in two Midwest fields [21]. Early studies usually used the bands of remote sensing images and their mathematical transformations (i.e., logarithm, reciprocal) to build predictive models.

With the development of advanced sensors, the use of hyperspectral and multispectral data has significantly enhanced soil property analysis. Hyperspectral data, with their numerous narrow spectral bands, enable detailed soil property analysis, while multispectral data, with fewer but broader bands, are more suitable for efficient large-scale mapping. Both types of data play complementary roles in remote sensing, with hyperspectral data being particularly useful for detailed analysis and multispectral data for broader spatial coverage [22,23]. Especially since the launch of Sentinel-2 in 2015, its broad spectral range, high spatial resolution (10–60 m), and short revisit time (5 days) have greatly improved the spatial estimation of soil properties by enabling frequent and detailed observations of soil reflectance characteristics [22,24,25,26,27]. However, soil property prediction based on remote sensing data with moderate spatial resolution, such as Sentinel-2, has primarily focused on large-scale areas or has used remote sensing data as auxiliary variables in soil mapping studies. For example, Peng et al. used Sentinel-1, Sentinel-2, and Landsat-8 along with DEM data and machine learning methods for SOM mapping in hilly and mountainous regions, and their results showed that Sentinel-2 data have a significant impact on cropland SOM prediction in hilly and mountainous areas [28]. Ji et al. utilized Sentinel-2 remote sensing data as predictors for predicting SOC content at varying spatial resolutions in Germany, demonstrating that hybrid models, particularly at 100 m resolution, provided more accurate predictions for national-scale SOC monitoring [29]. These studies often leverage the broad spatial coverage of medium-resolution sensors to address large-area soil monitoring challenges. However, their application in small-scale or high-resolution studies remains limited due to spatial resolution constraints and the influence of vegetation cover. The Chinese Gaofen-1 satellite (GF-1), with its high-resolution multispectral sensors (spatial resolution of 2 m for panchromatic and 8 m for multispectral) and frequent revisit cycle (4 days), offers a valuable data source for soil property estimation and provides high-spatial-resolution soil data for precision agriculture and farmland management.

Although remote sensing has the advantages of rapid data acquisition, extensive area monitoring, and data simplification [25,30], its application in modeling the spatial distribution of soil properties is limited by spatial resolution and plant disturbance. The proximal soil sensing through vis–NIR provides accurate soil property estimation but lacks the capability for large-area monitoring. Previous research has primarily focused on individual approaches for soil measuring and mapping, such as solely relying on vis–NIR spectra or remote sensing data. However, there is a lack of comprehensive research exploring the synergistic utilization of these two data sources for spatial estimation of soil properties.

To the best of our knowledge, few studies have investigated the combined use of field vis–NIR spectroscopy and remote sensing data, particularly applying this approach to spatially estimate soil attributes at field scale. By integrating the high-resolution information derived from field vis–NIR spectroscopy and the wide coverage and timeliness offered by remote sensing data, we aim to introduce a novel approach that overcomes the limitations of individual techniques and provides accurate and extensive area monitoring. Therefore, in this study, we directly integrated in situ proximal vis–NIR spectroscopy with remote sensing data (GF-1) for spatial estimation of SOM and TN. Unlike traditional approaches that use remote sensing data as auxiliary variables in predictive models, our method establishes a direct fusion framework, leveraging the high spectral resolution of vis–NIR for precise local measurements while utilizing GF-1’s broad spatial coverage for large-area mapping. By synergistically combining these complementary datasets and employing machine learning algorithms, we aim to enhance both the spatial continuity and estimation accuracy of SOM and TN, addressing the limitations of single-source data.

The main objectives are as follows: (i) to compare the model performance based on bare soil GF-1 data and multi-temporal GF-1data in different crop growth stages using linear and nonlinear models, (ii) to compare the model performance using GF-1 data and the combination of GF-1 data and field vis–NIR spectra data, and (iii) to optimize the best spatial predictive model for digital soil mapping.

This paper is structured as follows: firstly, we introduce the materials and methods of this study, including the study area, sampling methods, physicochemical and spectral testing, and modeling approaches. Then, we present the main research results, focusing on the prediction results based on GF-1 remote sensing data, the prediction results from the fusion of GF-1 and field vis–NIR data, and a comparison between the two. We also cover the spatial estimation results based on the fusion model. Then, we discuss and analyze the main findings, comparing them with those of previous studies. The final section summarizes the key conclusions of this research. The study process is shown in Figure 1.

2. Materials and Methods

2.1. Study Area and Soil Sampling

The study area (114°0′50″–114°1′15″E, 34°6′30″–34°6′40″N) is located in Xuchang, Henan Province, a critical food production region (i.e., wheat, corn) in China (Figure 2). This region has a warm temperate continental monsoon climate with an annual rainfall of between 671.1 mm and 736.0 mm and a mean annual temperature of 14.6 °C. The predominant soil type in the study area is fluvo-aquic soil, classified as Calcaric Cambisols under the World Reference Base for Soil Resources (WRB). The soil is characterized by a loam texture, which contributes to its favorable water retention capacity and permeability properties. In this study, 240 topsoil samples (0–20 cm) were collected in October, 2019, before wheat planting based on grid sampling with a spacing of 20 m using a cuboid sampling tool (10 × 10 × 20 cm). After removing stones and roots, all the soil samples were air-dried, ground, and sieved to less than 2 mm in the laboratory for further analysis [13].

2.2. Data Acquisition and Treatment

2.2.1. Chemical Analysis and Spectra Measurement

SOM and TN were analyzed at Zhejiang Academy of Agricultural Sciences (located in Hangzhou, China), a nationally certified laboratory, using the H₂SO₄–K₂Cr₂O₇ oxidation method for SOM and the semi-micro Macro Kjeldahl method for TN, following standard analytical protocols. The vis–NIR spectra were measured in field and laboratory conditions using a Fieldspec^® ProFR spectrometer and a high-intensity contact probe with its own light source (Malvern Panalytical Ltd., Malvern, UK). In field conditions, three vertical locations (i.e., bottom, middle, and top) were selected for spectral measurement from each soil sample (10 × 10 × 20 cm). Ten spectra were scanned at each selected location, and a total of 30 spectra were averaged to represent the final spectra of each soil sample. To maintain the quality of the data, we calibrated the spectrometer using a Spectralon panel with 99% reflectance every 10 soil samples. In laboratory conditions, the spectra were measured on the air-dried, ground, and sieved soil samples in a petri dish. Three points were selected for measurement, with the final results being the average of the results of the three points.

2.2.2. Spectra Preprocessing

The vis–NIR spectra were first reduced to 400–2450 nm to eliminate the influence of noise at the edge and then down-sampled to a resolution of 10 nm to improve the calculation efficiency and reduce collinearity. Then, the reflectance spectra (R) were transformed into absorbance with log(1/R) and smoothed using the Savitzky–Golay algorithm (SG, a window size of 15 and polynomial of order 2) to further reduce the noise and enhance the signal [31]. Direct standardization (DS) [32] was used to transform field spectra into laboratory spectra to reduce the influence of water and other environmental factors. After spectral preprocessing, principal component analysis (PCA) was applied for dimensionality reduction [33]. The first four principal components were then interpolated onto the study area using ordinary kriging in ArcGIS. All these processes were conducted in R 4.2.1 [34] and Matlab (The MathWorks Inc., Natick, MA, USA).

2.2.3. Remote Sensing Data Acquisition

The remote sensing data used in this study were GF-1 satellite data (

Table 1. Parameters of GF-1 satellite data.

Satellite	Band		Band Range (nm)	Resolution (m)
GF-1	Panchromatic image		450~900	2
	Multispectral image	1 (blue)	450~520	8
		2 (green)	530~590	8
		3 (red)	630~690	8
		4 (near infrared)	770~890	8

). The GF-1 satellite is the first satellite of China’s high-resolution earth observation system with high spatial and temporal resolution. It has been widely used in land monitoring, disaster management, environmental protection, and various other areas. We collected four periods of GF-1 data on different growth stages of wheat (December 27, February 21, and March 14) and the bare soil (October 17).

2.2.4. Remote Sensing Data Preprocessing and Spectra Indices Calculation

The preprocessing of GF-1 data included radiometric calibration, atmospheric correction, geometric correction, and fusion panchromatic and multispectral images. The resulting multispectral data with a spatial resolution of 2 m were used in this study. The preprocessing was conducted in ENVI 5.3.

We calculated the spectral indices (

Table 2. Definitions of spectral indices and their calculations.

Index	Definition	Reference
NDVI	${\displaystyle \frac{r_{\mathrm{N}\mathrm{I}\mathrm{R}}-r_{\mathrm{R}\mathrm{e}\mathrm{d}}}{r_{\mathrm{N}\mathrm{I}\mathrm{R}}+r_{\mathrm{R}\mathrm{e}\mathrm{d}}}}$	[36]
TVI	${({\displaystyle \frac{r_{\mathrm{N}\mathrm{I}\mathrm{R}}-r_{\mathrm{R}\mathrm{e}\mathrm{d}}}{r_{\mathrm{N}\mathrm{I}\mathrm{R}}+r_{\mathrm{R}\mathrm{e}\mathrm{d}}}}+0.5)}^{{\displaystyle \frac{1}{2}}}\times 100$	[37]
EVI	${\displaystyle \frac{2.5\times {(r}_{\mathrm{N}\mathrm{I}\mathrm{R}}-r_{\mathrm{R}\mathrm{e}\mathrm{d}})}{r_{\mathrm{N}\mathrm{I}\mathrm{R}}+{6\times r}_{\mathrm{R}\mathrm{e}\mathrm{d}}-{7.5\times r}_{\mathrm{R}\mathrm{e}\mathrm{d}}+1}}$	[38]
SAVI	${\displaystyle \frac{(r_{\mathrm{N}\mathrm{I}\mathrm{R}}-r_{\mathrm{R}\mathrm{e}\mathrm{d}})\times 1.5}{r_{\mathrm{N}\mathrm{I}\mathrm{R}}+r_{\mathrm{R}\mathrm{e}\mathrm{d}}+0.5}}$	[39]
GNDVI	${\displaystyle \frac{r_{\mathrm{N}\mathrm{I}\mathrm{R}}-r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}}{r_{\mathrm{N}\mathrm{I}\mathrm{R}}+r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}}}$	[40]
GRVI	${\displaystyle \frac{r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}-r_{\mathrm{R}\mathrm{e}\mathrm{d}}}{r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}+r_{\mathrm{R}\mathrm{e}\mathrm{d}}}}$	[41]
BI	${\displaystyle \frac{\sqrt{{(r}_{\mathrm{R}\mathrm{e}\mathrm{d}}\times r_{\mathrm{R}\mathrm{e}\mathrm{d}})+{(r}_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}\times r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}})}}{2}}$	[42]
BI2	${\displaystyle \frac{\sqrt{{(r}_{\mathrm{R}\mathrm{e}\mathrm{d}}\times r_{\mathrm{R}\mathrm{e}\mathrm{d}})+{(r}_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}\times r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}})+{(r}_{\mathrm{N}\mathrm{I}\mathrm{R}}\times r_{\mathrm{N}\mathrm{I}\mathrm{R}})}}{3}}$	[42]
RI	${\displaystyle \frac{{(r}_{\mathrm{R}\mathrm{e}\mathrm{d}}\times r_{\mathrm{R}\mathrm{e}\mathrm{d}})}{r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}\times r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}\times r_{\mathrm{G}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{n}}}}$	[43]
V	${\displaystyle \frac{r_{\mathrm{N}\mathrm{I}\mathrm{R}}}{r_{\mathrm{R}\mathrm{e}\mathrm{d}}}}$	[44]

), which mainly included vegetation indices closely related to soil organic matter and some indices related to soil brightness [25,30]. Moreover, we calculated the change rate (SI_r) and coefficient of variation (SI_cv) of spectral indices in different periods as the variables for soil estimation. The spectra indices were calculated using the ‘rgdal’ package [35] in R 4.2.1.

$${\mathrm{S}\mathrm{I}}_{\mathrm{r}}={\displaystyle \frac{{{\mathrm{S}\mathrm{I}}_{t_2}-\mathrm{S}\mathrm{I}}_{t_1}}{{\mathrm{S}\mathrm{I}}_{t_1}}}$$

(1)

$${\mathrm{S}\mathrm{I}}_{\mathrm{c}\mathrm{v}}={\displaystyle \frac{\mathrm{s}\mathrm{d}\left(\mathrm{S}\mathrm{I}\right)}{\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}\left(\mathrm{S}\mathrm{I}\right)}}$$

(2)

where SI_r represents the change rate of spectra indices in different periods,${\mathrm{S}\mathrm{I}}_{t_1}$ and ${\mathrm{S}\mathrm{I}}_{t_2}$ represent the spectra indices in different periods, and SI_cv represents the coefficient of variation in spectra indices for different periods.

2.3. Estimation Model

In this study, we used partial least squares regression (PLSR) and random forest (RF) for spatial estimation of SOM and TN. Partial least squares regression (PLSR) is one of the most commonly used multiple linear calibration algorithms for spectral calibration and prediction [45]. It has the advantage of eliminating the problem of multiple collinearities of the independent variables [46].

Random forest (RF) is a combined tree prediction method [47] that contains a series of random classification regression tree models. The final regression or classification was achieved by assembling these tree models. RF does not need to screen variables and can deal with classified and continuous variables simultaneously. Compared with other models, RF can produce lower estimation deviation and variance. In the PLSR model, the main parameter is the number of components (ncomp). In RF, the main parameters include the number of trees (ntree), the minimum number of terminal nodes (node size), and the number of covariates that are randomly selected at each tree (mtry) [47,48]. All of these parameters were optimized through 10-fold cross-validation using “caret” package in R 4.2.1.

We also used ordinary kriging interpolation (OK) for spatial analysis of SOM and TN and compared the results with the best model based on the combination of GF-1 data and field vis–NIR spectra data.

2.4. Model Development and Performance Evaluation

Before model construction, we randomly split the data into calibration (two-thirds) and validation (one-third) sets using a random sampling method. To improve the model’s robustness and quantify the prediction uncertainty, we used 50 bootstraps on the calibration data [49].

Lin’s concordance correlation coefficient (${\rho }_c$) [50], the coefficient of determination (R²), root mean square error (RMSE), the ratio of performance to interquartile distance (RPIQ), the standard deviation of the error (SDE), and the mean error (ME) were used to evaluate and compare the performance of the models. Additionally, the estimation uncertainty was calculated using their 90% confidence intervals.

$$\mathrm{S}\mathrm{D}\mathrm{E}=\sqrt{\sum _{I=1}^n{\displaystyle \frac{{{(\widehat{y}}_i-\overline{y})}^2}{(n-1)}}}$$

(3)

$$\mathrm{M}\mathrm{E}=\sum _{i=1}^n{\displaystyle \frac{{(\widehat{y}}_i-\overline{y})}{n}}$$

(4)

$$U_j={\displaystyle \frac{S_{upper,j}-S_{lower,j}}{{\overline{S}}_j}}$$

(5)

where $\widehat{y}$ is the predicted value of soil properties;$\overline{y}$ is the mean value of the measured value of soil properties, n represents the number of soil samples; $U_j$ represents the uncertainty of the j-th grid of the spatial estimation; $S_{upper,j}$ and $S_{lower,j}$ are the upper confidence interval and lower confidence interval of the j-th grid, respectively; and ${\overline{S}}_j$ is the mean value of 50 bootstraps of the j-th grid.

3. Results

3.1. Descriptive Statistics of SOM and TN

The statistics of SOM and TN are shown in Figure 3 and

Table 3. Statistics of SOM and TN for different datasets.

Properties	Dataset	N	Max.	Median	Min.	Mean	Standard Deviation
SOM	All	240	24.1	18.1	9.1	18.15	2.62
	calibration	160	24.1	18.05	12.7	18.16	2.6
	validation	80	24	18.15	9.1	18.15	2.67
TN	All	240	1.74	1.33	0.82	1.34	0.14
	calibration	160	1.73	1.33	0.82	1.34	0.14
	validation	80	1.74	1.34	1.03	1.35	0.15

. The SOM ranged from 9.10 to 24.10 g/kg, and its mean value was approximate to the median value. The variance of SOM was small, with a coefficient of variance (CV) value of 14.48%. As for TN, the range was from 0.82 to 1.74 g/kg, and its variance was also slight, with a CV of 10.54%. From the violin plots of SOM and TN, we found that the calibration and validation datasets were similar to that of the whole dataset, indicating that the data division was representative.

3.2. Estimation of SOM and TN Using Remote Sensing Data

In this study, we selected the GF-1 data of bare soil, which closely matched the sampling data, along with three periods of GF-1 data at the different growth stages of wheat to estimate SOM and TN. We compared the model performance of SOM and TN based on the bands and spectra indices of four periods and the change rate (SI_r) and coefficient of variation (SI_cv) of the bands and spectral indices in different periods using a PLSR and RF model. The model performance based on bare soil GF-1 data was not good, which may result from the non-tillage system in the study area (Figure 4). Although the crop harvest was completed in October, the soil surface was still covered with corn straw, which affected the acquisition of bare soil image data. Compared to the estimation results of SI_r and SI_cv of different periods, the best model performance was achieved in the model based on the SI_cv of data from four periods. This result indicates that the growth difference between different growth stages of crops plays an important role in estimating soil properties.

According to ${\rho }_c$, RMSE, and RPIQ, the RF models had better model performance than the PLSR models in estimating SOM and TN (Figure 4). When considering SDE, the RF models were better than the PLSR models, while the overfitting problem of RF was greater than that of the PLSR model (

Table 4. Model performance based on GF-1 data for SOM and TN (in g/kg).

Properties		SOM		TN
Methods		PLSR	RF	PLSR	RF
Calibration	${\rho }_c$	0.57	0.96	0.31	0.95
	RMSE	2.05	0.69	0.14	0.04
	RPIQ	1.90	5.65	1.43	4.96
	ME	0	−0.01	0	0
	SDE	2.06	0.69	0.15	0.04
Cross-validation	${\rho }_c$	0.33	0.43	0.11	0.10
	RMSE	3.02	2.46	0.17	0.18
	RPIQ	1.42	1.66	1.16	1.04
	ME	−0.24	−0.19	0.00	−0.01
	SDE	2.99	2.45	0.17	0.18
Validation	${\rho }_c$	0.44	0.53	0.16	0.31
	RMSE	2.29	1.98	0.16	0.14
	RPIQ	1.56	1.79	1.06	1.20
	ME	0.02	0.00	0.05	0.04
	SDE	2.29	1.98	0.15	0.13

). The estimation result of SOM was acceptable, with a ${\rho }_c$ of 0.53 and RPIQ of 1.79, which showed that GF-1 data could achieve quantitative analysis for SOM. As for TN, the estimation using GF-1data was poor, with a ${\rho }_c$ of 0.31 and RPIQ of 1.20. Fertilization management is necessary during the growing process of wheat, and the content of nitrogen is greatly affected by fertilization management. Therefore, it is difficult to estimate TN using remote sensing data.

3.3. Field Spectra Transformation and PCA Analysis

The field spectra fluctuate considerably compared with the laboratory spectra, especially around 1400 nm and 1900 nm (Figure 5a). We used DS for the transformation from field to laboratory spectra, which has been proven to be an effective method for reducing the effect of environmental factors [12]. Before DS transformation, we used KS [51] to select the representative transformation samples from field spectra, and the result showed that the best number of transformation samples was 55 in this study (Figure 6). The DS-transferred spectra were similar to laboratory spectra, and their absolute differences were generally lower than 0.1, which indicates good transformation (Figure 5b).

The DS-transferred spectra were then analyzed using PCA. The PCA analysis showed that the cumulative contribution of the first four principal components reached 99%, of which the contributions of the first two principal components were larger, which were 64.55% and 30.93%, respectively (

Table 5. Principal component analysis of DS transferred vis-NIR spectra.

	Standard Deviation	Variance Contribution (%)	Cumulative Contribution (%)
PC1	2.39	64.55	64.55
PC2	1.65	30.93	95.48
PC3	0.53	3.20	98.68
PC4	0.24	0.65	99.32

). We reserved the first four principal components for subsequent analysis to maintain more spectral information and improve the estimation accuracy.

3.4. Estimation of SOM and TN Through the Synergistic Use of Field Spectra and GF-1 Data

We combined GF-1 data and the first four PCs of field spectra for the spatial estimation of SOM and TN using the PLSR and RF model (

Table 6. Estimation of SOM and TN and their uncertainties based on a combination of field spectra and GF-1 using PLSR and RF (g/kg).

Properties	Methods	${\mathit{\rho }}_{\mathit{c}}$	RMSE	RPIQ	ME	SDE
SOM	PLSR	0.72 (0.66~0.78)	1.93 (1.60~2.27)	1.84 (1.54~2.14)	0.32 (0.06~0.59)	1.91 (1.60~2.22)
	RF	0.63 (0.57~0.70)	1.78 (1.63~1.93)	1.99 (1.83~2.15)	−0.04 (−0.30~0.25)	1.78 (1.64~1.93)
TN	PLSR	0.72 (0.57~0.88)	0.11 (0.07~0.15)	1.59 (1.08~2.09)	0.02 (0.00~0.05)	0.11 (0.06~0.15)
	RF	0.56 (0.44~0.68)	0.11 (0.10~0.13)	1.51 (1.32~1.69)	0.03 (0.02~0.05)	0.11 (0.09~0.12)

Note: The data in brackets are 90% confidence intervals.

). The estimation accuracies of SOM and TN were generally improved following the synergistic use of the above two techniques (Table 4 and Table 6). The ${\rho }_c$ of SOM improved from 0.53 to 0.72, RMSE decreased from 1.98 g/kg to 1.78 g/kg, RPIQ increased from 1.79 to 1.99, and SDE was also reduced. As for TN, the ${\rho }_c$ improved from 0.31 to 0.72; RMSE and SDE decreased from 0.14 g/kg and 0.12 to 0.11 g/kg and 0.11, respectively; and RPIQ increased from 1.20 to 1.59. Compared with the PLSR and RF models, we found that the RF model was better than the PLSR model for SOM considering all the five model evaluation indexes as well as the model uncertainty. Additionally, the PLSR model was better than the RF model for TN, while the model uncertainty was larger than that of the RF model. Generally, the SOM and TN estimations were improved and could be successfully estimated through the synergistic use of proximal soil sensing and remote sensing.

We analyzed the significance of each variable for SOM and TN (Figure 7). In general, the PCs of field spectra played an essential role in estimating SOM and TN. This explains why the estimation accuracy was vastly improved after adding field spectra in the model (Table 3 and Table 4). PC1 and PC3 of field spectra, B4, NDVI, EVI, SAVI, GNDVI, and V of GF-1 were found to be significant, all exceeding 40% for SOM. As for TN, the significance of GF-1 data was relatively low, around 20%. The PCs of field spectra were found to be highly significant, among which PC1, PC3, and PC4 were the main driving factors, especially PC3 (above 90%).

3.5. Spatial Analysis of SOM and TN

We used the best predictive models of SOM and TN for spatial analysis in the study area and compared the results with the ordinary kriging interpolation (OK) of SOM and TN (Figure 8). The spatial estimation based on the synergistic use of remote sensing and proximal soil sensing had a similar distribution trend with the ordinary kriging interpolation, and it had higher estimation accuracy than the latter (Figure 9). Moreover, it expressed more detailed information than ordinary kriging interpolation (Figure 8). The distribution of SOM was generally low in the west and high in the east, which is roughly a block distribution in the north–south direction. The area with the highest SOM content (SOM > 20.00 g/kg) was located in the southeast of the field, and the area with the lowest SOM content (SOM < 14.27 g/kg) was located in the west and a banding area in the west-central area of the field. The TN distribution was similar to that of SOM and was also low in the west and high in the east on the whole. Less detailed information on TN was available compared to that of SOM because the GF-1 data played a less important role in the estimation model of TN (Figure 7).

The spatial estimation uncertainties of SOM and TN were expressed by calculating the ratio of the 90% confidence intervals and the corresponding average value of the point (the 90% confidence intervals were calculated using 50 bootstraps) (Figure 8). The uncertainty of SOM was relatively low, with most areas less than 0.06 and only a few small areas larger than 0.10. The estimation uncertainty of soil TN was slightly higher than SOM, with most areas of the field between 0.03 and 0.19. The region with uncertainty greater than 0.19 accounted for only a small part of the area, and the area greater than 0.37 was mainly located at the edge of the field.

4. Discussion

Remote sensing data can be analyzed using physically based or empirical methods to derive soil properties. For soil attributes, remote sensing data are mainly used for deriving soil moisture, texture, or soil salinity with acceptable accuracy [17]. For SOM and TN, the feasibility was generally low, which was consistent with the findings of our study (Figure 4 and Table 4) [17]. Moreover, remote sensing data were mainly used as secondary sources in digital soil mapping for soil attributes in large areas [52,53,54,55]. With the development of the spatial and spectral resolution of the recent multispectral satellite sensors, the application of multispectral satellite data, particularly Sentinel data, for soil property estimation has increased significantly [24,56]. While Sentinel data are not suitable for small areas like in this study, Gholizadeh et al. utilized 18 spectral indices for soil property estimation, demonstrating good performance in predicting soil organic carbon (SOC) across different locations, with a prediction RMSE of less than 0.24 and an RPD exceeding 1.60 [25]. Based on the spectral indices used by Gholizadeh et al. [25], we selected 10 indices that were compatible with our dataset. Unlike Sentinel-2, which offers 10 bands, including shortwave infrared (SWIR), the GF-1 satellite provides only 4 bands (red, green, blue, and near-infrared). As a result, we referred to these indices and used our data to determine 10 spectral indices in this study.

It has been proved by many researchers that SOM and TN can be estimated using vis–NIR spectra [4,13]. However, these estimations have mostly been made in lab conditions. In field conditions, spectra are largely affected by factors such as the water content, particle size, etc.; therefore, algorithms are needed to eliminate these effects [12,57,58]. In this study, we used DS algorithms. After transmission, the field spectra closely resembled laboratory spectra, aligning with previous studies [12]. To further analyze the mechanism of vis–NIR spectra, we analyzed the eigenvectors of the four PCs (Figure 10); the loadings of PC1 had prominent peaks around 600 nm and 1923 nm and small peaks around 1420 nm and 2214 nm, and the loadings increased in the range of 2300–2450 nm. These characteristics are mainly related to iron oxides, clay minerals, and organic matter [14,59,60,61]. Loadings around 600 nm are affected by hematite, and loadings around 1420 nm and 2214 nm are characteristic of clay minerals such as kaolinite and montmorillonite [62]. Loadings around 1420 nm and 1923 nm are also due to the OH- groups of water, and loadings in the range of 2300–2450 nm are dominated by the effect of CH3 groups in organic matter [63,64]. The characteristics of PC2 loadings were mainly observed around 1100 nm, 1400 nm, and 1914 nm. These loadings around 1100 nm are due to the aromatic cyclic hydrocarbon in organic matter. The characteristics of PC3 loadings were mainly around 480 nm, 1124 nm, 1420 nm, 1916 nm, and 2208 nm. Loadings around 480 nm were primarily affected by goethite, loadings around 2208 nm were related to kaolinite, and other characteristics were similar to the first two components. The characteristics of PC4 loadings were mainly around 640 nm, 1100 nm, 1390 nm, 1955 nm, 2212 nm, and 2287 nm. Loadings around 640 nm may be due to the effect of organic matter [33]. Loadings around 1955 nm and 2287 nm are mainly related to phenols and aliphatic compounds in organic matter.

The combined use of different sensors or data has been proven to be an effective approach for soil assessment [45,65]. Previous studies have combined vis–NIR spectra with remote sensing data through digital soil mapping, using vis–NIR spectra as a covariate [66,67]. For example, Zhou et al. combined vis–NIR spectra with remote sensing data and digital elevation model data for digital mapping of SOM [67]. Their results showed that vis–NIR spectra and vegetation factors were strongly correlated with SOM, which was similar to our study (Figure 8). In this study, we only combined vis–NIR spectra with remote sensing data directly and attained satisfactory results for soil mapping (Table 6 and Figure 10). In future work, we will consider exploring different fusing methods, temporal information, spatial information, and spectral information of proximal and remote sensing data for improving soil mapping accuracy.

5. Conclusions

In this study, we investigated the synergistic use of proximal soil sensing and remote sensing for the spatial estimation of SOM and TN. We compared the single period of GF-1 data with the change rate and coefficient of variation in different periods of GF-1 data. We found that the best estimation for SOM and TN came from the coefficient of variation in four periods of GF-1 data. Moreover, the RF model had better estimation accuracy than that of the PLSR model. We compared the combination of proximal soil sensing and remote sensing with remote sensing alone. We found that the estimations for SOM and TN based on the combination were more accurate than those using remote sensing data alone. Additionally, the best estimates for SOM and TN came from the RF model based on the combination of variation in four periods of GF-1 data and the first four PCs of field spectra. We used these models for the spatial analysis of SOM and TN and compared them with ordinary kriging interpolation results, finding that they had a similar distribution trend. Moreover, the RF models contained more detailed information and had higher estimation accuracy. Therefore, the synergistic use of proximal soil sensing and remote sensing was an effective approach for rapid spatial analysis of soil properties.