A method for predicting soil organic matter content based on spectral coupling effect

Abstract

The application is suitable for the technical field of soil organic matter detection, and provides a soil organic matter content prediction method based on spectral coupling effect, which reduces the coupling effect of soil physical properties on spectrum and improves the space-time transferability of the SOM prediction model. Based on satellite hyperspectral images and soil physical variables such as soil moisture, soil surface roughness and soil bulk density, a soil spectrum correction strategy based on information decomposition is established. The results show that the soil spectrum correction based on the fourth order polynomial and XG-Boost algorithm has good precision and generalization ability. In addition, when the soil spectrum correction strategy is used, the precision of the SOM prediction model and the generalization ability of the model after migration are significantly improved. Compared with the direct migration prediction of the model, the RMSE of the SOM prediction result is reduced by 57.90% and 60.27% respectively by using the soil spectrum correction strategy based on the fourth order polynomial and XG-Boost. The work provides a new research paradigm for the prediction of soil property parameters in other regions.

Description

Technical Field

[0001] This invention belongs to the field of soil organic matter detection technology, and in particular relates to a method for predicting soil organic matter content based on spectral coupling effect. Background Technology

[0002] Soil (the terrestrial ecosystem) is the largest carbon pool in terrestrial ecosystems, forming the global carbon cycle together with the hydrosphere, atmosphere, biosphere, geosphere, and lithosphere. Even small changes in the soil carbon pool can significantly alter atmospheric carbon dioxide concentrations, thus impacting the global carbon cycle and climate. The vast majority of carbon stored in soil is organic carbon, the carbon component of organic matter. Soil organic matter (SOM) is also a major source of biological nutrients and energy in soil, and its content is frequently used as an important indicator for assessing soil fertility. Therefore, accurately understanding soil organic matter content and its spatial distribution is crucial for promoting sustainable agricultural development, enhancing soil carbon sequestration potential, and moderating global climate change.

[0003] Remote sensing is a low-cost, high-precision, and real-time method for Earth observation, characterized by its multi-angle, multi-time-period, and large-area capabilities. Currently, the ability of hyperspectral remote sensing to predict and map soil element content (SOM) has been confirmed in numerous studies. With the rapid growth of remote sensing data and the urgent need for large-scale soil surveys, research on soil element content prediction based on remote sensing is gradually shifting from building high-precision prediction models to establishing prediction models with strong spatiotemporal transferability. Imaging spectroscopy is the most important data source, and its characteristic response to soil chemical composition is a crucial foundation for SOM content prediction based on hyperspectral remote sensing. However, imaging spectra are not solely influenced by soil composition but comprehensively reflect the physical and chemical properties of the soil within the ground sample; the physical and chemical properties of the soil exhibit a coupling effect on the spectral response. Studies have shown that soil physical properties such as soil moisture (SM) and surface roughness (e.g., root mean square height RMSH) significantly influence the scattering contribution of spectral reflectance to hyperspectral data on SOM content. Near-infrared spectroscopy is highly sensitive to small amounts of water or hydroxyl groups, easily resulting in irregular radiation characteristics. As SM content increases until saturation, soil reflectance initially decreases, then increases due to specular reflection effects. Furthermore, the increase in RMSH enhances light scattering and transmission at the soil surface, thereby reducing reflectivity, especially in the visible and near-infrared wavelength range. Simultaneously, long-term high-intensity mechanized farming increases soil bulk density (SBW). The changes in SM, SBW, and spectral reflectivity exhibit a complex relationship. Generally, increases in SM, SBW, and RMSH reduce spectral reflectivity, demonstrating a coupling effect. Notably, the impact of SOM content on soil spectra is far weaker than that on soil physical properties. Due to the uncertainties in satellite revisit cycles and soil physical conditions, noise interference from soil physical properties in the spectrum significantly limits the accuracy and spatiotemporal portability of remote sensing-based SOM evaluation models, which is a problem urgently needing to be addressed.

[0004] When mapping and predicting target information, the reliability and completeness of data are often key to the generalization ability of the model. To mitigate the influence of soil physical properties on hyperspectral data and improve the spatiotemporal transferability of SOM content prediction models, researchers have fused long-term series of hyperspectral data to reduce the model's sensitivity to spectral differences caused by changes in soil physical properties. Ge et al. (2022) attempted to introduce soil physical parameters as input variables into the elemental content prediction model. Pan et al. (2022) established SOM content prediction models based on different SM ranges. These innovative improvements at the macroscale have mostly been successful, but in-depth development and analysis of satellite hyperspectral images at the pixel scale are still lacking. Considering the reflectance differences caused by soil physical properties, Minasny et al. (2011) developed a spectral correction model based on the EOP method to eliminate the influence of SM. Castaldi et al. (2015) synthesized dry soil spectra by calculating the statistical variability of dry and wet soil, thereby improving the prediction accuracy of the model. Although these methods have corrected hyperspectral data to some extent, they are mostly based on a single physical parameter and ignore the coupled response of different soil physical properties to the spectrum. Due to the lack of soil physical data from satellite-ground synchronous experiments, the potential of hyperspectral correction methods that comprehensively consider multiple soil physical properties has not yet been fully explored.

[0005] Therefore, developing a soil spectral correction method to mitigate the coupling effect of surface properties on soil pixel spectral density is a long-term solution to improve the spatiotemporal transferability of SOM prediction models. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting soil organic matter content based on spectral coupling effect, aiming to solve the problems mentioned in the background art.

[0007] This invention is implemented as follows: a method for predicting soil organic matter content based on spectral coupling effect. Before establishing a high-precision, highly spatiotemporally transferable SOM content prediction model, a soil spectral correction method was studied to alleviate the coupling effect of surface properties on soil pixel spectra. Figure 1First, empirical relationships between satellite hyperspectral data and three main soil physical property parameters (SM, RMSH, and SBW) were established using parameter estimation equations. These empirical relationships were then used to correlate soil physical parameters with satellite hyperspectral imagery, resulting in three sets of simulated soil spectral data based on SM, RMSH, and SBW. Next, based on the simulated spectra, soil pixel spectra, and ground spectra, a soil pixel spectral correction model was constructed using multi-order polynomials and various machine learning models to separate soil physicochemical information from the pixel spectral data. Finally, a soil spectral accuracy (SOM) prediction model was constructed using XG-Boost based on the original and corrected soil spectral data. The spatiotemporal transferability of the spectral correction model and the SOM prediction model established based on the soil sample from site 1 was evaluated using soil sample from site 2.

[0008] This invention provides a method for predicting soil organic matter (SOM) content based on spectral coupling effects. By mitigating the coupling effect of soil physical properties on the spectrum, the spatiotemporal transferability of the SOM prediction model is improved. Based on satellite hyperspectral imagery and soil physical variables such as soil moisture (SM), soil surface roughness (root mean square height, RMSH), and soil bulk density (SBW), a soil spectral correction strategy based on information decomposition is established. Two important grain-producing areas in Northeast China are selected as study areas to verify the performance and portability of the spectral correction model and the SOM content prediction model. The results show that the soil spectral correction based on fourth-order polynomials and the XG-Boost algorithm has good accuracy and generalization ability, with residual prediction deviation (RPD) exceeding 1.4 for almost all bands. Furthermore, when the soil spectral correction strategy is adopted, the accuracy of the SOM prediction model and the generalization ability after model transfer are significantly improved. The SOM prediction accuracy based on the XG-Boost corrected spectrum is the highest, with a coefficient of determination (R²) of 1.4. 2 The accuracy of the prediction of the transferred model is 0.76, the root mean square error (RMSE) is 5.74 g / kg, and the RPD is 1.68. 2 The RMSE and RPD were 0.72, 6.71 g / kg, and 1.53, respectively. Compared with the direct migration prediction of the model, the RMSE of the SOM prediction results was reduced by 57.90% and 60.27%, respectively, by adopting a soil spectral correction strategy based on fourth-order polynomials and XG-Boost. The performance comparison shows the advantage of considering soil physical properties in regional-scale SOM prediction. Attached Figure Description

[0009] Figure 1 A flowchart for predicting SOM content using hyperspectral data;

[0010] Figure 2 An overview map of the study area;

[0011] Figure 3A schematic diagram of soil sampling and topsoil parameter measurement;

[0012] Figure 4 The graph shows the correlation coefficients of soil pixel spectra, soil surface spectra, and the reflectance of the two sets of spectra.

[0013] Figure 5 Spectral characteristics of soils with different physical properties;

[0014] Figure 6 To fit soil physical parameters R to soil pixel spectra based on a multi-parameter estimation model 2 ;

[0015] Figure 7 To simulate soil spectra based on empirical equations of SM(a), RMSH(b), and SBW(c);

[0016] Figure 8 To derive the accuracy of the soil spectral correction validation set based on multi-order polynomial coefficient regression;

[0017] Figure 9 To study the accuracy of the validation set for soil spectral correction based on machine learning models;

[0018] Figure 10 The correlation coefficients of two sets of spectral reflectance obtained by using the XG-Boost model and the fourth-order polynomial model to correct the soil pixel spectrum;

[0019] Figure 11 In this context, 'a' represents the Pearson correlation coefficient between SOM content and spectral reflectance in each band; 'b' represents the RMSECV (unit: g / kg) of the multiple regression under different CARS iterations.

[0020] Figure 12 A diagram of the algorithm code for building a SOM prediction model using XG-Boost;

[0021] Figure 13 Scatter plot of SOM content predicted and measured based on four types of spectral data;

[0022] Figure 14 To calibrate soil pixel spectra using the XG-Boost model and the fourth-order polynomial model, and the RPD of the two calibration models;

[0023] Figure 15 The XG-Boost model is built using Site 1 data, where a is based on the original pixel spectrum, b is based on the ground spectrum, c is based on the fourth-order polynomial correction spectrum, and d is a scatter plot of the measured and predicted SOM content from Site 2 data.

[0024] Figure 16In this context, 'a' and 'b' represent the contribution rates of soil properties (SM, RMSH, and SBW) to the estimated SOM bias of Site 1 and Site 2, respectively (where "random" indicates the portion that cannot be explained by these three variables). Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0027] One embodiment of this invention provides a method for predicting soil organic matter content based on spectral coupling effects. Before establishing a high-precision, spatiotemporally transferable SOM content prediction model, a soil spectral correction method was studied to alleviate the coupling effect of surface properties on soil pixel spectra. Figure 1 First, empirical relationships between satellite hyperspectral data and three main soil physical property parameters (SM, RMSH, and SBW) were established using parameter estimation equations. These empirical relationships were then used to correlate soil physical parameters with satellite hyperspectral imagery, resulting in three sets of simulated soil spectral data based on SM, RMSH, and SBW. Next, based on the simulated spectra, soil pixel spectra, and ground spectra, a soil pixel spectral correction model was constructed using multi-order polynomials and various machine learning models to separate soil physicochemical information from the pixel spectral data. Finally, a soil spectral accuracy (SOM) prediction model was constructed using XG-Boost based on the original and corrected soil spectral data. The spatiotemporal transferability of the spectral correction model and the SOM prediction model established based on the soil sample from site 1 was evaluated using soil sample from site 2.

[0028] In the study area, such as Figure 2 As shown (where a represents the geographical locations of Site 1 and Site 2; b represents the soil parameter measurement points of Site 1; c represents the topsoil sampling points of Site 2; d represents the soil surface during the "bare soil period" of Site 1; and e represents the soil surface during the "bare soil period" of Site 2), site 1 is located within the Black Soil Cultivated Land Protection Area in Heilongjiang Province, Northeast China (131°30′–132°03′ E, 46°36′–46°49′ N), covering an area of ​​1095 km². 2The region has a temperate monsoon climate with an annual precipitation of approximately 614 mm. According to the World Reference Base for Soil Resources (WRB), the arable land is primarily black calcareous soil, with a sedimentary layer beneath the topsoil. This layer is sticky, has poor permeability, and frequently forms saucer-shaped pools of water on the surface during rainfall. The dark organic matter in the clay particles leaches downwards, forming a heavy sedimentary layer that enhances surface water retention. The arable land surface is covered with a layer of black humus over 10 cm thick. The soil has extremely high fertility, is rich in organic matter, and is suitable for crop growth.

[0029] Site 2 is located in Changchun, Jilin Province, Northeast China (125°24′-125°43′E, 44°36′-44°46′N), covering an area of ​​713 km². 2 The terrain is flat, with an elevation ranging from 189 to 237 meters. Due to its geographical location and atmospheric circulation, the region has a temperate continental monsoon climate with relatively high precipitation, a frost-free period of approximately 135 days, and an average annual precipitation of about 580 mm. The basin has abundant water resources, resulting in relatively plentiful agricultural water resources, and the SOM content prediction model exhibits strong spatial heterogeneity. The soil in this region is predominantly brown soil with a fertile cultivated layer; the main crops are maize and rice. The soil type, surface characteristics, and other environmental factors at Site 2 differ significantly from those at Site 1, validating the spatiotemporal transferability of the SOM content prediction model in this method.

[0030] Specifically, the method includes the following steps:

[0031] Step 1: Soil sampling and topsoil parameter measurement:

[0032] On October 29, 2022, a total of 104 soil samples were collected at Site 1. Figure 2 b). On April 14, 2023, 40 soil samples were collected at Site 2. Figure 2 c). Eighty soil samples from Site 1 served as the training set for the soil spectral correction model and the SOM prediction model, while the remaining 24 samples served as the validation set. Simultaneously, 40 soil samples from Site 2 were used to verify the spatiotemporal transferability of the spectral correction model and the SOM prediction model. All soil samples were collected from the cultivated land portion of the study area during the "bare soil period." First, a 3D laser scanner (Trimble TX8, maximum standard range: 120 meters; scanning speed: 1 million points / second) was installed at the midpoint of each side of the quadrat to scan the soil surface structure (…). Figure 3a) After scanning, sampling was performed to ensure the natural state of the soil surface structure within the sampling area. Next, nine subsamples were collected from each 30×30m quadrat using a ring scalpel (5cm depth, 200mL volume). The latitude and longitude of the midpoint of the quadrat were recorded using real-time kinematics (RTK) measurement technology.

[0033] After transportation to the laboratory, the SM and SBW of nine subsamples in each quadrat were obtained by weighing and drying, and the average value of the subsamples was calculated to represent the overall level of the quadrat. Then, the nine subsamples were mixed into a composite sample, ground, and sieved to a size ≤0.2 mm for subsequent spectral measurements and SOM content testing. SOM content was determined using the potassium dichromate heating method. Soil spectral reflectance was measured in a darkroom using an ASD Fieldspec4 spectrometer. To ensure the SBW of each sample was identical, soil samples were placed in disposable petri dishes (60 mm in diameter) for spectral measurements. Each soil sample was measured 10 times, and the average value was taken as the soil surface spectral data. The soil surface point cloud data obtained from three-dimensional laser scanning (e.g., ...) was analyzed. Figure 3 (As shown in b) The points are stitched, cut, and filtered to establish a three-dimensional relative coordinate system. The processed point cloud density is greater than 3 points / cm². 3 The relative coordinate system accuracy of the point cloud is less than 2mm. The Z-coordinate of the point cloud data within the sample quadrat is extracted, and the standard deviation is calculated as the RMSH of the quadrat.

[0034] Step 2: Hyperspectral image data acquisition and preprocessing:

[0035] The hyperspectral image data for Resource 1-02D ​​(ZY1-02D) was obtained from the Institute of Aerospace Information, Chinese Academy of Sciences. Based on the soil sampling times for the two regions, images generated on October 29, 2022 (Site 1) and April 14, 2023 (Site 2) were selected as the data source. All images showed cloud cover of less than 1%, consistent with the characteristics of the "bare soil period." The hyperspectral images have a spatial resolution of 30m, with a total of 166 spectral channels, ranging from 400 to 2500 nm (Table 1). The sensor is affected by strong noise in the 400–450 nm and 2460–2500 nm wavelength ranges, and by atmospheric water vapor absorption in the 1290–1408 nm and 1828–1963 nm wavelength ranges. Therefore, the 450–1290 nm, 1408–1828 nm, and 1963–2460 nm bands were selected as the spectral bands. The raw reflectance data was obtained by performing stripe removal, geometric correction, and atmospheric correction on the image using Radiometric Calibration and FLAASH in environmental visualization. The bidirectional reflectance distribution function (BRDF) effect of the image was corrected by calculating the zenith and azimuth angles of the sun (and satellite). A kernel-driven BRDF model was used to normalize the reflectance of ZY1-02D, reducing the influence of observation geometry on reflectance.

[0036] Table 1. Parameters of the hyperspectral camera on the ZY1-02D ​​satellite

[0037]

[0038] Step 3, Spectral Correction Strategy:

[0039] Image pixel spectra comprehensively reflect the physical properties (e.g., SM, RMSH, and SBW) and chemical composition of soil within ground plots. Spectral correction aims to separate reflectance characteristics attributable to soil physicochemical properties from pixel spectral data, thereby mitigating the coupling effect of soil physicochemical properties on the spectrum. First, empirical relationships between satellite hyperspectral data and soil physical parameters SM, RMSH, and SBW are established band-by-band using linear, exponential, power-law, and logarithmic parameter estimation equations. These parameter estimation methods for fitting the relationship between soil physical properties and spectral reflectance have been validated in several studies.

[0040] By correlating soil physical parameters with satellite hyperspectral data using empirical relationships, three sets of simulated soil spectral data based on SM, RMSH, and SBW were obtained. Soil ground spectra measured from dried and ground soil samples were considered "pure spectra" reflecting only soil chemical composition information. Based on this, a spectral correction model was constructed using pixel spectra and the simulated spectra of the three sets of soil physical parameters as input, and ground spectra as the training target. Through multi-order polynomials and various machine learning algorithms, the correction relationship between pixel spectra and ground spectra was established, stripping away reflectance information attributable to soil physical properties from the pixel spectra. The multi-order polynomial equation is as follows:

[0041]

[0042] In the formula, R G R is the ground spectral reflectance in a certain wavelength band. SM R is the spectral reflectance based on SM simulation. RMSH R represents the spectral reflectance based on RMSH simulations. SBW R represents the spectral reflectance based on SBW simulations. P Let be the spectral reflectance of a pixel, i be the polynomial order, and a be the spectral reflectance of a pixel. i b i c i d i and e are the regression coefficients, respectively.

[0043] The machine learning model used:

[0044] (1) Competitive Adaptive Reweighted Sampling (CARS):

[0045] CARS was used to extract sensitive bands corresponding to the Root Mean Square Error (SOM) from hyperspectral data. CARS mimics the "survival of the fittest" principle in Darwinian evolution. Through adaptive weighted sampling, wavelengths with larger absolute coefficients in the PLS model were selected, while wavelengths with smaller weights were removed, resulting in multiple subsets of wavelength variables. Next, cross-validation was used to select the wavelength subset with the smallest root mean square error as the optimal subset. CARS can effectively retain the best wavelength combinations related to the measured characteristics.

[0046] (2) Limiting Gradient Boosting (XG-Boost):

[0047] XG-Boost is an ensemble learning model based on the Boosting strategy, which combines multiple CART trees into a strong learner. As an ensemble algorithm framework, it supports parallel gradient boosting of the base learners, thus significantly improving the training speed of the model. It uses Newton's method to find the extremum of the loss function and expands it into a second-order loss function using Taylor's formula. The loss function is optimized using first-order and second-order gradient functions to reduce model complexity. Simultaneously, regularization reduces the probability of overfitting, significantly improving the model's generalization ability.

[0048] (3) Model validation:

[0049] This method selects the coefficient of determination (R²), root mean square error (RMSE), and residual prediction deviation (RPD) as evaluation indicators, which are expressed as follows:

[0050]

[0051]

[0052] RPD = SD / RMSE (4)

[0053] Where n is the number of samples; y i and Y i These are measured values ​​and predicted values, respectively. SD is the average of the measured values; SD is the standard deviation.

[0054] Step 4: Describe soil physical parameters and SOM content:

[0055] The statistical results of soil physical parameters and SOM content are shown in Table 2. The average SM, RMSH, and SBW of site 1 were 0.25 cm. 3 / cm 3 2.49cm and 0.98g / cm 3 The coefficients of variation (CVs) were 31.99%, 30.92%, and 15.31%, respectively. The moderately high CVs and SDs indicate the combined influence of structural and anthropogenic factors on soil surface physical properties, exhibiting strong spatial heterogeneity. SOM content varied significantly within the range of 25.84–75.97 g / kg, with a standard deviation of 10.51 g / kg and a coefficient of variation of 24.30%. Soil properties at site 2 differed significantly from those at site 1. The average SM, RMSH, and SBW were 0.37 cm⁻¹. 3 / cm 3 3.65cm and 1.13g / cm 3The SOM content at Site 2 was significantly higher than that at Site 1, indicating greater variability. The SOM content at Site 2 ranged from 27.40 to 72.97 g / kg, with an average of 41.57 g / kg, which was lower than that at Site 1.

[0056] Table 2. Soil physical parameters and SOM content statistics in the two locations

[0057]

[0058] Step 5: The influence of soil physical properties on soil spectra:

[0059] To verify the reliability of the ZY1-02D ​​hyperspectral image, the soil pixel spectrum was compared with the soil ground spectrum. Figure 4 Although soil pixel spectra are similar in shape to soil surface spectra, they exhibit some noise and relatively low smoothness, especially in the VNIR wavelength range. Furthermore, the spectral reflectance of soil pixels is slightly lower than laboratory measurements. Spearman correlation coefficients (SCCs) and Pearson correlation coefficients (PCCs) between soil pixel reflectance and soil surface reflectance in each band were calculated. The results show that the PCCs of both sets of spectral data are less than 0.5 in most wavelength ranges, while the SCCs in the visible and shortwave infrared bands are generally greater than 0.5, indicating a possible nonlinear relationship between pixel spectral reflectance and ground spectral reflectance within the same wavelength range. To further reveal the influencing factors of pixel spectra, the differences in soil reflectance under different physical property gradients were compared. With the increase of SM, soil spectral reflectance decreased significantly, especially in the wavelength ranges of 500–1300 nm and 1450–1700 nm. Figure 5 Soil spectral reflectance decreased relatively little with increasing SBW. RMSH had the most significant impact on soil spectra, with reflectance decreasing significantly with increasing RMSH. In summary, the coupling effect of various soil physical properties on the spectrum is a major reason for the discrepancy between the two sets of spectral data, severely limiting the acquisition of "pure spectra" of soil by imaging spectrometers. Therefore, it is necessary to separate soil physicochemical information from pixel spectral data to improve the accuracy of hyperspectral remote sensing SOM prediction.

[0060] Step 6: Empirical relationship between satellite hyperspectral imagery and soil physical properties:

[0061] Regression was performed on empirical coefficients based on field data and soil pixel spectra to determine the relationship between soil reflectance and SM, RMSH, and SBW. Figure 6Among the fitting equations for SM and soil reflectance, the exponential equation showed the best fit. Except for the wavelength range of 2000–2500 nm, the fitting results were good, with R² ranging from 0.49 to 0.68. In the 2000–2500 nm wavelength range, the fit between SM and reflectance was poor, possibly due to the absorption of spectral characteristics caused by hydroxyl groups in the soil by clay minerals. Higher clay mineral content correlated with stronger soil water retention capacity. The results showed that the exponential equation had the best fit in the wavelength range of 450–1800 nm, with R² ranging from 0.50 to 0.69. The power exponential equation had the best fit in the wavelength range of 2000–2500 nm. Across the entire wavelength range, among the three sets of soil physical parameters, RMSH showed the strongest fit with soil reflectance and had the most significant impact on the soil spectrum. Of the four equations, the logarithmic equation showed the best fit, with R² ranging from 0.55 to 0.69. In general, the empirical relationship between soil reflectance and SM is exponential in the range of 450–1800 nm, the empirical relationship with RMSH is logarithmic, and the empirical relationship with SBW is power-law in the range of 2000–2500 nm. Based on the empirical relationship between soil physical parameters and soil spectra, three sets of soil reflectance data were simulated ( Figure 7 Within the 2000–2500 nm range, the soil reflectance simulated by SM showed a basically consistent trend, indicating that the influence of SM on reflectance in this wavelength range was suppressed by other factors, and the spectral characteristics were not obvious. Apart from this, the other simulated soil spectra showed significant differences. A soil spectral correction model was constructed using the soil spectra obtained from simulations based on empirical equations of soil physical properties.

[0062] Step 7, Soil Spectral Correction Model:

[0063] An empirical coefficient model and a machine learning model were used to establish the correction relationship between soil pixel spectra and soil "pure spectra". Using the original pixel spectra and three sets of soil spectra simulated based on SM, RMSH, and SBW as input spectral data, and the subsurface soil spectra as the training target, a soil spectral correction model was constructed band by band. The accuracy of the soil spectral correction model based on multi-order polynomials improved with increasing order, and its RPD and RMSE were optimal for all bands at the fourth order. Figure 8 Excessively high orders can make empirical equations overly complex, leading to overfitting, reduced adaptability to new data, and decreased accuracy. The fourth-order polynomial model has an RPD above 1.5 in all bands, indicating that it has a good correction effect on the soil spectrum.

[0064] Four machine learning algorithms—Support Vector Machine Regression (SVR), Extreme Learning Machine (ELM), Backpropagation Neural Network (BPNN), and XG-Boost—were used to construct soil spectral correction models in the same manner. The optimal soil spectral correction model was determined by comparing the mapping capabilities of different machine learning algorithms to the coupling relationships between multiple soil spectra. Figure 9 The calibration results show significant differences in the accuracy of the four machine learning algorithms for soil spectral calibration. Accuracy fluctuations near the 1000nm wavelength may be caused by other noise in the spectral data. Otherwise, all soil spectral calibration results are good, with overall accuracy higher than the multinomial model. As a representative algorithm of ensemble learning models, XG-Boost shows the best spectral calibration performance, with R0 values ​​across all bands. 2 All scores were above 0.6, and RPD scores were all above 1.6. ELM was the next best, while SVM and BPNN performed poorly.

[0065] Based on the accuracy of soil spectral correction, the soil spectral correction results of the fourth-order polynomial model and the XG-Boost model were further analyzed. The results show that the soil spectrum corrected using the XG-Boost model is smoother and closer to the spectral shape of the soil's "pure spectrum" than that corrected using the fourth-order polynomial model. Figure 10 The correlation coefficients (PCCs) between the corrected soil pixel spectra and the soil "pure spectrum" were calculated. It can be seen that the PCCs for most wavelengths are above 0.8, a significant improvement compared to the correlation between the original pixel spectra and the ground spectrum. Therefore, after spectral correction, the spectral response caused by soil physical properties in the soil pixel spectra is alleviated, and the proportion of soil chemical composition response signals in the pixel spectral data is significantly increased. Regarding the spectral morphology of the soil pixel spectral correction results and their correlation with the soil "pure spectrum," the XG-Boost model's correction results are slightly better than the fourth-order polynomial model. However, the improvement in accuracy of hyperspectral SOM prediction by these methods requires further analysis through modeling.

[0066] Step 8: Prediction accuracy of SOM content based on different spectral data:

[0067] Four types of soil spectral data—pixel spectrum, fourth-order polynomial corrected spectrum, XG-Boost corrected spectrum, and ground-based spectrum—were used to establish SOM content prediction models. To reduce data dimensionality and improve model computational efficiency, feature extraction was first performed on the spectral data. The sensitive bands for SOM were determined using the Pearson correlation coefficient threshold. The correlation coefficient distributions between the four sets of soil spectral data and soil organic matter content showed a relatively consistent trend. Specifically, the correlation coefficient decreased with increasing wavelength before 800 nm and increased with increasing wavelength after 800 nm. Figure 11a) Bands with an absolute correlation coefficient greater than 0.5 were selected as the sensitive bands for SOM. In four spectral datasets—pixel spectrum, fourth-order polynomial-corrected spectrum, XG-Boost corrected spectrum, and ground-based spectrum—the sensitive spectral bands corresponding to SOM were concentrated in 628–1023 nm, 524–1223 nm, 542–1560 nm, and 550–1762 nm, respectively. CARS was used to further extract the optimal feature subset with the least redundant information from the sensitive bands. The optimal number of iterations for CARS was determined by RMSECV of multivariate regression. Figure 11 b). The bands listed in Table 3 are the spectral bands selected by CARS for SOM inversion modeling and validation analysis.

[0068] Table 3. Feature frequency band statistics based on CARS

[0069]

[0070] Four spectral bands selected by CARS and SOM content were used as input data for the model. The XG-Boost algorithm was employed to construct an SOM prediction model. Figure 12 The results show that these two spectral correction methods significantly improve the prediction accuracy of SOM based on the original pixel spectra. Among all SOM prediction results, the prediction accuracy of ground spectral data is the highest. RMSEP and RPD were 0.79, 4.89 g / kg and 1.97 g / kg, respectively. Figure 13 When the raw pixel spectral data is used as model input data, based on The prediction accuracy of the assessment is 0.64. A soil spectral correction strategy based on a fourth-order polynomial is employed, resulting in a prediction accuracy of... The RMSEP decreased by 2.28 g / kg, and the RPD increased by 0.38. The soil spectral correction strategy based on the XG-Boost model significantly improved the prediction accuracy of SOM. The accuracy of the corrected spectrum in predicting soil molecular structure (SOM) was close to that of the ground spectrum, indicating that mitigating the coupling effect of soil properties on soil pixel spectra can effectively improve the prediction accuracy of hyperspectral SOM.

[0071] Regarding the transferability between the soil spectral correction model and the SOM prediction model:

[0072] Soil pixel spectral correction methods based on empirical coefficients and machine learning models offer a new strategy for improving the accuracy of SOM prediction based on hyperspectral imagery. Two soil spectral correction methods based on different models each have their advantages and disadvantages. The XG-Boost-based method significantly improves SOM prediction accuracy, but its correction process and principle are difficult to express mathematically. While the fourth-order polynomial model-based method shows a weaker improvement in SOM prediction accuracy, expressing the improved method using coefficient equations makes it more suitable for generalization. High transferability of the soil spectral correction method is a key prerequisite for constructing a SOM prediction model with strong generalization ability. To verify its spatiotemporal transferability, 40 sets of soil pixel spectra from Site 2 and ground experimental data were imported into two spectral correction models. The spectral correction results show that the corrected soil spectrum closely matches the shape of the soil's "pure spectrum" (…). Figure 14 Compared to the soil spectrum corrected by the fourth-order polynomial, the soil spectrum corrected by the XG-Boost model is smoother. From the accuracy of the model transfer experiment results, the soil spectral correction model based on XG-Boost has good transfer performance, with RPD above 1.4 for each band. Machine learning algorithms, represented by XG-Boost, are significantly superior to coefficient models in terms of computational ability to establish coupling relationships between multiple soil spectra and adaptability to new data. This is because this ensemble learning model comprehensively utilizes all feature values ​​of each soil sample point and iteratively adjusts the tree weights to explore the optimal solution for the coupling relationship between soil physical properties and pixel spectra.

[0073] The poor spatiotemporal transferability of traditional SOM prediction models is mainly due to their poor applicability to different spatiotemporal spectral data. Evaluating the effect of soil spectral correction methods on improving the spatiotemporal transferability of SOM prediction models is a direct basis for proving the effectiveness of spectral correction methods. The transferability of the SOM prediction model established using Site 1 data was evaluated using soil samples and spectral data from Site 2. The SOM prediction model based on ground spectra showed the best transferability, with an RMSEP increase of only 0.39 g / kg. Figure 15However, the transferability of the SOM prediction model based on the original pixel spectrum is very poor because changes in surface physical properties lead to deviations in spectral reflectance. Transferability verification showed an increase in RMSEP of 8.05 g / kg and a decrease in RDP of 44.04%. Two soil spectral correction strategies significantly improved the transferability of the prediction model based on the original pixel spectrum, with RPD values ​​exceeding 1.4 for both models. The SOM prediction model based on the XG-Boost corrected spectrum exhibited greater portability. Compared to the transferability verification of the pixel-based model, RMSEP decreased by 60.27%, and RPD increased by 150.82%. These results demonstrate the effectiveness of the soil spectral correction method and the feasibility of predicting soil organic matter content using corrected satellite hyperspectral data. The soil organic matter prediction model based on corrected satellite hyperspectral data can be used at two sites with different soil types, soil physical properties, soil organic matter content, and spatiotemporal characteristics. The core of this soil spectral correction method is to comprehensively consider the coupling effect of various soil properties on spectral reflectance to restore the true spectral characteristics of the research target. For different research objectives, the main factors influencing the target's spectral response can be analyzed based on the actual environment and imaging conditions. Therefore, this method is not limited to SOM prediction and can provide valuable insights for soil property prediction based on satellite hyperspectral data.

[0074] The influence of soil physical properties on the prediction bias of SOM content:

[0075] The results show that the soil spectral correction method can significantly improve the prediction accuracy of soil organic matter content and the spatiotemporal transferability of pixel spectra. In other words, soil properties (SM, RMSH, and SBW) are likely the main factors contributing to the SOM estimation error based on the original pixel spectra. The error dependencies of the original pixel spectra and two sets of corrected spectral data on SM, RMSH, and SBW were studied, and their contributions to the SOM content prediction bias were estimated using a stepwise regression method. The results show that the cumulative contribution rate of these three soil properties to the SOM prediction results based on the original pixel spectra is greater than 70%. Figure 16Therefore, soil physical properties are the main source of error in SOM prediction. SM contributes the most to SOM bias, followed by RMSH and SBW. This is related to the fact that the pixel spectrum responds most significantly to SM within the sensitive wavelength range of SOM, and may also be related to the stronger spatial heterogeneity of SM compared to RMSH and SBW. The stronger the spatial heterogeneity of soil physical properties, the greater the difference in their influence on pixel spectra, and the greater the SOM prediction bias. The spectral correction strategy significantly reduced the bias of SM, RMSH, and SBW in SOM prediction. Soil spectral correction based on XG-Boost significantly reduced the SOM prediction bias caused by soil physical properties compared to soil spectral correction based on fourth-order polynomials. This result fundamentally explains the higher prediction accuracy and stronger spatiotemporal transferability of the SOM prediction model based on XG-Boost corrected spectra. Although the random error increased relatively, the accuracy of the total bias prediction result decreased significantly. Therefore, the spectral correction strategy did not introduce more error sources, but only increased the relative contribution of other error factors to the SOM prediction bias, such as the uncertainty of hyperspectral image processing and the uncertainty of field data acquisition. From the relative contribution rates of soil physical properties to SOM prediction bias before and after spectral correction, SM showed the largest decrease in contribution rate, exceeding 10%, followed by RMSH. The spectral correction methods based on polynomials and XG-Boost reduced the average relative contribution rate of RMSH at the two sites by 10% and 14.5%, respectively. Although SBW had the smallest contribution rate to reducing SOM prediction bias, the decrease exceeded 6%. By comparing the improvement effects of different input variables on SOM prediction bias, the improvement effect of soil physical properties on the accuracy of SOM content prediction was SM > RMSH > SBW, which is consistent with the spatial heterogeneity of the three soil physical properties within the VNIR range and the order of soil spectral sensitivity to them. Therefore, soil physical properties with strong spatial heterogeneity and sensitive spectral response should be given priority in soil spectral correction.

[0076] Potential and limitations of soil spectral correction models:

[0077] The soil spectral correction method is designed to address the coupling effect of surface physical properties on hyperspectral imagery and is applicable to remote sensing image processing for predicting various soil chemical compositions. This method suppresses the sensitivity of spectral data to SM, RMSH, and SBW, reducing the possibility of SOM prediction results getting trapped in local optima. Another advantage is the establishment of empirical relationships between SM, RMSH, and SBW and the hyperspectral soil reflectance from the ZY1-02D ​​satellite spectrum, improving the method's generalization ability and application efficiency. Furthermore, this method has two potential applications: 1) it enables the integration of optical and radar remote sensing for soil physicochemical property estimation; 2) it addresses the problem of spatiotemporal heterogeneity of spectral data caused by uncertainties in surface physical conditions during multi-source remote sensing data fusion. Although the soil spectral correction model can recover most of the "pure spectral" characteristics of soil, some uncertainties remain. The applicability of this method on airborne hyperspectral sensors or other hyperspectral satellites requires further evaluation with more data. Moreover, this experiment only considered the influence of soil properties within the top 5 cm on the spectrum, while the spectrum may have different sensing depths for different soil properties. Although spectroscopy can only directly detect SM changes in the shallow soil layer (approximately 0–2 cm), this depth varies under different SBW and RMSH conditions. The vertical heterogeneity of SM and SBW is likely a major factor contributing to soil spectral correction errors. Employing a stratified strategy to establish spectral correction models for soil properties at different depths, or assigning different weights to soil physical properties at different depths, can maximize the model's effectiveness within a specified depth and range.

[0078] This method, for the first time, utilizes "pure spectra" obtained by soil spectral correction considering SM, RMSH, and SBW for SOM content prediction, demonstrating its good spatiotemporal transferability. Strategies to mitigate the influence of soil physical properties on spectral coupling can provide a paradigm for future remote sensing-based soil element content prediction. Soil spectral correction of the entire hyperspectral image requires real-time, high-resolution regional-scale soil physical parameter data. The high sensitivity of synthetic aperture radar (SAR) remote sensing to soil physical parameters allows for the acquisition of rich information in practical applications. Future research could combine the advantages of hyperspectral imaging and radar remote sensing to improve the prediction accuracy of soil physicochemical parameters. Considering the heterogeneity in type and dimension of optical and radar data, a new approach combining the two types of data is proposed. Meanwhile, due to the time difference between radar and hyperspectral imaging, corresponding surface physical properties also change, especially SM, which has a significant impact on weather. Therefore, combining optical and radar data to eliminate the bias caused by temporal phase may be the best strategy to solve this problem. Furthermore, since radar sensors have better penetration than optical sensors, quantifying the vertical heterogeneity of SM and SBW using radar remote sensing may be key to further improving the accuracy of soil spectral correction. These strategies reduce errors in large-scale soil physical and chemical parameter surveys to support regional strategic arrangements for sustainable agricultural development.

[0079] In summary, soil spectral correction models based on fourth-order polynomials and XG-Boost were constructed using satellite and ground-based hyperspectral data, as well as soil physical property data, to mitigate the coupling effect of soil physical properties on pixel spectra. Data from two sites were used to evaluate the performance of the soil spectral correction models and their impact on the accuracy and spatiotemporal transferability of SOM prediction models. The main conclusions are as follows:

[0080] (1) The spectral reflectance of soil pixels and the spectral reflectance of soil surface exhibit a nonlinear relationship. Differences in surface properties are the main factor causing the deviation between the two spectral data. RMSH has the most significant impact on soil pixel spectra, followed by SM, and then SBW.

[0081] (2) The fourth-order polynomial model and the XG-Boost model have good soil spectral correction accuracy. The soil spectral correction model based on XG-Boost takes into account all features, continuously adjusts the tree weights, avoids the results from getting trapped in local optima, and has higher accuracy and stronger spatiotemporal transferability.

[0082] (3) Soil spectral correction significantly alleviated the coupling effect of soil physical properties on soil pixel spectra, effectively improving the accuracy of the SOM prediction model. More importantly, it greatly enhanced the spatiotemporal transferability of the pixel-based SOM prediction model. Considering the response of satellite hyperspectral imaging to soil physical properties helps to understand its role in predicting soil organic matter content. This work provides a new research paradigm for the prediction of soil property parameters in other regions.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting soil organic matter content based on spectral coupling effect, characterized in that, Includes the following steps: Step 1: Establish an empirical relationship between satellite hyperspectral data and three soil physical parameters SM, RMSH, and SBW using parameter estimation equations; Step 2: Using empirical relationships, soil physical parameters are correlated with satellite hyperspectral images to obtain three sets of simulated soil spectral data based on SM, RMSH, and SBW. Step 3: Based on simulated spectra, soil pixel spectra, and ground spectra, construct a soil pixel spectral correction model using multi-order polynomials and machine learning models to separate soil physicochemical information from pixel spectral data; Step 4: Based on the original soil spectral data and the corrected soil spectral data, construct the SOM prediction model using XG-Boost; In step 3, the multi-order polynomial equation is: （1）； In the formula, R G R is the ground spectral reflectance in a certain wavelength band. SM R is the spectral reflectance based on SM simulation. RMSH R represents the spectral reflectance based on RMSH simulations. SBW R represents the spectral reflectance based on SBW simulations. P Let be the spectral reflectance of a pixel, i be the polynomial order, and a be the spectral reflectance of a pixel. i b i c i d i and e are the regression coefficients, respectively; In step 3, the machine learning models used include CARS and XG-Boost; The coefficient of determination R², root mean square error (RMSE), and residual prediction deviation (RPD) were selected as evaluation indicators, and are expressed as follows: （2）； （3）； （4）； in, It is the number of samples; and These are measured values ​​and predicted values, respectively. It is the average of the measured values; It is the standard deviation.

2. The method for predicting soil organic matter content based on spectral coupling effect according to claim 1, characterized in that, In step 1, soil sampling and topsoil parameter measurement include the following specific steps: Step 1.1: Use a 3D laser scanner, install it at the midpoint of each side of the sample plot, scan the soil surface structure, and take samples after scanning to ensure the natural state of the soil surface structure in the sampling area. Step 1.2: In each 30 × 30 m quadrat, collect 9 subsamples using a ring cutter, and record the latitude and longitude of the midpoint of the quadrat using real-time kinematic measurement technology; Step 1.3: After transportation to the laboratory, the SM and SBW of the nine subsamples in each quadrat were obtained by weighing and drying, and the average value of the subsamples was calculated to represent the overall level of the quadrat. Then, the nine subsamples were mixed into a composite sample, ground and sieved to a size ≤0.2 mm. The SOM content was determined by the potassium dichromate heating method, and the soil spectral reflectance was measured in a dark room using an ASD Fieldspec4 spectrometer. The soil samples were placed in disposable petri dishes for spectral measurement. Each soil sample was measured 10 times, and the average value was taken as the soil ground spectral data. Step 1.4: The soil surface point cloud data obtained from 3D laser scanning is stitched, cut, and filtered to establish a 3D relative coordinate system. The processed point cloud density is greater than 3 points / cm². 3 The relative coordinate system accuracy of the point cloud is less than 2mm. The Z coordinate of the point cloud data within the sample quadrat is extracted, and the standard deviation is calculated as the RMSH of the quadrat.

Images