Method and system for soil total nitrogen anti-cloud and fog remote sensing mapping for continuous cloud and fog area

Abstract

The application provides a soil total nitrogen anti-cloud and mist remote sensing mapping method and system for a continuous cloud and mist area, and belongs to the field of remote sensing monitoring and soil resource investigation. The method comprises the following steps: acquiring synthetic aperture radar images, multispectral images and digital elevation data of a target area; constructing a spectrum, SAR and terrain covariant candidate library after pretreatment; combining soil total nitrogen measured values, performing initial screening through a Pearson correlation coefficient and removing collinearity through a variance inflation factor, to obtain an optimal spectrum-SAR-terrain SST covariant combination; taking the combination as input characteristics, training an inversion model by using an extreme gradient boosting algorithm; and inputting full-coverage SST grid data of the target area into the model, to generate a soil total nitrogen spatial distribution map through pixel-by-pixel prediction. The application effectively overcomes the limitation of low availability of optical images in a continuous cloud and mist area, realizes high-precision and anti-interference soil total nitrogen remote sensing mapping, and has cross-region migration applicability, thereby providing a scientific basis for precision agriculture management.

Description

Technical Field

[0001] This invention relates to the field of information technology, and more specifically, to a method and system for remote sensing mapping of soil total nitrogen resistance to fog in continuous foggy areas. Background Technology

[0002] Soil total nitrogen is a crucial indicator for assessing soil fertility and guiding precision agricultural management, and remote sensing technology provides an efficient means for spatial mapping of its distribution. Currently, soil mapping techniques based on optical remote sensing imagery are widely used, typically employing spectral indices to construct inversion models. However, in areas with continuous cloud and fog, optical images are easily affected by cloud cover and fog shading, resulting in low image usability and significant data gaps. This makes it difficult to guarantee the accuracy and applicability of traditional optical remote sensing-based mapping methods. Therefore, there is an urgent need to develop a remote sensing method that can overcome cloud and fog interference and achieve high-precision, stable mapping even under continuous cloud and fog conditions. Summary of the Invention

[0003] In view of the above problems, the present invention provides a method for remote sensing mapping of total nitrogen in soil to resist fog in areas of continuous cloud and fog, the method comprising:

[0004] Acquire multi-source remote sensing data and digital elevation data of the target area, wherein the multi-source remote sensing data includes synthetic aperture radar image data and multispectral image data;

[0005] The acquired raw data are preprocessed, including geometric correction, radiometric correction, cropping, resampling and spatial registration, so that the data are unified to the same spatial resolution and geographic coordinate system.

[0006] A candidate library of spectral covariates is constructed based on the preprocessed multispectral image data; a candidate library of synthetic aperture radar covariates is constructed based on the preprocessed synthetic aperture radar image data; and a candidate library of topographic covariates is constructed based on the digital elevation data.

[0007] Obtain the measured values ​​of total nitrogen in the soil from multiple sampling points in the target area;

[0008] The correlation coefficient between each candidate covariate and the measured value of total nitrogen in the soil was calculated. The candidate covariates were initially screened based on the magnitude of the correlation coefficient, and the candidate covariates with strong correlation to total nitrogen in the soil were retained.

[0009] The variance inflation factor is calculated for all candidate covariates retained in the initial screening. Candidate covariates with multicollinearity are successively eliminated based on the condition that the variance inflation factor is greater than a set threshold, until the variance inflation factor of the remaining candidate covariates is not greater than the set threshold, thus obtaining the optimal combination of spectral-synthetic aperture radar-topography covariates.

[0010] The soil total nitrogen inversion model is trained using the optimal combination of spectral-synthetic aperture radar-topographic covariates at each sampling point as input features and the measured value of soil total nitrogen at the corresponding sampling point as target variable.

[0011] The raster data of the optimal spectral-synthetic aperture radar-topographic covariate combination covering the entire target area is input into the trained soil total nitrogen inversion model, and the predicted value of soil total nitrogen is calculated pixel by pixel to generate a spatial distribution map of soil total nitrogen in the target area.

[0012] In addition, the present invention provides a soil total nitrogen resistance remote sensing mapping system for continuous cloud and fog areas, which is used to store and execute the above-described methods.

[0013] Based on the above, the present invention provides a practical and effective solution and achieves the following significant technical effects:

[0014] First, this invention fully utilizes the characteristics of synthetic aperture radar (SAR) signals, which are not obstructed by clouds and have all-weather penetration capabilities, and coordinates them with spectral data affected by clouds and fog and digital elevation data that does not rely on optical signals to construct a spectral-SAR-topography (SST) multi-source covariate system. This effectively makes up for the perception shortcomings of relying solely on optical remote sensing, fundamentally overcomes the technical bottleneck of low image availability and severe data loss in continuous cloud and fog areas, and realizes stable mapping under cloud and fog conditions.

[0015] Secondly, by employing the PCC-VIF two-step screening method, this invention effectively eliminates multicollinearity and information redundancy among covariates while evaluating the correlation between various environmental covariates and total soil nitrogen. This significantly reduces the dimensionality of the model input features and substantially improves the fitting efficiency and numerical stability of the subsequent machine learning model. Simultaneously, the three types of covariates achieve information complementarity from multiple dimensions, including vegetation growth status, surface structure characteristics, and topographic environmental control, comprehensively characterizing the spatial distribution pattern of total soil nitrogen.

[0016] Third, the Extreme Gradient Boosting (XGBoost) algorithm is employed. By iteratively integrating weak learners and explicitly introducing L1 and L2 regularization terms, it accurately captures the complex nonlinear mapping relationship between multidimensional environmental covariates and total soil nitrogen, effectively preventing model overfitting. Validated by examples, the model's prediction accuracy significantly outperforms conventional algorithms such as partial least squares regression, support vector regression, and random forests, greatly improving inversion accuracy and generalization ability.

[0017] Fourth, the mapping framework of this invention has good cross-regional adaptability and maintains high prediction accuracy in multiple different geographical regions. It can provide reliable and efficient technical support for soil fertility evaluation, precision nutrient management and sustainable agricultural development in areas with continuous cloud cover. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method provided by the present invention; Figure 2 The above are the frequency distribution histograms and box plots of total nitrogen content in soil from the total sample set (n=425) in this embodiment of the invention. Figure 3 The above are the frequency distribution histograms and box plots of total nitrogen content in soil from the calibration sample set (n=340) in this embodiment of the invention. Figure 4 This is a histogram and box plot showing the frequency distribution of total nitrogen content in soil of the verification sample set (n=85) in this embodiment of the invention. Figure 5 The correlation coefficient of the topographic index; Figure 6 The variance inflation factor (VIF) diagnostic results of the spectral, SAR, and topographic covariates retained after initial screening by PCC in this embodiment of the invention; Figure 7 This is a scatter plot comparing the estimated and measured values ​​of total nitrogen in soil using the XGBoost model with only spectral covariates as input in this embodiment of the invention. Figure 8 This is a scatter plot comparing the estimated and measured values ​​of total nitrogen in soil using the XGBoost model with radar (SAR) covariates as input only, as an embodiment of the present invention. Figure 9 This is a scatter plot comparing the estimated and measured values ​​of total soil nitrogen in the XGBoost model with only topographic covariates as input in this embodiment of the invention. Figure 10 This is a scatter plot comparing the estimated and measured values ​​of total nitrogen in soil using the XGBoost model with the spectral-radar-topographic SST covariate as input in an embodiment of the present invention. Figure 11 This is a schematic diagram of the soil total nitrogen remote sensing mapping model based on spectral and radar (SAR) covariates in an embodiment of the present invention. Figure 12 This is a schematic diagram of the soil total nitrogen remote sensing mapping model based on spectral and topographic covariates in an embodiment of the present invention; Figure 13 This is a schematic diagram of the soil total nitrogen remote sensing mapping model based on radar and topographic covariates in an embodiment of the present invention; Figure 14 This is a beehive diagram of the optimal SST covariates for the SHAP values ​​of the XGBoost soil total nitrogen inversion model in this embodiment of the invention. Figure 15 This is a chart showing the sorting of the average absolute SHAP values ​​of each SST covariate and the percentage of their contribution from different sources in this embodiment of the invention. Detailed Implementation

[0019] like Figures 1-15 As shown in the figure. This technical solution provides a method and system for remote sensing mapping of soil total nitrogen resistance to fog in continuous foggy areas. The technical details are explained in detail below with reference to a specific application scenario.

[0020] Continuous fog regions refer to areas where dense fog occurs frequently or for extended periods near the ground, while cloud cover exists in the mid-to-high atmosphere. These regions are commonly found in coastal areas, mountainous areas, and areas with high forest cover. These areas are influenced by a combination of factors, including topographic relief, ocean-land moisture transport, vegetation evapotranspiration, and local circulation, resulting in frequent and prolonged fog occurrences and limited opportunities for clear-sky remote sensing observations.

[0021] In areas with continuous cloud cover and fog, optical remote sensing sensors are easily affected by cloud cover and fog, resulting in scattering and absorption, making it difficult to reliably acquire effective surface reflectance information. This leads to problems such as low image availability, insufficient temporal phase, and reduced mapping accuracy in soil total nitrogen retrieval models built based on clear-sky optical imagery. This invention utilizes synthetic aperture radar remote sensing data, which is less affected by cloud cover and fog, and combines multispectral optical data with topographic data to construct a multi-source covariate system. A limiting gradient boosting algorithm is then used to establish the retrieval model, thereby improving the applicability of remote sensing retrieval and mapping of soil total nitrogen in areas with continuous cloud cover and fog.

[0022] The total nitrogen in soil as described in this invention refers to the sum of all forms of nitrogen in the soil, measured by mass ratio in grams per kilogram. A continuous cloud and fog region refers to a geographical area where dense fog frequently occurs near the ground and is covered by clouds in the mid-to-high altitudes. Cloud and fog resistant remote sensing mapping refers to the use of synthetic aperture radar signals with cloud and fog penetration capabilities, in conjunction with optical signals and topographic signals, to invert and map the spatial distribution of total nitrogen in soil under cloud and fog conditions.

[0023] This invention provides a method for remote sensing mapping of total nitrogen in soil to resist fog in areas with continuous cloud cover, comprising the following steps.

[0024] Step S1: Acquisition and Preprocessing of Multi-Source Remote Sensing Data Physical and Logical Structure: This step involves three types of raw data sources. The first type is synthetic aperture radar (SAR) imagery data, using C-band dual-polarization SAR imagery from the Sentinel-1 satellite. Polarization methods include co-polarization (VV) and cross-polarization (VH), with a revisit period of 6 days and an initial spatial resolution of 5 meters in the range direction and 20 meters in the azimuth direction. The second type is multispectral imagery data, using multispectral imagery from the Sentinel-2 satellite, containing 13 spectral bands covering the spectral range from 400 nm to 2500 nm, encompassing the visible, near-infrared, and shortwave infrared regions. The revisit period is 5 days. Bands 2, 3, 4, and 8 have a spatial resolution of 10 meters; bands 5, 6, 7, 8A, 11, and 12 have a spatial resolution of 20 meters; and bands 1, 9, and 10 have a spatial resolution of 60 meters. The third type is digital elevation data, using ASTER global digital elevation model data, with a spatial resolution of 30 meters.

[0025] Connections and signal flow: The above three types of data are input into the data preprocessing flow. The preprocessed data is then uniformly output to the covariate construction flow.

[0026] The preprocessing workflow is as follows. For synthetic aperture radar (SAR) image data, orbit correction, radiometric calibration, speckle filtering, terrain correction, and conversion to decibel values ​​are performed sequentially. The speckle filtering uses a Refined Lee filter with a 7×7 filter window. This generates a backscattering coefficient image with a pixel size of 30m×30m, expressed in decibels. Finally, the image is cropped and mosaicked.

[0027] For multispectral imagery data, Level-2A data with cloud cover less than 5% and approximately synchronized with soil sampling time were selected. During preprocessing, bands 1, 9, and 10, which are associated with water vapor noise, were removed, and the remaining bands were resampled to a spatial resolution of 30 meters.

[0028] For digital elevation data, terrain-derived indices were calculated using SAGA GIS software version 7.8.2.

[0029] The above preprocessing unifies all data to the same spatial resolution of 30m×30m and the same geographic coordinate system, eliminating spatial misalignment and scale differences between multi-source data, and providing a spatially consistent data foundation for subsequent covariate construction.

[0030] Step S2: Construction of a candidate library of spectral covariates Physical and logical structure: This step constructs a candidate library of spectral covariates based on the preprocessed multispectral image data. The candidate library includes reflectance values ​​for each band of the multispectral image, as well as 15 vegetation spectral indices calculated based on the multispectral image.

[0031] The spectral indices of the 15 vegetation types are calculated using the following formula.

[0032] The formula for calculating the Normalized Difference Vegetation Index (NDVI) is:

[0033]

[0034] B8 represents the near-infrared reflectance, and B4 represents the red light reflectance. This index reflects vegetation greenness and growth vitality.

[0035] The formula for calculating the first normalized vegetation index (NDVI1) is:

[0036]

[0037] B8A represents the narrow near-infrared reflectance. This index uses a narrow near-infrared band instead of the standard near-infrared band, making it more sensitive to the response of vegetation canopy structure.

[0038] The formula for calculating the first normalized difference vegetation index, red-edge NDVIRE1, is as follows:

[0039]

[0040] B5 represents the red-edge band reflectance. This index utilizes the high sensitivity of the red-edge band to vegetation chlorophyll content to enhance the monitoring ability of vegetation physiological status.

[0041] The formula for calculating the second normalized difference vegetation index, red-edge NDVIRE2, is as follows:

[0042]

[0043] B6 represents the reflectivity of the red-edge band 2.

[0044] The formula for calculating the third normalized difference vegetation index, red-edge NDVIRE3, is as follows:

[0045]

[0046] B7 represents the red-edge 3-band reflectivity.

[0047] The formula for calculating the Enhanced Vegetation Index (EVI) is as follows:

[0048]

[0049] B2 represents the blue light reflectance. This index corrects for the effects of atmospheric aerosols by incorporating the blue light band and avoids signal saturation issues in areas with high vegetation cover.

[0050] The formula for calculating the Difference Vegetation Index (DVI) is as follows:

[0051]

[0052] This index reflects differences in vegetation density by measuring the difference between the near-infrared and red light bands.

[0053] The formula for calculating the simple ratio index SR is:

[0054]

[0055] This index enhances vegetation signals by using the ratio of near-infrared to red light bands and is sensitive to chlorophyll content.

[0056] The formula for calculating the luminance index (BI) is:

[0057]

[0058] This index integrates information from red and near-infrared bands to reflect the overall brightness characteristics of the Earth's surface.

[0059] The formula for calculating the Soil-Regulated Vegetation Index (SAVI) is as follows:

[0060]

[0061] This index reduces the impact of soil background on the vegetation index by introducing a soil adjustment factor of 0.5.

[0062] The formula for calculating the soil-adjusted vegetation index (OSAVI) is as follows:

[0063]

[0064] The index uses an optimized soil modifier of 0.16 to further reduce background soil disturbance.

[0065] The formula for calculating the Soil-Regulated Vegetation Index (MSAVI) has been modified as follows:

[0066]

[0067] This index automatically optimizes soil correction intensity under different vegetation cover by using an adaptive soil adjustment factor.

[0068] The formula for calculating the Green Normalized Difference Vegetation Index (GNDVI) is as follows:

[0069]

[0070] B3 represents the green light reflectance. This index uses the green light band instead of the red light band, making it more sensitive to changes in chlorophyll content.

[0071] The formula for calculating the Normalized Green Index (NDGI) is as follows:

[0072]

[0073] This index reflects vegetation greenness through the normalized difference between the green and red light bands.

[0074] The formula for calculating the Cosmic Spectral Response Index (COSRI) is as follows:

[0075]

[0076] This index integrates information from four bands: blue, green, red, and near-infrared, reflecting the comprehensive spectral characteristics of vegetation and soil.

[0077] The aforementioned spectral covariates describe the growth status, greenness, cover, and soil background characteristics of surface vegetation from different perspectives. In areas with continuous cloud cover, the quality of some spectral bands deteriorates when optical imagery is affected by clouds and fog, limiting the accuracy of soil total nitrogen inversion relying solely on spectral covariates. Therefore, it is necessary to introduce synthetic aperture radar covariates, which are unaffected by clouds and fog, as a supplement.

[0078] Step S3: Construction of the Synthetic Aperture Radar Covariate Candidate Library

[0079] Physical and Logical Structure: This step constructs a candidate library of synthetic aperture radar (SAR) covariates based on preprocessed SAR image data. First, four polarization combination parameters are constructed based on the backscattering coefficients of co-polarized (VV) and cross-polarized (VH). Then, the gray-level co-occurrence matrix method is used to extract texture parameters from each polarization combination image, resulting in a total of 20 SAR covariates. These four polarization combination parameters (VHplusVV, VHdiffVV, VHmulVV, VHdivVV), together with the original co-polarized signal (VV) and cross-polarized signal (VH), constitute the basic set of polarization parameters for the SAR covariate candidate library. VV represents the vertically polarized co-polarized backscattering signal, and VH represents the horizontally transmitted and vertically received cross-polarized backscattering signal; both directly reflect the backscattering intensity of ground objects under the two polarization modes. The four polarization combination parameters derived from VH and VV further characterize the nonlinear interactions between polarizations.

[0080] The four polarization combination parameters are constructed as follows:

[0081] VHplusVV: Represents the sum of VH polarization and VV polarization, reflecting the total backscattering intensity.

[0082] VHdiffVV: Represents the difference between VH polarization and VV polarization, reflecting the scattering difference between the polarizations.

[0083] VHmulVV: Represents the product of VH polarization and VV polarization, reflecting the joint scattering characteristics between polarizations.

[0084] VHdivVV: This represents the quotient of VH polarization and VV polarization, reflecting the relative change in the polarization ratio.

[0085] For the above polarization combination images, seven texture parameters were extracted using the gray-level co-occurrence matrix (GLCM) method. The GLCM describes the joint distribution probability of gray levels between two pixels separated by a certain distance and direction in an image. Let the GLCM be... ,in and These represent the gray levels of two pixels respectively. Given the total number of gray levels, the calculation formulas for each texture parameter are as follows.

[0086] Contrast: Reflects the sharpness of an image and the depth of its texture grooves. The formula for calculation is:

[0087]

[0088] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels. A higher contrast value indicates deeper image texture and higher clarity.

[0089] Dissimilarity: Reflects the degree of local variation in texture; the calculation formula is:

[0090]

[0091] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels. The larger the dissimilarity value, the more drastic the local changes in the texture.

[0092] Homogeneity: Reflects the uniformity of texture; the formula for calculation is:

[0093]

[0094] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels. A higher uniformity value indicates a more uniform texture.

[0095] Entropy: Reflects the randomness and complexity of a texture; the formula for calculation is:

[0096]

[0097] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels. The higher the entropy value, the greater the randomness and complexity of the texture.

[0098] Mean: Reflects the average gray level of the texture, calculated using the following formula:

[0099]

[0100] in, The gray level of the first pixel. Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence This represents the total number of gray levels. The mean value reflects the average intensity of gray levels within the texture region.

[0101] Variance: Reflects the degree of gray-level dispersion in a texture; the calculation formula is:

[0102]

[0103] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence The gray level of the first pixel. for The mean, This represents the total number of gray levels. The larger the variance value, the higher the degree of gray level dispersion in the texture.

[0104] Correlation: Reflects the linear dependence of textures; the calculation formula is:

[0105]

[0106] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. for The mean, for The mean, for variance for variance This represents the total number of gray levels. A higher correlation value indicates a stronger linear dependence of the texture.

[0107] For the same polarization (VV) image, the gray-level co-occurrence matrix method was also used to extract seven texture parameters. Since a VV polarization image was used, the gray-level variables are denoted as follows: and The calculation formulas for each texture parameter are as follows:

[0108] VV polarization contrast (VV_CON) reflects the sharpness of an image and the depth of its texture grooves. The calculation formula is as follows:

[0109]

[0110] in, For the gray level pairs in the gray-level co-occurrence matrix of VV polarization image The joint probability distribution of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels.

[0111] VV polarization anisotropy (VV_DIS) reflects the degree of local variation in texture, and is calculated using the following formula:

[0112]

[0113] in, For the gray level pairs in the gray-level co-occurrence matrix of VV polarization image The joint probability distribution of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels.

[0114] VV polarization uniformity (VV_HOM) reflects the uniformity of the texture, and is calculated using the following formula:

[0115]

[0116] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels.

[0117] The VV polarization entropy value (VV_ENT) reflects the randomness and complexity of the texture, and is calculated using the following formula:

[0118]

[0119] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. This represents the total number of gray levels.

[0120] VV polarization mean (VV_MEA) reflects the average gray level of the texture, and is calculated using the following formula:

[0121]

[0122] in, The gray level of the first pixel. Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence This represents the total number of gray levels.

[0123] VV polarization variance (VV_VAR) reflects the degree of grayscale dispersion of the texture, and is calculated using the following formula:

[0124]

[0125] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence The gray level of the first pixel. for The mean, This represents the total number of gray levels.

[0126] VV polarization correlation (VV_COR) reflects the linear dependence of texture, and is calculated using the following formula:

[0127]

[0128] in, Gray level pairs in the gray-level co-occurrence matrix The probability of occurrence and These represent the gray levels of two pixels respectively. for The mean, for The mean, for variance for variance This represents the total number of gray levels.

[0129] The four polarization combination parameters mentioned above, along with the seven texture parameters for each polarization combination, constitute the candidate library of synthetic aperture radar (SAR) covariates. SAR signals can penetrate clouds and part of the vegetation canopy, are unaffected by weather conditions, and their backscattering coefficient is sensitive to the vertical structure of the vegetation canopy and surface roughness. In areas with continuous cloud cover, SAR covariates can effectively compensate for the lack of surface information perception when optical signals are absent.

[0130] Step S4: Construction of the candidate library of terrain covariates

[0131] Physical and Logical Structure: This step constructs a candidate library of topographic covariates based on the preprocessed digital elevation data. The candidate library of topographic covariates includes 12 topographic factors: elevation (Ele), topographic humidity index (TWI), runoff accumulation (FA), flow direction (FD), curvature (Cur), planar curvature (Plc), profile curvature (Prc), slope (Slo), slope variability (SOS), aspect (Asp), aspect variability (SOA), and topographic relief (RA).

[0132] The topographic covariates mentioned above were calculated from the ASTER digital elevation model using SAGA GIS software version 7.8.2. Elevation directly reflects the influence of altitude on the redistribution of hydrothermal conditions. Slope reflects the degree of surface inclination, controlling surface runoff velocity and soil erosion intensity. Aspect reflects the influence of surface orientation on solar radiation reception. Slope variability reflects the rate of change of slope in space, indicating the severity of topographic relief. Topographic moisture index reflects the control of topography on soil moisture. Runoff accumulation reflects the convergence characteristics of surface runoff. Curvature, planar curvature, and profile curvature reflect the degree of curvature of the surface morphology. Aspect variability reflects the spatial rate of change of slope aspect. Topographic relief reflects the relative elevation difference of the terrain.

[0133] The aforementioned topographic covariates characterize the control of topography on soil formation and development from different dimensions. Topography significantly affects soil hydrothermal conditions, geological erosion, and material migration processes by altering regional temperature, precipitation, and humidity, and is one of the key controlling factors of soil total nitrogen content. In areas with continuous cloud cover, the topographic covariates are independent of optical signals and can provide stable environmental background information for soil total nitrogen retrieval.

[0134] Step S5: PCC-VIF covariate screening

[0135] Physical and logical structure: This step employs a two-step screening method combining Pearson correlation coefficient and variance expansion factor to select the optimal combination of covariates from the candidate libraries of spectral covariates, synthetic aperture radar covariates, and topographic covariates. This step requires measured values ​​of total soil nitrogen as the dependent variable.

[0136] Connections and Signal Flow: The input to this step is the values ​​of all candidate covariates in the three covariate candidate libraries at each sampling point, and the measured value of total soil nitrogen at the corresponding sampling point. The output is the optimal combination of spectral-synthetic aperture radar-topography covariates, denoted as the SST covariate combination, which is then output to the model training step.

[0137] The first step is preliminary correlation screening. Pearson correlation coefficients are calculated between each candidate covariate from the spectral covariate candidate library, the synthetic aperture radar covariate candidate library, and the topographic covariate candidate library and the measured values ​​of total soil nitrogen. (Pearson correlation coefficient) The calculation formula is:

[0138]

[0139] in Indicates the first At the sampling point The values ​​of each covariate, Indicates the first The measured values ​​of total nitrogen in the soil at each sampling point were obtained. For the first The mean of each covariate across all sampling points. The value represents the average of the measured total nitrogen in the soil across all sampling points. This represents the total number of sampling points. Calculate the Pearson correlation coefficient. Sort the variables by absolute value of the correlation coefficient from largest to smallest at a significance level of 0.01. Retain the 7 covariates with the largest absolute values ​​of their correlation coefficients from the spectral covariate candidate pool: B2, B3, B5, B6, B7, BI, and NDGI. Retain the 7 covariates with the largest absolute values ​​of their correlation coefficients from the synthetic aperture radar covariate candidate pool: VH_ENT, VH_HOM, VV_ENT, VV_HOM, VH, VHplusVV, and VHdiffVV. Retain the 5 covariates with the largest absolute values ​​of their correlation coefficients from the topographic covariate candidate pool: Ela, TWI, Slo, SOS, and RA.

[0140] The second step is collinearity diagnosis. The variance inflation factor (VIF) is calculated for all 19 covariates retained from the initial screening in the first step. The formula for calculating the variance inflation factor is:

[0141]

[0142] in VIF is the coefficient of determination obtained by performing a linear regression on the i-th covariate with respect to all other covariates except itself. The larger the VIF value, the more severe the multicollinearity between the covariate and other covariates. VIF = 10 is set as the threshold for determining multicollinearity. The covariate with the highest VIF value among those with a VIF value greater than 10 is successively removed. After each removal, the VIF value of the remaining covariates is recalculated until the VIF value of all remaining covariates is no greater than 10. After the above successive removal, 7 covariates are finally retained to form the optimal SST covariate combination, including elevation Ela, topographic humidity index TWI, slope variability SOS, the sum of VH polarization and VV polarization VHplusVV, the difference between VH polarization and VV polarization VHdiffVV, multispectral image second band B2, and normalized green index NDGI.

[0143] The PCC-VIF two-step screening method effectively eliminates information redundancy between covariates while evaluating the correlation between each covariate and total soil nitrogen, reduces the dimensionality of the model input, and is conducive to improving the fitting speed and generalization performance of subsequent machine learning models.

[0144] Step S6: Obtaining measured values ​​of total nitrogen in soil

[0145] Physical and logical components: This step involves field sampling and laboratory analysis. Pre-defined sampling points are located within the target area. A handheld GPS device guides the user to these points to collect surface soil samples and determine the total nitrogen content.

[0146] Based on considerations of land use type, soil parent material, and accessibility, sampling points were pre-defined using the fishing net tool in ArcGIS Pro software version 3.6. During sampling, a handheld GPS device was used to guide the user to the pre-defined sampling point locations, and the actual sampling points were located and their coordinates recorded within a 200-meter radius. Topsoil samples from 0 to 20 cm depths were collected using a five-point sampling method during satellite transit days, with each sample weighing approximately 0.5 kg. Stones, grass, and other foreign objects were initially removed from the soil on-site, and the samples were stored in sealed polyethylene bags. In the laboratory, the soil samples were air-dried, passed through a 0.149 mm nylon sieve, and then placed in dry Kjeldahl flasks. The total nitrogen content in the soil samples was determined using the Kjeldahl method, yielding measured values ​​of total nitrogen in grams per kilogram (g / kg) for each sampling point.

[0147] The obtained measured values ​​of total nitrogen in the soil are used for both PCC correlation analysis in step S5 and target variables for model training in step S7.

[0148] Step S7: XGBoost Model Training

[0149] Physical and Logical Structure: This step uses the values ​​of the optimal SST covariate combinations at each sampling point as input features and the measured values ​​of total soil nitrogen at the corresponding sampling points as target variables. The ultimate gradient boosting algorithm is used to train the soil total nitrogen inversion model. The ultimate gradient boosting algorithm is an ensemble learning algorithm based on gradient boosting decision trees. It iteratively integrates multiple weak learners to gradually correct prediction errors and construct a strong evaluator.

[0150] Connections and Signal Flow: The inputs to this step are the values ​​of the optimal SST covariate combination output from step S5 at each sampling point, and the measured values ​​of total soil nitrogen at the corresponding sampling points output from step S6. The output is the trained XGBoost soil total nitrogen inversion model.

[0151] Suppose the training set contains n samples, each sample consisting of an m-dimensional SST covariate feature vector x_i=(x_{i1},x_{i2},...,x_{im}) and the corresponding measured soil total nitrogen value y_i. The XGBoost model predicts the samples by ensemble K decision trees, and the formula for calculating the predicted value is:

[0152]

[0153] Where f_k is the prediction function of the k-th decision tree, each decision tree maps the input features to the weights of the leaf nodes, and the sum of the predictions of all decision trees is the final output of the model.

[0154] The goal of training the XGBoost model is to minimize the following objective function:

[0155]

[0156] The first item The loss function measures the predicted value. Compared with measured values The fitting error between them is used as the mean square error as the loss function in this invention; the second term This is a regularization term used to control model complexity and prevent overfitting. It represents the regularization term for each decision tree. The calculation formula is:

[0157]

[0158] in The number of leaf nodes in the decision tree. For the first The weight of each leaf node, The penalty coefficient for the number of leaf nodes. The L2 regularization coefficient is... These are the L1 regularization coefficients. The L1 regularization term. The L2 regularization term makes the model tend to use fewer features. This makes the model tend to use smaller weight values. Combining both methods can effectively prevent the model from overfitting.

[0159] Before model training, all sampling points are divided into training and validation sets in an 8:2 ratio. The training set samples are used for learning model parameters, while the validation set samples are used to evaluate the model's generalization performance. Five-fold cross-validation and grid search are used for hyperparameter tuning. The training set is randomly divided into five equal subsets. Each time, four subsets are used to train the model, and the remaining subset is used to validate the model. This process is repeated five times, and the average performance is used as the evaluation result for that set of hyperparameters. An exhaustive search is performed on the predefined hyperparameter combination space, and the hyperparameter combination with the best validation performance is selected as the final model configuration. The hyperparameter search range is as follows: number of trees n_estimators: 50 to 500; maximum tree depth max_depth: 3 to 9; learning rate learning_rate: 0.01, 0.1, or 0.3; sample sampling rate subsample and feature sampling rate colsample_bytree: 0.7 to 1.0; L1 regularization coefficient reg_alpha: 0, 0.1, or 1; L2 regularization coefficient reg_lambda: 1, 1.5, or 2. After joint optimization using grid search and five-fold cross-validation, the optimal combination of hyperparameters was determined as follows: 200 trees, maximum tree depth of 6, learning rate of 0.1, sample sampling rate of 0.8, feature sampling rate of 0.8, L1 regularization coefficient of 0.1, and L2 regularization coefficient of 1.

[0160] After the above hyperparameter tuning and model training, the XGBoost model can effectively fit the complex nonlinear mapping relationship between SST covariates and total nitrogen in soil. The coefficient of determination R² on the validation set reaches 0.82, and the relative analysis error RPIQ reaches 3.03, showing good prediction accuracy and generalization performance.

[0161] Compared to the graph neural network-weighted XGBoost method GNNW-XGBoost, the covariate design in this invention has a clear physical mechanism basis. First, based on the physicochemical environment of soil formation and evolution in continuous cloud and fog regions, this invention identifies and defines the categories and applicable scope of environmental variables affecting the spatial distribution of total nitrogen in soil. Second, based on the theory of soil-vegetation particle radiative transfer and scattering, it constructs environmental covariates that can characterize soil, vegetation, and topographic information. This physical mechanism constraint ensures that the selection of covariates has a clear basis in soil formation processes and electromagnetic scattering mechanisms, aiming to solve the problem of unstable inversion of total nitrogen in soil when optical remote sensing images are obscured by clouds and fog, rather than simply relying on data-driven feature selection.

[0162] Step S8: Spatial Distribution Prediction and Mapping

[0163] Physical and Logical Structure: This step inputs the optimal SST covariate combination raster data covering the entire target area into the trained XGBoost soil total nitrogen inversion model, calculates the predicted soil total nitrogen value pixel by pixel, and generates a spatial distribution map of soil total nitrogen in the target area. This distribution map is stored in raster format, with each pixel measuring 30 meters × 30 meters. The pixel value is the predicted soil total nitrogen content at that location, expressed in grams per kilogram.

[0164] Connections and Signal Flow: The input to this step is the raster data of the entire target region determined in step S5, based on the optimal SST covariate combination, and the XGBoost model trained in step S7. The output is a spatial distribution map of total nitrogen in the soil of the target region.

[0165] For each 30m × 30m pixel within the study area, the raster values ​​corresponding to the seven SST covariates at that pixel are extracted to form a 7-dimensional feature vector. This vector is then input into the trained XGBoost model, which outputs the predicted total nitrogen value of the soil at that pixel location, which is assigned to that pixel. This process is repeated pixel-by-pixel throughout the entire study area to ultimately generate a full-coverage spatial distribution map of total nitrogen in the soil.

[0166] The generated spatial distribution map can intuitively reflect the spatial distribution pattern and content status of total nitrogen in the soil within the target area. A specific application example of this invention selects a region in eastern my country as the study area, with a total area of ​​approximately 1700 square kilometers. The southern part of this study area is mainly hilly, while the northern part is flat, consisting mainly of plains and depressions. The climate type is temperate monsoon, with a multi-year average temperature of approximately 14.7 degrees Celsius and an average annual precipitation of approximately 671.1 millimeters. Land use in the study area is mainly arable land, primarily used for growing wheat, corn, soybeans, and peanuts. The soil types are mainly brown soil and alluvial soil. This region has significant topographic relief, abundant water vapor, and is prone to cloud and fog formation and accumulation, making it a typical continuous cloud and fog region, suitable for verifying the effectiveness of the technical solution of this invention.

[0167] Within the study area, a total of 425 soil samples were collected according to the method described in step S6. The average total nitrogen content of the overall sample set was 1.13 g / kg, ranging from 0.38 to 3.58 g / kg, with a median of 1.11 g / kg. The average total nitrogen content of the calibration and validation sets was 1.13 and 1.12 g / kg, respectively, with medians of 1.11 and 1.09 g / kg, consistent with the characteristics of the overall data. The coefficients of variation for the three sample sets ranged from 28.25% to 34.30%, indicating a moderate degree of spatial variability and uneven distribution of soil total nitrogen within the study area. High total nitrogen values ​​were mainly distributed in the hilly areas of the southern part of the study area, while the content was relatively low in the northern plains, with a decreasing trend in total nitrogen values ​​from south to north. This distribution map can provide spatial decision support for precision agricultural management and soil nutrient regulation.

[0168] This invention also provides a soil total nitrogen resistance remote sensing mapping system for continuous cloud and fog areas, including a data acquisition and preprocessing module, a covariate construction module, a covariate screening module, a model training module, and a spatial mapping module.

[0169] The data acquisition and preprocessing module is configured to acquire synthetic aperture radar image data, multispectral image data, and digital elevation data of the target area, and perform geometric correction, radiometric correction, cropping, resampling, and spatial registration on each raw data. It outputs preprocessed multi-source remote sensing data and digital elevation data with unified spatial resolution and geographic coordinate system. The specific operation of this module is consistent with step S1.

[0170] The signal input of the covariate construction module is connected to the signal output of the data acquisition and preprocessing module. It is configured to construct a spectral covariate candidate library based on preprocessed multispectral image data, a synthetic aperture radar (SAR) covariate candidate library based on preprocessed SAR image data, and a topographic covariate candidate library based on digital elevation data. The specific operation of this module is consistent with steps S2 to S4. This module outputs three types of covariate candidate libraries to the covariate filtering module.

[0171] The signal input of the covariate screening module is connected to the signal output of the covariate construction module and the input of the measured soil total nitrogen value, respectively. It is configured to calculate the correlation coefficient between each candidate covariate and the measured soil total nitrogen value for initial screening. Then, the variance inflation factor is calculated for the candidate covariates retained in the initial screening. Candidate covariates exhibiting multicollinearity are successively eliminated based on the variance inflation factor exceeding a set threshold, and the optimal SST covariate combination is output. The specific operation of this module is consistent with step S5.

[0172] The signal input terminal of the model training module is connected to the signal output terminal of the covariate selection module and the input terminal of the measured soil total nitrogen value, respectively. It is configured to use the values ​​of the optimal SST covariate combination at each sampling point as input features and the measured soil total nitrogen value at the corresponding sampling point as the target variable, and to train the soil total nitrogen inversion model using the limiting gradient boosting algorithm. The specific working method of this module is the same as that of steps S6 to S7.

[0173] The signal input terminal of the spatial mapping module is connected to the signal output terminal of the model training module, and receives the optimal SST covariate combination raster data covering the entire target area. It is configured to input the raster data pixel-by-pixel into the soil total nitrogen inversion model to calculate the predicted soil total nitrogen value, generating and outputting a spatial distribution map of soil total nitrogen in the target area. The specific operation of this module is consistent with step S8.

[0174] The present invention also includes a process for verifying and assessing the uncertainty of the above-mentioned mapping model.

[0175] Accuracy Validation: The model accuracy was evaluated using three metrics: coefficient of determination (R²), root mean square error (RMSE), and relative analytical error (RPIQ). The formula for calculating the coefficient of determination (R²) is as follows:

[0176]

[0177] in For the first Measured total nitrogen values ​​in soil from one validation sample. These are the model's predicted values. To verify the mean of the measured values ​​of the sample. The closer R² is to 1, the better the model fit.

[0178] The formula for calculating the root mean square error (RMSE) is:

[0179]

[0180] A smaller RMSE value indicates a smaller model prediction error.

[0181] The formula for calculating the relative analysis error RPIQ is:

[0182]

[0183] Where Q_3 and Q_1 are the upper and lower quartiles of the measured values ​​of the validation samples, respectively, and Q_3-Q_1 is the interquartile range. A larger RPIQ value indicates better model prediction performance; generally, an RPIQ greater than 2 indicates that the model has good predictive ability.

[0184] SHAP Interpretability Analysis: The SHAP method is used to quantify the contribution of each input covariate to the model output. Based on the Shapley value in game theory, SHAP calculates the marginal contribution of each covariate to the prediction result across all possible combinations of features, obtaining the SHAP value of that covariate for each sample's predicted value. A positive SHAP value indicates that the covariate has a positive effect on the prediction result, while a negative SHAP value indicates a negative effect. The mean of the absolute SHAP values ​​of each covariate across all samples yields a global ranking of the covariate's contribution to the model, thus identifying the dominant factors influencing the spatial distribution of total nitrogen in soil.

[0185] Uncertainty Assessment: The standard deviation of the model's predictions was assessed using 50 repeated calculations. Specifically, the training set was divided into five parts using a stratified approach. Five sub-models were obtained through 5-fold cross-validation. This sample partitioning and model training process was repeated 10 times, resulting in a total of 50 models. For each sampling point, the standard deviation of the 50 predicted values ​​was calculated as a quantitative indicator of the prediction uncertainty at that point. A spatial distribution map of uncertainty in the study area was generated to identify regions with low prediction confidence, providing a basis for subsequent sampling strategy optimization and model improvement.

[0186] The soil total nitrogen cloud and fog resistance remote sensing mapping framework of this invention can be transferred to new target areas for application.

[0187] When migrating the framework to a new region, the corresponding covariates for the target region must be calculated according to a unified data preparation and covariate construction process, instead of directly using the covariate values ​​from the original region. The specific method is as follows.

[0188] First, acquire Sentinel-1 synthetic aperture radar remote sensing imagery of the target area, as well as Sentinel-2 multispectral optical remote sensing imagery that is time-matched or time-similar to this radar imagery. Simultaneously, acquire ground elevation data for the target area.

[0189] Secondly, preprocessing is performed on various types of raw data, including geometric correction, radiometric correction, cropping, resampling to 30-meter spatial resolution, and spatial registration, so that all data are unified to the same spatial resolution and geographic coordinate system.

[0190] Then, according to the calculation formulas set in steps S2, S3, and S4, the synthetic aperture radar covariate is calculated based on radar data, the spectral covariate is calculated based on optical data, and the terrain covariate is calculated based on elevation data.

[0191] Next, the covariates calculated in the new region are organized according to the same variable type in the model training phase, namely Ela, TWI, SOS, VHplusVV, VHdiffVV, B2, NDGI, the same calculation method, the same spatial resolution of 30 meters, the same data format, and the same numerical standardization method.

[0192] Finally, the processed SST covariate raster data is input into the trained XGBoost soil total nitrogen inversion model, and the predicted value of soil total nitrogen is calculated pixel by pixel to generate a spatial distribution map of soil total nitrogen in the target area and mapping products.

[0193] The present invention has at least the following technical effects: First, by introducing synthetic aperture radar covariates, its ability to penetrate clouds and fog is fully utilized, effectively solving the problem of difficulty in soil total nitrogen inversion caused by the low availability of optical images in continuous cloud and fog areas.

[0194] Second, by constructing an SST covariate system that includes three sources: spectral, SAR, and topographic, the synergistic utilization of multi-source remote sensing information was realized, and the spatial distribution pattern of total nitrogen in the soil was characterized from multiple dimensions such as vegetation growth status, surface structure characteristics, and topographic control.

[0195] Third, the PCC-VIF two-step screening method eliminates the redundancy of multicollinearity information among covariates, reduces the dimensionality of the model input, and improves the model's fitting efficiency and numerical stability.

[0196] Fourth, by adopting the XGBoost ensemble learning algorithm and introducing L1 and L2 regularization terms, the complex nonlinear mapping relationship between soil total nitrogen and multidimensional environmental covariates was accurately fitted while effectively preventing overfitting.

[0197] Fifth, the mapping framework constructed by this invention has been verified in many typical regions around the world, and has good cross-regional applicability, which can provide reliable technical support for soil fertility evaluation, precision agricultural management and soil health monitoring in areas with continuous cloud cover.

[0198] Further explanation is needed regarding the following: First, a sampling point refers to the geographic coordinates of soil samples actually obtained through field sampling within the target area, where the total nitrogen content of the soil was measured. The training set refers to a portion of the sampling points proportionally extracted from all sampling points for model parameter learning. The validation set refers to a portion of the sampling points proportionally extracted from all sampling points for evaluating the model's generalization performance. A pixel refers to the smallest spatial unit in remote sensing imagery or raster data; in this invention, the unified pixel size is 30 meters × 30 meters. Raster data refers to a spatial data matrix composed of pixels arranged in rows and columns, with each pixel storing a numerical value. The weak learner in the extreme gradient boosting algorithm refers to a classification and regression tree; each decision tree divides samples into different leaf nodes by thresholding the input features, and each leaf node outputs a weight value as a prediction contribution. Regularization refers to the technique of adding a penalty term to the objective function to limit model complexity.

[0199] Second, regarding the difference between this invention and the GNNW-XGBoost method, the model of this invention does not simply adopt the GNNW-XGBoost method, but is a physical mechanism-machine learning integrated model constructed to meet the needs of remote sensing inversion of total nitrogen in soil in continuous cloud and fog regions. This model integrates the theory of soil formation process, the theory of radiative transfer and scattering of soil-vegetation particles, and the limiting gradient boosting algorithm. Specifically, firstly, based on the physicochemical environment of soil formation and evolution in continuous cloud and fog regions, the categories of environmental variables affecting the spatial distribution of total nitrogen in soil and their applicable scope are identified and defined; secondly, based on the theory of radiative transfer and scattering of soil-vegetation particles, environmental covariates that can characterize soil, vegetation, and topographic information are constructed; finally, multiple weak learners are constructed using the limiting gradient boosting algorithm, and the prediction results of each weak learner are summarized through ensemble learning to form a high-precision soil total nitrogen inversion and mapping model. Compared with GNNW-XGBoost, the core difference of this invention lies in its emphasis on the physical environment constraints and radar remote sensing information adaptation under continuous cloud and fog conditions. The covariate design has a clear basis for soil formation process and electromagnetic scattering mechanism, aiming to solve the problem of unstable inversion of total nitrogen in soil when optical remote sensing images are obscured by clouds and fog.

[0200] Third, regarding the correlation coefficient in step S5 The significance test, at the 0.01 significance level, refers to the value when the calculated Pearson correlation coefficient is significantly lower than the mean. When the value is less than 0.01, the correlation between the covariate and the measured value of total nitrogen in the soil is considered to be statistically significant. The null hypothesis that the two are not related is rejected, and the covariate is retained accordingly. The calculation of the value is based on Distribution, test statistic The calculation formula is ,in The correlation coefficient, Let be the total number of sampling points. This statistic follows a sequence with degrees of freedom. of distributed.

[0201] Fourth, regarding the loss function and regularization term of the XGBoost objective function in step S7, as described in step S7 of the specific implementation method, the following supplementary explanation is provided. The first term in the objective function is the loss function, which uses the mean squared error as the loss function. For the... For each sample, the loss function is... The calculation formula is Used to measure model predictions Compared with measured values The deviation between them. The second term in the objective function is a model complexity penalty term, used to prevent overfitting. For the... A decision tree, with a complexity penalty term The number of leaf nodes and their weights are both determined by the leaf node count and weights, and the calculation formula is as follows:

[0202] ,in This represents the total number of leaf nodes in the decision tree. For the first The weight values ​​output by each leaf node. This is a penalty coefficient for the number of leaf nodes, used to control the growth scale of the tree. The L2 regularization coefficient is used to constrain the sum of squares of the weights of the leaf nodes. This is the L1 regularization coefficient, used to constrain the sum of the absolute values ​​of the weights of the leaf nodes. By introducing this regularization term, the model tends to select decision trees with simple structures during training, effectively improving the model's generalization ability.

[0203] Fifth, regarding the specific calculation method of the SHAP value, the following supplementary explanation is provided. The SHAP value is calculated based on the Shapley value principle in game theory, for a specific covariate. In a certain sample SHAP value on The calculation method is as follows: consider all possible subsets of covariates. ,in Given the set of all covariates, calculate the result in the set containing the covariates. and without covariates The difference between the model predictions in the two cases is calculated by weighting the values ​​by subset size, using the following formula:

[0204] ,in For subset The number of covariates, The total number of all covariates. For using only subsets The covariates in the model are the predicted values ​​of the model output. When When the value is positive, it indicates a covariate. On this sample, the presence of this covariate positively contributes to the prediction results, meaning it increases the predicted value of total soil nitrogen; when When the value is negative, it indicates a covariate. On this sample, the covariate has a negative inhibitory effect on the prediction results, meaning its presence reduces the predicted value of total nitrogen in the soil. The global contribution of this covariate can be obtained by taking the absolute value of the SHAP values ​​for all samples and then averaging them; a larger contribution indicates that the covariate is more important in the model's predictions.

[0205] Sixth, regarding the numerical standardization process for pixel-by-pixel prediction in step S8, the following supplementary explanation is provided. Before inputting the SST covariates into the XGBoost model, all covariate values ​​need to be standardized. Standardization uses the Z-score standardization method. There are 1 covariates, whose original values ​​are given in each cell. Standardized values The calculation formula is ,in This is the mean of the covariate across all sampling points in the training set. This represents the standard deviation of the covariate across all sampling points in the training set. The mean and standard deviation parameters on the training set are calculated and saved during model training and directly applied to the covariate values ​​of all pixels during spatial mapping. This standardization process ensures that different covariates have the same numerical dimensions, preventing certain covariates with large value ranges from dominating the model training process.

[0206] Seventh, regarding the speckle filtering principle of synthetic aperture radar imagery, the Refined Lee filter is an adaptive speckle filtering algorithm. Its working principle is as follows: a speckle filter is set centered on each pixel. The filtering window is defined, and the local mean and local variance of all pixels within the window are calculated. The ratio of the local variance to the variance of the entire image is used to determine whether the region is uniform, non-uniform, or a point target region. A higher filtering intensity is used in uniform regions to effectively suppress speckle noise, a moderate filtering intensity is used in non-uniform regions to preserve texture information, and a lower filtering intensity is used in point target regions to maintain spatial resolution. This filtering method effectively suppresses the inherent speckle noise of synthetic aperture radar images while preserving the edge and texture information of ground features to the greatest extent, which is beneficial for the accurate extraction of subsequent texture parameters.

[0207] Eighth, regarding the explanation of water vapor noise-related bands in multispectral imagery, the bands removed refer to bands 1, 9, and 10 in the Sentinel-2 imagery. Band 1 is an aerosol band, band 9 is a water vapor band, and band 10 is a cirrus cloud band. These three bands are primarily used for atmospheric correction and are significantly affected by water vapor absorption, resulting in low signal-to-noise ratios for their surface reflectance signals, making them unsuitable for quantitative inversion of surface parameters. After removing these three bands, the remaining 10 high-quality bands are used for subsequent spectral covariate construction.

[0208] Ninth, regarding the physical mechanism basis of the covariate construction in this invention, the design of the spectral covariate is based on the vegetation spectral response mechanism, namely, the total nitrogen content of the soil affects the plant growth status, while the chlorophyll content of plant leaves, canopy structure, and coverage significantly change their reflectance spectral characteristics. By measuring the reflectance at different wavelengths and constructing spectral indices, the soil nitrogen status can be indirectly inverted. The design of the synthetic aperture radar covariate is based on the microwave scattering mechanism, namely, C-band microwave signals can penetrate clouds and part of the vegetation canopy, and its backscattering coefficient is sensitive to the vertical structure of the canopy, leaf water content, and surface roughness, all of which are closely related to the soil fertility level. The design of the topographic covariate is based on the soil genesis theory, namely, topography determines the direction of soil development and the spatial distribution pattern of nutrients by controlling the redistribution of water and heat conditions and the process of material migration. The integration of the above three physical mechanisms gives the SST covariate system clear physical interpretability.

[0209] Tenth, this invention employs the XGBoost algorithm instead of a deep learning algorithm. Firstly, the number of samples in soil total nitrogen inversion modeling is usually limited, and deep learning models are prone to overfitting on finite datasets. The XGBoost model, by introducing L1 and L2 regularization terms and using cross-validation for tuning, effectively avoids this problem. Secondly, the training time of the XGBoost model is much shorter than that of deep learning models. This invention uses grid search combined with five-fold cross-validation to complete hyperparameter optimization within a reasonable time. Thirdly, the XGBoost model has good interpretability; the SHAP method can quantify the contribution of each covariate to the prediction results, facilitating the analysis of the dominant factors influencing the spatial distribution of soil total nitrogen.

[0210] It should be further noted that, regarding the model performance verification of this invention, step S7 of the specific implementation method states that the coefficient of determination R² on the validation set reaches 0.82 and the relative analysis error RPIQ reaches 3.03. To verify the effectiveness of the XGBoost model constructed in this invention, it is compared with four conventional machine learning models, including partial least squares regression, artificial neural networks, support vector regression, and random forest. All comparison models use the same SST covariates as in this invention as input features, the same measured values ​​of total soil nitrogen as target variables, and the same methods for dividing the training and validation sets.

[0211] The partial least squares regression model had a determination coefficient (R²) of 0.48, a root mean square error (RMSE) of 0.22, and a relative analysis error (RPIQ) of 1.65. This model performed the worst among all the comparative models, indicating that the relationship between the SST covariate and total soil nitrogen is not a simple linear one, and that the linear regression model is insufficient to fully characterize the mapping between the two.

[0212] The artificial neural network model has a coefficient of determination (R²) of 0.52, a root mean square error (RMSE) of 0.21, and a relative analytical error (RPIQ) of 1.73. The support vector regression model has a coefficient of determination (R²) of 0.63, a root mean square error (RMSE) of 0.18, and a relative analytical error (RPIQ) of 2.02. These two nonlinear models show some improvement over the partial least squares regression model, indicating that nonlinear models can better fit the complex relationship between covariates and total soil nitrogen, but they are still insufficient to fully describe this nonlinear mapping.

[0213] The random forest model has a determination coefficient (R²) of 0.76, a root mean square error (RMSE) of 0.15, and a relative analytical error (RPIQ) of 2.42. This model performs best among all conventional models, exhibiting high prediction accuracy and generalization ability.

[0214] The model constructed using the extreme gradient boosting algorithm in this invention performs best among all compared models, with a coefficient of determination (R²) of 0.82, a root mean square error (RMSE) of 0.12, and a relative analytical error (RPIQ) of 3.03. Compared to the partial least squares regression model, the coefficient of determination is improved by approximately 70.8%; compared to the random forest model, the coefficient of determination is improved by approximately 7.9%.

[0215] The above comparison results fully verify the superiority of the extreme gradient boosting algorithm used in this invention in remote sensing inversion modeling of total nitrogen in soil. By iteratively integrating weak learners and introducing L1 and L2 regularization terms, this algorithm can more accurately fit the complex nonlinear mapping relationship between SST covariates and total nitrogen in soil, outperforming conventional models such as partial least squares regression, artificial neural networks, support vector regression, and random forests in both prediction accuracy and generalization ability.

[0216] In summary, this invention constructs a multi-source SST covariate system integrating spectral, synthetic aperture radar (SAR), and topographic information. It then uses the PCC-VIF two-step screening method to eliminate information redundancy among covariates and employs the extreme gradient boosting algorithm to establish a soil total nitrogen inversion model. This achieves high-precision, interference-resistant remote sensing mapping in continuous cloud and fog regions. This framework fully utilizes the penetration advantage of SAR signals, which are unaffected by cloud and fog, effectively overcoming the technical bottlenecks of low image availability and reduced mapping accuracy in traditional optical remote sensing under cloud and fog conditions. Implementation examples demonstrate that this invention achieves good prediction accuracy and generalization performance in multiple different geographical regions, providing reliable technical support for soil fertility evaluation, precise nutrient management, and sustainable agricultural development in continuous cloud and fog regions. The above descriptions are merely preferred embodiments of this invention and are not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for remote sensing mapping of total nitrogen in soil to resist fog in areas with continuous cloud cover, characterized in that, The method includes: Acquire multi-source remote sensing data and digital elevation data of the target area, wherein the multi-source remote sensing data includes synthetic aperture radar image data and multispectral image data; The acquired raw data are preprocessed, including geometric correction, radiometric correction, cropping, resampling and spatial registration, so that the data are unified to the same spatial resolution and geographic coordinate system. A candidate library of spectral covariates is constructed based on the preprocessed multispectral image data; a candidate library of synthetic aperture radar covariates is constructed based on the preprocessed synthetic aperture radar image data; and a candidate library of topographic covariates is constructed based on the digital elevation data. Obtain the measured values ​​of total nitrogen in the soil from multiple sampling points in the target area; The correlation coefficient between each candidate covariate and the measured value of total nitrogen in the soil was calculated. The candidate covariates were initially screened based on the magnitude of the correlation coefficient, and the candidate covariates with strong correlation to total nitrogen in the soil were retained. The variance inflation factor is calculated for all candidate covariates retained in the initial screening. Candidate covariates with multicollinearity are successively eliminated based on the condition that the variance inflation factor is greater than a set threshold, until the variance inflation factor of the remaining candidate covariates is not greater than the set threshold, thus obtaining the optimal combination of spectral-synthetic aperture radar-topography covariates. The soil total nitrogen inversion model is trained using the optimal combination of spectral-synthetic aperture radar-topographic covariates at each sampling point as input features and the measured value of soil total nitrogen at the corresponding sampling point as target variable. The raster data of the optimal spectral-synthetic aperture radar-topographic covariate combination covering the entire target area is input into the trained soil total nitrogen inversion model, and the predicted value of soil total nitrogen is calculated pixel by pixel to generate a spatial distribution map of soil total nitrogen in the target area.

2. The method for remote sensing mapping of total nitrogen in soil against cloud and fog in continuous cloud and fog regions according to claim 1, characterized in that, The synthetic aperture radar image data is C-band dual-polarization synthetic aperture radar image data, and the polarization modes include co-polarization and cross-polarization.

3. The method for remote sensing mapping of total nitrogen in soil against fog in continuous foggy areas according to claim 1, characterized in that, The multispectral image data includes visible light band, near-infrared band, and short-wave infrared band.

4. The method for remote sensing mapping of total nitrogen in soil against fog in continuous foggy areas according to claim 1, characterized in that, The resampling process unifies all data to the same cell size.

5. The method for remote sensing mapping of total nitrogen in soil against cloud and fog in continuous cloud and fog regions according to claim 1, characterized in that, The preprocessing of the synthetic aperture radar image data includes, in sequence: orbit correction, radiometric calibration, speckle filtering, terrain correction, and conversion to decibel values. The speckle filtering uses a refined Lee filter with a preset window size. The preprocessing of the multispectral image data includes: removing bands related to water vapor noise and resampling the remaining bands to a uniform spatial resolution.

6. The method for remote sensing mapping of total nitrogen in soil against fog in continuous foggy areas according to claim 1, characterized in that, The candidate library of spectral covariates includes reflectance of each band of multispectral images and vegetation spectral indices calculated based on multispectral images. The vegetation spectral indices include normalized vegetation index, first normalized vegetation index, first normalized differential vegetation index (red edge), second normalized differential vegetation index (red edge), third normalized differential vegetation index (red edge), enhanced vegetation index, difference vegetation index, simple ratio index, brightness index, soil-adjusted vegetation index, optimized soil-adjusted vegetation index, modified soil-adjusted vegetation index, green normalized difference vegetation index, normalized green index, and comprehensive spectral response index.

7. The method for remote sensing mapping of total nitrogen in soil against fog in continuous foggy areas according to claim 2, characterized in that, The method for constructing the candidate library of synthetic aperture radar covariates is as follows: First, based on the backscattering coefficients of co-polarization and cross-polarization, a variety of polarization combination parameters are constructed. The polarization combination parameters include the sum of co-polarization and cross-polarization, the difference between co-polarization and cross-polarization, the product of co-polarization and cross-polarization, and the quotient of cross-polarization and co-polarization. Then, the gray-level co-occurrence matrix method is used to extract a variety of texture parameters from each polarization combination image. The texture parameters include contrast, dissimilarity, uniformity, entropy, mean, variance, and correlation. The above-mentioned polarization combination parameters and the texture parameters of each polarization combination together constitute the candidate library of synthetic aperture radar covariates.

8. The method for remote sensing mapping of total nitrogen in soil against fog in continuous foggy areas according to claim 1, characterized in that, The candidate library of topographic covariates includes elevation, topographic humidity index, runoff accumulation, flow direction, curvature, plan curvature, profile curvature, slope, slope variability, aspect, aspect variability, and topographic relief.

9. The method for remote sensing mapping of total nitrogen in soil against fog in continuous foggy areas according to claim 1, characterized in that, The initial screening of candidate covariates based on correlation coefficients involves: calculating the Pearson correlation coefficient between each candidate covariate in the spectral covariate candidate library, the synthetic aperture radar covariate candidate library, and the topographic covariate candidate library and the measured value of total nitrogen in the soil; sorting the candidates by absolute value of the correlation coefficient from largest to smallest at a preset significance level; and retaining the top-ranked candidate covariates from each candidate library. The threshold value for the variance expansion factor is a preset collinearity determination threshold.

10. A soil all-nitrogen cloud-resistant remote sensing mapping system for continuous cloud and fog regions, characterized in that, include: The data acquisition and preprocessing module is used to acquire synthetic aperture radar image data, multispectral image data and digital elevation data of the target area, and to perform geometric correction, radiometric correction, cropping, resampling and spatial registration on each raw data, and output multi-source remote sensing data and digital elevation data with unified spatial resolution and geographic coordinate system after preprocessing. The covariate construction module has its signal input terminal connected to the signal output terminal of the data acquisition and preprocessing module. It is used to construct a spectral covariate candidate library based on the preprocessed multispectral image data, a synthetic aperture radar covariate candidate library based on the preprocessed synthetic aperture radar image data, and a topographic covariate candidate library based on the digital elevation data. It also outputs the three types of covariate candidate libraries to the covariate filtering module. The covariate screening module has its signal input terminal connected to the signal output terminal of the covariate construction module and the input terminal of the measured soil total nitrogen value, respectively. It is used to calculate the correlation coefficient between each candidate covariate and the measured soil total nitrogen value for initial screening. Then, it calculates the variance expansion factor of the candidate covariates retained in the initial screening and successively removes candidate covariates with multicollinearity based on the condition that the variance expansion factor is greater than a set threshold. The optimal combination of spectral-synthetic aperture radar-topography covariates is output. The model training module has its signal input terminal connected to the signal output terminal of the covariate screening module and the input terminal of the measured soil total nitrogen value, respectively. It is used to train and output the soil total nitrogen inversion model by using the value of the optimal spectral-synthetic aperture radar-topographic covariate combination at each sampling point as the input feature and the measured soil total nitrogen value at the corresponding sampling point as the target variable, and by using the extreme gradient boosting algorithm. The spatial mapping module has its signal input end connected to the signal output end of the model training module, and receives raster data of the optimal spectral-synthetic aperture radar-topographic covariate combination covering the entire target area. It is used to input the raster data pixel by pixel into the soil total nitrogen inversion model to calculate the predicted value of soil total nitrogen, and generate and output the spatial distribution map of soil total nitrogen in the target area.

Images