A soil heavy metal remote sensing estimation method and system based on GSWR-XGB model

Abstract

The application discloses a soil heavy metal remote sensing estimation method and system based on a GSWR-XGB model, and belongs to the technical field of heavy metal estimation. The application compensates for the spectral defects of a single data source through multi-source data spectrum compensation, constructs a GSWR-XGB model, introduces an elevation and multi-center point dynamic weighted Gaussian kernel mechanism, breaks through the spatial modeling limitation of a fixed bandwidth, deeply fuses spatial weights and spectral characteristics, enhances the description ability of the model to spatial heterogeneity, and realizes accurate estimation from sparse samples to continuous spatial distribution.

Description

Technical Field

[0001] This invention relates to the field of heavy metal estimation technology, and more specifically to a remote sensing estimation method and system for heavy metals in soil based on the GSWR-XGB model. Background Technology

[0002] Soil is the fundamental natural resource for agricultural production and the cornerstone of sustainable global ecosystem development. However, soil heavy metal pollution, especially cadmium (Cd) pollution, has become a serious problem in the environmental and public health fields because cadmium is toxic to plants and can cause nephrotoxicity in humans through the food chain. Large-scale surveys show that 16.1% of soil samples exceeded the threshold set by the Ministry of Environmental Protection, with cadmium pollution accounting for 7% of recorded pollution cases. Furthermore, due to the scarcity of arable land, some cadmium-contaminated farmland continues to be cultivated. This increasingly serious situation highlights the necessity of establishing accurate and efficient methods for monitoring soil cadmium pollution.

[0003] Currently, while geochemical methods based on field sampling offer high accuracy, they are often costly and difficult to implement for large-scale, continuous assessments. The rapid development of hyperspectral remote sensing technology, combined with machine learning methods, provides a new technical approach for large-scale soil pollutant detection. However, existing technologies still have several key limitations. Existing technologies, such as the Zhuhai-1 hyperspectral satellite sensor, cannot capture the crucial shortwave-infrared (SWIR) bands required for soil heavy metal retrieval; while widely used multispectral data such as Landsat 8 and Sentinel-2, although covering the SWIR bands, lack sufficient spectral resolution for high-precision retrieval. Therefore, a single remote sensing data source cannot simultaneously meet the dual requirements of spectral coverage and resolution for high-precision retrieval of large-scale soil cadmium pollution.

[0004] Besides the limitations of spectral data itself, the inherent spatial heterogeneity of cadmium pollution distribution significantly increases the difficulty of inversion. This heterogeneity mainly stems from differences in land use patterns, such as mining, fertilization, and industrial emissions. Models such as Geographically Weighted Regression (GWR) can incorporate spatial information to some extent by fitting local regressions at each geographic point. However, GWR assumes that spatial influences depend only on a fixed bandwidth, which may lead to an underestimation of spatial interactions between variables in highly heterogeneous regions. On the other hand, advanced ensemble learning methods such as Extreme Gradient Boosting (XGB), while improving prediction accuracy, cannot fundamentally solve the problem of spatial autocorrelation. XGB, by training multiple decision trees sequentially and integrating regularization and optimal splitting strategies, possesses strong nonlinear fitting capabilities but still lacks a spatial modeling mechanism. The recently proposed GW-XGBoost model attempts to introduce the spatial weights calculated by GWR as additional features into XGBoost. While promising, it has not yet overcome the limitations imposed by the fixed bandwidth in traditional GWR and has not been validated for application to soil heavy metal estimation using large-scale remote sensing imagery.

[0005] Therefore, how to propose a remote sensing estimation method and system for soil heavy metals based on the GSWR-XGB model to achieve high-precision remote sensing estimation of soil cadmium pollution is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model, in order to solve the problems of incomplete spectral information from a single remote sensing data source and the difficulty of traditional models in handling spatial heterogeneity in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: On one hand, this invention discloses a remote sensing estimation method for heavy metals in soil based on the GSWR-XGB model, comprising: Acquire multi-source remote sensing images of the cultivated area, and collect soil samples within the cultivated area and screen out bare soil samples. Obtain the multi-source enhancement features of the bare soil sample; The heavy metal content of each bare soil sample is obtained using a pre-trained GSWR-XGB model. The GSWR-XGB model generates geospatial information weights for the bare soil sample points through the GSWR mechanism. The multi-source enhanced features and the geospatial information weights are input into the XGB model to output the heavy metal content of each bare soil sample. The heavy metal content of each bare soil sample was obtained by Kriging spatial interpolation to obtain a spatial distribution map of heavy metal content in the cultivated area.

[0008] Preferably, the multi-source remote sensing images include Zhuhai-1 hyperspectral images and Landsat 8 multispectral images.

[0009] Preferably, soil samples are collected within the cultivated area and bare soil samples are screened out, including: Soil samples that meet the preset vegetation index threshold are selected as bare soil samples based on the vegetation index of the corresponding pixels in the Landsat 8 multispectral image. The vegetation indexes include the Normalized Difference Vegetation Index, the Soil-Adjusted Vegetation Index, and the Improved Soil-Adjusted Vegetation Index.

[0010] Preferably, the multi-source enhancement features include the spectral features of the Zhuhai-1 hyperspectral image obtained through correlation analysis and the spectral features of the Landsat 8 multispectral image.

[0011] Preferably, the geospatial information weights of bare soil sample points are generated through the GSWR mechanism, including: Obtain the geographic coordinates and elevation data of the bare soil sample points, and establish a three-dimensional rectangular coordinate system; Multiple center points were randomly selected, and the weighted Euclidean distance between each bare soil sample point and each center point was calculated. The weighted Euclidean distance is converted into geospatial information weights using a Gaussian kernel function.

[0012] Preferred weighted Euclidean distance The calculation formula is as follows: ; In the formula, and These represent the sampling point and the center point at the th... Coordinates in spatial dimensions These are the weight parameters.

[0013] Preferred geospatial information weights The conversion formula is as follows: ; In the formula, This is the bandwidth parameter.

[0014] A remote sensing estimation method for heavy metals in soil based on the GSWR-XGB model also includes a model training step, comprising: The measured heavy metal content, multi-source enhancement characteristics, and geospatial information generated based on the GSWR mechanism were obtained from bare soil samples. Using the multi-source enhancement features and geospatial information weights of bare soil samples as inputs to the XGB model, and the measured soil heavy metal content as the target, the model parameters are optimized through cross-validation and random search to obtain the trained GSWR-XGB model.

[0015] On the other hand, this invention discloses a soil heavy metal remote sensing estimation system based on the GSWR-XGB model, used to implement the aforementioned soil heavy metal remote sensing estimation method based on the GSWR-XGB model, comprising: The data acquisition module is used to acquire multi-source remote sensing images of the cultivated area, and at the same time collect soil samples in the cultivated area and screen out bare soil samples. The feature acquisition module is used to acquire the multi-source enhanced features of the bare soil sample; The model prediction module is used to obtain the heavy metal content of each bare soil sample using a pre-trained GSWR-XGB model. The GSWR-XGB model generates geospatial information weights for the bare soil sample points through the GSWR mechanism. The multi-source enhancement features and the geospatial information weights are input into the XGB model together, and the heavy metal content of each bare soil sample is output. The spatial interpolation module is used to obtain a spatial distribution map of heavy metal content in the soil of cultivated areas by interpolating the heavy metal content of each bare soil sample through Kriging space. The model training module is used to take the multi-source enhancement features and geospatial information weights of the bare soil sample as input, and the measured heavy metal content of the bare soil sample as the training target. The model parameters are optimized through cross-validation and random search to obtain the trained GSWR-XGB model.

[0016] As can be seen from the above technical solutions, compared with the prior art, this invention discloses a method and system for remote sensing estimation of soil heavy metals based on the GSWR-XGB model. By acquiring multi-source remote sensing data and performing spectral compensation and feature screening, it overcomes the dual deficiencies of a single data source in terms of spectral coverage and resolution, and improves the correlation between input features and heavy metal content. By introducing elevation information and a multi-center point dynamically weighted Gaussian kernel spatial weight generation mechanism, it breaks through the dependence of traditional geographic weighted regression models on fixed bandwidth, enabling the calculation of spatial weights to adapt to local spatial heterogeneity and avoid overfitting or information loss. By deeply integrating geospatial information weights as additional features with multi-source enhancement features into the XGB model, it achieves joint learning of spectral features and spatial distribution features, enhancing the model's ability to characterize the spatial distribution patterns of soil heavy metals. Finally, by using Kriging spatial interpolation, the predicted values ​​of sparse sample points are expanded into a continuous spatial distribution map of the entire cultivated area, providing a reliable technical means for large-scale soil heavy metal pollution monitoring. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart of the method provided by the present invention; Figure 2 This is a schematic diagram showing the total cadmium content of soil samples. Figure 3 A schematic diagram illustrating the correlation between various Landsat 8 data features and cadmium content; Figure 4 A schematic diagram illustrating the correlation between various data characteristics of Zhuhai No. 1 and cadmium content; Figure 5 A comparison chart of predicted and actual measured values ​​of soil cadmium content under different data conditions; Figure 6 This diagram illustrates the performance evaluation of the GSWR-XGB model and the effectiveness of the MA method. Figure 7 This is a spatial distribution map of soil cadmium content based on Landsat 8 multispectral data. Figure 8 This is a spatial distribution map of soil cadmium content based on Zhuhai-1 hyperspectral data; Figure 9 This is a spatial distribution map of soil cadmium content based on multi-source enhanced image data. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention discloses a remote sensing estimation method for heavy metals in soil based on the GSWR-XGB model. It combines an enhanced spatial heterogeneity module and multi-source image enhancement technology, utilizing Landsat 8 shortwave infrared data to "fill" the spectral gaps in the shortwave infrared range of the Zhuhai-1 hyperspectral image. The overall process is as follows: Figure 1 As shown, the specific steps include: S1. Acquire multi-source remote sensing images of the cultivated area (Landsat 8 multispectral image and Zhuhai-1 hyperspectral image), and collect soil samples in the cultivated area and screen out bare soil samples.

[0021] This embodiment uses a city as an example to collect soil samples, with a total sampling area of ​​approximately 600 square kilometers. A total of 105 soil samples were collected (83 from rice-growing areas and 22 from other agricultural areas), with only the topsoil (0-20 cm) collected. A handheld GPS device was used to record and photograph the geographic coordinates and soil type of each sampling point. The sampling method was as follows: First, a central point was selected. Second, sub-samples were collected at 10-meter intervals on both sides of this point, collecting 500 grams of soil from each sub-sample. The three sub-samples were then combined into a composite sample. The soil samples were air-dried, finely crushed, and sieved through a 100-mesh sieve. Acid digestion (AQ) was used for digestion before cadmium determination, and the total cadmium content was determined by inductively coupled plasma mass spectrometry (ICP-MS).

[0022] The acquisition of Landsat 8 multispectral images and Zhuhai-1 hyperspectral images was synchronized with the soil sample collection time.

[0023] Download Landsat 8 data (Collection 2, Tier 1, Level 2) from Google Earth. Landsat 8 Collection 2 data is the official reprocessing of Landsat Level 1 data, including several key corrections (radiometric correction, geometric correction, atmospheric correction, orthorectification, and geographic projection) to improve image quality and geolocation accuracy. It spans the coastal zone (B1), visible light (blue band (B2), green band (B3), and red band (B4)), near-infrared (B5), and shortwave infrared (B6 and B7), with a resolution of 30 meters for all bands. NDVI and other indicators no longer require additional corrections. Download Zhuhai-1 hyperspectral imagery data from its official service platform. Zhuhai-1 hyperspectral imagery data has 32 bands, a spectral resolution of 2.5 nm, a wavelength range of 400 to 1000 nm, and a spatial resolution of 10 meters. Radiometric and atmospheric corrections were performed using preprocessing software provided by the service platform to obtain reflectance images, and orthorectification of the Zhuhai-1 image was performed with Landsat 8 imagery as a reference. To ensure the reliability and rigor of the experimental results, high-precision downscaling is challenging; therefore, the Zhuhai-1 imagery was resampled to 30 meters to achieve the same spatial resolution as the Landsat 8 imagery.

[0024] Traditional hyperspectral data processing methods involve performing various mathematical transformations on the spectral data, combining correlation analysis to select characteristic bands, and then modeling. This embodiment uses the traditional method to process the Zhuhai-1 hyperspectral data. To suppress background noise and enhance the spectral characteristics of the Zhuhai-1 hyperspectral data, the original spectral curves were subjected to Savitzky-Golay smoothing (SG) with a polynomial order of 3 and a window size of 5. Furthermore, 11 spectral transformations were performed on the smoothed SG spectrum, including reciprocal transformation (RT), anti-reciprocal transformation (AT), reciprocal first derivative (RTFD), anti-reciprocal first derivative (ATFD), first derivative (FD), reciprocal second derivative (RTSD), anti-reciprocal second derivative (ATSD), second derivative (SD), multiplicative scattering correction (MSC), continuous removal (CR), and the preferred 1.75th order fractional derivative (1.75 FD) transformation.

[0025] The accumulation of heavy metals in soil has a profound impact on soil quality degradation, and Geographic Environmental Factors (GEFs) are widely used in SHM mapping. Considering that spectral indices and environmental ancillary variables constructed using Landsat 8 spectral information can effectively characterize GEFs, this embodiment uses spectral indices and environmental ancillary variables to characterize GEFs. The spectral indices used include the Soil Adjusted Vegetation Index (SAVI), its modified types (Modified Soil Adjusted Vegetation Index (MSAVI)), and the Normalized Difference Soil Organic Carbon (NDSOC).

[0026] SAVI's formula is as follows: ; In the above formula, R represents the red band, NIR represents the near-infrared band, and L is 0.5.

[0027] Environmental ancillary variables include the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI), both related to vegetation health. For indicators of land surface characteristics, such as Land Surface Brightness (LSB), Land Surface Greenness (LSG), and Land Surface Wetness (LSW), the formulas for LSB, LSG, and LSW are as follows: ; ; ; in It is near-infrared. This represents the average reflectivity of the B1-B7 bands. It's the green band. This is shortwave infrared band B6. To fully reflect surface brightness, all non-thermal infrared bands (B1-B7) are used in the LSB calculation. This embodiment uses a simplified vegetation indicator to calculate the LSG, which shows the difference between the near-infrared and green light bands. For the surface moisture index, the difference between the near-infrared and shortwave infrared bands is also used as an indicator because these two bands are more sensitive to moisture and reflect surface moisture conditions.

[0028] Regarding soil properties, the Soil Brightness Index (BI) and the Normalized Difference Moisture Index (NDMI) were selected to assess soil erosion and moisture conditions. Furthermore, to monitor soil salinization, the Soil Salinity Index (SI) and the Soil Salinity Remote Sensing Monitoring Index (SDI) were used, effectively reflecting soil salinity. Finally, the Clay Index (CI) was chosen to assess soil composition.

[0029] Due to spatial resolution and temporal differences between ground soil samples (point scale) and Landsat 8 data (30 meters), some bare soil areas during the sampling period may appear as vegetated areas in the imagery. Therefore, the spectral signals at these locations primarily reflect the vegetation canopy rather than the underlying soil surface. Including these samples in subsequent processing may lead to errors in feature selection and model fitting. Therefore, using all samples is inappropriate. To ensure the accuracy and reasonableness of the experimental results, this embodiment used more stringent criteria: NDVI index between 0.05 and 0.2; SAVI and MSAVI indices between 0.02 and 0.1, used to identify bare soil pixels in the imagery. Based on this criterion, 41 bare soil samples were selected from 105 soil samples for modeling.

[0030] S2. Obtain multi-source enhancement features of bare soil samples.

[0031] Total cadmium content was analyzed in 105 soil samples, and the results are as follows: Figure 2 As shown, Figure 2 The horizontal axis represents the total cadmium content, and the vertical axis represents the number of samples. Figure 2 The figure shows the number of soil samples across different ranges of total cadmium content. The red dashed line represents the average total cadmium content in the soil samples, the green dashed line represents the SD range near the average, and the blue line describes the distribution characteristics of total cadmium content in the soil samples. Additionally, the legend includes further statistical information on the total cadmium content of 105 soil samples.

[0032] according to Figure 2 The study showed that the average cadmium content in the soil of the study area was 0.33 mg / kg, with the highest content reaching 1.34 mg / kg. The average cadmium content of all soil samples exceeded the background value (0.126 mg / kg) for the province where the city is located by 161.9%, with 45.7% of the samples exceeding the national background value (0.10 mg / kg). The coefficient of variation for cadmium content was 0.61, indicating significant variability and spatial heterogeneity in soil cadmium content. Given its obvious non-normal distribution pattern, this study employed nonlinear Spearman correlation coefficients for subsequent analysis. The analysis results collectively indicate significant cadmium pollution in the soil of the study area. This analysis underscores the urgency and necessity of conducting large-scale, non-destructive, rapid, and accurate soil cadmium content monitoring in this region.

[0033] Spearman correlation analysis revealed a relatively mixed correlation between spectral indices, band information, and soil cadmium. Figure 3 To establish the correlation between various Landsat 8 data characteristics and cadmium content, Figure 4The correlation between various Zhuhai No. 1 data characteristics and cadmium content is shown. The direction of the ellipse indicates the sign of the correlation, and its width and color intensity reflect the strength of the correlation. The significance level is represented by... This indicates that p < 0.01. (× indicates p < 0.05, × indicates p > 0.05). This may be due to the low spatial resolution and limited spectral information of the Landsat 8 data. In this embodiment, the correlation coefficient threshold for feature selection was adjusted to improve model accuracy and ensure that the selected features are statistically significant (p < 0.05). Accordingly, the feature selection results for modeling are shown in Table 1.

[0034] The feature selection in this embodiment includes two methods. i) The Single-source Approach (SA) combines features constructed from a single satellite image data with correlation analysis. Furthermore, for Zhuhai-1 hyperspectral data, if the highest correlation of multiple spectral transforms overlaps in the same band, the band with the second highest correlation is selected to avoid collinearity. ii) The Multi-source Enhancement Approach (MA) utilizes Landsat 8 multispectral data to compensate for the limitations of spectral coverage in Zhuhai-1 hyperspectral data, thereby including all features selected in SA. In other words, it combines features from two images obtained based on the SA method to obtain multi-source enhanced features.

[0035] Table 1 Feature selection results (SA: single-source method; MA: multi-source enhancement method)

[0036] S3. Obtain the heavy metal content of each bare soil sample using a pre-trained GSWR-XGB model; the GSWR-XGB model generates geospatial information weights for bare soil sample points through the GSWR mechanism, inputs the multi-source enhanced features and geospatial information weights into the XGB model, and outputs the heavy metal content of each bare soil sample.

[0037] GSWR-XGB is an enhanced model built upon GWR and XGB. This model further considers the influence of elevation dimension (https: / / www.earthdata.nasa.gov), extending the two-dimensional geographic parameters in GWR to three-dimensional geospatial information (GSI), and dynamically increasing the number of center points. Specifically, ArcGIS is used to convert imagery from a geographic coordinate system to a projected coordinate system, obtaining coordinates in meters, and then combining this with elevation to establish a spatial rectangular coordinate system. Then, the spatial parameters for each sample are obtained by calculating the weighted Euclidean distance between sample points within the bandwidth (bw) and randomly selected center points (not limited to a single center point). The formula for calculating the spatial parameters is as follows: ; In the above formula, and They represent the sampling point and the center point in the i-th spatial dimension, respectively. The coordinates of ). The weight, a key adjustable parameter, determines the relative importance of the centroid to the remaining sample points within the bounding square (bw). Through this weighting mechanism, the model can dynamically identify the centroids with the strongest data heterogeneity, better adapting to the spatial heterogeneity between data points (indirectly identifying which centroids best express spatial heterogeneity through "random centroid combination + weighted distance calculation + model accuracy evaluation"). If the model accuracy significantly improves after adding a certain set of centroids, it indicates that these centroids can effectively represent the spatial heterogeneity in the data.

[0038] The Gaussian kernel function and distance are used to calculate the GSI and incorporate it into the model. Spatial correlation is utilized to capture spatial heterogeneity between samples, enhancing the model's fit to the distribution patterns of heavy metal content. The specific formula for Gaussian kernel processing is as follows: ; In the above formula, the `bw` parameter plays a crucial role in defining the spatial range of sampling points influenced by the center point. Specifically, when the distance between a sampling point and the center point is small, the GSI will approach 1, indicating a high spatial correlation between the sampling point and the center point; conversely, when the distance between them is large, the value will decrease, indicating a low spatial correlation between the sampling point and the center point. Weight control reduces the rate of decrease. Furthermore, the `Random_times` parameter is introduced into the model to control the number of searches for random center point combinations. As `Random_times` increases, the model can search for more center point combinations, thereby reducing the uncertainty caused by a single random sampling and improving the stability of center point selection.

[0039] During model training, the measured cadmium content of each bare soil sample point was used as the target variable, and the multi-source enhancement features and geospatial information weights were used as inputs.

[0040] The GWR model underwent 10 rounds of 4-fold cross-validation using Matlab's adaptive kernel function. The GSWR-XGB and XGB models have more parameters and a wider parameter range, especially the central parameter of GSWR-XGB. First, 10 rounds of 4-fold cross-validation were performed, combined with 5000 random searches. Then, within the parameter overlap range of the optimal fold in each cross-validation, another 4-fold cross-validation and 1000 random searches were performed. Experimental results are expressed as average performance, with model parameters based on the best performance under the second cross-validation. Therefore, the results of the GSWR-XGB model were obtained through over 200,000 modeling iterations, effectively improving the reliability of the model results. The computational platform consisted of MATLAB R2022b and Python 3.9 open-source libraries. The standard parameter settings (except for the first four parameters) of the GSWR-XGB and XGB models are consistent, as shown in Table 2.

[0041] Table 2. Definitions and ranges of parameters

[0042] S4. The heavy metal content of each bare soil sample was obtained by Kriging spatial interpolation to obtain a spatial distribution map of heavy metal content in the cultivated area.

[0043] The embodiments of the present invention further evaluate the effectiveness of the GSWR-XGB model and MA in estimating soil cadmium content.

[0044] The model evaluation metrics are as follows: The coefficient of determination (R²) is a key indicator for evaluating model quality. Root mean square error (RMSE) and mean absolute error (MAE) are commonly used to quantify model error. It is worth noting that MAE is less affected by extreme values ​​than RMSE. Therefore, a well-performing model typically has a high R² value and low RMSE and MAE values.

[0045] ; ; ; In the above equation, This represents the actual measured value. Indicates the predicted value. This represents the average of all measurements. Additionally, n represents the sample size.

[0046] Table 3 shows the model performance for each method. Figure 5 The comparison shows the predicted and actual cadmium content values ​​of soil based on the optimal model under different data conditions: (a) Zhuhai No. 1 hyperspectral data, (b) Landsat 8 multispectral data, and (c) multi-source enhanced data. Figure 6 Tables (a)-(c) demonstrate the effectiveness of MA. Figure 6 Tables (d)-(f) demonstrate the effectiveness of the GSWR-XGB model.

[0047] Table 3 Model Performance

[0048] Table 3 and Figure 5 The results show that the model's overall accuracy is poor in the SA (Simplified, Average) scenario. The model accuracy based on the Zhuhai-1 satellite data is better than that based on the Landsat 8 data. In contrast, the GSWR-XGB model outperforms the traditional model, highlighting its stability and generalization ability. In the MA (Multiple Actions) scenario, the overall accuracy of the model improves, with the GSWR-XGB model significantly outperforming the traditional model.

[0049] Figure 6 The performance of the GSWR-XGB model and the MA method are evaluated, including: (a) evaluating the effectiveness of MA using R²; (b) evaluating the effectiveness of MA using RMSE; (c) evaluating the effectiveness of MA using MAE; (d) evaluating the effectiveness of the GSWR-XGB model using R²; (e) evaluating the effectiveness of the GSWR-XGB model using RMSE; and (f) evaluating the effectiveness of the GSWR-XGB model using MAE. The first three subplots evaluate the effectiveness of MA using metrics of the best model (GSWR-XGB) in each method. The latter three subplots evaluate the effectiveness of the GSWR-XGB model using metrics of the base model in MA. The upper half of each subplot represents the metric type, the x-axis represents the item, and the y-axis represents the metric value.

[0050] Specifically: (1) In the SA method, the nonlinear model performed well in the modeling results of Zhuhai No. 1 and Landsat 8, especially the GSWR-XGB model, whose R² exceeded 0.7 and RMSE and MAE were less than 0.1. Therefore, the model performance ranking is GSWR-XGB>XGB>GWR. In addition, the soil cadmium estimation model constructed using Zhuhai No. 1 hyperspectral data is more accurate than the model constructed using Landsat 8 multispectral data. (2) In the comparison between the MA method and the SA method ( Figure 6In (a)-(c)), the three indicators R², RMSE and MAE improved by 20%, 25% and 29% respectively. This shows that multi-source augmented data can improve estimation accuracy compared with single-source image data. (3) In the MA method, the nonlinear model also shows better performance. The GSWR-XGB model is compared with the traditional model ( Figure 6 (d)-(f)) The results show that R², RMSE, and MAE were improved by an average of 33%, 135%, and 146%, respectively. The model performance ranking remains GSWR-XGB > XGB > GWR. Overall, this confirms that multi-source augmented data is more effective than single-source image modeling, and that the GSWR-XGB model is applicable and accurate in both single-source and multi-source augmented data modeling.

[0051] Therefore, the spatial distribution map of soil cadmium content was subsequently plotted using the GSWR-XGB model. Figure 7 , Figure 8 , Figure 9 These are spatial distribution maps obtained based on Landsat 8 multispectral imagery, Zhuhai-1 hyperspectral imagery, and multi-source enhanced imagery, respectively. Color depth represents soil cadmium content.

[0052] Estimation results of Landsat 8 multispectral data and Zhuhai-1 hyperspectral data ( Figure 7 and Figure 8 The differences are mainly concentrated in the southwest and central regions. In contrast, such as Figures 7-9 As shown, the differences between the multi-source enhanced data and Landsat 8 and Zhuhai-1 data are mainly concentrated in the central and southwestern regions, respectively. A comprehensive analysis of the three maps indicates that the Zhuhai-1 data may contain errors in estimating the southwestern region, possibly due to differences and limitations in its spectral coverage. Multi-source enhanced imagery effectively mitigates these issues. Furthermore, existing research indicates that areas with severe cadmium pollution in the study area are mainly distributed in the central and southwestern regions. In contrast, other areas have lower cadmium levels. This finding is consistent with the results of this study. Figure 9 This further demonstrates the effectiveness of the multi-source enhancement method.

[0053] On the other hand, embodiments of the present invention also disclose a soil heavy metal remote sensing estimation system based on the GSWR-XGB model, used to implement the aforementioned soil heavy metal remote sensing estimation method based on the GSWR-XGB model, comprising: The data acquisition module is used to acquire multi-source remote sensing images of the cultivated area, and at the same time collect soil samples in the cultivated area and screen out bare soil samples. The feature acquisition module is used to acquire the multi-source enhanced features of the bare soil sample; The model prediction module is used to obtain the heavy metal content of each bare soil sample using a pre-trained GSWR-XGB model. The GSWR-XGB model generates geospatial information weights for the bare soil sample points through the GSWR mechanism. The multi-source enhancement features and the geospatial information weights are input into the XGB model together, and the heavy metal content of each bare soil sample is output. The spatial interpolation module is used to obtain a spatial distribution map of heavy metal content in the soil of cultivated areas by interpolating the heavy metal content of each bare soil sample through Kriging space. The model training module is used to take the multi-source enhancement features and geospatial information weights of the bare soil sample as input, and the measured heavy metal content of the bare soil sample as the training target. The model parameters are optimized through cross-validation and random search to obtain the trained GSWR-XGB model.

[0054] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0055] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing estimation method for heavy metals in soil based on the GSWR-XGB model, characterized in that, include: Acquire multi-source remote sensing images of the cultivated area, and collect soil samples within the cultivated area and screen out bare soil samples. Obtain the multi-source enhancement features of the bare soil sample; The heavy metal content of each bare soil sample is obtained using a pre-trained GSWR-XGB model. The GSWR-XGB model generates geospatial information weights for the bare soil sample points through the GSWR mechanism. The multi-source enhanced features and the geospatial information weights are input into the XGB model to output the heavy metal content of each bare soil sample. The heavy metal content of each bare soil sample was obtained by Kriging spatial interpolation to obtain a spatial distribution map of heavy metal content in the cultivated area.

2. The method for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model according to claim 1, characterized in that, The multi-source remote sensing images include Zhuhai-1 hyperspectral images and Landsat 8 multispectral images.

3. The method for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model according to claim 2, characterized in that, Soil samples were collected from the cultivated area, and bare soil samples were screened out, including: Soil samples that meet the preset vegetation index threshold are selected as bare soil samples based on the vegetation index of the corresponding pixels in the Landsat 8 multispectral image. The vegetation indexes include the Normalized Difference Vegetation Index, the Soil-Adjusted Vegetation Index, and the Improved Soil-Adjusted Vegetation Index.

4. The method for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model according to claim 2, characterized in that, The multi-source enhancement features include the spectral features of the Zhuhai-1 hyperspectral image obtained through correlation analysis and the spectral features of the Landsat 8 multispectral image.

5. The method for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model according to claim 1, characterized in that, Geospatial information weights for bare soil sample points are generated using the GSWR mechanism, including: Obtain the geographic coordinates and elevation data of the bare soil sample points, and establish a three-dimensional rectangular coordinate system; Multiple center points were randomly selected, and the weighted Euclidean distance between each bare soil sample point and each center point was calculated. The weighted Euclidean distance is converted into geospatial information weights using a Gaussian kernel function.

6. The method for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model according to claim 5, characterized in that, Weighted Euclidean distance The calculation formula is as follows: ； In the formula, and These represent the sampling point and the center point at the th... Coordinates in spatial dimensions These are the weight parameters.

7. The method for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model according to claim 6, characterized in that, Geospatial Information Weight The conversion formula is as follows: ； In the formula, This is the bandwidth parameter.

8. The method for remote sensing estimation of heavy metals in soil based on the GSWR-XGB model according to claim 5, characterized in that, It also includes model training steps, including: Using the multi-source enhancement features and geospatial information weights of bare soil samples as inputs to the XGB model, and the measured soil heavy metal content as the target, the model parameters are optimized through cross-validation and random search to obtain the trained GSWR-XGB model.

9. A soil heavy metal remote sensing estimation system based on the GSWR-XGB model, used to implement the soil heavy metal remote sensing estimation method based on the GSWR-XGB model as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire multi-source remote sensing images of the cultivated area, and at the same time collect soil samples in the cultivated area and screen out bare soil samples. The feature acquisition module is used to acquire the multi-source enhanced features of the bare soil sample; The model prediction module is used to obtain the heavy metal content of each bare soil sample using a pre-trained GSWR-XGB model. The GSWR-XGB model generates geospatial information weights for the bare soil sample points through the GSWR mechanism. The multi-source enhancement features and the geospatial information weights are input into the XGB model together, and the heavy metal content of each bare soil sample is output. The spatial interpolation module is used to obtain a spatial distribution map of heavy metal content in the soil of cultivated areas by interpolating the heavy metal content of each bare soil sample using Kriging spatial interpolation.

10. A soil heavy metal remote sensing estimation system based on the GSWR-XGB model according to claim 9, characterized in that, It also includes a model training module, which takes the multi-source enhancement features and geospatial information weights of the bare soil sample as input, the measured heavy metal content of the bare soil sample as the training target, and optimizes the model parameters through cross-validation and random search to obtain the trained GSWR-XGB model.

Images