Soil heavy metal inversion method and system integrating satellite remote sensing and near-end sensing

Abstract

The invention discloses a soil heavy metal inversion method and system integrating satellite remote sensing and near-end sensing, and the method comprises the steps: collecting a soil sample, and measuring the soil heavy metal content and a soil visible light-near infrared spectrum; obtaining a time sequence multispectral image of a research area, calculating a spectral index, and selecting and screening bare soil pixels through a threshold value to obtain a bare soil image; performing spectrum correction on the bare soil image; obtaining a joint dictionary and a sparse coefficient through sparse representation and dictionary learning, and reconstructing a hyperspectral image of the bare soil image; converting the one-dimensional spectral data into a two-dimensional spectrogram by using continuous wavelet transform, extracting spectral features in combination with a 2D-CNN algorithm, and constructing a soil heavy metal inversion model; and using the trained inversion model to predict the soil heavy metal content of the research area based on the reconstructed hyperspectral image. According to the method, satellite remote sensing and near-end sensing are integrated to obtain a large-scale accurate soil heavy metal content distribution map, deep features are extracted in combination with a 2D-CNN algorithm, and the inversion model precision and model efficiency are improved.

Description

Technical Field

[0001] This invention belongs to the field of soil information extraction, and more specifically, relates to a method and system for inverting heavy metals in soil by integrating satellite remote sensing and near-end sensing. Background Technology

[0002] Soil, as a vital ecological and material foundation for social development, plays a crucial role in mitigating the impacts of global climate change, protecting ecosystems, safeguarding human health, and ensuring food security. However, intensified human activities, including rapid urbanization, industrialization, and agricultural development, have threatened soil environmental safety, leading to continuous and worsening soil pollution. Heavy metal pollution in soil is particularly prominent. Heavy metals are highly toxic, widely distributed, and have long half-lives. They can migrate to agricultural products through bioaccumulation and then enter the human body through the food chain, posing a long-term threat to food security and human health. Therefore, accurate assessment of heavy metal content in soil is essential for soil environmental safety monitoring, the development of a global carbon cycle, and sustainable agricultural development.

[0003] Currently, soil heavy metal content determination still relies primarily on traditional methods, which are complex, time-consuming, and labor-intensive, making large-area monitoring impossible. Remote sensing technology is time-saving and labor-saving, and possesses rich spectral information, facilitating the acquisition of more potential ground cover information and providing a new and effective means for soil heavy metal monitoring. However, multispectral satellite spectral resolution is insufficient, and the spectral channels of sensors are limited, lacking key spectral bands for estimating heavy metals and making it difficult to extract weak features related to soil heavy metals. While near-end sensing laboratory spectra can provide rich spectral information, they are limited by sampling points, making it impossible to map the distribution of soil heavy metals. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method and system for soil heavy metal inversion that integrates satellite remote sensing and near-end sensing. By integrating satellite remote sensing and near-end sensing, a large-scale and accurate distribution map of soil heavy metal content is obtained, and deep features are extracted by combining a 2D-CNN model, further improving the accuracy and efficiency of the inversion model.

[0005] According to a first aspect of the present invention, a method for inverting heavy metals in soil by integrating satellite remote sensing and near-end sensing is provided, comprising the following steps: Soil samples were collected, and the content of heavy metals in the soil and the visible-near-infrared spectral data of the soil were measured and preprocessed. Time-series multispectral images of the study area were acquired, and the normalized vegetation index, normalized fire index, and normalized soil index were calculated. Bare soil pixels were selected by thresholding to obtain bare soil images. The spectrum of the bare soil image is extracted, and the spectral correction of the bare soil image is performed based on the laboratory spectrum using a direct normalization algorithm. A joint dictionary and sparse coefficients are obtained through sparse representation and dictionary learning. The multispectral bare soil image is then reconstructed using the joint dictionary and sparse coefficients to obtain the reconstructed hyperspectral image. One-dimensional spectral data was converted into two-dimensional spectrograms using continuous wavelet transform. A 2D-CNN model was then used to extract spectral features and construct a soil heavy metal inversion model. The trained inversion model was then used to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

[0006] Based on the above technical solution, the present invention can also be improved as follows.

[0007] Optionally, the measurement of soil heavy metal content includes: After collecting soil samples, the samples were air-dried, sieved, and ground. The heavy metal content in the soil was measured using the ICP-MS method.

[0008] Optional, measurements of soil visible-near-infrared spectral data include: The visible-near-infrared spectrum of soil was measured using a portable ground object spectrometer. Multiple reflectance spectral curves were measured for each soil sample. After removing outliers from the reflectance spectral curves, the arithmetic mean was performed to obtain the actual reflectance spectral data of the soil sample. The 350-399nm and 2401-2500nm spectral bands with large noise were removed. The SG filtering algorithm was used to smooth the soil spectrum.

[0009] Optionally, the acquisition of time-series multispectral images of the study area and the calculation of the spectral indices required for extracting bare soil images include: Time-series multispectral images of the study area were acquired, and radiometric calibration, atmospheric correction, and cloud removal were performed on the spectral images. Spectral indices required for bare soil images were calculated and extracted from the processed images, including: Normalized Difference Vegetation Index (NDVI), Normalized Burn Index (NDI), and Normalized Soil Index (NSSI). Among them, the NDVI reflects the land cover vegetation status, the NDI is used to detect crop residue cover in farmland, and a higher NDI value indicates higher crop residue cover in farmland; the NDI reflects the information of bare soil on the land surface.

[0010] Optionally, the step of selecting and filtering bare soil pixels by threshold to obtain a bare soil image includes: Regions with thresholds of 0 < spectral index, normalized vegetation index < 0.25, normalized combustion index < 0.075, and normalized soil index > 0 are marked as bare soil pixels with no vegetation cover and little straw residue; the median reflectance of each pixel in each band is calculated, and finally a stable image of bare soil is synthesized.

[0011] Optionally, the extraction of the spectrum from the bare soil image and the spectral correction of the bare soil image based on the laboratory spectrum using a direct normalization algorithm include: Based on the spectral response function, laboratory spectral data is resampled to obtain multispectral data that matches the bands of bare soil images. Select spectral transfer sets with different sample sizes from the data, select subsets of resampled laboratory spectral and image spectral data, and use a direct normalization algorithm to determine the spectral transfer matrix; After converting the residual matrix into a baseline difference matrix and averaging and centering the resampled laboratory spectrum and bare soil image spectrum, the bias transition matrix is ​​solved using partial least squares. Based on the bias transition matrix, the residual matrix is ​​calculated.

[0012] Optionally, the method of obtaining the joint dictionary through sparse representation and dictionary learning includes: The multispectral and hyperspectral data were normalized separately to construct a joint spectral data matrix; Initialize the number of dictionary atoms, and then initialize the hyperspectral dictionary and multispectral dictionary based on the number of dictionary atoms; After normalizing the multispectral dictionary and the hyperspectral dictionary respectively, a joint dictionary is constructed; Initialize sparsity constraints and iteration counts, use the K-SVD algorithm in the hyperspectral space to train the joint dictionary, and separate the multispectral dictionary and hyperspectral dictionary from the joint dictionary.

[0013] Optionally, the reconstructed hyperspectral image is obtained by reconstructing the multispectral bare soil image using a joint dictionary and sparse coefficients. This includes: based on the shared sparsity coefficients of the hyperspectral and multispectral dictionaries, the reconstructed hyperspectral image is obtained using the following formula:

[0014] Where χ represents the hyperspectral image; Represented as sparsity coefficients; Represented as a hyperspectral dictionary.

[0015] Optionally, the step of using continuous wavelet transform to convert one-dimensional spectral data into a two-dimensional spectrogram, combining it with a 2D-CNN model to extract spectral features and construct a soil heavy metal inversion model, and using the trained inversion model to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image includes: The Mexican hat wavelet function is used as the basis function. The decomposition scale is selected to perform multi-scale decomposition on the reconstructed hyperspectral data based on bare soil image to obtain wavelet coefficients at multiple scales. The scale wavelet coefficients are then connected in parallel row by row through the scale dimension fusion method to construct a two-dimensional wavelet coefficient matrix. Before training, the Kennard-Stone algorithm was used to divide the sample data into training and validation sets proportionally. The two-dimensional wavelet coefficient matrix was used as input, and deep feature extraction was performed through a 2D-CNN model. The root mean square error was used as the loss function in the optimization process. The optimal model obtained from the training was applied to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

[0016] According to a second aspect of the present invention, a soil heavy metal retrieval system integrating satellite remote sensing and near-end sensing is provided, comprising: The sample collection module is used to collect soil samples, measure the content of heavy metals in the soil samples and the visible-near-infrared spectral data of the soil, and perform preprocessing. The bare soil image acquisition module is used to acquire time-series multispectral images of the study area, calculate the spectral indices required for extracting bare soil images, select and filter bare soil pixels through thresholds, and obtain bare soil images. The spectral correction module is used to extract the spectrum from the bare soil image and perform spectral correction on the bare soil image based on the laboratory spectrum using a direct normalization algorithm; The hyperspectral image reconstruction module is used to obtain a joint dictionary and sparse coefficients through sparse representation and dictionary learning, and to reconstruct the multispectral bare soil image using the joint dictionary and sparse coefficients to obtain the reconstructed hyperspectral image. The soil heavy metal inversion module is used to convert one-dimensional spectral data into two-dimensional spectrograms using continuous wavelet transform, and to extract spectral features using a 2D-CNN model to construct a soil heavy metal inversion model. The trained inversion model is then used to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

[0017] The technical effects and advantages of this invention are as follows: This invention provides a method and system for inverting soil heavy metals by integrating satellite remote sensing and near-end sensing. This method reconstructs hyperspectral images in the 400-2400 nm range by integrating satellite remote sensing and near-end sensing, solving the problem of insufficient spectral resolution and improving model stability. Furthermore, this invention overcomes the problem of insufficient spectral resolution in existing multispectral data inversion of soil heavy metals. In the soil heavy metal inversion integrating satellite remote sensing and near-end sensing, one-dimensional spectral data is converted into two-dimensional spectrograms. Deep learning methods are used to capture deep features in the spectral data, and a 2D-CNN model is used to extract these deep features, resulting in a large-scale, accurate distribution map of soil heavy metal content, thereby further improving the accuracy and efficiency of the inversion model. Attached Figure Description To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart of a method for inverting heavy metals in soil by integrating satellite remote sensing and near-end sensing, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the accuracy of soil heavy metal As inversion based on an optimal model using reconstructed hyperspectral images, as provided in an embodiment of the present invention. 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] Understandably, given the deficiencies in the background technology, this invention proposes a method for inverting heavy metals in soil by integrating satellite remote sensing and near-end sensing, specifically as follows: Figure 1 As shown, it includes the following steps: Step 1: Collect soil samples, measure the content of heavy metals in the soil and the visible-near-infrared spectral data of the soil in the laboratory, and perform preprocessing. Laboratory measurements of soil heavy metal content in soil samples include: After collecting 155 soil surface samples in the study area, the soil samples were air-dried, sieved, and ground. The content of heavy metal As in the soil was measured using the ICP-MS method.

[0021] Measuring and preprocessing visible-near-infrared spectral data includes: The ASD portable ground object spectrometer was used to measure the reflectance spectrum of each soil sample. Five reflectance spectral curves were obtained for each soil sample. After removing outliers, the arithmetic mean was calculated to obtain the actual reflectance spectral data of the soil sample.

[0022] Due to the presence of noise interference in the spectrum, preprocessing is necessary. Specifically, considering the significant influence of moisture on the near-infrared band edges and their low signal-to-noise ratio, the noisy 350-399nm and 2401-2500nm spectral bands are removed. Furthermore, to reduce noise during measurement, the Savizky-Golay (SG) filter method is used to smooth the soil spectrum, with a polynomial order of 2 and a window size of 7nm.

[0023] Step 2: Acquire time-series multispectral images of the study area. After radiometric calibration, atmospheric correction and cloud removal, calculate spectral indices including normalized vegetation index, normalized combustion index and normalized soil index. Select bare soil pixels by threshold selection to obtain bare soil images. In this embodiment, all available L2A-level Sentinel-2 SR (surface reflectance) data for the study area were collected using the Google Earth Engine (GEE) platform. This data has undergone radiometric calibration and atmospheric correction. Images with cloud cover less than 50% were selected. For each image, Cloud Probability data was used to obtain the probability of each pixel being covered by clouds, and a threshold of 30 was set for cloud removal processing.

[0024] The spectral indices required for extracting bare soil images were calculated from the processed images, including the Normalized Difference Vegetation Index (NDVI), Normalized Burn Index (NBR2), and Normalized Bare Soil Index (NDSI). Normalized Difference Vegetation Index (NDVI) reflects the vegetation status of land cover. A higher NDVI value indicates higher vegetation cover. Therefore, NDVI is used to mask vegetation. In this invention, based on the specific vegetation cover of the study area and previous experience, a threshold of 0.25 is set, meaning pixels with an NDVI greater than 0.25 are considered to have vegetation cover. The NDVI calculation formula is: (1) In the formula, and These represent the reflectance of the near-infrared and red bands of the Sentinel-2 image, respectively, corresponding to bands 8 and 4 of the Sentinel-2 image.

[0025] Normalized Burning Index (NBR2) is used to detect crop residue (straw) cover in farmland. A higher NBR2 value indicates higher crop residue cover. Therefore, crop residue is masked based on NBR2. In this invention, a threshold of 0.075 is set, meaning pixels with an NBR2 greater than 0.075 are considered to have crop residue cover. The NBR2 calculation formula is: (2) In the formula, and respectively represent the near-infrared band and short-wave infrared reflectance corresponding to the Sentinel-2 image, corresponding to bands 8 and 12 of the Sentinel-2 image respectively.

[0026] The Normalized Bare Soil Index (NDSI) reflects the bare soil information on the ground surface. This index can enhance the bare soil information and effectively extract bare soil pixels. Its calculation formula is: (3) In the formula, and respectively represent the short-wave infrared band and near-infrared reflectance corresponding to the Sentinel-2 image, corresponding to bands 11 and 8 of the Sentinel-2 image respectively.

[0027] Among them, the areas with thresholds of 0 < NDVI < 0.25, NBR2 < 0.075, and NDSI > 0 are marked as bare soil pixels with no vegetation cover and less straw residue. Considering that the heavy metal content will not change significantly in the short term, all eligible bare soil pixels in the study area from 2019 to 2023 were collected. The median reflectance of each pixel in each band was calculated. Finally, stable bare soil pixels were synthesized.

[0028] Step 3, extract the spectrum of the bare soil image, and perform spectral correction on the bare soil image based on the laboratory spectrum using the direct standardization algorithm; The performing spectral correction on the bare soil image based on the laboratory spectrum using the direct standardization algorithm includes: First, based on the official spectral response function given by the Sentinel-2 satellite, resample the laboratory spectrum to the bands corresponding to the Sentinel-2 image according to formula (4): (4) In the formula, is the resampled laboratory spectrum, R T is the spectral response function of the Sentinel-2 image, , m is the total number of bands of the multispectral data, k is the total number of bands of the hyperspectral data, is the laboratory spectrum, is the spectral response function corresponding to the λ-th band.

[0029] Secondly, based on the KS algorithm, the entire sample set was divided into several spectral transfer sample sets with varying numbers of samples (t=50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150). A subset of the laboratory spectral and image spectral data was selected, and the spectral transfer matrix was determined using the Direct Standardization (DS) algorithm. This matrix describes the deviation between the laboratory spectra and the bare soil image spectra, and its formula is as follows: (5) In the formula, To resample laboratory spectra, The spectrum of bare soil image, To reflect and The transition matrix of the deviation, This represents the residual matrix.

[0030] Finally, the residual matrix E is converted into the baseline difference matrix and then... and After average centering, the transition matrix B is solved by partial least squares, and then the residual matrix E is calculated based on the transition matrix B. Based on the determined parameter transfer matrix B and residual matrix E, the spectral correction of the bare soil image is completed according to equation (6): (6) In the formula, The image spectrum of the bare soil after correction.

[0031] For each subset with different sample sizes, the present invention also needs to perform the following steps in sequence: Establish the corresponding spectral transfer matrix; The image spectrum is corrected based on the spectral transfer matrix corresponding to each subset, and the cosine similarity θ between the corrected spectrum and the laboratory-measured spectrum is calculated. The performance of the direct normalization algorithm is evaluated, and the optimal subset is determined. The formula for calculating cosine similarity is as follows: (7) In the formula, To resample the laboratory spectrum, Here is the spectrum of the bare soil image, and n is the total number of bands. The image spectrum of the bare soil after correction.

[0032] The cosine similarity reaches its highest value when the sample size is 140. Therefore, this invention uses the transition matrix obtained by fitting 140 samples to correct the spectral density of Sentinel-2 images.

[0033] Step 4: Obtain a joint dictionary and sparse coefficients through sparse representation and dictionary learning; reconstruct the multispectral bare soil image using the joint dictionary and sparse coefficients to obtain the reconstructed hyperspectral image. It should be noted that sparse representation methods can recover signals in high-dimensional linear spaces using only a linear combination of a small number of basis vectors (or atoms) under the constraint of target sparsity. Based on sparse representation theory, the spectrum of hyperspectral data can be represented as: (8) In the formula, This represents the N-band hyperspectral curve. Let K be a set of basis vectors of size K in N-dimensional space, i.e., the hyperspectral dictionary. represents the sparsity coefficients corresponding to the hyperspectral data.

[0034] Multispectral data can be represented as: (9) In the formula, This represents the M-band multispectral curve. Let K be a set of basis vectors of size K in M-dimensional space, i.e., a multispectral dictionary. represents the sparsity coefficients corresponding to the multispectral data.

[0035] Assuming that multispectral data contains a large amount of hyperspectral information, then = The functional model for dictionary learning is defined as follows: (10) The multispectral and hyperspectral data were normalized separately, and then a joint spectral data matrix was constructed. , For the corrected image multispectral, For laboratory spectroscopy. Initialize the dictionary atom count K to 150, and then initialize the hyperspectral dictionary based on this atom count. and multispectral dictionary For the multispectral dictionary respectively and hyperspectral dictionary After normalization, construct a union dictionary. .

[0036] With a sparsity constraint T of 3 and an iteration count of 300, the K-SVD algorithm is used to train a combined dictionary based on the objective of formula (10) in the high multispectral space. The multispectral dictionary is then separated from the combined dictionary. and hyperspectral dictionary ; Based on the multispectral data to be reconstructed After normalizing the multispectral dictionary, the sparse representation coefficients of the multispectral data are obtained using the OMP method based on equation (11). : (11) Based on the assumption that the hyperspectral dictionary and the multispectral dictionary share sparsity coefficients, and combining the obtained sparsity coefficients and the hyperspectral dictionary, the hyperspectral curve is reconstructed using equation (12): (12) Where χ represents the hyperspectral image; Represented as sparsity coefficients; Represented as a hyperspectral dictionary.

[0037] Step 5: Use continuous wavelet transform to convert one-dimensional spectral data into two-dimensional spectrograms, combine 2D-CNN model to extract spectral features to construct soil heavy metal inversion model, and use the trained inversion model to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

[0038] The process involves using continuous wavelet transform to convert one-dimensional spectral data into a two-dimensional spectrogram, combining this with a 2D-CNN model to extract spectral features and construct a soil heavy metal inversion model. The trained inversion model is then used to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image, including: Using the Mexican hat wavelet function as the basis function, multi-scale decomposition was performed on reconstructed hyperspectral data based on bare soil images at decomposition scales of 21, 22, 23, 24, 25, and 26, resulting in wavelet coefficients at six scales. A scale-dimensional fusion method was then used to concatenate the wavelet coefficients from the six scales in parallel rows to construct a two-dimensional wavelet coefficient matrix. , where m represents the total number of bands in the reconstructed hyperspectral data, thereby realizing the conversion from one-dimensional spectral data to two-dimensional spectrogram; Before training, the Kennard-Stone algorithm was used to divide the sample data into training and validation sets in a 7:3 ratio. The two-dimensional wavelet coefficient matrix was used as input, and a 2D-CNN model was employed for deep feature extraction. The root mean square error (MSELoss) was used as the loss function during the optimization process. The accuracy of the optimal model for soil heavy metal As retrieval is as follows: Figure 2 As shown, the optimal model obtained through training is used to predict the soil heavy metal As content of the entire study area based on reconstructed hyperspectral images.

[0039] In summary, this invention provides a method for inverting soil heavy metals by integrating satellite remote sensing and near-end sensing. This method reconstructs hyperspectral images in the 400-2400 nm range by integrating satellite remote sensing and near-end sensing, solving the problem of insufficient spectral resolution and improving model stability. Furthermore, this invention overcomes the problem of insufficient spectral resolution in existing multispectral data inversion of soil heavy metals. In the soil heavy metal inversion integrating satellite remote sensing and near-end sensing, one-dimensional spectral data is converted into two-dimensional spectrograms. Deep learning methods are combined to capture deep features in the spectral data, and a 2D-CNN model is used to extract these deep features, resulting in a large-scale, accurate distribution map of soil heavy metal content, thereby further improving the accuracy and efficiency of the inversion model.

[0040] According to a second aspect of the present invention, a soil heavy metal retrieval system integrating satellite remote sensing and near-end sensing is provided, comprising: The sample collection module is used to collect soil samples, measure the content of heavy metals in the soil samples and the visible-near-infrared spectral data of the soil, and perform preprocessing. The bare soil image acquisition module is used to acquire time-series multispectral images of the study area, calculate the spectral indices required for extracting bare soil images, select and filter bare soil pixels through thresholds, and obtain bare soil images. The spectral correction module is used to extract the spectrum from the bare soil image and perform spectral correction on the bare soil image based on the laboratory spectrum using a direct normalization algorithm; The hyperspectral image reconstruction module is used to obtain a joint dictionary and sparse coefficients through sparse representation and dictionary learning, and to reconstruct the multispectral bare soil image using the joint dictionary and sparse coefficients to obtain the reconstructed hyperspectral image. The soil heavy metal inversion module is used to convert one-dimensional spectral data into two-dimensional spectrograms using continuous wavelet transform, and to extract spectral features using a 2D-CNN model to construct a soil heavy metal inversion model. The trained inversion model is then used to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

[0041] It is understood that the soil heavy metal inversion system integrating satellite remote sensing and near-end sensing provided by the present invention corresponds to the soil heavy metal inversion method integrating satellite remote sensing and near-end sensing provided in the foregoing embodiments. The relevant technical features of the soil heavy metal inversion system integrating satellite remote sensing and near-end sensing can be referred to the relevant technical features of the soil heavy metal inversion method integrating satellite remote sensing and near-end sensing, and will not be repeated here.

[0042] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0043] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0044] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0045] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inverting heavy metals in soil by integrating satellite remote sensing and near-end sensing, characterized in that, Includes the following steps: Soil samples were collected, and the content of heavy metals in the soil and the visible-near-infrared spectral data of the soil were measured and preprocessed. Time-series multispectral images of the study area were acquired, the spectral indices required for extracting bare soil images were calculated, and bare soil pixels were selected and filtered by thresholding to obtain bare soil images. The spectrum of the bare soil image is extracted, and the spectral correction of the bare soil image is performed based on the laboratory spectrum using a direct normalization algorithm. A joint dictionary and sparse coefficients are obtained through sparse representation and dictionary learning. The multispectral bare soil image is then reconstructed using the joint dictionary and sparse coefficients to obtain the reconstructed hyperspectral image. One-dimensional spectral data was converted into two-dimensional spectrograms using continuous wavelet transform. A 2D-CNN model was then used to extract spectral features and construct a soil heavy metal inversion model. The trained inversion model was then used to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

2. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, The measurement of soil heavy metal content includes: After collecting soil samples, the samples were air-dried, sieved, and ground. The heavy metal content in the soil was measured using the ICP-MS method.

3. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, The measurement of soil visible-near-infrared spectral data includes: The visible-near-infrared spectrum of soil was measured using a portable ground object spectrometer. Multiple reflectance spectral curves were measured for each soil sample. After removing outliers from the reflectance spectral curves, the arithmetic mean was performed to obtain the actual reflectance spectral data of the soil sample. The 350-399nm and 2401-2500nm spectral bands with large noise were removed. The SG filtering algorithm was used to smooth the soil spectrum.

4. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, The acquisition of time-series multispectral images of the study area and the calculation of the spectral indices required for extracting bare soil images include: Time-series multispectral images of the study area were acquired, and radiometric calibration, atmospheric correction, and cloud removal were performed on the spectral images. Spectral indices required for bare soil images were calculated and extracted from the processed images, including: Normalized Difference Vegetation Index (NDVI), Normalized Burn Index (NDI), and Normalized Soil Index (NSSI). The NDVI reflects the vegetation cover status of the land; the NDI is used to detect crop residue cover in farmland, with higher NDI values ​​indicating higher crop residue cover; and the NDI reflects information about bare soil on the land surface.

5. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, The step of selecting bare soil pixels by threshold to obtain a bare soil image includes: Regions with thresholds of 0 < spectral index, normalized vegetation index < 0.25, normalized combustion index < 0.075, and normalized soil index > 0 are marked as bare soil pixels with no vegetation cover and little straw residue; the median reflectance of each pixel in each band is calculated, and finally a stable image of bare soil is synthesized.

6. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, The extraction of spectra from bare soil images and the spectral correction of bare soil images based on laboratory spectra using a direct normalization algorithm include: Based on the spectral response function, laboratory spectral data is resampled to obtain multispectral data that matches the bands of bare soil images. Select spectral transfer sets with different sample sizes from the data, select subsets of resampled laboratory spectral and image spectral data, and use a direct normalization algorithm to determine the spectral transfer matrix; After converting the residual matrix into a baseline difference matrix and averaging and centering the resampled laboratory spectrum and bare soil image spectrum, the bias transition matrix is ​​solved using partial least squares. Based on the bias transition matrix, the residual matrix is ​​calculated.

7. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, The joint dictionary obtained through sparse representation and dictionary learning includes: The multispectral and hyperspectral data were normalized separately to construct a joint spectral data matrix; Initialize the number of dictionary atoms, and then initialize the hyperspectral dictionary and multispectral dictionary based on the number of dictionary atoms; After normalizing the multispectral dictionary and the hyperspectral dictionary respectively, a joint dictionary is constructed; Initialize sparsity constraints and iteration counts, use the K-SVD algorithm in the hyperspectral space to train the joint dictionary, and separate the multispectral dictionary and hyperspectral dictionary from the joint dictionary.

8. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, Reconstructing a multispectral bare soil image using a joint dictionary and sparse coefficients yields a reconstructed hyperspectral image. This image is based on the shared sparse coefficients of the hyperspectral and multispectral dictionaries, and is obtained using the following formula: Where χ represents the hyperspectral image; Represented as sparsity coefficients; Represented as a hyperspectral dictionary.

9. The method for soil heavy metal inversion integrating satellite remote sensing and near-end sensing according to claim 1, characterized in that, The process involves using continuous wavelet transform to convert one-dimensional spectral data into a two-dimensional spectrogram, combining this with a 2D-CNN model to extract spectral features and construct a soil heavy metal inversion model. The trained inversion model is then used to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image, including: The Mexican hat wavelet function is used as the basis function. The decomposition scale is selected to perform multi-scale decomposition on the reconstructed hyperspectral data based on bare soil image to obtain wavelet coefficients at multiple scales. The scale wavelet coefficients are then connected in parallel row by row through the scale dimension fusion method to construct a two-dimensional wavelet coefficient matrix. Before training, the Kennard-Stone algorithm was used to divide the sample data into training and validation sets proportionally. The two-dimensional wavelet coefficient matrix was used as input, and deep feature extraction was performed through a 2D-CNN model. The root mean square error was used as the loss function in the optimization process. The optimal model obtained from the training was applied to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

10. A soil heavy metal retrieval system integrating satellite remote sensing and near-end sensing, characterized in that, include: The sample collection module is used to collect soil samples, measure the content of heavy metals in the soil samples and the visible-near-infrared spectral data of the soil, and perform preprocessing. The bare soil image acquisition module is used to acquire time-series multispectral images of the study area, calculate the spectral indices required for extracting bare soil images, select and filter bare soil pixels through thresholds, and obtain bare soil images. The spectral correction module is used to extract the spectrum from the bare soil image and perform spectral correction on the bare soil image based on the laboratory spectrum using a direct normalization algorithm; The hyperspectral image reconstruction module is used to obtain a joint dictionary and sparse coefficients through sparse representation and dictionary learning, and to reconstruct the multispectral bare soil image using the joint dictionary and sparse coefficients to obtain the reconstructed hyperspectral image. The soil heavy metal inversion module is used to convert one-dimensional spectral data into two-dimensional spectrograms using continuous wavelet transform, and to extract spectral features using a 2D-CNN model to construct a soil heavy metal inversion model. The trained inversion model is then used to predict the soil heavy metal content of the entire study area based on the reconstructed hyperspectral image.

Images