A cross-platform inversion method for radiometric correction and chlorophyll content retrieval of multi-source remote sensing data

Abstract

This invention relates to the fields of quantitative remote sensing in agriculture and smart agriculture, and discloses a cross-platform method for radiometric correction and chlorophyll content retrieval from multi-source remote sensing data. This method simultaneously acquires ground-based measured data, UAV (unmanned aerial vehicle) data, and satellite remote sensing data; calculates vegetation indices and uses Pearson correlation analysis to screen sensitive characteristic variables; establishes a radiometric correction model using the ratio-mean method; performs radiometric normalization on satellite data based on UAV data to eliminate sensor response differences; constructs and optimizes a SPAD retrieval model based on Z-score standardization; applies this model to the corrected satellite data; and generates regional-scale chlorophyll distribution maps pixel by pixel, thereby formulating differentiated farmland management strategies. This invention solves the difficulties of multi-source data coordination, improves the accuracy of regional-scale crop chlorophyll monitoring, and provides scientific decision support for large-scale crop fertilization.

Description

Technical Field

[0001] This invention relates to the fields of agricultural quantitative remote sensing and smart agriculture technology, specifically a method for cross-platform inversion of radiometric correction and chlorophyll content from multi-source remote sensing data. Background Technology

[0002] As a vital global food crop, maize's growth status directly impacts food security and bioenergy supply. Chlorophyll content is a key physiological indicator characterizing crop photosynthetic capacity and nitrogen nutrition status. Rapid and accurate monitoring of the spatiotemporal dynamics of chlorophyll content is crucial for guiding field fertilization, integrated water and fertilizer management, and crop yield prediction. With the development of remote sensing technology, using multispectral imagery for physicochemical parameter inversion has become a mainstream method, but limitations still exist in practical applications.

[0003] Existing remote sensing technologies for crop chlorophyll monitoring rely on a single remote sensing platform, making it difficult to achieve a balance between monitoring range and spatial resolution. While satellite remote sensing offers wide-area coverage, its long revisit cycles and low spatial resolution make it prone to pixel mixing effects when monitoring fragmented plots or field details, leading to insufficient monitoring accuracy. Low-altitude UAV remote sensing, while possessing high spatial resolution and capable of capturing canopy information, has limited coverage per operation, making it difficult to directly extend to regional-scale macro-monitoring. Furthermore, multispectral remote sensing data contains high-dimensional spectral information; if directly used for modeling, strong correlations between bands can lead to severe information redundancy and multicollinearity problems, thus affecting the computational efficiency and predictive robustness of the inversion model.

[0004] When attempting collaborative monitoring using multi-source data from air and space, spectral inconsistencies between sensors on different platforms are a key bottleneck restricting the application of the technology. Because sensors on UAVs and satellites have different spectral response functions, and satellite data is easily affected by atmospheric transmission path radiation and changes in illumination conditions, there are systematic deviations in the spectral reflectance observations of the same ground feature on different platforms. This deviation causes the accuracy of optimized models trained on high-precision near-ground data (such as UAVs or ground-based measurements) to decrease when directly applied to satellite imagery. This prevents the model from generalizing across different spatial scales and makes it difficult to support continuous monitoring of large areas using high-precision data from small areas.

[0005] Furthermore, current research largely focuses on the construction and accuracy evaluation of chlorophyll content inversion algorithms, stopping at generating crop growth monitoring maps and lacking a supporting mechanism to translate monitoring results into specific agronomic management decisions. The existing technological system has not yet established an automated mapping logic from chlorophyll monitoring values ​​to differentiated fertilization rates, and cannot directly generate variable fertilization prescription maps to guide actual production based on the inversion results. This means that high-precision remote sensing data can only serve as an observation record, unable to directly empower refined farmland management, thus limiting the practical application value of remote sensing technology in achieving agricultural fertilizer reduction and efficiency improvement, and precise resource input. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a cross-platform inversion method for radiometric correction and chlorophyll content retrieval of multi-source remote sensing data. This method solves the problems that existing single remote sensing monitoring methods cannot simultaneously achieve high spatial resolution and large regional coverage, the differences in spectral response of different sensors lead to low cross-platform transfer accuracy of inversion models, and the difficulty in directly converting remote sensing monitoring results into differentiated farmland management decisions.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a cross-platform inversion method for radiometric correction and chlorophyll content retrieval of multi-source remote sensing data, comprising the following steps:

[0008] First, data acquisition and preprocessing are performed. Ground-based measured data, UAV remote sensing data, and satellite remote sensing data are acquired simultaneously within the target area. Based on the band settings of the UAV remote sensing data, corresponding bands with overlapping spectral ranges are selected from the satellite remote sensing data to establish the correspondence between the spectral channels of the multi-source data.

[0009] Preferably, the corresponding bands cover the blue, green, red, first red edge, second red edge, and near-infrared spectral regions; the satellite remote sensing data uses Sentinel-2A satellite multispectral imagery, and mapping relationships are established between each band of the satellite imagery and the corresponding bands of the UAV imagery.

[0010] Secondly, spectral feature parameters were constructed and screened. Canopy spectral reflectance and vegetation indices were calculated based on UAV remote sensing data. These vegetation indices included green chlorophyll index, canopy structure resistance index, normalized difference vegetation index, normalized pigment chlorophyll index, ratio vegetation index, conversion vegetation index, normalized difference water index, soil-regulated vegetation index, and difference vegetation index. Pearson correlation analysis was used to analyze the correlation between the band reflectance and vegetation indices of the UAV remote sensing data and the ground-measured data. Specifically, the Pearson product-moment correlation coefficient was used to quantitatively describe the linear dependence between each feature variable to be screened and the ground-measured data. All feature variables to be screened were ranked according to the absolute value of the correlation coefficient, and the top few variables with the highest correlation to the ground-measured data were selected as sensitive feature variables based on the ranking results.

[0011] Secondly, cross-platform spectral consistency radiometric correction is performed to eliminate systematic biases among multi-source sensors. A radiometric correction model using UAV remote sensing data as a benchmark is established using the ratio-mean method. The specific implementation logic is as follows: Within the radiometric correction sample area, the UAV digital orthophotos are pixel-aggregated to generate UAV simulated reflectance data that matches the pixel scale of the satellite remote sensing data. For each corresponding band, the arithmetic mean of the ratios of the UAV simulated reflectance values ​​to the original satellite image reflectance values ​​for all samples within the radiometric correction sample area is calculated by summing the ratios and dividing by the total number of samples. This yields the correction coefficient. Using this correction coefficient, a pixel-by-pixel linear transformation is performed on the satellite remote sensing data for the entire region, mapping the satellite remote sensing reflectance observations to the radiometric scale of the UAV remote sensing data, thus completing the radiometric normalization process.

[0012] Subsequently, the inversion model was constructed and evaluated. Sensitive feature variables were selected as input variables, and ground-measured data were used as response variables to construct a SPAD (chlorophyll content) inversion model. Before modeling, Z-score standardization was performed to eliminate dimensional differences between feature variables. The specific calculation logic of this standardization process is as follows: the arithmetic mean of each feature variable in the modeling dataset is calculated by summing the feature variable values ​​of all samples and dividing by the total number of samples. The standard deviation is calculated by summing the squares of the differences between the feature variable values ​​of each sample and the arithmetic mean, dividing the sum by the total number of samples, and taking the square root. The arithmetic mean and standard deviation are then used to dimensionlessly process the feature variables in the modeling dataset, validation set, and subsequent regional mapping data. When constructing the model, one or more of the following algorithms are used: partial least squares regression, random forest regression, and backpropagation neural network. Among them, the partial least squares regression algorithm establishes a linear regression equation by projecting high-dimensional feature variables onto a low-dimensional latent variable space; the random forest regression algorithm performs ensemble prediction by constructing multiple independent decision trees; and the backpropagation neural network algorithm simulates nonlinear mapping relationships using a topological structure containing input layers, hidden layers, and output layers. Statistical evaluation indicators (coefficient of determination, root mean square error, and mean absolute error) are used to assess model accuracy, and the model with the largest coefficient of determination and the smallest root mean square error and mean absolute error is selected as the optimal SPAD inversion model.

[0013] Finally, regional-scale mapping and application are performed. The accuracy-validated SPAD inversion model is transferred and applied to the radiometrically normalized satellite remote sensing data. The specific process includes masking and cropping the satellite remote sensing data using a vector boundary file of the maize planting area, constructing the image matrix to be inverted, extracting sensitive feature variables for each valid pixel and standardizing them, inputting the optimal SPAD inversion model to calculate the SPAD prediction value, and generating a regional-scale spatial distribution map of maize chlorophyll content.

[0014] Based on the spatial distribution map, this invention further generates differentiated farmland management strategies. First, the target SPAD reference value for the growth stage is determined. Then, the recommended pure nitrogen application rate for each pixel is calculated. The calculation logic is as follows: calculate the difference between the target SPAD reference value and the predicted SPAD value, and multiply this difference by a nitrogen sensitivity coefficient and a fertilizer demand discrimination coefficient. The fertilizer demand discrimination coefficient is a logical control variable; it takes a value of 1 when the target SPAD reference value is greater than the predicted SPAD value, indicating a nutrient deficit requiring topdressing; and a value of 0 when the target SPAD reference value is less than or equal to the predicted SPAD value, indicating sufficient nutrients and no need for topdressing. Finally, the distribution range of the recommended pure nitrogen application rate is divided into discrete management levels using the natural breakpoint method, generating a variable fertilizer prescription map.

[0015] This invention provides a cross-platform method for radiometric correction and chlorophyll content retrieval from multi-source remote sensing data. It offers the following advantages:

[0016] 1. This invention solves the contradiction between spatial resolution and coverage of a single remote sensing platform by constructing a collaborative inversion framework of air, space, and ground data and using Pearson correlation analysis to select sensitive feature variables. This method utilizes the high spatial resolution advantage of UAV low-altitude remote sensing to compensate for the mixed pixel defects of satellite remote sensing. At the same time, it improves the reliability and prediction accuracy of the chlorophyll content inversion model by eliminating redundant information and collinear interference from multispectral data through feature screening.

[0017] 2. This invention proposes a cross-platform spectral consistency radiometric correction technique based on the ratio mean method. It uses the UAV image after pixel aggregation as a radiometric reference to normalize the satellite image, thereby eliminating the differences in spectral response functions between different source sensors and the radiometric errors caused by atmospheric transmission paths. It solves the problem of the decrease in the accuracy of the inversion model when it is transferred between different remote sensing platforms, and enables the inversion model built based on small-area UAV data to be extended to large-area satellite remote sensing monitoring.

[0018] 3. By setting a target SPAD reference value and introducing a fertilization demand discrimination coefficient, this invention can automatically calculate the recommended fertilization amount and generate a variable fertilization prescription map. This method directly transforms quantitative remote sensing monitoring data into specific agronomic management schemes, realizes pixel-level diagnosis and discrete hierarchical management of nutrient deficiency in farmland, improves nitrogen fertilizer utilization efficiency, and provides data support for regional-scale agricultural production. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall process of a cross-platform inversion method for radiometric correction and chlorophyll content retrieval of multi-source remote sensing data according to the present invention.

[0020] Figure 2 This is a detailed flowchart of the cross-platform spectral consistency radiometric correction method in this invention;

[0021] Figure 3 This is a detailed flowchart of the construction, optimization, and evaluation of the SPAD inversion model in this invention;

[0022] Figure 4 This is a flowchart illustrating the logic of generating differentiated farmland management strategies based on SPAD prediction values ​​in this invention. Detailed Implementation

[0023] The technical solutions in 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.

[0024] See attached document Figure 1 This invention provides a cross-platform inversion method for radiometric correction and chlorophyll content retrieval of multi-source remote sensing data, the method comprising the following steps:

[0025] First, the data acquisition and preprocessing step S1 is performed. In this step, according to the preset sampling scheme, ground-based measured data, UAV remote sensing data, and satellite remote sensing data are simultaneously acquired within the target area. The ground-based measured data consists of SPAD values ​​of leaf chlorophyll content collected during the critical growth period of maize. The UAV remote sensing data consists of low-altitude high-resolution images acquired by a UAV platform equipped with a multispectral imaging system, which are then processed through photogrammetry to generate digital orthophoto maps. The satellite remote sensing data consists of Sentinel-2A satellite multispectral images corresponding to the acquisition time of the UAV remote sensing data. After acquiring the satellite remote sensing data, based on the band settings of the UAV remote sensing data, corresponding bands with overlapping spectral ranges are selected from the satellite remote sensing data. These corresponding bands cover the blue, green, red, red-edge, and near-infrared spectral regions.

[0026] Next, step S2, spectral feature parameter construction and screening, is performed. In this step, the canopy spectral reflectance corresponding to each sampling point is extracted based on the digital orthophoto map, and various vegetation indices are calculated using the spectral reflectance. These vegetation indices include the green chlorophyll index, canopy structure resistance index, normalized difference vegetation index, normalized pigment chlorophyll index, ratio vegetation index, conversion vegetation index, normalized difference water index, soil-regulated vegetation index, and difference vegetation index. After calculating the vegetation indices, Pearson correlation analysis is used to calculate the linear correlation between the reflectance of each band and each vegetation index and the ground-measured data, and sensitive feature variables are screened based on the correlation analysis results.

[0027] Subsequently, a cross-platform spectral consistency radiometric correction step S3 is performed. This step aims to eliminate spectral inconsistencies between UAV remote sensing data and satellite remote sensing data caused by differences in sensor response and atmospheric transport paths. This step uses the ratio-mean method, with the reflectance of the UAV remote sensing data as a benchmark, to calculate correction coefficients for corresponding bands of the UAV and satellite remote sensing data. The calculated correction coefficients are then used to perform radiometric normalization on the corresponding bands of the Sentinel-2A satellite imagery, ensuring that the corrected satellite image reflectance and the UAV image reflectance are on the same metric scale.

[0028] Next, step S4, the inversion model construction and evaluation, is performed. In this step, the selected sensitive feature variables are used as input variables, and the ground-measured data are used as response variables to construct the SPAD inversion model. The algorithm used to construct the inversion model includes at least one of partial least squares regression, random forest regression, or backpropagation neural network. After the model is constructed, statistical evaluation indicators are used to assess the model's prediction accuracy and stability. The evaluation indicators include the coefficient of determination, root mean square error, and mean absolute error.

[0029] Finally, step S5, regional-scale mapping and application, is performed. In this step, the optimal SPAD inversion model, validated for accuracy, is transferred and applied to the radiometrically corrected Sentinel-2A satellite imagery. SPAD predictions are calculated pixel-by-pixel, generating a regional-scale spatial distribution map of maize chlorophyll content. Based on the high and low chlorophyll content distribution characteristics reflected in the spatial distribution map, differentiated farmland management strategies are generated. For areas with high chlorophyll content, fertilization recommendations to maintain the current nutrient status are generated; for areas with low chlorophyll content, fertilization recommendations to increase water and fertilizer input are generated, achieving precision agricultural management within the region.

[0030] In the specific implementation of the experimental area setup and sampling strategy, the target agricultural planting area was first spatially divided, defining the core experimental area A for model construction, and the first independent validation area B and the second independent validation area C for evaluating the model's cross-spatial generalization performance. The core experimental area A, the first independent validation area B, and the second independent validation area C do not overlap spatially to ensure the independence of the validation data. Within the core experimental area A, 66 fixed monitoring points were deployed using a stratified random sampling method; within the first independent validation area B and the second independent validation area C, 48 fixed monitoring points were deployed respectively. The spatial distribution of all monitoring points must cover the edge and central areas of the field to ensure that the sampling data includes the heterogeneous characteristics of soil and crops under different growth environments, thereby improving the spatial representativeness of the samples.

[0031] To optimize the growth cycle of maize, four key growth stages were selected as data collection points: seedling stage, jointing stage, tasseling stage, and silking stage. At each monitoring point, three maize plants with uniform growth and no signs of disease or pests were selected as observation subjects. For each selected maize plant, the third fully unfolded leaf was located, and the measurement area was selected from 30% to 70% of the leaf base. Using a portable chlorophyll meter, five measurement points were symmetrically selected along both sides of the leaf midrib within the measurement area. After measurement, the arithmetic mean of the five measurement points was calculated as the SPAD value of the single maize plant; then, the arithmetic mean of the SPAD values ​​of the three maize plants was calculated, and this arithmetic mean was used as the measured relative chlorophyll content of the leaves at the current growth stage of the monitoring point.

[0032] Before performing the measurement operations, the portable chlorophyll meter was calibrated using a standard calibration plate to ensure the consistency of all measurement data benchmarks. After data acquisition, the raw data was cleaned to remove outliers caused by leaf physical damage or sudden environmental disturbances, resulting in a standard ground-based measured dataset for subsequent modeling and validation. During the model building phase, the dataset collected in the core experimental area A was randomly divided into a 2:3 ratio as the modeling training set, with the remaining 1 / 3 used as the internal validation set. All data collected in the first independent validation area B and the second independent validation area C were used as the external independent test set to verify the applicability of the inversion model in different spatial domains.

[0033] In the specific implementation of low-altitude remote sensing data acquisition by UAVs, a quadcopter UAV was selected as the flight platform, and a multispectral imaging system was carried out to perform aerial surveying tasks. The multispectral imaging system was configured to simultaneously capture spectral information in the blue, green, red, red-edge, and near-infrared bands.

[0034] To ensure the stability and consistency of imaging radiometric quality, all aerial survey missions are conducted under clear, unclouded weather conditions with wind speeds less than 3 m / s. The flight operation window is strictly limited to 11:00 AM to 1:00 PM daily to maintain the solar altitude angle within 60° ± 5°, thereby ensuring uniform illumination of the crop canopy and minimizing the interference of shadows on spectral reflectance.

[0035] In terms of flight mission parameter configuration, the UAV's flight altitude relative to the ground was set to 60 meters, resulting in a ground spatial resolution of 0.04 meters for the acquired multispectral imagery. Simultaneously, the forward overlap rate was set to 70%, and the lateral overlap rate to 80%, to meet the feature point matching requirements of subsequent image stitching processing. Before each flight, four standard diffuse reflection calibration plates, made of polytetrafluoroethylene (PTFE), were deployed in a flat, open area within the test zone. These plates covered a grayscale range of 2% to 80% of reflectivity and were used for radiometric correction of subsequent images.

[0036] After acquiring the original aerial imagery, photogrammetry software was used to perform aerial triangulation and regional network adjustment, generating a digital orthophoto map covering the entire test area. By introducing ground control points for geometric correction, the planar geometric error of the digital orthophoto map was controlled to within 2 cm. Subsequently, image processing tools were used to crop the corrected imagery, retaining the effective data range including the core test area and independent verification areas.

[0037] To address soil background interference in the imagery, a supervised classification method was employed to classify the digital orthophoto image pixels, identifying and separating soil pixels from vegetation pixels. A mask was generated based on the classification results to remove soil background areas from the image, retaining only the spectral reflectance information of the maize canopy. Based on the geographic coordinates of each monitoring point, the average canopy reflectance value within the corresponding area was extracted from the image and used as the UAV remote sensing observation data for that point in the corresponding band, for subsequent model building and data analysis.

[0038] In the specific implementation of satellite data acquisition and screening, Sentinel-2A satellite multispectral imagery data covering the target study area was acquired. The acquisition time of the satellite imagery was matched with the acquisition time of the UAV remote sensing data to ensure that the two data sources reflect the spectral characteristics of the same crop growth state. The data was downloaded through the European Space Agency's Copernicus Data Sharing Platform, and L2A-level product data that had undergone geometric and atmospheric correction was selected to reduce the impact of atmospheric scattering and absorption on surface reflectivity.

[0039] To ensure the comparability of satellite remote sensing data and UAV remote sensing data in the spectral dimension, a rigorous band selection and matching procedure was performed. Based on the band settings of the multispectral sensor on the UAV, six bands with center wavelengths similar to those of the UAV sensor and overlapping spectral ranges were selected from the 13 available bands of the Sentinel-2A satellite as the corresponding analysis bands.

[0040] The specific band matching configuration is as follows: the blue light band of the Sentinel-2A satellite (band number B2, center wavelength 490nm) is selected to correspond to the blue light band of the UAV (center wavelength 450nm); the green light band of the Sentinel-2A satellite (band number B3, center wavelength 560nm) is selected to correspond to the green light band of the UAV (center wavelength 555nm); and the red light band of the Sentinel-2A satellite (band number B4, center wavelength 665nm) is selected to correspond to the red light band of the UAV (center wavelength 660nm).

[0041] For the red-edge and near-infrared regions that are sensitive to changes in chlorophyll content, the first red-edge band of the Sentinel-2A satellite (band number B5, center wavelength 705nm) was selected to correspond to the first red-edge band of the UAV (center wavelength 720nm); the second red-edge band of the Sentinel-2A satellite (band number B6, center wavelength 740nm) was selected to correspond to the second red-edge band of the UAV (center wavelength 750nm); and the near-infrared band of the Sentinel-2A satellite (band number B8, center wavelength 842nm) was selected to correspond to the near-infrared band of the UAV (center wavelength 840nm).

[0042] Among the selected satellite bands, bands B2, B3, B4, and B8 have a spatial resolution of 10 meters, while bands B5 and B6 have a spatial resolution of 20 meters. This band selection strategy established a spectral channel correspondence between satellite data and UAV data, minimizing systematic spectral biases introduced by inconsistencies in the center wavelengths of different sensors. This provides fundamental data support for subsequent cross-platform data fusion and correction.

[0043] In the specific implementation of the vegetation index calculation system, multiple vegetation indices are constructed based on canopy spectral reflectance data extracted from UAV multispectral imagery or satellite imagery to quantitatively characterize the biophysical properties of crops. To standardize the variable symbols in the calculation formulas, [the following is defined / set]. Represents the reflectivity of the blue light band. Represents the reflectivity in the green light band. Represents the reflectivity in the red light band. Represents the reflectivity of the first red-edge band. Represents the reflectivity of the second red-edge band. This represents the reflectivity in the near-infrared band. For UAV data, the above bands correspond to center wavelengths of 450nm, 555nm, 660nm, 720nm, 750nm, and 840nm, respectively; for Sentinel-2A satellite data, the above bands correspond to center wavelengths of 490nm, 560nm, 665nm, 705nm, 740nm, and 842nm, respectively.

[0044] Based on the band reflectance variable defined above, nine vegetation indices related to chlorophyll content were calculated.

[0045] First, calculate the green chlorophyll index. The calculation formula is as follows:

[0046] ;

[0047] Calculate the drag index of the canopy structure The calculation formula is as follows:

[0048] ;

[0049] Next, the normalized difference vegetation index was calculated. The calculation formula is as follows:

[0050] ;

[0051] Calculate the normalized pigment chlorophyll index The calculation formula is as follows:

[0052] ;

[0053] Then, the ratio vegetation index is calculated. The calculation formula is as follows:

[0054] ;

[0055] Calculate the conversion vegetation index The calculation formula is as follows:

[0056] ;

[0057] Subsequently, the normalized differential water index was calculated. The calculation formula is as follows:

[0058] ;

[0059] Calculate the soil-regulated vegetation index The calculation formula is as follows:

[0060] ;

[0061] Finally, the difference vegetation index is calculated. The calculation formula is as follows:

[0062] ;

[0063] The above nine vegetation indices cover a combination of visible light, red edge, and near-infrared bands, reflecting crop growth from different dimensions such as pigment absorption, canopy structure, and background modulation, and serving as input feature variables for subsequent correlation analysis and model construction.

[0064] In the specific implementation of feature selection based on Pearson correlation analysis, in order to eliminate multicollinearity among input variables and improve the computational efficiency of the inversion model, a sensitivity assessment is performed on each vegetation index calculated in the aforementioned spectral feature parameter construction step, as well as the raw band reflectance extracted in the UAV low-altitude remote sensing data acquisition step. The aforementioned spectral features are defined as the set of feature variables to be screened, and the ground-measured SPAD values ​​are defined as the target response variables.

[0065] The Pearson product-moment correlation coefficient is used to quantitatively describe the linear dependence between each feature variable to be screened and the target response variable. This is applicable to features including... The dataset of the nth sample is used to calculate the nth sample. Correlation coefficients between each characteristic variable and the SPAD value Correlation coefficient The calculation formula is:

[0066] ;

[0067] in, Indicates the first The sample at the th The specific numerical values ​​of each characteristic variable; Indicates the first Measured SPAD values ​​for each sample; Indicates the first The arithmetic mean of the characteristic variables; Represents the arithmetic mean; Indicates to From 1 to Summation operation is performed on all terms; The scaling factor represents the product of the standard deviation of the characteristic variable and the standard deviation of the SPAD value. It is used to standardize the numerator, thereby eliminating the influence of the variable's dimensions and order of magnitude on the correlation coefficient calculation results. This indicates the total number of samples used for correlation analysis.

[0068] Through the above calculations, the correlation coefficients between all spectral characteristic variables and SPAD values ​​were obtained. All characteristic variables were ranked according to the absolute values ​​of the calculation results; the closer the absolute value is to 1, the stronger the linear correlation between the characteristic variable and the chlorophyll content of maize leaves. Based on the ranking results, the top few variables with the highest correlation to SPAD values ​​were selected, while variables with weak correlation to SPAD values ​​or exhibiting redundant information were removed. Finally, the optimal subset of sensitive features for constructing the SPAD inversion model was determined. This optimization process was performed independently for each growth stage to ensure that the selected features accurately reflect the crop's physiological state at specific growth stages.

[0069] See attached document Figure 2 In the specific implementation of the ratio-mean correction model, to address the radiometric inconsistency caused by differences in sensor spectral response functions and imaging environments in multi-source remote sensing data, a mathematical transformation relationship was established using UAV low-altitude remote sensing data as the radiometric benchmark. Areas with uniform spatial distribution and uniform land cover types within the core experimental area were selected as radiometric correction sample areas to ensure that UAV imagery and satellite imagery have corresponding spectral information of land cover within these areas.

[0070] Because the spatial resolution of UAV imagery is significantly higher than that of satellite imagery, a spatial scale matching operation is performed before numerical calculations. High-resolution UAV digital orthophotos are pixel-aggregated within selected sample areas, and the arithmetic mean of the reflectance of all UAV pixels corresponding to a satellite image pixel spatial range is calculated. This generates simulated UAV reflectance data that matches the pixel scale of the satellite imagery. A formula for calculating the ratio mean is constructed for each corresponding spectral band.

[0071] Regarding the first For each spectral band, the radiometric correction coefficient for that band is calculated using the ratio-mean method. Radiation correction factor The calculation formula is:

[0072] ;

[0073] in, Indicates the first The radiometric correction factor for each spectral band, which is a dimensionless ratio factor. The index number indicates the spectral band, corresponding to blue light, green light, red light, first red edge, second red edge, and near-infrared band; This represents the total number of samples used to build the calibration model; This represents the sample index number, with values ​​ranging from 1 to... ; Indicates the first The sample at the th The simulated reflectance values ​​of the UAV in each spectral band are average values ​​after spatial aggregation. Indicates the first The sample at the th Raw reflectance values ​​of Sentinel-2A satellite imagery in each spectral band; Indicates all The spectral response ratios of each sample were summed to gather information on radiation differences across the entire test area.

[0074] It is important to note that the radiation correction factor... The radiation correction coefficient is a dynamic variable constrained by the spatiotemporal environment. Its value depends not only on the sensor's own response characteristics but also on geographical environmental factors such as the solar altitude angle, atmospheric transport path, and surface environment at the imaging time. Therefore, when the inversion task is switched to different geographical regions or different imaging times, the aforementioned geographical environmental factors change, and the radiation correction coefficient needs to be recalculated based on synchronous observation samples under that specific spatiotemporal environment. This is to eliminate radiation errors introduced by environmental differences and ensure the adaptability and accuracy of the calibration model.

[0075] Through the above calculation process, the set of correction coefficients for all corresponding bands is obtained sequentially. This set of coefficients quantifies the systematic radiation deviation between the UAV sensor and the satellite sensor in each specific band, providing a definite multiplicative correction factor for subsequent band-by-band radiation normalization of satellite imagery across the entire region.

[0076] In the specific implementation of radiometric normalization processing for satellite imagery, the radiometric correction coefficients for each band, calculated in the ratio-mean correction model construction step, are used to adjust the radiance values ​​of the Sentinel-2A satellite multispectral imagery covering the entire region pixel by pixel. This processing aims to uniformly map the reflectance observations of satellite imagery to the radiometric scale of UAV low-altitude remote sensing data, thereby eliminating systematic radiometric differences between multi-source data and ensuring the cross-platform transfer accuracy of the inversion model.

[0077] The satellite imagery to be processed is set to be located in space by OK The column is composed of a pixel matrix. For the first... Each spectral band is used to traverse every image coordinate in the image matrix. Apply the corresponding radiation correction factor Perform a linear transformation calculation. Corrected satellite image reflectance. The calculation formula is:

[0078] ;

[0079] in, Indicates the first Under each spectral band, located at the image coordinates The reflectance of Sentinel-2A satellite imagery at a location after radiometric correction; Indicates the first The radiation correction coefficients corresponding to each spectral band; Indicates the first Under each spectral band, located at the image coordinates Raw reflectance values ​​from Sentinel-2A satellite imagery at the location; This represents the row index number of the pixel in the image matrix, with a value ranging from 1 to... ; This represents the column index number of the pixel in the image matrix, with a value ranging from 1 to... ; The index number indicates the spectral band, corresponding to blue light, green light, red light, first red edge, second red edge, and near-infrared band.

[0080] The above calculations are performed until all selected spectral bands are traversed, generating a complete radiometrically normalized satellite image dataset. This radiometrically normalized satellite image dataset possesses spectral properties similar to those of high-resolution UAV imagery, enabling direct support for the transfer and application of SPAD inversion models trained on UAV data at the regional scale. This lays the data foundation for the subsequent generation of large-scale maize chlorophyll content distribution maps.

[0081] See attached document Figure 3 In the specific implementation of the modeling dataset construction, a basic data matrix for training and validating the SPAD inversion model was constructed based on ground-measured data and UAV remote sensing data collected in the core experimental area A. The optimal sensitive feature subset determined in the feature optimization step based on Pearson correlation analysis was retrieved, and the sensitive feature values ​​of each monitoring point in the core experimental area A at the corresponding growth stage were extracted as input feature variables for the model. Simultaneously, the measured mean values ​​of the relative chlorophyll content of leaves at each corresponding monitoring point obtained in the experimental area division and ground sampling point layout steps were retrieved as the target response variable for the model.

[0082] The total number of samples in core experimental area A is set as follows: The total number of sensitive feature variables selected is .definition Index of feature variables ( ),definition Index of the sample ( Construct the original feature dataset and the target dataset.

[0083] To eliminate the negative impact of differences in dimensions and orders of magnitude between different feature variables on the model's convergence speed and accuracy, the Z-score normalization method is used to perform dimensionless processing on the original feature data. For each feature variable... Calculate the mean and standard deviation, and transform all sample data for this feature. Standardized feature values. The calculation formula is:

[0084] ;

[0085] in, Indicates the first The sample at the th Standardized feature values ​​for each feature variable; Indicates the first The sample at the th The specific numerical values ​​of each characteristic variable; Indicates standard deviation; It represents the arithmetic mean.

[0086] After standardization, a standard feature matrix is ​​generated. The standard feature matrix With the inclusion of measured SPAD values The target vectors are aligned by row index and merged to form a complete modeling dataset. Based on a pre-defined partitioning strategy, a stratified random sampling algorithm is used to divide the modeling dataset into a training set and an internal validation set. The training set contains approximately 2 / 3 of the sample data from core test area A, used for iterative optimization of model parameters; the internal validation set contains the remaining approximately 1 / 3 of the sample data from core test area A, used to monitor the model's fit and adjust hyperparameters during training to prevent overfitting. The final dataset format is configured to directly adapt to the input interface requirements of subsequent machine learning algorithms.

[0087] In the specific implementation of constructing a SPAD inversion model with multi-mechanism collaboration, based on the inclusion of feature values... Compared with the measured SPAD value Using the training set, three different mathematical mechanisms—Partial Least Squares Regression (PLSR), Random Forest (RF), and Backpropagation Neural Network (BPNN)—were employed to establish inversion models for maize chlorophyll content. These three algorithms represent linear dimensionality reduction regression, ensemble decision tree regression, and nonlinear mapping regression, respectively, aiming to comprehensively capture the complex response relationship between spectral features and SPAD values.

[0088] First, a partial least squares regression (PLSR) model is constructed. This algorithm addresses the multicollinearity problem among independent variables by projecting high-dimensional feature variables into a low-dimensional latent variable space. Using the training set data, latent factors that maximize the explanation of spectral feature variation while having the largest covariance with the SPAD value are identified, and a linear regression equation is established. For the [missing information - likely a specific model or process], the model is then... For each sample, the predicted value of the PLSR model The calculation formula is:

[0089] ;

[0090] in, Indicates the PLSR model for the first Predicted values ​​of SPAD for each sample; This represents the intercept constant term in the regression model; This represents the total number of sensitive characteristic variables; Indicates the first The regression coefficients corresponding to each feature variable are determined by an optimization process that minimizes the sum of squared predicted residuals. Indicates the first The sample at the th The standardized feature values ​​for each feature variable.

[0091] Secondly, a Random Forest (RF) regression model is constructed. This algorithm is based on the Bagging ensemble learning strategy, constructing multiple independent decision trees to make predictions together. During training, a Bootstrap sampling method is used to generate multiple training subsets, and at each node split, a subset of features is randomly selected for optimization. The final prediction result of the RF model is the arithmetic mean of the predictions from all decision trees. For the first... For each sample, the predicted value from the RF model is... The calculation formula is:

[0092] ;

[0093] in, Indicates the random forest model for the th Predicted values ​​of SPAD for each sample; This represents the total number of decision trees constructed in the random forest; This represents the index number of the decision tree, with a value ranging from 1 to... ; Indicates the first Each sample is composed of The input feature vector consists of several sensitive feature variables; Indicates the first Each decision tree is used to process the input feature vector. The single-tree prediction value output after regression analysis; Indicates to From 1 to Summing all terms.

[0094] Finally, a backpropagation neural network (BPNN) model is constructed. This algorithm employs a topology containing an input layer, hidden layers, and an output layer to simulate the nonlinear mapping relationship between spectral features and SPAD values. The network calculates the prediction error through forward propagation and uses gradient descent to adjust the network weights and biases in reverse. The number of neurons in the hidden layer is set to [value missing]. Regarding the first For each sample, the predicted value of the BPNN model The calculation formula is:

[0095] ;

[0096] in, This indicates that the BPNN model is for the first... Predicted values ​​of SPAD for each sample; This represents the linear activation function of the output layer; Indicates the first The connection weights from hidden layer neurons to output layer neurons; This represents the nonlinear activation function of the hidden layer (using either the Sigmoid or TanH function). Indicates the input layer's first... The feature variable is passed to the hidden layer. The connection weights of each hidden layer neuron; Indicates the first Bias terms for each hidden layer neuron; This represents the bias term of the output layer neurons; Indicates to From 1 to Summation operation is performed on all terms; Indicates to From 1 to Summing all terms.

[0097] During the training of all three models, a partitioned internal validation set was used for hyperparameter tuning. For the PLSR model, cross-validation was used to determine the optimal number of latent factors; for the RF model, the total number of decision trees was optimized. Minimum number of samples for node splitting; for BPNN models, determine the optimal number of iterations and the number of hidden layer neurons by monitoring the validation set error. Finally, three optimized inversion models are saved for subsequent accuracy evaluation and collaborative integration.

[0098] In the specific implementation of model accuracy evaluation, to quantitatively assess the inversion ability of three models—partial least squares regression, random forest, and backpropagation neural network—for maize chlorophyll content, an independent validation dataset not involved in the modeling process was used for testing. Spectral features from the validation set were input into each trained model to obtain the corresponding SPAD prediction values. The coefficient of determination (COP) was selected. ) and root mean square error ( As a core statistical evaluation indicator, it is used to measure the consistency and degree of deviation between model predictions and ground-based measured values.

[0099] First, calculate the coefficient of determination. This is used to characterize the extent to which the model explains the variation in the dependent variable. It is applicable to models containing... For the validation dataset of samples, the formula for calculating the coefficient of determination is:

[0100] ;

[0101] in, This represents the coefficient of determination, and its value typically ranges from 0 to 1. Indicates the first SPAD prediction values ​​obtained from the model for each sample; Indicates the first Measured SPAD values ​​for each sample; The sum of squares represents the regression sum, which characterizes the degree of variation of the model's predicted values ​​relative to the measured mean, i.e., the amount of variation that the model can explain. Represents the arithmetic mean; This represents the total sum of squares, which characterizes the overall degree of fluctuation in the measured data.

[0102] Secondly, calculate the root mean square error. This is used to measure the average level of model prediction error. The calculation formula is:

[0103] ;

[0104] in, This represents the root mean square error, and its unit is the same as the SPAD value. It represents the mean square error, which is the average of the squares of the differences between the predicted and measured values; Indicates to From 1 to Summing all terms.

[0105] At the same time, calculate the mean absolute error. This value directly reflects the average level of the absolute error between the predicted and actual values, and better reflects the actual situation of the predicted error. The calculation formula is:

[0106] ;

[0107] in, It represents the mean absolute error, and its unit is the same as the SPAD value unit; Indicates the first SPAD prediction values ​​obtained from the model for each sample; Indicates the first The measured SPAD value of each sample.

[0108] Through the above calculations, the three inversion models were obtained on different datasets. , and Numerical value. Based on The largest value and and The performance of the three models is ranked based on the comprehensive evaluation criterion of minimum numerical value. Only models that simultaneously possess high fitting accuracy (high...) are considered the best. and low prediction error (low and Only models that meet the requirements are considered the optimal inversion model for that reproductive period and are selected for subsequent large-scale remote sensing mapping tasks.

[0109] In the specific implementation of cross-space independent validation, to verify the generalization ability and transferability of the selected optimal SPAD inversion model in non-training areas, an independent validation area located outside the core test area A (denoted as core test area B) was selected as the test object. Sentinel-2A satellite imagery and corresponding ground-based SPAD measured data of core test area B at the same time phase were acquired. First, radiometric normalization was performed on the satellite imagery of core test area B, and sensitive spectral feature variables consistent with those in the modeling steps were extracted.

[0110] The feature data of core test area B are standardized using the statistical parameters of the training set. To ensure the consistency of the model input data distribution, the data in the validation set (core test area B) must be scaled using the mean and standard deviation of the training set (core test area A), rather than using the statistical parameters of the validation set itself. For the first feature data in core test area B... The sample at the th The original feature values ​​of each feature variable Its standardized value The calculation formula is:

[0111] ;

[0112] in, Indicates the first in the independent verification region The sample at the th The standardized values ​​of each feature variable will be directly used as input to the inversion model. Indicates the first in the independent verification region The sample at the th The original feature values ​​of each feature variable are extracted from the radiometrically normalized satellite imagery. Represents the arithmetic mean; It represents the standard deviation.

[0113] The standardized feature matrix is ​​input into the optimal inversion model to calculate the SPAD prediction value for each sample in the core test area B. To quantitatively evaluate the model's inference accuracy in the spatial dimension, the average relative accuracy (ARA) is used. (This is used as an assessment indicator.) The formula for calculating the average relative accuracy is:

[0114] ;

[0115] in, The average relative accuracy of the model is expressed as a percentage. This metric reflects the overall degree of agreement between the model's predictions and the actual values. This indicates the total number of samples used for testing in the independent validation region; Indicates the sample's index number; Indicates the first in the independent verification region The true measured values ​​of SPAD on the ground for each sample; Indicates the first in the independent verification region SPAD model predictions for each sample; Indicates to From 1 to Summing all terms.

[0116] If the calculated result If the numerical values ​​meet the preset threshold requirements for engineering applications, the inversion model is deemed to have universality across spatial scales and can be approved for spatial mapping of maize chlorophyll content across the entire region.

[0117] In the specific implementation of regional SPAD spatiotemporal distribution mapping, the optimal SPAD inversion model determined through accuracy evaluation and independent verification in the aforementioned steps is applied to the entire large-scale study area. Sentinel-2A satellite remote sensing images of the entire region during the monitoring period are retrieved. The generated vector boundary files of the maize planting areas are used to perform masking and cropping operations on the images, removing non-planting areas such as roads, water bodies, and buildings, retaining only effective pixels containing maize crops, thereby constructing the image matrix to be inverted.

[0118] For each valid pixel in the image matrix, extract pixels that are identical to those extracted during the model training phase. Several sensitive feature variables. To ensure the isomorphism of the input data and the model training data in the feature space distribution, the image data of the entire region must be standardized using the sample mean and standard deviation of the core test area A. The entire region image matrix is ​​set to contain... OK The column will be located in the first position. line, number The cells in the column are in the first The original values ​​of the feature variables are denoted as follows: Calculate the normalized feature value of this pixel. The standardized calculation formula is:

[0119] ;

[0120] in, Indicates the first in the image line, number The cells in the column are in the first The standardized feature values ​​of each feature variable are used as input components of the model. Indicates the first in the image line, number The cells in the column are in the first The original values ​​of each feature variable; Represents the arithmetic mean; Indicates standard deviation;

[0121] The standardized feature vectors are input into the optimal SPAD inversion model, and the predicted chlorophyll content is calculated pixel by pixel. The regression model function of the optimal inversion model is set as follows. Then the first line, number SPAD prediction values ​​for maize leaves in a single pixel The calculation formula is:

[0122] ;

[0123] in, This represents the calculated SPAD prediction value of the maize leaf at this pixel location; This represents the optimized regression model function (corresponding to one of partial least squares regression, random forest, or backpropagation neural network). This indicates the pixel corresponding to One sensitive feature variable.

[0124] The above calculation process is repeated, iterating through all valid pixels in the entire image matrix, to generate a two-dimensional SPAD numerical matrix consistent with the spatial resolution and coordinate system of the original image. This numerical matrix is ​​then assigned geospatial projection information and visualized using a pre-defined color mapping table, generating a thematic map of the spatial distribution of maize SPAD content during this growth stage. This thematic map visually demonstrates the spatial heterogeneity of maize growth within the study area, with high SPAD values ​​indicating sufficient nitrogen nutrition and low SPAD values ​​indicating potential nitrogen stress, providing fundamental data support for subsequent variable fertilization decisions.

[0125] See attached document Figure 4 In implementing differentiated farmland management strategies, a variable fertilization decision model for precision agriculture is constructed based on the generated regional SPAD spatiotemporal distribution thematic map. The core of this step lies in transforming continuously distributed crop chlorophyll monitoring values ​​into discrete agricultural machinery fertilization instructions, allocating nitrogen resources on demand according to the spatial heterogeneity of crop growth, in order to achieve the goal of homogenized yield by smoothing out peaks and valleys.

[0126] First, the target SPAD reference value for this growth period was determined. Based on the statistical results of measured data from high-yield plots in core experimental area A, the average measured SPAD value in the top 10% quantile was selected as the target SPAD reference value for this period, denoted as . This value represents the relative chlorophyll content level required for crops to achieve their potential maximum yield under current environmental conditions. (This refers to the first value in the imagery of the entire region.) line, number The columns of pixels are used to calculate the SPAD prediction values ​​of maize leaves. Calculate the theoretical nitrogen requirement for that location.

[0127] A nitrogen fertilizer recommendation algorithm based on spectral diagnostics is constructed. For each pixel location, the recommended pure nitrogen application rate is calculated. The calculation formula is:

[0128] ;

[0129] in, Indicates the first in the image line, number Recommended pure nitrogen application rates at the listed locations, in kilograms per hectare (kg / ha). This represents the preset target SPAD reference value for crops at this growth stage, which is a fixed constant. This indicates the first value obtained using the optimal inversion model. line, number SPAD prediction values ​​for maize leaves in a column of pixels; This represents the nitrogen sensitivity coefficient, which is determined through field gradient fertilization trials and characterizes the amount of pure nitrogen input required to increase the SPAD value by one unit. This represents the fertilizer requirement discrimination coefficient, used to handle situations where fertilization is not required. This indicates the degree of nutrient deficit between the current crop growth status and the target ideal status;

[0130] Discrimination coefficient for fertilizer demand Its value selection rule is defined as: when When this occurs, it indicates a nitrogen deficiency. The value is 1; when When the nitrogen content of the crop is sufficient or excessive, no additional fertilizer is needed. The value is set to 0 to ensure that the calculated amount of fertilizer is non-negative.

[0131] To adapt to the operating width and control precision of actual agricultural fertilization machinery, the calculated pixel-level continuous fertilization rate matrix was resampled and graded. The natural breakpoint method was used to further refine the matrix. The numerical distribution range is divided into 5 discrete management levels, corresponding to five decision categories: zero fertilization, low-volume topdressing, medium-volume topdressing, high-volume topdressing, and enhanced topdressing.

[0132] The final variable fertilizer prescription map is generated. The defined fertilization levels are mapped to a control command grid recognizable by the fertilizer applicator, with each grid representing an independent fertilization operation unit. This prescription map directly guides the variable fertilizer applicator to automatically adjust the opening or rotation speed of the fertilizer outlet based on GNSS positioning information during field operations. This increases fertilizer application in low SPAD areas to promote growth, while reducing or stopping fertilization in high SPAD areas to prevent excessive vegetative growth, delayed maturity, and environmental pollution, thus completing a closed-loop management system from remote sensing monitoring to agronomic decision-making.

Claims

1. A cross-platform inversion method for radiometric correction and chlorophyll content retrieval of multi-source remote sensing data, characterized in that, Includes the following steps: S1. Simultaneously acquire ground-measured data, UAV remote sensing data and satellite remote sensing data within the target area. Based on the band settings of the UAV remote sensing data, select corresponding bands with overlapping spectral ranges from the satellite remote sensing data and establish a correspondence relationship between the spectral channels of the multi-source data. S2. Calculate the canopy spectral reflectance and vegetation index based on the UAV remote sensing data, and use the Pearson correlation analysis method to analyze the correlation between the band reflectance of the UAV remote sensing data and the vegetation index and the ground measured data, and screen out sensitive feature variables. S3. A radiometric correction model is established using the ratio mean method, with the UAV remote sensing data as the reference. The correction coefficients of the UAV remote sensing data and the satellite remote sensing data in each corresponding band are calculated. The correction coefficients are used to perform radiometric normalization processing on the satellite remote sensing data to eliminate sensor response differences. S4. Using the sensitive feature variables as input variables and the ground measured data as response variables, construct a SPAD inversion model, and use statistical evaluation indicators to evaluate the prediction accuracy of the SPAD inversion model. S5. The SPAD inversion model, which has been verified for accuracy, is transferred and applied to the satellite remote sensing data after radiometric normalization. The SPAD prediction value is calculated pixel by pixel and a regional-scale spatial distribution map of maize chlorophyll content is generated. Based on the spatial distribution map, a differentiated farmland management strategy is generated. In step S3, the ratio mean method includes: Within the radiometrically corrected sample area, the UAV digital orthophoto imagery is subjected to pixel aggregation processing to generate UAV simulated reflectance data that matches the pixel scale of the satellite remote sensing data. For each corresponding band, the arithmetic mean of the ratio of the simulated UAV reflectance values ​​of all samples in the radiometric correction sample area to the original reflectance values ​​of the satellite image is calculated by summing the ratios of the simulated UAV reflectance values ​​of all samples in the radiometric correction sample area to the original reflectance values ​​of the satellite image and dividing the sum by the total number of samples, thus obtaining the correction coefficient. The correction coefficients are used to perform a pixel-by-pixel linear transformation calculation on the satellite remote sensing data for the entire region, and the observed reflectance values ​​of the satellite remote sensing data are mapped to the radiometric scale of the UAV remote sensing data. In step S4, the algorithm used to construct the SPAD inversion model includes one or more of the following: partial least squares regression algorithm, random forest regression algorithm, and backpropagation neural network algorithm. The partial least squares regression algorithm establishes a linear regression equation by projecting high-dimensional feature variables onto a low-dimensional latent variable space. The random forest regression algorithm performs ensemble prediction by constructing multiple independent decision trees. The backpropagation neural network algorithm uses a topological structure that includes an input layer, hidden layers, and an output layer to simulate nonlinear mapping relationships.

2. The method for cross-platform inversion of radiometric correction and chlorophyll content in multi-source remote sensing data according to claim 1, characterized in that, In step S1, the corresponding band covers the blue, green, red, first red edge, second red edge, and near-infrared spectral regions. The satellite remote sensing data is Sentinel-2A satellite multispectral imagery. The corresponding band selection includes: The blue light band of the Sentinel-2A satellite corresponds to the blue light band of the UAV; the green light band of the Sentinel-2A satellite corresponds to the green light band of the UAV; the red light band of the Sentinel-2A satellite corresponds to the red light band of the UAV; the first red edge band of the Sentinel-2A satellite corresponds to the first red edge band of the UAV; the second red edge band of the Sentinel-2A satellite corresponds to the second red edge band of the UAV; and the near-infrared band of the Sentinel-2A satellite corresponds to the near-infrared band of the UAV.

3. The method for cross-platform inversion of radiometric correction and chlorophyll content in multi-source remote sensing data according to claim 1, characterized in that, In step S2, the vegetation indices include the green chlorophyll index, canopy structure resistance index, normalized difference vegetation index, normalized pigment chlorophyll index, ratio vegetation index, conversion vegetation index, normalized difference water index, soil-regulated vegetation index, and difference vegetation index.

4. The method for cross-platform inversion of radiometric correction and chlorophyll content in multi-source remote sensing data according to claim 1, characterized in that, In step S2, the Pearson correlation analysis method includes: The Pearson product-moment correlation coefficient was used to quantitatively describe the degree of linear dependence between each feature variable to be screened and the ground-measured data. All feature variables to be screened are sorted according to the absolute value of the Pearson product-moment correlation coefficient; Based on the sorting results, the top few variables with the highest correlation to the ground-measured data are selected as the sensitive feature variables.

5. The method for cross-platform inversion of radiometric correction and chlorophyll content in multi-source remote sensing data according to claim 1, characterized in that, In step S4, Z-score normalization is performed before constructing the SPAD inversion model: The arithmetic mean of each feature variable in the modeling dataset is calculated by summing the feature variable values ​​of all samples and dividing by the total number of samples. The standard deviation is calculated by summing the squares of the differences between the feature variable values ​​of each sample and the arithmetic mean, dividing the sum of squares by the total number of samples, and taking the square root. The arithmetic mean and the standard deviation are used to perform dimensionless processing on the feature variables in the modeling dataset, validation set, and subsequent regional mapping data.

6. The method for cross-platform inversion of radiometric correction and chlorophyll content in multi-source remote sensing data according to claim 1, characterized in that, In step S4, the statistical evaluation indicators include the coefficient of determination, root mean square error, and mean absolute error. The model selection principle is to select the model with the largest coefficient of determination and the smallest root mean square error and mean absolute error as the optimal SPAD inversion model.

7. The method for cross-platform inversion of radiometric correction and chlorophyll content in multi-source remote sensing data according to claim 6, characterized in that, In step S5, generating a regional-scale spatial distribution map of maize chlorophyll content includes: The satellite remote sensing data is masked and cropped using the vector boundary file of the corn planting area to construct the image matrix to be inverted. For each valid pixel in the image matrix, the sensitive feature variable is extracted; The effective pixel-sensitive feature variables are standardized using the arithmetic mean and standard deviation of the modeling dataset; The standardized feature vectors are input into the optimal SPAD inversion model to calculate the SPAD prediction values.

8. The method for cross-platform inversion of radiometric correction and chlorophyll content in multi-source remote sensing data according to claim 1, characterized in that, In step S5, generating differentiated farmland management strategies includes: Determine the target SPAD reference value for the current reproductive period; The calculation of the recommended pure nitrogen application rate for each pixel location includes the following steps: Calculate the difference between the target SPAD reference value and the predicted SPAD value, and multiply the difference by the nitrogen sensitivity coefficient and the fertilizer demand discrimination coefficient; The recommended pure nitrogen application rate numerical distribution range was divided into discrete management levels using the natural breakpoint method to generate a variable fertilizer prescription map. The fertilizer demand discrimination coefficient is set to 1 when the target SPAD reference value is greater than the SPAD predicted value, and is set to 0 when the target SPAD reference value is less than or equal to the SPAD predicted value.

Images