Satellite image-based unmanned aerial vehicle multispectral image radiometric correction method

Abstract

The application discloses a kind of unmanned aerial vehicle multispectral image radiometric correction method based on satellite image, comprising the following steps: 1) obtain unmanned aerial vehicle multispectral image and the satellite image of the same area similar to unmanned aerial vehicle flight time, uniform to same coordinate system;2) select a variety of uniform ground objects as sampling area, extract corresponding band satellite image reflectivity and unmanned aerial vehicle multispectral image DN value, as sampling data;3) for each spectral band, with satellite image reflectivity as dependent variable, unmanned aerial vehicle multispectral image DN value as independent variable, establish regression model;4) evaluate regression model goodness of fit, when the determination coefficient R 2 <0.8, return to the second step, resample or increase sampling area, when the determination coefficient R 2 >0.8, evaluation passes;5) each band regression model is applied to corresponding band unmanned aerial vehicle multispectral image original image for radiometric correction, and each pixel is calculated.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing technology, specifically to a method for processing images from unmanned aerial vehicles (UAVs). Background Technology

[0002] Unmanned Aerial Vehicles (UAVs) multispectral remote sensing technology can rapidly and non-destructively acquire surface spectral information, and has become one of the core tools for modern agriculture, ecological environment monitoring, and resource surveys. By accurately measuring crop canopy reflectance, key physiological parameters such as leaf area index (LAI) and chlorophyll content (LCC) can be further retrieved, thus serving precision agricultural management processes such as growth assessment, nutrient diagnosis, and yield prediction.

[0003] However, the Digital Number (DN) values ​​directly acquired by UAV sensors are subject to complex influences from various factors, including sunlight conditions, atmospheric transmission, and the sensor's own radiometric response characteristics, and are not the true surface reflectance. Therefore, performing high-precision radiometric correction on the original imagery to convert its DN values ​​into physically meaningful and spatiotemporally comparable surface reflectance is the primary prerequisite and key foundation for ensuring the reliability of subsequent quantitative remote sensing inversion results. The accuracy of radiometric correction directly determines the quality of vegetation indices and biophysical and chemical parameters, serving as a crucial bridge connecting remote sensing data with agricultural applications.

[0004] The currently common radiometric correction method is the empirical linear method based on ground calibration boards. This method involves deploying several calibration boards with known reflectance within the survey area before the UAV flight. After acquiring images during flight, the average DN value of each calibration board region is precisely extracted from the images. Subsequently, using the known reflectance of the calibration board as the dependent variable and its corresponding image DN value as the independent variable, a linear regression model is established using the least squares method. Finally, this model is applied to the entire image, converting the DN values ​​of all pixels into reflectance values.

[0005] Although the empirical linear method based on ground calibration boards is widely used in field-scale UAV remote sensing experiments, it still faces a series of significant problems and challenges in practical large-scale and operational applications. First, the calibration board-based method is highly dependent on fieldwork, requiring the transport, deployment, and retrieval of calibration boards before and after each flight, which is time-consuming and labor-intensive. Furthermore, the calibration boards themselves may exhibit changes in reflectivity due to dust, rain, or aging, introducing additional errors. Second, this method requires the calibration board to be clearly visible in the image and occupy a sufficient number of pixels, a condition that is difficult to meet in densely vegetated or complex terrain areas, limiting its applicability. Third, its correction effect is heavily dependent on the instantaneous illumination conditions during flight. For images acquired at different times and under different weather conditions, the calibration board needs to be redeployed and modeled, making it difficult to guarantee the consistency of radiometric data over long periods.

[0006] Patent application CN110070513A discloses a method for radiometric correction of remote sensing images using a "standard ground feature database." The core of this technology lies in pre-constructing a database containing standard reflectance spectra of various typical ground features (such as roads, bare soil, and buildings). During correction, these "standard ground features" are first automatically identified as target areas from the UAV or satellite imagery to be corrected using machine learning algorithms. Then, their theoretical reflectance is obtained by querying the database. Finally, based on the correspondence between the DN value of the target area image and the theoretical reflectance, a radiometric correction model is established and applied to the entire image.

[0007] While this technology avoids the need for on-site calibration boards, the construction and maintenance of the "standard ground feature database" upon which it relies is itself a costly and complex task. Ground feature spectra are significantly affected by factors such as materials, aging, humidity, and season, exhibiting spatiotemporal variability. This can lead to discrepancies between the "standard" reflectance in the database and the reflectance of ground features in the actual scene. Furthermore, the accuracy of machine learning in identifying ground features directly affects the reliability of sampling points, introducing additional uncertainties.

[0008] Therefore, there is an urgent need for a radiometric correction method that can reduce dependence on external hardware, lower operational complexity, and ensure high correction accuracy and spatiotemporal consistency, so as to support the effective application of UAV remote sensing technology in large-scale, routine agricultural monitoring. Summary of the Invention

[0009] To achieve the above objectives and overcome the shortcomings of existing technologies, this invention aims to provide a UAV multispectral image radiometric correction method based on satellite imagery, which can improve the accuracy and consistency of UAV image reflectance data.

[0010] The UAV multispectral image radiometric correction method of the present invention does not require the prior setting of a ground calibration board. It uses open and available satellite imagery to provide the true surface reflectance and establishes a regression model by extracting corresponding pixels of uniform ground features, thereby achieving fast and robust radiometric correction.

[0011] This invention discloses a radiometric correction method for UAV multispectral images based on satellite imagery.

[0012] Figure 1 This is a technical flowchart illustrating one embodiment of the present invention.

[0013] The first step is data preparation. This involves acquiring satellite imagery data and multispectral imagery data of the UAV (unmanned aerial vehicle) to be calibrated, and ensuring spatial consistency through a unified coordinate system and georegistration.

[0014] Subsequently, a sampling strategy was implemented. Typical uniform land cover types such as bare soil, roads, and uniform vegetation were selected in the overlapping area of ​​the two types of images. Multiple (e.g., more than 3) sampling areas with significant brightness differences were extracted in each land cover type. The reflectance values ​​of the satellite image and the DN values ​​of the UAV image were then obtained for the corresponding bands to construct matching data pairs for modeling.

[0015] Next, based on these sampling data pairs, a univariate linear regression model and / or a univariate quadratic regression model are established band by band between satellite image reflectance values ​​and UAV image DN values, and the coefficient of determination R is used to determine the relationship between these models. 2 Evaluate the model's goodness of fit; if it does not reach the threshold, return to resampling for optimization.

[0016] Image correction is performed. After evaluation, regression models for each band are applied to the original UAV multispectral imagery. Radiometric correction is performed pixel-by-pixel through raster calculation to generate UAV multispectral reflectance images with physical meaning.

[0017] Optionally, accuracy assessment and verification are performed. The accuracy of the correction results is verified using measured data from a ground-based spectrometer, and the root mean square error (RMSE) and coefficient of determination (R²) are calculated. 2 The correction effect is evaluated using indicators such as ( ).

[0018] Ultimately, the corrected high-quality reflectance image data can be directly used for calculations of ground spectrometer measurement data, such as the calculation of vegetation indices like the Normalized Difference Vegetation Index (NDVI) and the Normalized Red Edge Index (NDRE), and further support the accurate inversion of key agricultural parameters such as Leaf Area Index (LAI) and Chlorophyll Content (LCC).

[0019] This invention discloses a radiometric correction method for UAV multispectral images based on satellite imagery, comprising the following steps:

[0020] (1) Acquire UAV multispectral images and satellite images of the same area with similar flight times to the UAV, and unify the satellite images and UAV multispectral images to the same coordinate system;

[0021] (2) Select multiple uniform ground features as sampling areas in the same area of ​​the two images, and extract the reflectance values ​​of the corresponding bands and the DN values ​​of the UAV multispectral images as sampling data;

[0022] (3) Based on the sampling data in step (2), establish a regression model between the satellite image reflectance value and the UAV multispectral image DN value for each spectral band;

[0023] (4) Evaluate the goodness of fit of the regression model. When the coefficient of determination R... 2 When the coefficient of determination R0.8 is less than 0.8, return to step two and resample. 2 When the value is greater than 0.8, the evaluation is passed;

[0024] (5) Apply the regression models of each band that have passed the evaluation in step 4) to the original UAV multispectral images of the corresponding bands for radiometric correction, calculate pixel by pixel, and generate radiometrically corrected UAV multispectral reflectance images.

[0025] Furthermore, in step (1) above, the time difference between the acquisition of UAV multispectral images and satellite images does not exceed 7 days, and both are acquired under clear sky and low cloud cover conditions.

[0026] Furthermore, in step (1) above, after acquiring satellite images of the same area with a flight time similar to that of the UAV, it is necessary to unify the satellite images and UAV images to the same coordinate system. This can be achieved by performing a reprojection operation using ArcMap software or by writing a Python script file.

[0027] Furthermore, in step (2) above, at least two uniform land cover types are selected within the same geographical area of ​​satellite imagery and UAV imagery, and multiple sampling areas with significant brightness differences are selected for each land cover type.

[0028] Furthermore, in step (2) above, the above-mentioned uniform land cover types include at least two of the following: bare soil, hardened road surface, and uniform vegetation cover area, and more than three sampling areas are selected for each land cover type.

[0029] Furthermore, in step (2) above, the sampling area should avoid interference areas such as the edges of ground features (affected by composite pixels) and cloud shadows to ensure pixel purity.

[0030] Furthermore, in step (3) above, for each spectral band, a regression model is established between the satellite image reflectance as the dependent variable and the UAV image DN value as the independent variable.

[0031] Furthermore, in step (3) above, the regression model adopts a univariate linear regression model and / or a univariate quadratic regression model.

[0032] The univariate linear regression model is as follows:

[0033] (Equation 1)

[0034] The univariate quadratic regression model is as follows:

[0035] (Equation 2)

[0036] in, This indicates the surface reflectance of satellite imagery in a specific band; This represents the original digital quantization value (DN value) of the UAV multispectral image in the corresponding band. , and These are the model coefficients (constants) obtained by fitting using the least squares method.

[0037] Furthermore, in step (3) above, a univariate linear regression model and / or a univariate quadratic regression model are used to fit the relationship between the satellite image reflectance value and the UAV DN value. This is mainly based on the following considerations: 1) The univariate linear regression model is the most commonly used model in radiometric correction and has a clear physical meaning. It corresponds to the ideal linear characteristics of the photoelectric response of the sensor within its normal operating range. It has the advantages of fewer parameters, high stability, and ease of interpretation and implementation. It is especially suitable for situations where the sensor response is linear, the ground object brightness range is moderate, and the atmospheric conditions are stable. 2) As a simple nonlinear model, the univariate quadratic regression model can effectively capture the small curve trends caused by the slight nonlinear response of the sensor, atmospheric scattering, or ground object reflection characteristics. It is suitable for data distribution situations where the response curve has a slightly convex or concave shape.

[0038] Furthermore, in step (3) above, a simpler and more robust linear model can be preferred in practical applications; the univariate quadratic regression model is used as an alternative to deal with slight nonlinearities in individual bands or specific scenarios, thereby enhancing the applicability and reliability of the method while ensuring the accuracy of the correction.

[0039] Furthermore, in step (4) above, the index for evaluating the goodness of fit is the coefficient of determination R. 2 Requires R 2 Not less than 0.8. When R 2 If R is greater than 0.8, proceed to the next image correction step; when R... 2If the value is less than 0.8, return to step (2) to resample or increase the sampling area to reduce the error caused by manual sampling.

[0040] Optionally, the corrected UAV multispectral reflectance image can be compared with the measured spectral data from a ground-based spectrometer to verify the correction accuracy.

[0041] Furthermore, the above verification steps include: comparing the corrected reflectance image with the synchronously acquired ground spectrometer measurements, and calculating the root mean square error (RMSE) and the coefficient of determination (R²). 2 wait.

[0042] Furthermore, the above method is applicable to image radiometric correction for various UAV multispectral systems.

[0043] Furthermore, the present invention also discloses the application of the above-mentioned corrected UAV multispectral reflectance imagery in remote sensing of crops.

[0044] Furthermore, the above applications include calculating vegetation indices such as the Normalized Difference Vegetation Index (NDVI) and the Normalized Red Edge Index (NDRE) based on the corrected UAV multispectral reflectance images, and further retrieving vegetation parameters such as the Leaf Area Index (LAI) and the Chlorophyll Content (LCC).

[0045] Furthermore, the present invention also discloses a computer-readable storage medium, characterized in that: the computer-readable storage medium includes a stored computer program, wherein, when the computer program is run by a processor, it controls the device where the storage medium is located to execute the UAV multispectral image radiometric correction method for the satellite image, thereby realizing regression model modeling and UAV multispectral image raw image conversion.

[0046] Furthermore, this invention also discloses a UAV multispectral image radiometric correction system based on satellite imagery, comprising: a data acquisition module for acquiring satellite imagery data and UAV multispectral imagery data; a sampling and feature extraction module for selecting sampling areas on the imagery and extracting satellite imagery reflectance values ​​and UAV multispectral imagery DN values; a model building and evaluation module for establishing and evaluating band reflectance regression models; an image correction processing module for applying the band reflectance regression models to the UAV multispectral imagery data to generate corrected imagery data; an accuracy verification module for evaluating the accuracy of the correction results; and a data application module for calculating vegetation indices and inverting vegetation parameters.

[0047] Compared with existing technologies, this invention innovatively utilizes open satellite imagery data to achieve automated radiometric correction of UAV images. By eliminating the need for complex preprocessing such as atmospheric correction of satellite images and fine spatial registration of UAV images, it effectively overcomes the dependence of traditional methods on hardware deployment and complex preprocessing, significantly simplifies the operation process, and demonstrates broad applicability. It can effectively meet the UAV multispectral image correction needs of different land cover types, seasons, and regions.

[0048] This invention utilizes readily available satellite imagery data as a radiometric benchmark. By selecting uniform ground object samples, a quantitative relationship model is established between the data and the original UAV multispectral imagery. This enables efficient and robust radiometric correction of UAV multispectral imagery, providing a high-quality, stable, reliable, and spatiotemporally comparable surface reflectance data foundation for long-term, operational monitoring of agricultural remote sensing.

[0049] This invention employs a multi-band separate modeling strategy, fully considering the response differences of each band, thereby ensuring reliable accuracy with a high degree of consistency between the corrected reflectance and the reflectance of satellite images.

[0050] This invention is applicable to various scenarios such as vegetation parameter inversion, agricultural monitoring, and ecological environment assessment.

[0051] Meanwhile, the standardized process of this invention is easy to integrate into existing UAV image processing systems, supporting the commercial application of large-scale, multi-temporal data.

[0052] Furthermore, this invention provides a low-cost radiation correction scheme that eliminates the need for additional purchases of irradiance sensors or frequent deployment of calibration boards. Correction can be completed solely using free satellite imagery data, making it practical and scalable. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating the technical process of the present invention.

[0054] Figure 2 This represents the univariate linear regression modeling result of one embodiment of the present invention.

[0055] Figure 3 This represents the univariate quadratic regression modeling result of one embodiment of the present invention.

[0056] Figure 4 This image shows a comparison of grayscale images of a UAV multispectral images before and after radiometric correction, according to an embodiment of the present invention. The left image is the original UAV multispectral image, and the right image is the radiometrically corrected UAV multispectral image.

[0057] Figure 5 This diagram illustrates the accuracy verification results of one embodiment of the present invention. Detailed Implementation

[0058] To better understand this invention, the following embodiments are provided in conjunction with the accompanying drawings. It should be understood that the embodiments of this invention are for illustrative purposes only and not for limiting the invention; the scope of protection of this invention is defined solely by the claims. The embodiments provided are merely preferred embodiments and are not intended to limit the invention in any way. Those skilled in the art can make changes, substitutions, or modifications based on the content of this invention to form different implementation methods. However, any changes and modifications, or equivalent substitutions, made to the method of this invention without departing from the inventive concept are within the scope of protection of this invention.

[0059] Example 1: Multispectral Image Acquisition by UAV

[0060] The imagery was captured on July 10, 2025, by a drone in the 856 Farm area of ​​Heilongjiang Province. The sensor used was a DJI Mavic 3 Multispectral Edition from Shenzhen DJI Innovation Technology Co., Ltd. The imagery includes four bands: green (G), red (R), red edge (RE), and near-infrared (NIR or nir). The spatial resolution of the imagery is approximately 4 cm, and it is stored as a GeoTIFF format DN value image.

[0061] The drone (DJI Mavic 3 Multispectral Edition) is equipped with a Real-Time Kinematic (RTK) module, which can directly provide centimeter-level positioning information (longitude, latitude, and elevation) for acquired multispectral imagery. RTK is a real-time positioning technology based on the Global Navigation Satellite System (GNSS). Through carrier phase differential correction between the base station and the rover, it can acquire three-dimensional coordinates (longitude, latitude, and elevation) with centimeter-level or even millimeter-level accuracy in real time. The RTK module enables the drone imagery itself to have high-precision spatial reference information, thus eliminating the need for complex and time-consuming manual georegistration steps when overlaying and analyzing with satellite imagery. This greatly improves the automation and overall efficiency of data processing, laying a reliable geometric foundation for radiometric correction and subsequent quantitative inversion.

[0062] Example 2: Sentinel-2 L2A Class Surface Reflectance Image Acquisition

[0063] This invention uses Sentinel-2 satellite imagery as an example, directly employing Sentinel-2 L2A-level surface reflectance products as the radiometric benchmark. This product is free, open-access global data that has undergone rigorous atmospheric correction, possesses clear physical meaning and high spatiotemporal consistency, and is itself an authoritative, dynamically updated "global benchmark library." This invention eliminates the need to build and maintain an independent spectral library or employ complex ground feature identification algorithms. It only requires manually selecting spectrally stable, uniform regions (such as bare soil or uniform vegetation) from spatiotemporally matched satellite and UAV imagery to establish regression relationships. This fundamentally avoids the accuracy maintenance challenges of self-built ground feature libraries and the uncertainties of machine learning identification, making the correction process simpler, more robust, lower in cost, and easier to promote and apply in operational, large-scale agricultural remote sensing monitoring.

[0064] Download the Sentinel-2 L2A-level surface reflectance product covering the Qixing Farm area in Heilongjiang Province from the official website of the European Space Agency's Copernicus Open Access Hub (in this example, the downloaded image is a cloud-free Sentinel-2 L2A-level satellite image from July 7, 2025). This product includes atmospherically corrected surface reflectance data, and its B3 (green), B4 (red), B5 (red edge), and B8 (near-infrared) bands are selected as the radiometric reference.

[0065] Example 3: Preprocessing of satellite and UAV multispectral images

[0066] First, DJI Terra software from Shenzhen DJI Innovations Technology Co., Ltd. was used to preprocess the original UAV multispectral imagery, including image stitching. Then, the Sentinel-2 satellite imagery and the UAV imagery were unified to the same coordinate system. In the ArcMap software environment, the "Projected Grid" tool was used to convert the coordinate system of the UAV imagery to the WGS84 Universal Transverse Mercator (UTM) projection, which is consistent with the Sentinel-2 imagery, thus completing the unification of spatial references.

[0067] Since RTK was used to calculate the precise 3D coordinates of the measurement center during UAV multispectral image acquisition, no additional georegistration is required here.

[0068] Example 4: Implementation of Sampling Strategy

[0069] In the overlapping area of ​​the two sets of images in coordinate system one, based on the texture and spectral characteristics of the Sentinel-2 image, three types of ground features with stable spectral characteristics and uniform spatial distribution were manually selected as sampling targets: (1) cement roads in the field (high reflectivity); (2) fallow bare soil (medium reflectivity); (3) uniformly growing crop canopy (low reflectivity, high reflectivity in the near-infrared band).

[0070] For each type of land cover, avoid interference areas such as land cover edges (affected by composite pixels) and cloud shadows, and select 3 to 5 polygonal areas with significant brightness differences as sampling areas to ensure that the total number of sampling blocks for each band is no less than 20.

[0071] Each sampling area must cover at least 3 pixels in the Sentinel-2 image to ensure spectral representativeness. Subsequently, the average value of all pixels in the corresponding band (G, R, RE, NIR) within each sampling area is extracted to obtain the reflectance value of the Sentinel-2 image and the original DN value of the UAV image, forming a sample pair for modeling.

[0072] Example 5: Regression Modeling: Univariate Linear Regression Model

[0073] Use Python to write scripts to perform regression analysis independently for each spectral band.

[0074] Each band was evaluated using a regression model, the formula of which is as follows.

[0075] (Equation 1)

[0076] In the formula, For Sentinel 2 reflectivity, For the DN value of the drone, , These are the model coefficients (constants) obtained by fitting using the least squares method.

[0077] Figure 2 This represents the univariate linear regression modeling results for the four bands: red, green, nir, and rededge.

[0078] Taking the red band as an example, the red band reflectance of Sentinel-2 extracted from all sampling areas is used as the dependent variable (y), and the red band DN value of the UAV is used as the independent variable (x). A univariate linear regression model is fitted using the least squares method. The obtained model is:

[0079] (Equation 3)

[0080] Calculate the coefficient of determination R of the model. 2 It is 0.9562.

[0081] Similarly, regression models are established for the remaining bands.

[0082] The model for the green band is:

[0083] (Equation 4)

[0084] Calculate the coefficient of determination R of the model. 2 It is 0.9451.

[0085] The model for the near-infrared (NIR) band is as follows:

[0086] (Equation 5)

[0087] Calculate the coefficient of determination R of the model. 2 It is 0.9225.

[0088] The model for the rededge band is:

[0089] (Equation 6)

[0090] Calculate the coefficient of determination R of the model. 2 It is 0.8204.

[0091] Example 6: Regression Modeling: Univariate Quadratic Regression Model

[0092] Use Python to write scripts to perform regression analysis independently for each spectral band.

[0093] Each band was evaluated using a regression model, the formula of which is as follows.

[0094] (Equation 2)

[0095] In the formula, For Sentinel 2 reflectivity, For the DN value of the drone, , and These are the model coefficients (constants) obtained by fitting using the least squares method.

[0096] Figure 3 This represents the univariate quadratic regression modeling results for the four bands: red, green, nir, and rededge.

[0097] Taking the red band as an example, the red band reflectance of Sentinel-2 extracted from all sampling areas is used as the dependent variable (y), and the red band DN value of the UAV is used as the independent variable (x). A univariate quadratic regression model is fitted using the least squares method. The obtained model is:

[0098] (Equation 7)

[0099] Calculate the coefficient of determination R of the model. 2 It is 0.9575.

[0100] Similarly, regression models are established for the remaining bands.

[0101] The model for the green band is:

[0102] (Equation 8)

[0103] Calculate the coefficient of determination R of the model. 2 It is 0.9453.

[0104] The model for the near-infrared (NIR) band is as follows:

[0105] (Equation 9)

[0106] Calculate the coefficient of determination R of the model. 2 It is 0.9247.

[0107] The model for the rededge band is:

[0108] (Equation 10)

[0109] Calculate the coefficient of determination R of the model. 2 It is 0.8248.

[0110] In this embodiment, it has been verified that, among the two types of models mentioned above, the goodness of fit (Ri) between the univariate linear regression model and the univariate quadratic regression model is... 2 The values ​​are very close, and both achieve a coefficient of determination higher than 0.8. This indicates that in the correspondence between the UAV multispectral data and the Sentinel-2 L2A reflectance data in this embodiment, the linear relationship is dominant, and the slight nonlinear component is not significant.

[0111] In this embodiment, a simpler and more robust linear model is preferred, while a univariate quadratic regression model is used as an alternative.

[0112] Example 7: Evaluating the goodness of fit of the model

[0113] The threshold (R) set according to the present invention 2 >0.8), all band models in this embodiment passed the evaluation.

[0114] If, during the evaluation process, a certain band R... 2 If the threshold is not reached, you need to go back to the previous step, re-examine and increase or adjust the sampling area.

[0115] Example 8: Image Correction

[0116] After obtaining the regression models for each band that have passed the evaluation, i.e., based on the modeling results of Example 5, the formula is determined.

[0117] (Equation 1)

[0118] coefficients in and Then, a script-based automated process using the Python programming language was employed to complete the radiometric correction of the entire original UAV imagery.

[0119] The process begins by reading each band of the raw UAV multispectral imagery from the Geospatial Data Abstraction Library (GDAL library) and loading it into memory as a digital matrix (DN value matrix). Then, based on pre-stored band correspondences, the coefficients from the regression models for each band are called. and The correction process is accomplished by performing a linear transformation on the DN value matrix of each band in memory, that is, calculating each pixel value (DN) in the matrix according to the formula. This calculation can be efficiently implemented using vectorized operations of array computing libraries such as Numerical Python (NumPy) without explicit loops, significantly improving the processing efficiency of large-scale image data. After the calculation is completed, the script rewrites the newly generated reflectance matrices for each band according to the spatial reference information and data structure of the original image and generates a new multi-band GeoTIFF format reflectance image file.

[0120] This method achieves full automation and batch processing capabilities in the calibration process through scripting, ensuring the consistency and repeatability of the processing flow, and can flexibly adapt to the rapid processing needs of different types of UAV sensors and image data from different regions.

[0121] Figure 4This image shows a comparison of grayscale images of UAV multispectral images before and after radiometric correction, containing four bands: red, green, red edge, and near-infrared (NIR). The left image is the original UAV multispectral image, and the right image is the radiometrically corrected UAV multispectral image. The UAV multispectral image in the image is from a field in Qixing Farm, Heilongjiang Province, collected on July 10, 2025. In this embodiment, a univariate linear regression model is used to perform radiometric correction on the UAV multispectral image. From the visual effect analysis of the single-band grayscale images, the UAV images before and after radiometric correction using the method of this invention maintain consistency in the spatial structure, texture features, and relative brightness levels of ground features, with limited visual differences. This is because the core of radiometric correction is to complete a quantitative conversion of radiometric calibration and atmospheric compensation. Its essence is to systematically convert the relative grayscale value (DN value) output by the sensor, which is affected by illumination and instrument response, into an absolute surface reflectance with a clear physical meaning by establishing a deterministic relationship with the standard reflectance reference (Sentinel 2 image). This process does not alter the "content" of the image, but rather endows it with standardized "physical meaning" and "units of measurement." After correction, each pixel value of the image is converted into surface reflectance that can be directly used for quantitative calculations and model inversion. This ensures the scientific validity and direct comparability of the data in temporal comparisons, multi-source fusion, and accurate inversion of vegetation parameters, laying a solid data foundation for moving from qualitative interpretation to quantitative analysis.

[0122] Example 9: Accuracy Evaluation and Verification

[0123] To evaluate the accuracy of the radiometric correction results in this embodiment, the corrected UAV reflectance image obtained in Example 8 was compared and analyzed with the simultaneously acquired ground-based spectrometer-measured reflectance data covering the sampling area. The root mean square error (RMSE) and coefficient of determination (R²) were calculated. 2 Quantitative evaluation is conducted using indicators such as )

[0124] While the UAV was in flight, a high-performance portable ground spectrometer, the RS-8800 from Spectral Evolution (USA), was used to simultaneously measure the ground spectrometer of the multispectral image acquisition area. Measurements were conducted on pre-selected uniform target areas such as bare soil, roads, and uniform crop canopies, recording hyperspectral reflectance data. During the measurement, a high-precision RTK module was used to record the positions of ground measurement points, thus establishing a spatial correspondence between the measured ground reflectance and image pixels. These measured reflectance data, serving as "ground truth values," were used for quantitative comparison with the corrected UAV image reflectance, thereby calculating accuracy evaluation indicators such as RMSE and R², and objectively verifying the radiometric correction effect.

[0125] like Figure 5As shown, the corrected reflectance values ​​(Predicted Values) for each band exhibit good consistency with the ground-measured values ​​(True Values): the coefficients of determination (R²) for the red-edge band and the near-infrared (nir) band reach 0.8589 and 0.8568, respectively, indicating a strong correlation; the R² values ​​for the red and green bands are also higher than 0.76. Meanwhile, the root mean square error (RMSE) for all bands is below 0.021, with the red-edge band exhibiting the smallest error (RMSE = 0.0092). In summary, the correction results obtained using this method are close to the ground truth, with errors controlled at a low level, meeting the data accuracy requirements for quantitative remote sensing applications such as vegetation parameter inversion.

[0126] The accuracy evaluation results show that the method of the present invention has high correction accuracy and reliability. Verification shows that the corrected reflectance of each band in this embodiment is highly consistent with the measured values ​​on the ground, indicating that the correction method is effective and achieves the expected accuracy.

[0127] The reflectance data of the image corrected using this method can be used as high-precision baseline data for subsequent vegetation index calculations and parameter inversion. Verification with ground-measured data shows that the correction results obtained using this method are reliable and can meet the needs of quantitative remote sensing applications.

Claims

1. A method for radiometric correction of UAV multispectral images based on satellite imagery, comprising the following steps: 1) Acquire UAV multispectral imagery and satellite imagery of the same area around the time of the UAV flight, and unify the satellite imagery and UAV multispectral imagery to the same coordinate system; 2) Select multiple uniform ground features as sampling areas within the same geographical area of ​​the two images, and extract the satellite image reflectance values ​​and UAV multispectral image DN values ​​of the corresponding bands as sampling data; 3) Based on the sampling data in step 2), for each spectral band, a regression model is established between the satellite image reflectance value and the UAV multispectral image DN value, with satellite image reflectance as the dependent variable and UAV multispectral image DN value as the independent variable. 4) Evaluate the goodness of fit of the regression model. When the coefficient of determination R... 2 When the coefficient of determination R0.8 is less than 0.8, return to step two to resample or increase the sampling area. 2 When the value is greater than 0.8, the evaluation is passed; 5) Apply the regression models of each band that passed the evaluation in step 4) to the original UAV multispectral images of the corresponding bands for radiometric correction, calculate pixel by pixel, and generate radiometrically corrected UAV multispectral reflectance images.

2. The UAV multispectral image radiometric correction method based on satellite imagery according to claim 1, characterized in that: Step 2) Selecting multiple uniform land cover types as sampling areas within the same geographical area of ​​the two images specifically involves selecting at least two uniform land cover types from bare soil, hardened road surfaces, and uniform vegetation cover areas within the same geographical area of ​​the satellite image and the UAV multispectral image, and selecting more than three sampling areas with significant brightness differences for each land cover type.

3. The UAV multispectral image radiometric correction method based on satellite imagery according to claim 1, characterized in that: In step 3), the regression model adopts a univariate linear regression model and / or a univariate quadratic regression model.

4. The UAV multispectral image radiometric correction method based on satellite imagery according to claim 1, characterized in that, It also includes the following steps: 6) Compare the corrected UAV multispectral reflectance image with the measured spectral data from the ground spectrometer to verify the correction accuracy.

5. The UAV multispectral image radiometric correction method based on satellite imagery according to claim 4, characterized in that: Step 6) verifies the correction accuracy by calculating the root mean square error (RMSE) and the coefficient of determination (R²).

6. The UAV multispectral image radiometric correction method based on satellite imagery according to any one of claims 1 to 5, characterized in that: Step 1) is implemented by performing reprojection using ArcMap software or by writing a Python script file; in Step 5), specifically, ArcMap software or a Python script file is used to perform raster conversion to obtain radiometrically corrected UAV multispectral reflectance images.

7. The application of UAV multispectral reflectance images obtained by the corrected satellite imagery-based UAV multispectral image radiometric correction method according to any one of claims 1 to 6 in remote sensing of crops.

8. The application according to claim 7, characterized in that: Based on the corrected UAV multispectral reflectance image, vegetation indices and inverted vegetation parameters are calculated. The vegetation indices include the Normalized Difference Vegetation Index (NDVI) and the Normalized Red Edge Index (NDRE). The vegetation parameters include the Leaf Area Index (LAI) and the Chlorophyll Content (LCC).

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, it controls the device where the storage medium is located to execute the radiometric correction method for UAV multispectral images based on satellite imagery as described in any one of claims 1 to 8, thereby realizing regression model modeling and conversion of UAV multispectral imagery into original images.

10. A UAV multispectral image radiometric correction system based on satellite imagery, comprising: A data acquisition module is used to acquire satellite imagery data and UAV multispectral imagery data; A sampling and feature extraction module is used to select sampling areas on image data and extract satellite image reflectance values ​​and UAV multispectral image DN values; A model building and evaluation module is used to build and evaluate band reflectivity regression models; An image correction processing module is used to apply the reflectance regression model of each band to the UAV multispectral image data to generate corrected image data; A precision verification module is used to evaluate the precision of the calibration results; One data application module is used for vegetation index calculation and vegetation parameter inversion.

Images