Xgboost suspended solids concentration aerial remote sensing inversion method based on characteristic wave band selection

Abstract

The application discloses an XGBoost suspended matter concentration aviation remote sensing inversion method based on feature band selection, and relates to the technical field of environmental monitoring.The XGBoost model used in the method has stronger nonlinear fitting capability and generalization capability, and is not prone to overfitting; compared with other complex machine learning models, the method provides highly concise and low-redundancy input for the model through the early feature band selection and feature index construction, which not only significantly improves the training efficiency and inversion precision of the model, but also makes the model more robust to noise, so that the method can finally realize rapid, accurate and large-range suspended matter concentration mapping of a complex water body environment, effectively overcomes the time and space limitations of traditional monitoring methods, and solves the problems of weak model generalization capability and high calculation cost in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring technology, and specifically relates to an XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection. Background Technology

[0002] Suspended solids are a core parameter of aquatic ecosystems, directly affecting water transparency, phytoplankton primary productivity, and ecosystem health. Traditional monitoring relies on shipborne sampling and laboratory analysis, which suffers from limited spatial and temporal coverage, high costs, and poor timeliness, making it difficult to meet the needs of large-scale dynamic monitoring. Remote sensing technology, with its advantages of large-scale synchronous observation and periodic coverage, has become the core means of suspended solids monitoring. In particular, airborne remote sensing, with its high spatial resolution and flexibility, provides important support for refined monitoring of nearshore, river, and lake water bodies.

[0003] Current water surface concentration extraction methods are mostly based on empirical models, semi-analytical models, and machine learning models. Statistical regression establishes linear or nonlinear relationships between band reflectance and suspended matter concentration. Its advantage is computational simplicity, but it relies on a large amount of measured data, has weak generalization ability, and is easily affected by regional hydrological conditions. Combining radiative transfer equations such as absorption-scattering coefficients with water optical properties, physical mechanisms such as the Gordon model simplify modeling. While these models have a physical basis, parameter calibration is complex, and they are sensitive to assumptions about water optical properties, resulting in insufficient applicability in complex water bodies. Neural networks and support vector machines invert concentration through nonlinear mapping; however, these models face the problem of high-dimensional data redundancy: hyperspectral data typically contains hundreds of bands, and strong correlations between bands lead to the curse of dimensionality, increasing computational burden and potentially introducing noise that reduces accuracy. Therefore, to eliminate the curse of dimensionality caused by multi-channel redundancy in remote sensing imagery, proposing an XGBoost airborne remote sensing inversion method for suspended matter concentration based on feature band selection is particularly important. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection, thereby solving the aforementioned technical problems.

[0005] An airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection using XGBoost includes the following steps:

[0006] Perform multispectral remote sensing data preprocessing to generate UAV remote sensing reflectance products;

[0007] Using the UAV remote sensing reflectance products and measured suspended matter concentration data, the correlation coefficient is calculated to determine the optimal characteristic band;

[0008] Construct a suspended matter index using the data from the optimal characteristic bands;

[0009] An XGBoost model is constructed that takes the spectral values ​​of the characteristic bands and the suspended matter index as inputs and the suspended matter concentration as output.

[0010] The XGBoost model is trained using training data;

[0011] The trained XGBoost model was applied to the UAV remote sensing reflectance product to retrieve the suspended matter concentration.

[0012] Preferably, the preprocessing of multispectral remote sensing data specifically involves: performing band registration, orthophoto stitching, and radiometric calibration on the spectral data collected by the UAV.

[0013] Preferably, when calculating the correlation coefficient using the UAV remote sensing reflectance product and measured suspended matter concentration data, the average reflectance of multiple pixels around the sampling point is selected as the reflectance data corresponding to the sampling point.

[0014] Preferably, the method for determining the optimal characteristic band is as follows:

[0015] Calculate the correlation coefficient between the reflectance data of each sampling point and the measured suspended matter concentration in different spectral bands;

[0016] The two bands with the highest absolute values ​​of correlation coefficients were selected as characteristic bands.

[0017] Preferably, the method for constructing the suspended matter index is as follows:

[0018] When the remote sensing reflectance of both characteristic bands is positively correlated with the measured suspended matter concentration, or both are negatively correlated, the following formula is used for calculation: ,

[0019] in, The suspended solids index, The reflectivity of the band is greater than Reflectivity of the band.

[0020] Preferably, the method for constructing the suspended matter index is as follows:

[0021] When the remote sensing reflectance of two characteristic bands is positively correlated with the measured suspended matter concentration in one band and negatively correlated in the other, the following formula is used for calculation: ,

[0022] in, The suspended solids index, The band with positive correlation. It is a negative correlation band.

[0023] Preferably, the input of the XGBoost model is The vector contains the spectral values ​​of characteristic band 1, the spectral values ​​of characteristic band 2, and the suspended matter index. The output of XGBoost is a suspended matter concentration.

[0024] Preferably, the structural parameters of the XGBoost model are set as follows: the maximum tree depth is 8, the number of trees is 100, the feature sampling ratio of each tree is 0.7, the subsample ratio is 0.7, the learning rate is 0.01, the L1 regularization coefficient is 0, and the L2 regularization coefficient is 1.

[0025] Preferably, when training the XGBoost model using training data, the feature segment spectral values ​​and the suspended matter index are used as training data.

[0026] Preferably, when applying the trained XGBoost model to the UAV remote sensing reflectance product to invert the suspended matter concentration, the characteristic band spectral value and the calculated suspended matter index are input into the trained XGBoost model for each pixel to obtain the spatial distribution of suspended matter concentration.

[0027] The beneficial effects of this invention are as follows: By accurately identifying the most sensitive feature information to the response of suspended solids concentration from a large number of bands, the data dimensionality and computational complexity of subsequent models are greatly reduced. Using these two selected feature bands to construct a suspended solids index, and employing a method of combining the information from the two most relevant bands in ratio or other forms, the response signal to changes in suspended solids concentration can be enhanced, and interference from water background, atmosphere, and other factors can be suppressed. Compared to traditional linear / nonlinear regression models, the XGBoost model used in this method has stronger nonlinear fitting and generalization capabilities and is less prone to overfitting. Compared to other complex machine learning models, this method provides the model with highly condensed and low-redundancy input through the prior feature band selection and feature index construction. This not only significantly improves the model's training efficiency and inversion accuracy but also makes the model more robust to noise. Ultimately, it enables rapid, accurate, and large-scale mapping of suspended solids concentration in complex aquatic environments, effectively overcoming the spatiotemporal limitations of traditional monitoring methods and solving the problems of weak model generalization ability and high computational cost commonly found in existing remote sensing inversion technologies. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A schematic diagram illustrating the steps of an airborne remote sensing inversion method for XGBoost suspended matter concentration based on feature band selection provided by this invention;

[0030] Figure 2 A schematic diagram illustrating the feature band selection of an XGBoost airborne remote sensing inversion method for suspended matter concentration based on feature band selection provided by this invention;

[0031] Figure 3 Spatial distribution map of suspended matter concentration when applying the XGBoost airborne remote sensing inversion method for suspended matter concentration based on feature band selection provided by the present invention under clear weather conditions.

[0032] Figure 4 Spatial distribution map of suspended matter concentration when applying the XGBoost airborne remote sensing inversion method for suspended matter concentration based on feature band selection provided by the present invention under cloudy conditions.

[0033] Figure 5 A scatter plot comparing the inverted values ​​and measured values ​​of suspended matter concentration in an application of the XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection provided by this invention under cloudy conditions. Detailed Implementation

[0034] The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and arrangements of specific examples are described below. Of course, these are merely examples and are not intended to limit the present invention.

[0035] The embodiments of the invention will now be described in detail with reference to the accompanying drawings.

[0036] like Figure 1 As shown, an airborne remote sensing inversion method for suspended matter concentration based on feature band selection using XGBoost includes the following steps:

[0037] Perform multispectral remote sensing data preprocessing to generate UAV remote sensing reflectance products;

[0038] Using the UAV remote sensing reflectance products and measured suspended matter concentration data, the correlation coefficient is calculated to determine the optimal characteristic band;

[0039] Construct a suspended matter index using the data from the optimal characteristic bands;

[0040] An XGBoost model is constructed that takes the spectral values ​​of the characteristic bands and the suspended matter index as inputs and the suspended matter concentration as output.

[0041] The XGBoost model is trained using training data;

[0042] The trained XGBoost model was applied to the UAV remote sensing reflectance product to retrieve the suspended matter concentration.

[0043] In its specific implementation, this invention first uses a multispectral sensor mounted on a drone to take aerial photographs of the target water area, acquiring raw image files containing multiple bands. Five to six sets of images are then precisely registered to eliminate spatial positional deviations between bands. Subsequently, orthophoto stitching is performed based on the registered images to generate a complete orthophoto map with a unified coordinate system and original digital values. Using calibration gray boards with different reflectance levels placed synchronously on-site, a quantitative relationship between the image DN value and the actual surface reflectance is established. This relationship is then applied to radiometric calibration of the entire image, ultimately generating a drone-based remote sensing reflectance product for analysis. After acquiring high-quality reflectance data, the coordinate data of pre-deployed ground water sampling points (which can be GPS coordinates) is used. These ground water sampling points include sampling points in sunny, non-shaded areas, sunny, shaded areas, and cloudy areas.

[0044] The following steps are used to process sampling points in the shaded area on a sunny day:

[0045] (1) Collect 5-10 sets of data on both sides of the dividing line at the junction of shadow and non-shadow to form data pairs.

[0046] (2) Match the reflectance of different bands of the data pair band by band to obtain the band reflectance correlation between the shadow area and the non-shadow area. This relationship is called the relationship R between the reflectance after shadow correction and the reflectance before shadow correction in this method.

[0047] (3) The reflectance of the shadow area band is corrected using the correction relationship R to obtain the corrected reflectance value of the shadow area.

[0048] The specific criteria for determining the shaded area on a sunny day are as follows:

[0049] A: The shooting time was during the day. The overall brightness histogram of the image shows a wide or bimodal distribution, with high contrast and clear shadows that can be identified.

[0050] B: In response to the aforementioned shadow effects, a dynamic threshold is set for the near-infrared band of the image. Any pixel whose near-infrared band reflectance is less than the dynamic threshold is initially marked as a candidate shadow pixel. The threshold can be determined based on the statistical distribution of the near-infrared band reflectance of the entire image.

[0051] C: Calculate the band ratio of the reflectance of the blue-green band to the red band of the candidate shadow pixels. Since this ratio will be abnormally high in the shadow of a sunny day, a ratio threshold is set here. Pixels in the candidate shadow pixels that meet the band ratio greater than the ratio threshold are identified as shadow pixels.

[0052] D: Perform morphological processing on the set of pixels identified as shadow pixels. First, use opening operations to remove isolated noisy bright spots, and then use closing operations to fill the small holes inside the shadow, forming a smooth, connected sunny shadow area.

[0053] The following steps are used to process sampling points in non-shaded areas on sunny days:

[0054] (1) Perform glare correction on the image. Taking advantage of the fact that the reflectivity of the water surface in the near-infrared band should be close to zero in theory, set a dynamic threshold, identify pixels with abnormally high reflectivity in the near-infrared band as glare pollution areas, and use the mean value of the pixels of the adjacent pure water body for interpolation replacement, thereby eliminating glare noise.

[0055] (2) The ratio of blue light band to near-infrared band is used to effectively identify water shadows and generate shadow masks. In subsequent calculations, these contaminated pixels are directly removed to avoid interference with correlation analysis and model training.

[0056] The criteria for determining non-shaded areas on sunny days are: the image histogram shows a wide or bimodal distribution with high contrast; clear ground or water surface shadows are identified, such as low reflectivity areas in the near-infrared band; and all pixels not covered by the shadow areas on sunny days are included.

[0057] The following steps are used to process sampling points on cloudy days:

[0058] (1) All bands are subjected to decorrelation stretching. Through principal component analysis transformation, the original highly correlated multispectral band data are projected onto a new orthogonal coordinate system. Histogram equalization is performed independently in each principal component direction to stretch its dynamic range. Finally, it is inversely transformed back to the original color space.

[0059] (2) After the data is enhanced, the spectral differences between different water components are significantly amplified. Then, correlation analysis, screening of sensitive bands, and construction of 3D vectors are performed to capture the effective information that was submerged by the original weak signal.

[0060] The criteria for determining whether an image is cloudy are as follows: the image was taken during the day; the overall brightness histogram of the image shows a narrow, single-peak distribution with low contrast; and no clear, directional shadow areas were found through edge detection or shadow detection algorithms.

[0061] To pinpoint the exact location on the reflectivity product, and to reduce random noise and sensor error, the average reflectivity within a 5x5 pixel window surrounding each sampling point was selected as the representative reflectivity data for that point. Next, a band-by-band correlation analysis was performed between these representative reflectivity data and precisely measured suspended matter concentrations in the laboratory. By calculating correlation coefficients and other indicators, the two bands with the highest absolute correlation coefficients were selected as the optimal characteristic bands. Figure 2 The diagram shows a schematic of characteristic band selection. The horizontal axis represents the average representative reflectance of different bands corresponding to the water sampling point locations on the aerial remote sensing image. The vertical axis represents the suspended solids concentration precisely measured in the laboratory. Figure 2 Based on the mid-band correlation scatter plot results, this study selected bands 1 and 3 as feature bands. This core step, compared to existing techniques, yields significant advantages: by accurately identifying the most sensitive features to suspended particulate matter (SPM) concentration from a large number of bands, it greatly reduces data dimensionality and the computational complexity of subsequent models. Using these two selected feature bands to construct a SPM index, combining the information from the two most correlated bands by ratio or other means enhances the response signal to changes in SPM concentration and suppresses interference from water background, atmosphere, and other factors. Finally, the reflectance values ​​of the two selected feature bands and the constructed SPM index are used as input features to train the XGBoost model. This model is then used to perform pixel-by-pixel concentration inversion on the entire remote sensing image, outputting the SPM concentration, as shown below. Figure 3 and Figure 4 The spatial distribution map of suspended solids concentration is shown. Compared with traditional linear / nonlinear regression models, the XGBoost model used in this method has stronger nonlinear fitting and generalization capabilities and is less prone to overfitting. Compared with other complex machine learning models, this method provides the model with highly condensed and low-redundancy inputs through the prior feature band selection and feature index construction. This not only significantly improves the model's training efficiency and inversion accuracy but also makes the model more robust to noise. Ultimately, it can achieve rapid, accurate, and large-scale mapping of suspended solids concentration in complex aquatic environments, effectively overcoming the spatiotemporal limitations of traditional monitoring methods and solving the problems of weak model generalization ability and high computational cost that are common in existing remote sensing inversion technologies.

[0062] More specifically, the preprocessing of multispectral remote sensing data includes: band registration, orthophoto stitching, and radiometric calibration of the spectral data collected by the UAV.

[0063] Because multispectral cameras on drones typically have independent sensors for each band, differences in sensor position, as well as minute vibrations and attitude changes during flight, can cause pixel-level misalignment between images of different bands. During band registration, corresponding points in each band's image are automatically identified, and geometric transformation algorithms are applied to align all band images onto a unified pixel grid, ensuring that any pixel in the image precisely corresponds to the same location on the ground in all bands. Next, during orthophoto stitching, multiple independently captured images with overlapping areas, already band-registered, are combined using coordinate and attitude data recorded by the drone. Calculations based on photogrammetry principles are then performed to eliminate perspective distortion caused by camera tilt and terrain undulations. These images are then stitched together into an orthophoto map covering the entire study area with unified true geographic coordinates and scale. Finally, during radiometric calibration, standard grayscale plates with known reflectances are laid out on the ground before aerial photography. The average DN value of these grayscale plates is extracted from the generated orthophoto and a linear regression relationship is established with their known standard reflectances. This relationship is then applied to every pixel of the entire image, thereby converting the physically meaningless DN value into a standardized surface reflectance with clear physical meaning.

[0064] More specifically, when calculating the correlation coefficient using the UAV remote sensing reflectance product and measured suspended matter concentration data, the average reflectance of multiple pixels around the sampling point is selected as the reflectance data corresponding to the sampling point.

[0065] To improve the robustness and generalization of the XGBoost model, the sampling point data and spectral data cover three main operating scenarios: sunny days, sunny-shaded areas, and cloudy days. This allows XGBoost to simultaneously consider the spectral characteristics of water bodies corresponding to different suspended solids concentrations under these three conditions.

[0066] Using high-precision coordinate data recorded during ground water sampling, precise spatial positioning is achieved on the pre-processed UAV remote sensing reflectance product to locate the central pixel corresponding to each sampling point. Using this central pixel as a base point, an N×N pixel window (e.g., 3×3 or 5×5) is defined, and the reflectance values ​​of all pixels within this window are extracted. Then, for each spectral band, the arithmetic mean of the reflectance values ​​of all pixels within the window is calculated independently. Finally, this set of calculated average reflectance values ​​encompassing all bands is used as a unique and stable spectral feature representing the location of the physical sampling point. This data is then paired with suspended matter concentration values ​​measured in the laboratory to construct a dataset for model training and validation.

[0067] More specifically, the method for determining the optimal characteristic band is as follows:

[0068] Calculate the correlation coefficient between the reflectance data of each sampling point and the measured suspended matter concentration in different spectral bands;

[0069] The two bands with the highest absolute values ​​of correlation coefficients were selected as characteristic bands.

[0070] This process first requires a paired dataset, where each dataset contains two parts: one part is the reflectance value of a water sample point across all spectral bands, i.e., the average value of surrounding pixels obtained in the previous step; the other part is the corresponding suspended solids concentration value, precisely measured in the laboratory. This dataset is then processed band by band. Taking the first band as an example, we construct a vector from the reflectance values ​​of all sample points in that band, and another vector from the corresponding suspended solids concentration values. Then, we calculate the correlation coefficient between these two vectors, such as the Pearson correlation coefficient. For example, when using the Pearson correlation coefficient as a metric, this coefficient ranges from -1 to +1, and its absolute value indicates the strength of the linear correlation between the reflectance and suspended solids concentration in that band. We repeat this process for each spectral band, ultimately obtaining a list of correlation coefficients equal in length to the number of bands. Finally, we take the absolute value of all the correlation coefficients in this list and sort them in descending order. Then we select the two original bands corresponding to the top two and establish them as the "best characteristic bands" for subsequent modeling.

[0071] More specifically, the method for constructing the suspended matter index is as follows:

[0072] When the remote sensing reflectance of both characteristic bands is positively correlated with the measured suspended matter concentration, or both are negatively correlated, the following formula is used for calculation: ,

[0073] in, The suspended solids index, The reflectivity of the band is greater than Reflectivity of the band.

[0074] More specifically, the method for constructing the suspended matter index is as follows:

[0075] When the remote sensing reflectance of two characteristic bands is positively correlated with the measured suspended matter concentration in one band and negatively correlated in the other, the following formula is used for calculation: ,

[0076] in, The suspended solids index, The band with positive correlation. It is a negative correlation band.

[0077] More specifically, the input to the XGBoost model is The vector contains the spectral values ​​of characteristic band 1, the spectral values ​​of characteristic band 2, and the suspended matter index. The output of XGBoost is a suspended matter concentration.

[0078] More specifically, the structural parameters of the XGBoost model are set as follows: the maximum tree depth is 8, the number of trees is 100, the feature sampling ratio of each tree is 0.7, the subsample ratio is 0.7, the learning rate is 0.01, the L1 regularization coefficient is 0, and the L2 regularization coefficient is 1.

[0079] More specifically, when training the XGBoost model using training data, the feature segment spectral values ​​and the suspended matter index are used as training data.

[0080] More specifically, when applying the trained XGBoost model to the UAV remote sensing reflectance product to invert the suspended matter concentration, the characteristic band spectral value and the calculated suspended matter index are input into the trained XGBoost model for each pixel to obtain the spatial distribution of suspended matter concentration.

[0081] Using the optimal feature bands established in the previous steps, such as bands A and B under clear or cloudy conditions, and the formula for constructing the suspended matter index, the entire UAV remote sensing reflectance product is raster-calculated. This means the system first generates a completely new single-band raster image, namely the spatial distribution map of the suspended matter index. The value of each pixel in this image is calculated from its reflectance values ​​in bands A and B using the suspended matter index formula. At this point, three spatially fully registered input data layers are available: the reflectance map of feature band A, the reflectance map of feature band B, and the newly generated suspended matter index map. Next, the system loads the trained and saved XGBoost model and iterates through every pixel of the entire image. For any pixel, the system simultaneously extracts its corresponding values ​​in these three layers and combines these three values ​​into an input vector of [reflectance A, reflectance B, suspended matter index value], which is then fed into the XGBoost model for prediction. The model immediately outputs a single numerical value, the predicted concentration of suspended matter in the area represented by that pixel. This process is performed pixel by pixel across the entire map, ultimately generating a completely new, registered single-band image of the same size as the original. However, the physical meaning of this image has fundamentally changed. The grayscale value of each pixel is no longer reflectance, but represents the specific numerical value of suspended matter concentration at that point. This achieves the transformation from spectral information to quantitative information on water parameters, resulting in the final spatial distribution map of suspended matter concentration. This distribution map will vary depending on the operating conditions, such as... Figure 3 This is a spatial distribution map of suspended solids concentration under clear weather conditions. Figure 4 This is a spatial distribution map of suspended solids concentration during application under cloudy conditions, and also... Figure 5 As shown, under cloudy conditions, the R-squared value of the XGBoost model inversion is 0.89, indicating high accuracy.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for airborne remote sensing inversion of suspended matter concentration based on characteristic band selection using XGBoost, characterized in that, Includes the following steps: Multispectral remote sensing data preprocessing is performed to generate UAV remote sensing reflectance products; using the UAV remote sensing reflectance products and measured suspended matter concentration data, correlation coefficients are calculated to determine the optimal characteristic band; a suspended matter index is constructed using the optimal characteristic band data; an XGBoost model is constructed with the spectral values ​​of the characteristic band and the suspended matter index as inputs and the suspended matter concentration as the output; the XGBoost model is trained using training data; The trained XGBoost model is applied to the UAV remote sensing reflectance product to retrieve the suspended matter concentration. The preprocessing steps include: determining whether the weather condition corresponding to the multispectral remote sensing data is sunny or cloudy; if the weather condition is sunny, a sunny data processing procedure is executed, which includes: identifying sunny shadow areas in the image; collecting data pairs on both sides of the boundary between the sunny shadow area and the sunny non-shadow area to establish a correction relationship for the reflectance of the shadow area; applying the correction relationship to correct the sampling point data within the shadow area to obtain the corrected reflectance data for the sunny shadow area; and performing glare correction and shadow removal on the sunny non-shadow area. If the weather condition is cloudy, a cloudy data processing procedure is executed, which includes: decorrelation stretching of all bands, which includes principal component analysis transformation of the data, histogram equalization in the principal component direction, and inverse transformation back to the original color space to enhance spectral differences, resulting in a cloudy corrected reflectance product.

2. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, The preprocessing of multispectral remote sensing data specifically involves: band registration, orthophoto stitching, and radiometric calibration of the spectral data collected by the UAV.

3. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, When calculating the correlation coefficient using the UAV remote sensing reflectance product and measured suspended matter concentration data, the average reflectance of multiple pixels around the sampling point is selected as the reflectance data corresponding to the sampling point.

4. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, The method for determining the optimal characteristic band is as follows: calculate the correlation coefficient between the reflectance data of each sampling point and the measured suspended matter concentration in different bands; select the bands with the highest absolute values ​​of the correlation coefficients as characteristic bands.

5. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, The method for constructing the suspended matter index is as follows: when the remote sensing reflectance of both characteristic bands is positively correlated with or negatively correlated with the measured suspended matter concentration, the following formula is used for calculation: in, The suspended solids index, The reflectivity of the band is greater than Reflectivity of the band.

6. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, The method for constructing the suspended matter index is as follows: when the remote sensing reflectance of two characteristic bands is positively correlated with the measured suspended matter concentration and negatively correlated with the concentration of one, the index is calculated using the following formula: ,in, The suspended solids index, The band with positive correlation. It is a negative correlation band.

7. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, The input to the XGBoost model is The vector contains the spectral values ​​of characteristic band 1, the spectral values ​​of characteristic band 2, and the suspended matter index. The output of XGBoost is a suspended matter concentration.

8. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 7, characterized in that, The structural parameters of the XGBoost model are set as follows: maximum tree depth is 8, number of trees is 100, feature sampling ratio of each tree is 0.7, subsample ratio is 0.7, learning rate is 0.01, L1 regularization coefficient is 0, and L2 regularization coefficient is 1.

9. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, When training the XGBoost model using training data, the feature segment spectral values ​​and the suspended matter index are used as training data.

10. The XGBoost airborne remote sensing inversion method for suspended matter concentration based on characteristic band selection according to claim 1, characterized in that, When the trained XGBoost model is applied to the UAV remote sensing reflectance product to invert the suspended matter concentration, the characteristic band spectral value and the calculated suspended matter index are input into the trained XGBoost model for each pixel to obtain the spatial distribution of suspended matter concentration.

Images