River flow measuring and calculating method based on multi-source remote sensing data and deep learning

Abstract

The invention discloses a river flow measuring and calculating method based on multi-source remote sensing data and deep learning, and belongs to the technical field of water resource monitoring. Comprising the following steps: obtaining an original remote sensing image, and carrying out geometric correction and atmospheric correction on the original remote sensing image to obtain a calibrated multispectral image; calculating a normalized difference water body index NDWI by using near infrared and green wave bands, and combining the normalized difference water body index NDWI with the calibrated multispectral image to form a multi-channel input tensor containing RGB, NDWI and elevation information as multi-source input data; the UNet model is improved; generating a water body probability graph by using the multi-source input data through an improved UNet model; and based on the water body probability graph, extracting a feature vector of the river, and predicting the river flow through an XGoos machine learning regression model. According to the invention, the efficiency and accuracy of water resource monitoring and management can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of water resources monitoring, and more specifically to a river flow measurement method based on multi-source remote sensing data and deep learning. Background Art

[0002] Accurate estimation of river flow is of great significance for water resource management, flood warning and ecological protection. Traditional river flow monitoring mainly relies on ground measurement stations, but due to the uneven spatial distribution of these measurement stations and the high cost of manpower and material resources, it is difficult to obtain river flow information in a timely and comprehensive manner. This limitation leads to challenges in water resource management and disaster prevention and mitigation, affecting the accuracy and reliability of the overall process.

[0003] Existing remote sensing technology provides a new approach for river flow estimation, but there are also some problems. First, the traditional NDWI (Normalized Difference Water Index) calculation method and model fusion are not sufficient. Usually, NDWI is directly calculated and used as a separate input channel, failing to fully utilize the correlation between NDWI and other spectral channels. Using NDWI as a preprocessing step may introduce additional calculation errors and data redundancy, and lacks an effective mechanism for dynamically learning and extracting NDWI features within the model, which limits the accuracy of water body segmentation. Secondly, the UNet model also shows certain limitations when dealing with complex scenes. Due to its relatively simple structure, it may not be able to fully capture the deep features in the data when faced with multi-source remote sensing data and complex backgrounds (such as high noise environments), especially in the fusion of elevation data and spectral indices. The UNet model lacks the ability to adapt to multimodal data, which limits its performance in accurately distinguishing water bodies from non-water bodies.

[0004] In the process of extracting river width, water surface area and water surface elevation from water body masks, existing technologies mainly rely on traditional image processing methods, such as morphological operations and skeleton extraction. However, when the river morphology is complex or the degree of curvature is high, these methods may not be able to accurately extract the centerline and cross-section of the river, resulting in errors in the calculation of river width. The calculation of water surface area is also limited by the spatial resolution of the image and the accuracy of the mask. At the same time, the difference in spatial resolution between elevation data and image data may introduce matching errors, affecting the accurate extraction of water surface elevation.

[0005] Finally, river flow prediction models also have shortcomings. Traditional models, such as simple linear regression or basic machine learning algorithms, may not be able to capture the complex nonlinear relationships between river flow and multiple characteristics. These models usually fail to fully utilize the latest machine learning algorithms (such as deep learning and ensemble learning) to improve prediction accuracy. In addition, the model may lack consideration of time series data and spatial heterogeneity, resulting in inaccurate and unstable predictions of flow changes.

[0006] Therefore, how to provide a river flow measurement method that can solve the above problems is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0007] In view of this, the present invention provides a river flow measurement method based on multi-source remote sensing data and deep learning, so as to solve the technical problems existing in the above-mentioned prior art.

[0008] In order to achieve the above object, the present invention provides the following technical solutions:

[0009] A river flow measurement method based on multi-source remote sensing data and deep learning includes the following steps:

[0010] Acquire an original remote sensing image, and perform geometric correction and atmospheric correction on the original remote sensing image to obtain a calibrated multispectral image;

[0011] The normalized difference water index NDWI is calculated using the near-infrared and green bands, and the normalized difference water index NDWI is combined with the calibrated multispectral image to form a multi-channel input tensor containing RGB, NDWI and elevation information as multi-source input data;

[0012] Improve UNet model;

[0013] Generate a water body probability map by using the multi-source input data through an improved UNet model;

[0014] Based on the water body probability map, the characteristic vector of the river is extracted, and the river flow is predicted using the XGBoos machine learning regression model.

[0015] Preferably, the geometric correction of the original remote sensing image comprises:

[0016] Using geometric correction function The original remote sensing image I raw The pixel coordinates (x, y) are converted to the corrected geographic coordinates (x', y'), and the formula is expressed as:

[0017]

[0018] Map the pixel coordinates (x, y) of the original remote sensing image to the geographic reference coordinates (x', y'),

[0019]

[0020] Among them, Θ includes the set of sensor attitude parameters, earth curvature parameters, terrain elevation data DEM, and systematic distortion parameters of the imaging system;

[0021] After geometric correction, the corrected remote sensing image is obtained, and the formula is:

[0022] I geo (x′, y′)=I raw (x, y);

[0023] In the formula, I geo It is the corrected remote sensing image.

[0024] Preferably, performing atmospheric correction on the corrected remote sensing image to obtain calibrated multispectral data includes:

[0025] Through the atmospheric correction function Eliminate the influence of atmospheric scattering and absorption on the radiation value received by the sensor, and convert the top of atmosphere radiation brightness I geo (λ) is converted to surface reflectivity I cal (λ), the formula is:

[0026]

[0027] Among them, Φ includes the set of atmospheric model parameters atmospheric pressure and solar zenith angle;

[0028] Atmospheric correction is achieved through the radiation transfer equation, which is:

[0029]

[0030] Among them, L path (λ) is the atmospheric path radiation, T v (λ) and T s (λ) are the atmospheric transmittance in the observation direction and the solar incident direction, E sun (λ) is the solar irradiance at the top of the atmosphere, θ s is the solar zenith angle, and π is a mathematical constant.

[0031] Preferably, the normalized difference water index NDWI is calculated using the near infrared and green bands, and the formula is:

[0032]

[0033] In the formula, I NIR is the near infrared channel, I G is the green channel, NDWI(x, y) represents the NDWI index at the x, y position, I G (x, y) represents the green channel value at the x, y position, I NIR (x, y) represents the infrared channel value corresponding to the x, y coordinate position.

[0034] Preferably, the improved UNet model comprises:

[0035] Each layer of the encoder part is represented as:

[0036]

[0037] l=1,2,...,L;

[0038] In the formula, is the convolution kernel weight matrix of the l-th layer encoder, E l-1 is the feature map of the l-1th layer encoder, is the bias term of the encoder at the lth layer.

[0039] Preferably, L=5.

[0040] The decoder part is reconstructed by upsampling and feature concatenation, expressed as:

[0041]

[0042] D l is the output feature map of the l-1 layer decoder, E l-1 is the feature map output of the l-1 layer encoder.

[0043] Preferably, the water body probability map P(x, y) is generated by using the multi-source input data through the improved UNet model:

[0044] P(x, y) = Sigmoid(W out *D0(x,y)+b out );

[0045] Among them, b out is the bias term of the output layer, W out is the convolution kernel weight matrix of the output layer, D0(x, y) is the final feature map output of the decoder, Indicates feature connection, and UpSample is an upsampling operation.

[0046] Preferably, after obtaining the water body probability map, it also includes generating a binary water body mask M(x, y) through a threshold T:

[0047]

[0048] Where P(x, y) is the water body probability map, and T represents the probability threshold for determining it as a water body.

[0049] Preferably, after obtaining the binary water body mask, a characteristic vector of the river is obtained according to the binary water body mask, wherein the characteristic vector includes: river width, water surface area and water surface elevation;

[0050] Among them, extracting the river width W includes:

[0051] Perform morphological thinning operations on the water body mask M(x, y) to extract the center line S(x, y) of the river;

[0052] At each pixel point on the center line S(x, y), calculate the cross section of the pixel point perpendicular to S(x, y);

[0053] For each cross section, calculate the intersection with the water mask M(x, y), measure the distance between the intersections, and get the river width w at the cross section. i ;

[0054] The river widths of all cross sections are counted and the average river width W is calculated:

[0055]

[0056] Where N represents the total number of cross sections;

[0057] Extract water surface area A includes:

[0058] Traverse the water mask M(x, y) and count the number of water pixels n water ;

[0059] According to the spatial resolution of the remote sensing image, the actual area s corresponding to a single pixel is calculated to obtain the water surface area A:

[0060] A=n water ×s;

[0061] Extracting the water surface elevation H includes:

[0062] Obtain the elevation data H of the corresponding area from the altimetry satellite data (x, y), at the position where the water body mask M(x, y) = 1, extract the corresponding elevation value to form a water body elevation set, and calculate the average water surface elevation:

[0063]

[0064] Among them, (x, y)∈Water represents the coordinate set of water body pixels.

[0065] Preferably, the extracted feature vector is input into the XGBoos machine learning regression model for traffic prediction, and its prediction formula is expressed as:

[0066] Q=h φ (F);

[0067] Among them, F = [W, A, H], is the extracted feature vector, h φ is the XGBoost regression model, and φ is the model parameter.

[0068] Preferably, the XGBoos machine learning regression model is trained by minimizing the mean square error loss function:

[0069]

[0070] Where M is the number of training samples, is the real flow value;

[0071] During the training process, the parameters of the combined improved UNet model and the XGBoos machine learning regression model are adjusted by segmentation loss and flow prediction loss Combined to get the total loss function:

[0072]

[0073] Among them, γ is the loss balance coefficient, α and β are regularization coefficients to prevent overfitting.

[0074] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a river flow measurement method based on multi-source remote sensing data and deep learning, which improves the accuracy of geometric correction and atmospheric correction, optimizes the calculation and model fusion of NDWI, enhances the adaptability of the UNet model to multimodal data, and improves the river feature extraction method, which helps to improve the accuracy of water body segmentation and flow prediction. At the same time, by introducing the most advanced machine learning regression model and making full use of the advantages of multi-source remote sensing data, it can provide more powerful technical support for water resources management, flood prevention and disaster reduction and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0076] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0077] Figure 2 It is a schematic diagram of the river width calculation principle at the cross section of the present invention. DETAILED DESCRIPTION

[0078] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0079] See also Figure 1 The embodiment of the present invention discloses a river flow measurement method based on multi-source remote sensing data and deep learning, comprising the following steps:

[0080] Acquire an original remote sensing image, and perform geometric correction and atmospheric correction on the original remote sensing image to obtain a calibrated multispectral image;

[0081] The normalized difference water index NDWI is calculated using the near-infrared and green bands, and the normalized difference water index NDWI is combined with the calibrated multispectral image to form a multi-channel input tensor containing RGB, NDWI and elevation information as multi-source input data;

[0082] Improve UNet model;

[0083] Generate a water body probability map by using the multi-source input data through an improved UNet model;

[0084] Based on the water body probability map, the characteristic vector of the river is extracted, and the river flow is predicted using the XGBoos machine learning regression model.

[0085] Specifically, in the field of remote sensing, multispectral images and elevation data are two extremely valuable and complementary data sources. Multispectral images reveal the material composition and characteristics of the surface by capturing spectral information from visible light to near-infrared range, and are an important tool for analyzing and monitoring surface elements such as water bodies, vegetation, and soil. Especially in water body monitoring, multispectral data can effectively distinguish water bodies from non-water areas by using spectral responses in different bands. On the other hand, elevation data usually comes from radar measurements and provides accurate three-dimensional morphological information of the surface. These data are crucial for understanding the elevation changes, terrain characteristics, and vertical distribution of water bodies on the surface, and are helpful for analyzing river water level changes and basin morphology.

[0086] In a specific embodiment, combining multispectral data with elevation data can give full play to the complementary advantages of multi-source data and achieve a more comprehensive and accurate analysis of surface features. The integration of multi-source data not only enhances the ability to identify and classify surface features, but also improves the accuracy and reliability of the analysis results. In the process of water body detection and analysis, multispectral data can be used to calculate spectral indices such as NDWI to clearly distinguish between water surface and other surface types; elevation data provides spatial location and height information of water bodies, assisting in the accurate measurement of river width, water surface area and water level elevation.

[0087] In a specific implementation, calibrated multispectral data is obtained by geometrically correcting and atmospherically correcting the original remote sensing image. NDWI is calculated using the near-infrared and green bands and combined with the calibrated RGB channels and elevation data to form a multi-channel input tensor containing RGB, NDWI, and elevation information. The improved UNet model uses this multi-source input data to generate an accurate water body probability map. On this basis, the key features of the river are extracted, and the river flow is accurately predicted through an advanced machine learning regression model. This process fully reflects the important application value of multi-source remote sensing data in water body analysis and water resources management, and helps to improve the efficiency and accuracy of water resources monitoring and management.

[0088] In a specific embodiment, first, the original remote sensing image I raw Perform geometric correction and atmospheric correction. Geometric correction function The pixel coordinates (x, y) of the original remote sensing image are converted to the corrected geographic coordinates (x', y'). The formula is:

[0089]

[0090] First, the original remote sensing image I raw Perform geometric correction and atmospheric correction. Geometric correction function It is used to correct the geometric distortion in the image and map the pixel coordinates (x, y) of the original remote sensing image to the geographic reference coordinates (x', y'). The formula is:

[0091]

[0092] Among them, Θ includes sensor attitude parameters (such as position, velocity, acceleration), earth curvature parameters, terrain elevation data (DEM), and systematic distortion parameters of the imaging system. After geometric correction, the corrected image I is obtained. geo , the formula is:

[0093] I geo (x′, y′)=I raw (x, y)

[0094] In a specific embodiment, atmospheric correction is then performed. The atmospheric correction function A is used to eliminate the influence of atmospheric scattering and absorption on the radiation value received by the sensor, and converts the atmospheric top radiance I geo (λ) is converted to surface reflectivity I cal (λ), the formula is:

[0095]

[0096] Among them, φ includes atmospheric model parameters (such as aerosol optical thickness, water vapor content, ozone concentration), atmospheric pressure, solar zenith angle, etc. Atmospheric correction is achieved through the radiation transfer equation:

[0097]

[0098] Among them, L path (λ) is the atmospheric path radiation, T v (λ) and T s (λ) are the atmospheric transmittance in the observation direction and the solar incident direction, E sun (λ) is the solar irradiance at the top of the atmosphere, θ s is the solar zenith angle, and π is a mathematical constant. Through geometric correction and atmospheric correction, the calibrated multispectral image I is obtained. cal .

[0099] Specifically, the above method first corrects the geometric distortion of the remote sensing image through the geometric correction function, and then accurately locates the image coordinates in the earth's geographic reference coordinate system. This is equivalent to correcting the error of the camera tilting in the first step. The second step is to map the pixel position captured by the camera to the position in the earth's coordinate system.

[0100] In one embodiment, the near infrared channel I is then used NIR and Green Channel I G The normalized difference water index (NDWI) was calculated using the formula:

[0101]

[0102] In the formula, I NIR is the near infrared channel, I G is the green channel, NDWI(x, y) represents the NDWI index at the x, y position, I G (x, y) represents the green channel value at the x, y position, I NIR (x, y) represents the infrared channel value corresponding to the x, y coordinate position. The pixel coordinates used here are used to retrieve the pixel value of the corresponding pixel position on the image.

[0103] In order to efficiently calculate NDWI in the model, a 4×1×1×1 convolution kernel W is defined NDWI :

[0104]

[0105] The numerator of the normalized difference water index NDWI is obtained through the convolution operation, and the formula is:

[0106] N num (x, y) = W NDWI *I cal (x, y);

[0107] The denominator of the normalized difference water index NDWI is calculated as:

[0108] N den (x, y) = I G (x,y)+I NIR (x, y)+∈;

[0109] Here, ∈ is a small constant that prevents the denominator from being zero.

[0110] In a specific embodiment, the normalized difference water index NDWI is calculated as:

[0111]

[0112] Where N num is the molecular part of the normalized difference water index NDWI, N den (x, y) is the denominator of the normalized difference water index NDWI.

[0113] In a specific embodiment, improving the UNet model includes:

[0114] Each layer of the encoder part is represented as:

[0115]

[0116] l=1,2,...,L;

[0117] In the formula, is the convolution kernel weight matrix of the l-th layer encoder, E l-1 is the feature map of the l-1th layer encoder, is the bias term of the encoder at the lth layer.

[0118] Specifically, L=5.

[0119] The decoder part is reconstructed by upsampling and feature concatenation, expressed as:

[0120]

[0121] l = L, L-1, ..., 1;

[0122] D l is the output feature map of the l-1 layer decoder, E l-1 is the feature map output of the l-1 layer encoder, 1 can take L-1 values.

[0123] Specifically, the water body probability map P(x, y) is generated by using the multi-source input data through the improved UNet model:

[0124] P(x, y) = Sigmoid(W out *D0(x,y)+b out );

[0125] Among them, b out is the bias term of the output layer, W out is the convolution kernel weight matrix of the output layer, D0(x, y) is the final feature map output of the decoder, indicating feature connection, and UpSample is the upsampling operation.

[0126] Specifically, after obtaining the water body probability map, it also includes generating a binary water body mask M(x, y) through a threshold T:

[0127]

[0128] Where P(x, y) is the water body probability map, and T represents the probability threshold for determining it as a water body.

[0129] According to the water mask, the river width W, water surface area A and water surface elevation H are extracted. The water mask M(x, y) is subjected to morphological thinning (thinning) operation to extract the center line S(x, y) of the river. At each pixel point of the center line S(x, y), the cross section perpendicular to S(x, y) is calculated. For each cross section, its intersection with the water mask M(x, y) is calculated, and the distance between the intersections is measured to obtain the river width w at the cross section. i , see Figure 2 As shown. The river widths of all cross sections are counted and the average river width W is calculated:

[0130]

[0131] Where N is the total number of cross sections.

[0132] Traverse the water body mask M(x, y) and count the number of water body pixels n water According to the spatial resolution of the remote sensing image (the actual ground size represented by each pixel), the actual area s corresponding to a single pixel is calculated.

[0133] A=nwater ×s

[0134] Obtain the elevation data H of the corresponding area from the altimetry satellite data (x, y), at the position where the water body mask M(x, y) = 1, extract the corresponding elevation value to form a water body elevation set, and calculate the average water surface elevation:

[0135]

[0136] Among them, (x, y)∈Water represents the coordinate set of water body pixels.

[0137] The extracted feature vector F = [W, A, H] is input into the selected XGBoos machine learning regression model for traffic prediction, and its prediction formula is expressed as:

[0138] Q=h φ (F)

[0139] Among them, h φ is the XGBoost regression model, and φ is the model parameter. The model is trained by minimizing the mean square error loss function:

[0140]

[0141] Where M is the number of training samples, is the actual flow value.

[0142] During the training process, the parameters of the UNet model and the regression model are jointly optimized, and the segmentation loss is combined through the total loss function. and flow prediction loss

[0143]

[0144] Among them, γ is the loss balance coefficient, α and β are regularization coefficients to prevent overfitting.

[0145] Specifically, in the inference phase, the input tensor is constructed using the new calibrated RGB image, the calculated NDWI, and the elevation data Predicting water body probability map p using the improved UNet model new , generate a new water mask M new , extract the new river width W new 、Water surface area A new and water surface elevation H new Finally, the new river flow Q is predicted by the regression model new :

[0146] Q new =h φ *(F new)

[0147] Among them, φ * These are the model parameters that are optimized after training.

[0148] Specifically, the original image is processed through geometric correction and atmospheric correction, NDWI is calculated using convolution operation and fused with RGB and elevation data, an improved UNet model is constructed for water segmentation, and then combined with the extracted river width, water surface area and elevation features, the SOTA machine learning regression model is used to achieve accurate river flow prediction. This process fully utilizes the advantages of multi-source remote sensing data and combines machine learning methods to improve the accuracy and reliability of water segmentation and flow prediction.

[0149] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0150] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A river flow measurement method based on multi-source remote sensing data and deep learning, characterized in that: The following steps are involved: Acquire an original remote sensing image, and perform geometric correction and atmospheric correction on the original remote sensing image to obtain a calibrated multispectral image; The normalized difference water index NDWI is calculated using the near-infrared and green bands, and the normalized difference water index NDWI is combined with the calibrated multispectral image to form a multi-channel input tensor containing RGB, NDWI and elevation information as multi-source input data; Improve UNet model; Generate a water body probability map by using the multi-source input data through an improved UNet model; Based on the water body probability map, the characteristic vector of the river is extracted, and the river flow is predicted using the XGBoos machine learning regression model.

2. A river flow measurement method based on multi-source remote sensing data and deep learning according to claim 1, characterized in that: The original remote sensing image is geometrically corrected, including: Using geometric correction function The original remote sensing image I raw The pixel coordinates (x, y) are converted to the corrected geographic coordinates (x', y'), and the formula is expressed as: Map the pixel coordinates (x, y) of the original remote sensing image to the geographic reference coordinates (x', y'), Among them, Θ is a set of sensor attitude parameters, earth curvature parameters, terrain elevation data DEM, and systematic distortion parameters of the imaging system; After geometric correction, the corrected remote sensing image is obtained, and the formula is: I geo (x',y')=I raw (x,y); In the formula, I geo It is the corrected remote sensing image.

3. A river flow measurement method based on multi-source remote sensing data and deep learning according to claim 2, characterized in that: Perform atmospheric correction on the corrected remote sensing image to obtain calibrated multispectral data, including: Through the atmospheric correction function Eliminate the influence of atmospheric scattering and absorption on the radiation value received by the sensor, and convert the top of atmosphere radiation brightness I geo (λ) is converted to surface reflectivity I cal (λ), the formula is: Among them, Φ includes the set of atmospheric model parameters atmospheric pressure and solar zenith angle; Atmospheric correction is achieved through the radiation transfer equation, which is: Among them, L path (λ) is the atmospheric path radiation, T v (λ) and T s (λ) are the atmospheric transmittance in the observation direction and the solar incident direction, E sun (λ) is the solar irradiance at the top of the atmosphere, θ s is the solar zenith angle, and π is a mathematical constant.

4. The method for calculating river flow based on multi-source remote sensing data and deep learning according to claim 1, characterized in that: The normalized difference water index NDWI is calculated using the near infrared and green bands, and the formula is: In the formula, I NIR is the near infrared channel, I G is the green channel, NDWI(x,y) represents the NDWI index at the x,y position, I G (x,y) represents the green channel value at the x,y position, I NIR (x,y) represents the infrared channel value corresponding to the x,y coordinate position.

5. The method for calculating river flow based on multi-source remote sensing data and deep learning according to claim 1, characterized in that: The improved UNet model includes: Each layer of the encoder part is represented as: l＝1,2,...,L； In the formula, is the convolution kernel weight matrix of the l-th layer encoder, E l-1 is the feature map of the l-1th layer encoder, is the bias term of the l-th layer encoder; The decoder part is reconstructed by upsampling and feature concatenation, expressed as: Where D l is the output feature map of the l-1 layer decoder, E l-1 is the feature map output by the l-1 layer encoder, is the bias term of the l-th layer decoder.

6. The method for calculating river flow based on multi-source remote sensing data and deep learning according to claim 1, characterized in that: The improved UNet model uses the multi-source input data to generate a water body probability map P(x, y): P(x,y)=Sigmoid(W out *D0(x,y)+b out ); Among them, b out is the bias term of the output layer, W out is the convolution kernel weight matrix of the output layer, D0(x,y) is the final feature map output of the decoder, Indicates feature connection, and UpSample is an upsampling operation.

7. The method for calculating river flow based on multi-source remote sensing data and deep learning according to claim 1, characterized in that: After obtaining the water body probability map, it also includes generating a binary water body mask M(x,y) through a threshold τ: Where P(x, y) is the water body probability map, and τ represents the probability threshold for determining it as a water body.

8. The method for calculating river flow based on multi-source remote sensing data and deep learning according to claim 7, characterized in that: After obtaining the binary water body mask, a feature vector of the river is obtained according to the binary water body mask, wherein the feature vector includes: river width, water surface area and water surface elevation; Among them, extracting the river width W includes: Perform morphological refinement on the water body mask M(x,y) to extract the center line S(x,y) of the river; At each pixel point on the center line S(x,y), calculate the cross section of the pixel point perpendicular to S(x,y); For each cross section, calculate the intersection with the water mask M(x,y), measure the distance between the intersections, and get the river width w at the cross section. i ; The river widths of all cross sections are counted and the average river width W is calculated: Where N represents the total number of cross sections; Extract water surface area A includes: Traverse the water mask M(x,y) and count the number of water pixels n water ; According to the spatial resolution of the remote sensing image, the actual area s corresponding to a single pixel is calculated to obtain the water surface area A: A=n water ×s; Extracting the water surface elevation H includes: Obtain the elevation data H of the corresponding area from the altimetry satellite data (x, y), at the position where the water body mask M(x, y) = 1, extract the corresponding elevation value to form a water body elevation set, and calculate the average water surface elevation: Among them, (x,y)∈Water represents the coordinate set of water body pixels.

9. A river flow measurement method based on multi-source remote sensing data and deep learning according to claim 8, characterized in that: The extracted feature vector is input into the XGBoos machine learning regression model for traffic prediction. The prediction formula is expressed as: Q=h φ (F); Among them, F = [W, A, H], is the extracted feature vector, h φ is the XGBoost regression model, and φ is the model parameter.

10. A river flow measurement method based on multi-source remote sensing data and deep learning according to claim 9, characterized in that: The XGBoos machine learning regression model is trained by minimizing the mean squared error loss function: Where M is the number of training samples, is the real flow value; During the training process, the parameters of the combined improved UNet model and the XGBoos machine learning regression model are calculated by segmentation loss. and flow prediction loss Combined to get the total loss function: Among them, γ is the loss balance coefficient, α and β are regularization coefficients to prevent overfitting.

Images