RCBAM model construction method for remote sensing inversion of river flow

Abstract

The invention discloses an RCBAM model construction method for remote sensing inversion of river flow. The method comprises the following steps: 1) collecting remote sensing image data; (2) high-precision DEM production; 3) obtaining a central line of a river channel; 4) river width and gradient acquisition; 5) river section shape extraction; 6) calculating an open-current area; and 7) estimating the river flow. According to the method, the section shape of the river channel is obtained based on a high-precision digital elevation model (DEM), and the section area of the river channel is dynamically calculated by combining instantaneous water surface information obtained by a remote sensing image, so that the problem that defects exist due to the fact that the overflowing area is obtained by depending on an empirical formula in a traditional BAM flow model is effectively solved, and river flow estimation considering the shape characteristics of the river channel is realized. And efficient and accurate river flow remote sensing inversion can be realized.

Description

Technical Field

[0001] The present invention relates to river flow remote sensing inversion technology, which belongs to the field of geographic information processing technology. In particular, it relates to a river flow inversion method for data-free areas based on river channel shape and Bayesian methods. Specifically, it relates to a method for constructing an RCBAM (River Channel Bayesian AMHG-Manning) model for remote sensing inversion of river flow, which can achieve efficient and accurate remote sensing inversion of river flow. Background Art

[0002] River flow, a crucial element in the hydrological cycle, is a fundamental and strategic data resource with profound impacts across numerous sectors. For example, in flood prevention and disaster reduction, river flow monitoring is crucial for early warning and prevention of disasters like floods and droughts. Timely and accurate flow data supports decision-making by management departments and reduces disaster losses. In water resource management, accurate river flow data is a crucial basis for rationally allocating water resources and planning the construction of water conservancy facilities.

[0003] Traditional river flow monitoring relies primarily on hydrological monitoring stations. However, due to factors such as geographic environment and economic development, the spatial distribution of hydrological monitoring stations is uneven. Remote, mountainous, or economically underdeveloped areas, in particular, have a sparse number of stations, resulting in widespread hydrological data scarcity.

[0004] Remote sensing, as a long-distance, non-contact information acquisition technology, offers the advantages of wide coverage, high observation frequency, and unrestricted geographic location. It can obtain near-real-time data on rivers worldwide, opening up new possibilities for river flow monitoring. Early studies estimated river flow by establishing area-flow relationships or width-flow relationships based on remotely sensed river area or width and measured flow. These methods, however, were simple and heavily reliant on empirical relationships between remotely sensed variables and field-measured flow. These relationships, however, were limited in accuracy and only applied to areas with available field-measured data, rendering them inapplicable in regions lacking such data.

[0005] In addition, some scholars use remote sensing observation data as input parameters of existing hydraulic equations (such as the Manning equation), or combine them with numerical hydraulic models to estimate river flow. The Bayesian AMHG-Manning (BAM) model is a typical representative of these models. It uses river width, slope and height data obtained from satellite remote sensing, and uses the Bayesian method to estimate river flow and quantify the uncertainty of the estimation results. BAM has shown certain advantages in river flow estimation. However, the important parameter flow area (A) in the algorithm is not well understood. it =A 0i +δA it ) is defective, the flow area Ait From the reference cross-sectional area A 0i and the changing flow area δA it The methods of obtaining the reference cross-sectional area by empirical formula and calculating by satellite altimetry data have the following defects: (1) Inaccuracy of the reference cross-sectional area: The reference cross-sectional area is set by using empirical formulas related to river width. The river width database has limited samples and cannot reflect the true characteristics of the complex and diverse rivers around the world. The empirical formula is inaccurate and difficult to correct uniformly; (2) Satellite altimetry data is limited and can only capture data on large rivers. It is not applicable to small and medium-sized rivers, and the data volume is relatively small. Summary of the Invention

[0006] To address the shortcomings of traditional BAM models in calculating flow area, this paper proposes a method for constructing a remotely sensed inversion (RCBAM) model for river flow that considers channel shape information. This model uses a proprietary, high-precision (5m) digital elevation model (DEM) to obtain the cross-sectional shape of the river channel. Combined with instantaneous water surface information captured by remote sensing imagery, the model dynamically calculates the flow area, thereby enabling flow estimation that takes channel shape into account.

[0007] The object of the present invention is achieved like this:

[0008] A method for constructing an RCBAM model for remote sensing inversion of river flow includes the following steps:

[0009] 1) Collect remote sensing image data:

[0010] Includes satellite foresight and forward-looking image pairs and time series optical remote sensing images.

[0011] 2) High-precision DEM generation:

[0012] The satellite front-view and back-view image pairs are used as input, and the digital surface model (DSM) is obtained through directional adjustment and stereo matching algorithms. The DSM is then filtered to remove artificial objects and vegetation to obtain high-precision numerical elevation model (DEM) data (5m).

[0013] Using satellite forward and rearward imagery combined with binocular stereo vision principles, a digital elevation model (DEM) is generated for the imaged area. Binocular stereo vision is an important form of machine vision. It is based on the principle of parallax and uses imaging equipment to obtain two images of the object from different positions. The method then calculates the positional deviation between corresponding points in the images to obtain the object's three-dimensional geometric information.

[0014] 3) Obtaining the river centerline:

[0015] The optical remote sensing image is used as input, the normalized water index (NDWI) is calculated to obtain a binary image of the water body, and the binary image of the water body is eroded to obtain the centerline of the river channel.

[0016] 4) Obtaining river width and slope:

[0017] Based on the river centerline, the vertical distance method is used to obtain the river width; the slope is obtained based on DEM data.

[0018] 5) River cross-section shape extraction:

[0019] In the DEM data, a section is cut perpendicular to the acquired river centerline to obtain the river section elevation data. Based on the river section elevation data, the discrete elevation points are linearly interpolated to obtain the river section shape curve.

[0020] 6) Calculation of flow area: The river width is matched with the river cross-section shape curve, and the Gauss-Kronrod integral is performed to obtain the flow area formed with the river cross-section under different river width conditions.

[0021] 7) The Manning formula and the multi-station hydraulic geometry method (AMHG) are used as the basic equations for flow estimation. The Bayesian method is used to perform logarithmic transformation on the Manning formula and the AMHG equation and construct a likelihood function. The Monte Carlo sampling method of Bayesian inference is used to extract samples from the specified distribution and calculate statistics to approximate the posterior distribution. The parameters used in the above formulas are: the reference cross-sectional area is set to the flow area corresponding to the minimum river width; the change in flow area is the difference between the instantaneous river width, the minimum river width, and the area formed by the river cross-section shape curve; the obtained river width, slope, and flow area are calculated according to Bayesian inference to obtain the flow estimation result.

[0022] In a further solution, the foreground and background image pairs used for DEM production in step 2) must be collected during the dry season of the river to maximize the inclusion of river channel shape information; the accuracy of the high-precision DEM is 5 meters. The specific steps of step 2) are:

[0023] 2-1 Use satellite image RPC parameters or ground control points to perform orientation adjustment on the image to obtain accurate orientation parameters. Use the accurate positioning model and orientation parameters to perform epipolar sampling on the foresight and backsight images of the image pair to obtain approximate epipolar images.

[0024] 2-2 Using the semi-global stereo matching algorithm (SGM), stereo matching and point intersection are performed on the approximate epipolar images to obtain a high-density 3D point cloud;

[0025] 2-3 Point cloud data registration generates a digital surface model (DSM) including ground buildings and vegetation, and then the DSM is filtered to remove artificial objects and vegetation to obtain a DEM.

[0026] In a further solution, in step 3), the normalized water index (NDWI) of the remote sensing image is calculated, and the OTSU adaptive threshold method is used to obtain a water body binary image, in which the water body pixel value is 1 and the non-water body pixel value is 0.

[0027]

[0028] Among them, ρ Green is the reflectivity of the green band, ρ NIR is the reflectivity in the near-infrared band.

[0029] Based on the binary image of the water body, a morphological thinning algorithm is used to continuously erode the edge of the water body while avoiding completely deleting the water body, ultimately leaving a center line with a width of 1 pixel.

[0030] For the image pixel coordinates (x, y) in the binary water image I (pixel value is 0 or 1, 1 represents water), the structural element B (m, n) (a small matrix with m and n as its coordinates) slides on the image. The erosion operation formula is:

[0031]

[0032] Repeat the erosion operation continuously until only a line with a width of 1 pixel remains in the water body. This line is the centerline of the river.

[0033] Step 3): Since the river width at different locations is inconsistent, corrosion judgment must be performed during corrosion. The judgment is divided into 8 cases: when it is judged that there is 1 data point around the pixel, the center point gradient is 7; when it is judged that there are 2 data points around the pixel, the center point gradient is 6; when it is judged that there are 3 data points around the pixel, the center point gradient is 5; when it is judged that there are 4 data points around the pixel, the center point gradient is 4; when it is judged that there are 5 data points around the pixel, the center point gradient is 3; when it is judged that there are 6 data points around the pixel, the center point gradient is 2; when it is judged that there are 7 data points around the pixel, the center point gradient is 1.

[0034] In a further solution, the river width is obtained using the perpendicular distance method described in step 4): starting from each point on the centerline, a search is performed in a direction perpendicular to the centerline to find two corresponding points on the water body boundary in the vertical direction. By calculating the distance between these two points, the river width at that location is obtained;

[0035] Assume that the centerline of the river is composed of a series of coordinate points (x i ,y i ),at point (x i, y i), the equation of the vertical line can be determined based on the slope of the tangent to the center line. If the slope of the tangent to the center line at this point is k, then the slope of the vertical line is -1 / k. The coordinates of the intersection of the equation of the line and the water body boundary (x j1 ,y j1 ) and (x j2 ,y j2 ), using the distance formula between two points: Calculate the river width W.

[0036] In step 4), the slope is obtained based on the DEM data: the slope is obtained by calculating the ratio of the elevation difference of pixels within or along the river channel to the horizontal distance within a certain distance;

[0037] The coordinates of the two pixels are (i, j) and (m, n), and their elevation values are Z i,j and Z m,n , then the elevation difference between them is ΔZ=Z i,j -Z m,n , horizontal distance Then the slope:

[0038]

[0039] In step 4): based on the long time series of optical images, obtain the long time series of river width and slope data of the selected section, with no less than 4 sections.

[0040] Further solution, step 5): the obtained river cross-section shape should include the entire river within the embankments on both sides of the river. Assuming that the coordinates of the river monitoring section are (x0, y0), a section is cut perpendicular to the river centerline obtained in step 3) in the DEM data obtained in step 2). Assuming that the section width is W, for each pixel (x, y) in the section, its elevation value Z(x, y) is directly obtained from the DEM data. In actual calculations, by traversing all pixels in the section, a series of elevation values Z1, Z2, ..., Z m .

[0041] Based on the profile elevation data, the discrete elevation points are processed by linear interpolation to draw a smooth profile line. Assume that two adjacent elevation points (x i ,Z i ) and (x i+1 ,Z i+1 ), the interval in the x direction is Δx, for x i and x i+1 For any point x between , its elevation Z can be calculated by linear interpolation formula:

[0042]

[0043] By performing such interpolation calculations on the points within the entire cross section, sufficiently dense elevation points are obtained, and then the river cross-section shape curve Z=f(x) is drawn.

[0044] In a further solution, in step 6), the original flow area calculation formula is as follows:

[0045] A it =A 0i +δAit

[0046] Among them, A it is the flow area, m 2 ; A 0i is the reference cross-sectional area, m 2 ; δA it is the change in flow area, which is calculated as follows:

[0047] First, the minimum river width length and long-term river width series obtained from remote sensing monitoring in step 4) are respectively aligned with the river cross-sectional shape line Z = f(x) drawn in step 5). The corresponding principle is: starting from the lowest point of the river channel f(x), the river width length is searched upward. When the river width length and the curves on both sides of f(x) intersect, it is considered to be the water surface at that time. Due to environmental factors, the two sides of the water surface may not be completely horizontal. Therefore, the intersection of the river width length and the two sides of the river channel shape is allowed to have a certain inclination angle and elevation difference. That is, the following conditions are met:

[0048]

[0049] Among them, h is the vertical height from bottom to top that meets the conditions, m; w is the river width, m; ∈ is the error, ∈ = 0.1m.

[0050] Use numerical iteration method to solve, from h min = min(f(x)), and gradually increase h; for each h, solve the equation f(x) = h to obtain the left and right intersection points x left (h) and x right (h); judge the following equation:

[0051]

[0052] When the above minimum value ≤∈, accept the solution. That is, find the corresponding position of river width and river channel.

[0053] After finding the corresponding relationship, the flow area formed by the river width and the cross-sectional shape is subjected to the Gauss-Kronrod integral, and the formula is as follows:

[0054]

[0055] Where A is the calculated flow area, m2 ; W is the river width, m; W / 2 is the scaling factor of the integration interval; m is the total number of nodes; W k is the weight corresponding to the integration node k; x k is the position of the integration node in the original interval [0,w].

[0056] Finally, the difference between the long-term river width and the area corresponding to the minimum river width is used to obtain the change in flow area:

[0057] δA it =A i -A min

[0058] Among them, δA it is the change in flow area, m 2 ; A i is the flow area corresponding to time i, m 2 ; A min is the flow area corresponding to the minimum river width in all time series, m 2 .

[0059] In step 6), the river width length obtained in the long time series is matched with the cross-sectional shape of the river channel. The elevation difference between the intersection points on the left and right sides does not exceed 10 cm. The flow area formed by the river width and the cross-sectional shape is calculated by Gauss-Kronrod integration, and the integration rule is G7-K15.

[0060] In a further solution, in step 7), the Manning formula and the multi-station hydraulic geometry method AMHG are used as the basic equations for flow estimation. The Bayesian method is used to perform logarithmic transformation on the Manning formula and the AMHG equation and construct the likelihood function. The Manning formula and the logarithmic transformation are respectively:

[0061]

[0062] Among them, Q t is the flow rate, m 3 / s; t is time; n is Manning coefficient, m -1 / 3 ·s;A it is the cross-sectional area: A 0i is the reference cross-sectional area, δA it is the change in flow area, m 2 ;W it is the river width, m; S it is the slope, dimensionless.

[0063] The AMHG equation and the logarithmic transformation are:

[0064]

[0065] log Wit =b i (log Q t -log Q c )+log W c +∈ g

[0066] Where Wit is the river width, m; Q t is the flow rate, m 3 / s;W c , Q c with b i is the AMHG parameter: Qc is the characteristic flow rate, m 3 / s; Wc is the global characteristic river width corresponding to Qc, m; bi is the width-discharge index, dimensionless; ∈ g is the AMHG error term, which is assumed to obey the normal distribution (∈g~N(0,σg)) and is dimensionless.

[0067] Prior distribution setting:

[0068] A 0i is the flow area corresponding to all minimum river widths in step 6), A 0i =A min ; logbi is defined by the regression model: logbi = 0.0216 + 0.4578 · SD (logWit) + ∈ b; Qt is set based on the global water model results and physical constraints (W min ×d min ×V min ≤Qt≤W max ×d max ×V max ) is a truncated normal distribution, where W min and W max is the minimum / maximum observed river width, m; d min =0.5m,d max =10m: minimum / maximum water depth; V min =0.5m / s,V max =5m / s: minimum / maximum flow velocity.

[0069] Joint posterior distribution:

[0070] p(Θ|x)∝f(x|θ)π(Θ)

[0071] Θ is a set of unknown parameters, including Q t 、n、A 0i 、W c , Q c , bi; x is the observed data (W it 、S it , δA it); f(x|Θ) is the joint likelihood function, which consists of the error terms of the Manning equation and the AMHG equation; π(Θ) is the product of the prior distribution of the parameters.

[0072] Using Bayesian Monte Carlo sampling, samples are drawn from a specified distribution and statistics (mean, standard deviation, and quantiles) are calculated to approximate the posterior distribution. The long-term river width and slope data from step 4) are combined with the long-term flow area data from step 6) using Bayesian Monte Carlo sampling to generate discharge estimates.

[0073] The advantages and beneficial effects of the present invention are:

[0074] This paper utilizes a proprietary high-precision digital elevation model (DEM) to fully account for the authenticity and spatial variability of river channel shape. Incorporating real-time water surface data, it develops a method for constructing an RCBAM model for remote sensing inversion of river flow. This method overcomes the uncertainty inherent in parameter acquisition in traditional BAM models and improves the accuracy of river flow estimation. Pearson correlation coefficient analysis with rainfall reveals that the correlation coefficients between RCBAM and BAM are 0.67 and 0.59, respectively, an improvement of 0.08 for RCBAM over BAM. This method enables efficient and accurate remote sensing inversion of river flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The present invention will be further described below with reference to the accompanying drawings and examples.

[0076] Figure 1 This is a DEM image of the Xibaxiaqu watershed produced based on GF-7 in the embodiment.

[0077] Figure 2 It is the center line of a section of the Xibaxiaqu River in December 2020 in the embodiment.

[0078] Figure 3 This is an example of the river water surface width at four sections of Xibaxiaqu in December 2020 in the embodiment.

[0079] Figure 4 This is an example image of the slope calculation at a section of Xibaxiaqu in December 2020 in the embodiment.

[0080] Figure 5 This is an example diagram of the cross-sectional shape of the river at a section of Xibaxiaqu in December 2020 in the embodiment.

[0081] Figure 6 This is an example diagram of the flow area (green shaded area) corresponding to the river width at a section of Xibaxiaqu in December 2020 in the embodiment.

[0082] Figure 7 2003-2022 is a river flow diagram of Xibaxiaqu in the embodiment.

[0083] Figure 8 This is a Pearson correlation analysis diagram of the rainfall data in the embodiment and the remote sensing monitoring results of the Xibaxia meander flow. DETAILED DESCRIPTION

[0084] Example 1:

[0085] A method for constructing an RCBAM model for remote sensing inversion of river flow includes the following steps:

[0086] 1) Collect remote sensing image data; this includes satellite fore- and aft-view image pairs and time series optical remote sensing images. In this example, the DEM production data comes from the China Resources Satellite Application Center's Gaofen-7 (GF-7), with a sub-meter spatial resolution. The water surface remote sensing data comes from the European Space Agency's Sentinel-2 satellite, with a spatial resolution of 10 meters. The Xibaxiaqu River was selected as a case study. This river is located in the Yarlung Zangbo River basin, a data-deficient area and is therefore representative.

[0087] 2) High-precision DEM generation: First, the satellite forward-view and back-view images of high-resolution images are used as input images. The images are directional adjusted using the satellite's RPC parameters or ground control points. Stereo matching is performed on the stereo pairs to obtain a digital surface model (DSM) containing ground buildings and vegetation. The DSM is then filtered to remove artificial objects and vegetation to obtain a DEM.

[0088] In this example, GF-7 is used as the input image, a semi-global stereo matching algorithm (SGM) is used to obtain DSM, and an iterative morphological filter algorithm (IMF) is used to remove objects and vegetation to obtain 5m high-precision digital elevation model DEM data. Figure 1 This is the DEM image of the Xibaxiaqu Basin produced based on GF-7.

[0089] 3) River Centerline Acquisition: The Normalized Difference Water Index (NDWI) is calculated based on the optical image. The OTSU adaptive thresholding method is used to obtain a binary image of the water body, where pixels with water are assigned a value of 1 and pixels without water are assigned a value of 0. The erosion operation is repeated until only a 1-pixel line remains within the water body. This line is the river centerline.

[0090] In this embodiment, NDWI is calculated using Sentinel-2 optical images to obtain a binary water body image, and the sliding window is set to 3×3. Figure 2 This is the centerline of a section of the Xibaxiaqu River in December 2020.

[0091] 4) River Width and Slope Determination: The river width is determined using the perpendicular distance method based on the water body image obtained in step 3) and the river centerline. Starting from each point on the centerline, a search is performed in a direction perpendicular to the centerline. Two corresponding points on the water body boundary are found in the vertical direction. The distance between these two points is then measured to determine the river width at that location.

[0092] In this embodiment, four river sections are selected, which extend perpendicularly outward from the center line of the river until they intersect with the boundaries of the river water body on both sides, and the width of the river is the middle of the two intersection points. Figure 3 This is the width of the river surface at four sections of Xibaxiaqu in December 2020.

[0093] Based on the DEM obtained in step 2), the slope is obtained by calculating the ratio of the elevation difference of pixels at a certain distance to the horizontal distance. In this embodiment, the slope is calculated by extending 500m upstream and downstream of Xibaxiaqu with the cross section as the center, that is, S = (127-110) / 1000 = 0.027. Figure 4 This is an example image of slope calculation at a section of Xibaxiaqu in December 2020.

[0094] 5) River Cross-Section Shape Extraction: At the river cross section, based on the DEM data generated in step 2), extract river cross-section elevation data perpendicular to the river centerline obtained in step 3). Based on this river cross-section elevation data, draw a river cross-section shape diagram. Figure 5 This is an example diagram of the river cross-section shape at a section of Xibaxiaqu in December 2020.

[0095] 6) Calculation of flow area: The river width obtained in step 4) is integrated with the cross-sectional shape line of the river channel obtained in step 5) by Gauss-Kronrod integration to calculate the cross-sectional area corresponding to the river width, and the flow area formed with the river channel cross section under different river width conditions is obtained. Figure 6 This is an example diagram of the flow area (green shaded area) corresponding to the river width at a section of Xibaxiaqu in December 2020.

[0096] 7) Inference and estimation of flow: The Manning formula and the multi-station hydraulic geometry method (AMHG) are used as the basic equations for flow estimation. The Bayesian method is used to perform logarithmic transformation on the Manning formula and the AMHG equation and construct a likelihood function. The Monte Carlo sampling method of Bayesian inference is used to extract samples from the specified distribution and calculate statistics to approximate the posterior distribution. The parameters used in the above method and formula are: the reference cross-sectional area is set to the flow area corresponding to the minimum river width; the change in the flow area is the difference between the instantaneous river width, the minimum river width, and the area formed by the river cross-sectional shape curve; based on the river width and slope obtained in step 4) and the flow area with river shape information obtained in step 6), flow estimation is performed. In this embodiment, the river flow of the Xibaxiaqu River for 20 years (2003-2022) is calculated. Figure 7 This is the river flow map of Xibaxiaqu from 2003 to 2022.

[0097] Verification results:

[0098] Since the Xibaxiaqu watershed is an area without data, a Pearson correlation analysis was conducted between the rainfall data and the remote sensing monitoring of the Xibaxiaqu runoff data ( Figure 8 ), the results showed that the river flow monitored by remote sensing had a significant positive correlation with rainfall, the correlation coefficients between RCBAM and BAM were 0.67 and 0.59, respectively, and the monitoring accuracy of river flow by RCBAM based on river channel shape was improved by 0.08.

[0099] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention and is not limiting. Although the present invention is described in detail with reference to the preferred arrangement scheme, ordinary technicians in this field should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for constructing an RCBAM model for remote sensing inversion of river flow, characterized by: include: 1) Collect remote sensing image data; Including satellite fore-and-aft image pairs and time series of optical remote sensing images; 2) High-precision DEM generation: Using satellite foreground and back-view image pairs as input, a digital surface model (DSM) is obtained through directional adjustment and stereo matching algorithms. The DSM is filtered to remove man-made objects and vegetation to obtain the numerical elevation model (DEM) data. 3) Obtaining the river centerline: Using the optical remote sensing image as input, the Normalized Difference Water Index (NDWI) is calculated to obtain a binary image of the water body. The binary image of the water body is then eroded to obtain the river centerline. 4) Obtaining river width and slope: Based on the river centerline, the river width is obtained using the vertical distance method; the slope is obtained based on DEM data; 5) River channel cross-sectional shape extraction: In the DEM data, a section is cut perpendicular to the acquired river channel centerline to obtain the river cross-sectional elevation data. Based on the river cross-sectional elevation data, the discrete elevation points are linearly interpolated to obtain the river channel cross-sectional shape curve; 6) Calculation of flow area: The river width is matched with the river cross-section shape curve, and Gauss-Kronrod integration is performed to obtain the flow area formed by the river cross-section under different river width conditions; 7) River flow estimation: The Manning formula and AMHG equations are used as the basic equations for flow estimation. The Bayesian method is used to perform logarithmic transformations on the Manning formula and AMHG equations and construct a likelihood function. The Monte Carlo sampling method based on Bayesian inference is used to draw samples from the specified distribution and calculate statistics to approximate the posterior distribution. The parameters used in the above formulas are: the reference cross-sectional area is set to the flow area corresponding to the minimum river width; the change in flow area is the difference between the instantaneous river width, the minimum river width, and the area formed by the river cross-sectional shape curve. The obtained river width, slope, and flow area are calculated using Bayesian inference to obtain the flow estimation result.

2. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: The data time of the foreground and background image pairs used for DEM production in step 2) must be in the dry season of the river flow to maximize the inclusion of river channel shape information; the accuracy of the high-precision DEM is 5m.

3. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: The specific steps of step 2) are: 2-1 Use satellite image RPC parameters or ground control points to perform orientation adjustment on the image to obtain accurate orientation parameters. Use the accurate positioning model and orientation parameters to perform epipolar sampling on the foresight and backsight images of the image pair to obtain approximate epipolar images. 2-2 Using a semi-global stereo matching algorithm, stereo matching and point intersection are performed on the approximate epipolar images to obtain a high-density 3D point cloud; 2-3 Point cloud data registration generates a DSM including ground buildings and vegetation, and then the DSM is filtered to remove man-made objects and vegetation to obtain a DEM.

4. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: In step 3), the NDWI is calculated and the OTSU adaptive threshold method is used to obtain a binary image of the water body, where the pixel value of the water body is 1 and the pixel value of the non-water body is 0. The erosion operation is repeated continuously. When only a line with a width of 1 pixel remains in the water body, this line is the centerline of the river channel.

5. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: In step 3), corrosion judgment is performed during corrosion, which is specifically divided into 8 situations for judgment: when it is judged that there is 1 data point around the pixel, the center point gradient is 7; when it is judged that there are 2 data points around the pixel, the center point gradient is 6; when it is judged that there are 3 data points around the pixel, the center point gradient is 5; when it is judged that there are 4 data points around the pixel, the center point gradient is 4; when it is judged that there are 5 data points around the pixel, the center point gradient is 3; when it is judged that there are 6 data points around the pixel, the center point gradient is 2; when it is judged that there are 7 data points around the pixel, the center point gradient is 1.

6. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: Based on the vertical distance method described in step 4), the river width is obtained: starting from each point on the centerline, a search is performed in a direction perpendicular to the centerline, and two corresponding points of the water body boundary are found in the vertical direction. By calculating the distance between these two points, the river width at that position can be obtained; in step 4), the slope is obtained based on the DEM data: the slope is obtained by calculating the ratio of the elevation difference between pixels at a certain distance in the river channel or at the edge of the river channel to the horizontal distance; in step 4), based on the long-term optical image series, the river width and slope data of the selected section are obtained for a long time series, and the number of sections is no less than 4.

7. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: The cross-sectional shape of the river channel obtained in step 5) should include the entire river channel within the embankments on both sides of the river.

8. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: The Gauss-Kronrod integral in step 6) is as follows: Where A is the calculated flow area, m 2 ; W is the river width, m; W / 2 is the scaling factor of the integration interval; m is the total number of nodes; W k is the weight corresponding to the integration node k; x k is the position of the integration node in the original interval [0,w]; The difference between the long-term river width and the area corresponding to the minimum river width is used to obtain the change in flow area: δA it = Yes i -IN min Among them, δA it is the change in flow area, m 2 ; A i is the flow area corresponding to time i, m 2 ; A min is the flow area corresponding to the minimum river width in all time series, m 2 .

9. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: In step 6), the river width length of the long time series obtained is matched with the cross-sectional shape of the river channel, and the elevation difference between the intersection points on the left and right sides does not exceed 10 cm.

10. The method for constructing an RCBAM model for remote sensing inversion of river flow according to claim 1, characterized in that: The Manning equation in step 7) and the logarithmic transformation are: Among them, Q t is the flow rate, m 3 / s; t is time; n is Manning coefficient, m -1 / 3 ·s;A it is the cross-sectional area: A 0i is the reference cross-sectional area, δA it is the change in flow area, m 2 ;W it is the river width, m; S it is the slope, dimensionless; The AMHG equation and the logarithmic transformation are: log W it =b i (log Q t -log Q c )+log W c +∈ g Among them, W it is the river width, m; Q t is the flow rate, m 3 / s;W c , Q c with b i is the AMHG parameter; Qc is the characteristic flow, m 3 / s; Wc is the global characteristic river width corresponding to Qc, m; b i is the width-flow index; ∈ g is the AMHG error term; A0 is the flow area corresponding to the minimum value of the river width in this section; δA it The area used is the difference between the instantaneous river width of the section, the minimum river width and the area formed by the river section shape curve.

Images