Natural resource measurement method and system based on remote sensing images

Abstract

The present invention discloses a natural resource measurement method and system based on remote sensing images, and relates to the technical field of remote sensing map processing. The method first trains a first learning model, a second learning model, and a third learning model respectively, establishes different learning models for different resource types, and improves the recognition efficiency of different resources. The volume data of the resource is determined based on the characteristic information of the same type of resource in multiple groups of remote sensing images, and the total amount of the resource is estimated. Furthermore, the present invention improves the efficiency of determining the first sampling point by fitting the first resource into a spline curve. At the same time, the present invention sets the second sampling point according to the contour line of the second resource, and improves the surface fitting accuracy of the second resource.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing map processing, and in particular to a natural resource measurement method and system based on remote sensing images. Background Art

[0002] Remote sensing images can be used to obtain geographic data in a large area. Classification and labeling of the geographic data of the images can segment the remote sensing images into different types of natural resources. For example, the Chinese patent application with publication number CN118658015A of the applicant discloses a map resource identification method, system and storage medium based on deep learning. The method trains different learning models to improve the recognition accuracy of map resources. After the map resource identification is completed, the relevant information needs to be further refined. The Chinese patent application with publication number CN118470550A discloses a natural resource asset data acquisition method and platform. The method obtains high-resolution images of the target area through drones, and constructs a time series remote sensing data set of the target area to monitor and analyze the spatial distribution and temporal and spatial changes of natural resources, and can comprehensively evaluate the total amount of natural resources. However, the evaluation of natural resources by this method is based on plane data, and the data acquisition accuracy is limited, so it is necessary to further improve it. Summary of the invention

[0003] In order to solve the defects of the above-mentioned prior art, the present invention proposes a natural resource measurement method and system based on remote sensing images. The method predicts the three-dimensional data of different natural resources according to the characteristic information of the same natural resource in multiple groups of corrected remote sensing images, thereby more accurately estimating the total amount of different natural resources.

[0004] In order to achieve the above object, the first aspect of the present invention provides a natural resource measurement method based on remote sensing images, comprising the following steps:

[0005] Step 1: The sample area has at least one or more of a building image, a first resource, and a second resource, obtains sample images of the sample area, and trains a first learning model, a second learning model, and a third learning model according to multiple groups of sample images;

[0006] Step 2: The camera unit collects multiple sets of remote sensing images of the target area, the ranging unit collects multiple sets of elevation images of the target area, the sensor unit detects the position parameters and attitude parameters of the camera unit, adjusts the remote sensing image based on the attitude parameters and the position parameters, and obtains the corrected remote sensing image;

[0007] Step 3: Identify the building image of the corrected remote sensing image according to the first learning model, and identify the first resource of the corrected remote sensing image according to the second learning model;

[0008] Step 4: Generate a center line of the first resource, fit the center line into a spline curve, extract at least one first sampling point of the spline curve, identify and correct the width data of the first sampling point in the remote sensing image and the elevation data of the first sampling point in the elevation image;

[0009] Step 5: generating depth data of the first sampling point by combining the width data of the first sampling point in multiple corrected remote sensing images and the elevation data in multiple elevation images, and estimating volume data of the first resource by the depth data of the multiple first sampling points;

[0010] Step 6: After deleting the building image and the first resource in the corrected remote sensing image, an intermediate image is generated, a second resource of the intermediate image is identified according to the third learning model, at least one boundary point of the second resource is extracted, shadow data of the boundary point is found, and height data of the boundary point is predicted according to the shadow data of the boundary point;

[0011] Step 7: extract at least one second sampling point of the second resource, predict the height data of the second sampling point according to the height data of multiple boundary points, and estimate the volume data of the second resource according to the height data and elevation data of the multiple second sampling points.

[0012] In the present invention, in step 1, the first resource is a river, and the second resource is a forest land.

[0013] In the present invention, in step 2, a rotation matrix is ​​constructed according to the posture parameters of the camera unit, a collinearity equation function is generated according to the rotation matrix and the position parameters of the camera unit, initial ground coordinates corresponding to the remote sensing image coordinates are preset, actual ground coordinates are iteratively generated according to the collinearity equation function, the corrected remote sensing image coordinates are calculated according to the actual ground coordinates, the pixel points of the remote sensing image are mapped to the corrected remote sensing image according to the corrected remote sensing image coordinates, and a coordinate correction ratio is generated according to the remote sensing image coordinates and the corrected remote sensing image coordinates.

[0014] In the present invention, in step 4, a smooth curve is fitted with the endpoints of the center line as control points. If the vertical distance d between at least one center line coordinate point on the center line and the smooth curve is greater than the distance threshold d', the maximum vertical distance d is set. max The corresponding center line coordinate points are used as new control points and the smooth curve is refitted until the vertical distances between all the center line coordinate points and the smooth curve are less than or equal to the distance threshold d', and the final smooth curve is used as the spline curve.

[0015] In the present invention, in step 4, the first sampling threshold is adjusted according to the coordinate correction ratio, any end point of the spline curve is taken as the first sampling point, and the next curve coordinate point is searched along the spline curve with the first sampling point as the starting point, the curvature of the curve coordinate point and the plane distance and elevation difference between the curve coordinate point and the previous first sampling point are extracted, and the cross-sectional change value of the curve coordinate point is calculated. If the cross-sectional change value is greater than the first sampling threshold, the curve coordinate point is taken as the first sampling point, and the next curve coordinate point is continued to be searched along the spline curve until the other end point of the spline curve is reached.

[0016] In the present invention, in step 5, the cross-sectional data of the first sampling point is calculated based on the width data and depth data of the first sampling point, and the cross-sectional function of the first resource is estimated based on the cross-sectional data of multiple first sampling points, and the integral of the cross-sectional function in the length direction of the spline curve is the volume data of the first resource.

[0017] In the present invention, in step 6, the pixel grayscale of the intermediate image is calculated, the shadow interval is extracted based on the shadow threshold, a boundary point is selected on the shadow interval and the second resource dividing line, the shadow data is the shadow width corresponding to the boundary point, the azimuth angle ε of the boundary point is obtained, and the altitude angle θ and azimuth angle α of the camera unit at the time of remote sensing image shooting and the altitude angle ω and azimuth angle β of the sun are obtained, and the altitude data of the boundary point is predicted according to the shadow width, ε, θ, α, β and ω.

[0018] In the present invention, in step 7, the second sampling threshold is adjusted based on the coordinate correction ratio, multiple groups of contour lines of the second resource are generated, the surface gradient of any contour line coordinate point on the contour line is calculated, the contour line coordinate point with the largest surface gradient is taken as the second sampling point, the next contour line coordinate point is searched along the contour line with the second sampling point as the starting point, the surface gradient of the contour line coordinate point and the plane distance between the contour line coordinate point and the previous second sampling point are extracted, the surface change value of the contour line coordinate point is calculated, and if the surface change value is greater than the second sampling threshold, the contour line coordinate point is taken as the second sampling point, and the next contour line coordinate point is continuously searched along the contour line until all the contour line coordinate points are found.

[0019] In the present invention, in step 7, the elevation data of the second sampling point is searched, a height function of the second resource is generated according to the height data of a plurality of second sampling points and the elevation data, and then the volume data of the second resource is estimated.

[0020] A second aspect of the present invention further provides a natural resource measurement system for implementing the natural resource measurement method based on remote sensing images, the natural resource measurement system comprising:

[0021] A camera unit configured to capture a remote sensing image of a target area;

[0022] a range-finding unit configured to collect an elevation image of a target area;

[0023] A sensing unit configured to detect position parameters and posture parameters of the camera unit;

[0024] an image processing unit configured to generate a sample image and a corrected remote sensing image;

[0025] A model generating unit is configured to train a first learning model, a second learning model and a third learning model according to a plurality of groups of sample images,

[0026] A resource identification unit configured to identify the building image of the corrected remote sensing image, the first resource, and the second resource according to the first learning model, the second learning model, and the third learning model;

[0027] a first analysis unit configured to estimate volume data of a first resource;

[0028] The second analysis unit is configured to estimate volume data of a second resource.

[0029] The implementation of the natural resource measurement method and system based on remote sensing images of the present invention has the following beneficial effects: the present invention confirms the volume data of the map resource based on the feature information of the same map resource in multiple sets of corrected remote sensing images, and realizes the estimation of the total amount of resources. Furthermore, the present invention reduces the interference pixels near the first resource in the target corrected image by fitting the first resource as a spline curve, thereby improving the recognition modeling efficiency. At the same time, the present invention sets the second sampling point according to the contour line of the second resource, thereby improving the volume fitting accuracy of the second resource. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flowchart of a natural resource measurement method based on remote sensing images of the present invention;

[0031] Figure 2 is a schematic diagram of a remote sensing image;

[0032] Figure 3 A simplified diagram of a remote sensing image, comprising a building image, a first resource, and a second resource;

[0033] Figure 4 A schematic diagram of a target area for photographing according to the present invention;

[0034] Figure 5 A schematic diagram of a spline curve fitted to the first resource of the present invention;

[0035] Figure 6 A schematic diagram of a first sampling point of a first resource of the present invention;

[0036] Figure 7is a schematic diagram of the second resource and the shaded area of ​​the present invention;

[0037] Figure 8 is a schematic diagram of a second sampling point of a second resource of the present invention;

[0038] Fig. 9 is a schematic diagram of width data of a first sampling point of the present invention;

[0039] Fig.10 is a cross-sectional schematic diagram of the first sampling point of the present invention;

[0040] Fig.11 Schematic diagram of boundary points of the shaded area of ​​the present invention;

[0041] Fig.12 It is a schematic diagram of the boundary point, the camera unit and the azimuth angle of the sun of the present invention;

[0042] Fig.13 Schematic diagram of the altitude angle of the camera unit and the sun of the present invention

[0043] Fig.14 is a cross-sectional schematic diagram of a second sampling point of the present invention;

[0044] Fig.15 is a block diagram of a natural resource measurement system of the present invention;

[0045] Fig.16 It is a block diagram of the model generation unit of the present invention. DETAILED DESCRIPTION

[0046] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0047] The existing natural resource measurement methods mainly rely on ground surveys and plane data sets, which is not only time-consuming and labor-intensive, but also difficult to fully and accurately reflect the true total amount of natural resources, especially in large-scale areas and complex environments. The natural resource measurement method of the present invention not only refers to the remote sensing data of the target area, but also introduces elevation data, which can estimate the volume data of various natural resources, further improving the efficiency of data acquisition and the accuracy of data estimation. Embodiment 1

[0048] Reference Figure 1 The natural resource measurement method based on remote sensing images described in this embodiment is used to estimate Figure 2 The different resource quantities of the remote sensing images shown are used to estimate the total amount of natural resources. The natural resource measurement method includes the following steps.

[0049] Step 1: The sample area has at least one or more of the building image, the first resource and the second resource. The sample image of the sample area is obtained, and the first learning model, the second learning model and the third learning model are trained according to the multiple groups of sample images. The building image of the present invention is an image of a man-made object with a regular set shape such as a building or a structure. The first resource is a river, and the second resource is a forest. Figure 2 The natural resources are simplified to Figure 3 , after simplification Figure 3 Three different resource types are shown.

[0050] Step 2: The camera unit collects multiple sets of remote sensing images of the target area, the distance measuring unit collects multiple sets of elevation images of the target area, the sensor unit detects the position parameters and attitude parameters of the camera unit, adjusts the remote sensing image based on the attitude parameters and position parameters, and obtains the corrected remote sensing image. Figure 4 , the camera unit captures the remote sensing image of the target area, the sensing unit is, for example, a gyroscope and an accelerometer, the sensing unit can detect the position parameters and attitude parameters of the camera unit, and the ranging unit is, for example, a laser radar. A zero elevation reference plane is preset to determine the height difference between the camera unit and the reference plane. The ranging unit measures the vertical distance between the remote sensing image coordinates and the camera unit through laser pulses, thereby obtaining the height difference between the remote sensing image coordinates and the reference plane. The height difference is the elevation data of the remote sensing image coordinates, and then the elevation image of the target area is obtained. The remote sensing image is matched with each coordinate point of the elevation image to form the three-dimensional surface data of the target area.

[0051] The lens and imaging system of the camera unit have inherent perspective distortions, which can cause different degrees of distortion in the remote sensing image. In order to facilitate subsequent resource identification, further correction is required. The present invention does not limit the correction algorithm. Referring to the Chinese patent application with publication number CN118658015A, a rotation matrix can be constructed according to the posture parameters of the camera unit, a collinearity equation function can be generated according to the rotation matrix and the position parameters of the camera unit, initial ground coordinates corresponding to the remote sensing image coordinates are preset, actual ground coordinates are iteratively generated according to the collinearity equation function, the corrected remote sensing image coordinates are calculated according to the actual ground coordinates, and the pixels of the remote sensing image are mapped to the corrected remote sensing image according to the corrected remote sensing image coordinates.

[0052] In this embodiment, the positions of the corresponding coordinates of the corrected remote sensing image and the remote sensing image change, thereby affecting the matching of the elevation data. As described in step 4 and step 7. The present invention further adjusts the first and second sampling thresholds according to the coordinate correction ratio of the previous and next images. The coordinate correction ratio is the ratio of the position adjustment of all coordinate points to the original position. For example, the remote sensing image coordinate (x 1i ,y 1i ) corresponds to the corrected remote sensing image coordinates (x 0i , y 0i), the number of pixels in the remote sensing image is I, and the coordinate correction ratio is .

[0053] Step 3: Identify the building image of the corrected remote sensing image according to the first learning model, and identify the first resource of the corrected remote sensing image according to the second learning model. The present invention does not limit the training method of the learning model, and the preferred training method can refer to the fourth embodiment.

[0054] Step 4: Generate a center line of the first resource, fit the center line into a spline curve, extract at least one first sampling point of the spline curve, identify the width data of the first sampling point in the corrected remote sensing image and the elevation data of the first sampling point in the elevation image. Figure 5 , the endpoints of the center line are used as control points to fit a smooth curve. If the vertical distance d between at least one center line coordinate point and the smooth curve is greater than the distance threshold d', the maximum vertical distance d max The corresponding center line coordinate points are used as new control points and the smooth curve is refitted until the vertical distances between all the center line coordinate points and the smooth curve are less than or equal to the distance threshold d', and the final smooth curve is used as the spline curve.

[0055] Reference Figure 6 , the two endpoints of the spline curve are the first sampling points. Take any endpoint of the spline curve as the starting point to search for the next curve coordinate point along the spline curve, extract the curvature of the curve coordinate point and the plane distance and elevation difference between the curve coordinate point and the previous first sampling point, and calculate the cross-sectional change value of the curve coordinate point. If the cross-sectional change value is greater than the first sampling threshold, take the curve coordinate point as the first sampling point and continue to search for the next curve coordinate point along the spline curve until the other endpoint of the spline curve is reached.

[0056] The smaller the density of the first sampling points, the smaller the amount of calculation, but the estimation accuracy is reduced. The present invention further adjusts the first sampling threshold according to the coordinate correction ratio. Specifically, the maximum value of the curvature is, for example, 0.01m -1 , the maximum value of the plane distance is, for example, 10m, and the maximum value of the elevation difference is, for example, 1m, then the initial first sampling threshold r1 is 0.1m, and the adjusted first sampling threshold R1= r1Δ. The smaller the coordinate correction ratio, the lower the matching degree of the coordinate points of the elevation image and the remote sensing image. At this time, it is necessary to lower the first sampling threshold and increase the sampling density to ensure the accuracy of the system.

[0057] Step 5: Generate the depth data of the first sampling point by combining the width data of the first sampling point in multiple corrected remote sensing images and the elevation data in multiple elevation images, and estimate the volume data of the first resource from the depth data of the multiple first sampling points. Calculate the cross-sectional data of the first sampling point based on the width data and depth data of the first sampling point, and estimate the cross-sectional function of the first resource based on the cross-sectional data of the multiple first sampling points. The integral of the cross-sectional function in the length direction of the spline curve is the volume data of the first resource. The present invention does not limit the method of calculating the cross-sectional data, and the specific embodiment can refer to the second embodiment.

[0058] Step 6: After deleting the building image and the first resource in the corrected remote sensing image, an intermediate image is generated. The second resource of the intermediate image is identified according to the third learning model, at least one boundary point of the second resource is extracted, the shadow data of the boundary point is found, and the height data of the boundary point is predicted according to the shadow data of the boundary point. The pixel grayscale of the intermediate image is calculated, and the shadow interval is extracted based on the shadow threshold. The shadow threshold is, for example, greater than 200. Figure 7 As shown, the shadow interval is the shadow formed by natural light at the boundary of the second resource. A boundary point with the largest shadow width is selected on the boundary line between the shadow interval and the second resource. The shadow data is the shadow width corresponding to the boundary point, and the azimuth angle ε of the boundary point is obtained, and the altitude angle θ and azimuth angle α of the camera unit at the time of remote sensing image shooting and the altitude angle ω and azimuth angle β of the sun are obtained, and the height data of the boundary point is predicted according to the shadow width, ε, θ, α, β and ω. The present invention does not limit the method for calculating the height data of the boundary point, and the specific embodiment can refer to Example 3.

[0059] Step 7: extract at least one second sampling point of the second resource, predict the height data of the second sampling point according to the height data of the plurality of boundary points, and estimate the volume data of the second resource according to the height data and elevation data of the plurality of second sampling points. Figure 8 , generate multiple groups of contour lines of the second resource, calculate the surface gradient of any contour line coordinate point on the contour line, and take the contour line coordinate point with the largest surface gradient as the second sampling point. Take the second sampling point as the starting point to search for the next contour line coordinate point along the contour line, and calculate the surface change value of the contour line coordinate point. The surface change value includes the surface gradient of the contour line coordinate point and the plane distance between the contour line coordinate point and the previous second sampling point. If the surface change value of the contour line coordinate point is greater than the second sampling threshold, take the contour line coordinate point as the second sampling point, and continue to search for the next contour line coordinate point along the contour line until all contour line coordinate points are found.

[0060] In this embodiment, the second resource is a three-dimensional surface, and the elevation data of the second sampling point is searched. The three-dimensional shape can be determined based on the height data and elevation data of multiple second sampling points, and then the volume data of the second resource is estimated. The maximum value of the surface gradient is, for example, 10%, and the maximum value of the plane distance is, for example, 20m, then the initial second sampling threshold r2 is 2m. Similar to step 4, the present invention adjusts the second sampling threshold based on the coordinate correction ratio. Specifically, the adjusted second sampling threshold R2=r2Δ. The smaller the coordinate correction ratio, the lower the matching degree of the coordinate points of the elevation image and the remote sensing image. At this time, it is necessary to lower the second sampling threshold, increase the sampling density, and ensure the accuracy of the system. The present invention does not limit the method of calculating the height data of the second sampling point, and the specific embodiment can refer to Example 3. Embodiment 2

[0061] This embodiment further discloses a method for determining a first sampling point and a method for calculating volume data of a first resource according to the first sampling point.

[0062] The cross-section change value includes the curvature n of the curve coordinate point, the plane distance D1 between the curve coordinate point and the previous first sampling point, and the elevation difference ΔH between the curve coordinate point and the previous first sampling point. The plane distance D1 is the distance between the curve coordinate point and the first sampling point along the spline curve direction, that is, the length of the spline curve between the curve coordinate point and the first sampling point. The elevation difference ΔH=|H'-H| is calculated based on the elevation data H corresponding to the curve coordinate point on the elevation image and the elevation data H' corresponding to the previous first sampling point. The specific algorithm formula is, for example, cross-section change value=n×D1×ΔH, or cross-section change value=n×q1+D1×q2+ΔH×q3, where q1, q2, and q3 are the influence weights of curvature, plane distance, and elevation difference, respectively.

[0063] Reference Fig. 9 , the width data of the first sampling point is obtained according to the number of pixels occupied by the first resource in the normal direction of the first sampling point, and the depth data of the first sampling point can be calculated according to the width data and elevation data of the same first sampling point of at least two corrected remote sensing images. The corrected remote sensing images at the maximum and minimum values ​​of the width data of the first sampling point are defined as the flood season image and the dry season image, respectively, and the flow loss in the flood season and the dry season is preset, for example, the flow loss is determined to be 80% of the maximum flow according to the climate data. Reference Fig.10, the width data of the same first sampling point of the two sets of corrected remote sensing images are L1 and L2 respectively, and the corresponding elevation data are H1 and H2 respectively, where L1>L2, H1>H2. The width of the riverbed at the first sampling point is L3, and the elevation data of the riverbed is H3. Since the cross section where the first sampling point is located is approximately an isosceles trapezoid, (H1-H2) (L1-L3) = (H1-H3) (L1-L2). Since the cross-sectional data in the flood season = flow loss / 80%, (H1-H3) (L1+L3) = (H1-H2) (L1+L2) / 0.8, the riverbed width and riverbed elevation data can be predicted accordingly. For any corrected remote sensing image of the first sampling point, the width data of the first resource is L and the elevation data is H, and the depth data is H-H3, then the cross-sectional data A'= (H-H3) (L+L3) / 2. Determine the cross-sectional data of each curve coordinate point of the spline curve according to the linear interpolation method and the cross-sectional data of the first sampling point, and establish the cross-sectional function A' (x, y) between the cross-sectional data A' and the spline curve coordinate point (x, y). The volume data S1 of the first resource = , where L4 is the length of the spline curve in the x-axis direction. Embodiment 3

[0064] This embodiment further discloses a method for determining a second sampling point and a method for calculating volume data of a second resource according to the second sampling point.

[0065] Calculate the pixel grayscale of the intermediate image and extract the shadow interval based on the shadow threshold. Fig.11 , select a boundary point on the dividing line between the shadow interval and the second resource, and the shadow data of the boundary point is the shadow width L5 of the shadow interval at the boundary point along the direction of sunlight. Fig.12 and Fig.13 , calculate the height data of the boundary point according to the shadow width L5 , where ε is the azimuth of the boundary point, α is the azimuth of the camera unit, θ is the altitude of the camera unit, β is the azimuth of the sun, and ω is the altitude of the sun.

[0066] Generate multiple groups of contour lines of the second resource, calculate the surface gradient of any contour line coordinate point on the contour line, take the contour line coordinate point with the largest surface gradient as the second sampling point, and use the second sampling point as the starting point to find the next contour line coordinate point along the contour line, and calculate the surface change value of the contour line coordinate point. The surface change value includes the plane distance D2 between the contour line coordinate point and the previous second sampling point and the surface gradient of the contour line coordinate point , surface change value = D2. Plane distance D2 is the distance between the contour coordinate point and the second sampling point along the contour line direction, that is, the length of the contour line between the contour coordinate point and the second sampling point. Surface gradient of the contour coordinate point (X0, Y0) , H0 is the elevation data of the contour line coordinate point (X0, Y0), is the partial derivative of the elevation data H0 in the X direction, is the partial derivative of the elevation data H0 in the Y direction. If any surface change value of the contour line coordinate point (X0, Y0) is greater than the second sampling threshold, the contour line coordinate point is used as the second sampling point, and the next contour line coordinate point is searched along the contour line until all the contour line coordinate points are found, and all the second sampling points of the second resource are determined.

[0067] Reference Fig.14 , given the height data of multiple boundary points, the height data of the second sampling point is determined by linear interpolation. The second sampling point corresponds to the top three-dimensional coordinates (X, Y, H6) of the second resource, and the second sampling point corresponds to the bottom three-dimensional coordinates (X, Y, H6-H7) of the second resource. H6 is the elevation data corresponding to the second sampling point (X, Y) in the elevation image, and H7 is the height data corresponding to the second sampling point (X, Y). In one embodiment, the height data of any coordinate point (X, Y) of the second resource is determined by the least squares method, and then the height function H (X, Y) is fitted, and the volume data of the second resource is Among them, X max is the maximum horizontal coordinate point of the second resource in the corrected remote sensing image, X min is the minimum horizontal coordinate point of the second resource in the corrected remote sensing image, Y max is the maximum ordinate point of the second resource in the corrected remote sensing image, Y min It is the minimum vertical coordinate point of the second resource in the corrected remote sensing image.

[0068] In another embodiment, the quadratic surface equation H6=a1X is fitted according to the top three-dimensional coordinates of the plurality of second sampling points. 2 +b1Y 2 +c1XY+d1X+e1Y+f1, solve the coefficients a1, b1, c1, d1, e1, f1, and fit the quadratic surface equation H6-H7=a2X according to the bottom three-dimensional coordinates of multiple second sampling points 2 +b2Y 2 +c2XY+d2X+e2Y+f2, solve the coefficients a2, b2, c2, d2, e2, f2. Calculate the height function a1X of the second sampling point 2 +b1Y 2 +c1XY+d1X+e1Y+f1-a2X 2 +b2Y 2+c2XY+d2X+e2Y+f2. . Embodiment 4

[0069] This embodiment further discloses training methods for the first learning model, the second learning model, and the third learning model.

[0070] The first learning model includes a backbone network module, a pyramid network module, and a full convolutional network module. When training the first learning model, an anchor point is set at the position of the sample image corresponding to the resource label. The resource label is, for example, the border of the building image. The first feature map of different levels of the sample image is extracted through the backbone network module, and the enhanced feature map of the first feature map is extracted through the pyramid network module. Then, the enhanced feature map is processed by the full convolutional network to obtain the return feature map. The return feature map has at least one bounding box. The intersection and union ratio of the anchor point and the bounding box is calculated, and the anchor point is classified into positive samples and negative samples according to the intersection and union ratio. Then, the classification loss of each anchor point is calculated by the first loss function, and the bounding box confidence between the bounding box and the border of the sample image is calculated. The first learning model is updated according to the classification loss and the bounding box confidence.

[0071] The second learning model uses a U-Net network, and the second learning model includes a backbone feature network module, an enhanced feature network module, and a prediction network module. The backbone feature network module obtains a multi-scale feature map through 3x3 convolution and downsampling operations, and the enhanced feature network module restores the detail features of the sample image through transposed convolution and upsampling operations. The prediction network module segments the sample image according to the detail features and identifies the segmented first resource.

[0072] The third learning model adopts a convolutional neural network, including a convolution layer, a pooling layer, and a fully connected layer. First, preset multiple reference shapes and reference pixels of the second resources, delete the building image and the first resource in the sample image, generate a sample intermediate image, extract the connected area of ​​the sample intermediate image, and calculate the shape similarity and pixel similarity of each connected area with the second resource. In this embodiment, for example, a preselected similarity is used to calculate the shape similarity, such as using the Euclidean distance to calculate the pixel similarity. The pseudo-probability label of the connected area relative to the second resource is calculated based on the shape similarity and the pixel similarity. The probability label of the connected area belonging to the second resource is obtained according to the third learning model. Finally, a cross entropy loss function is generated based on the probability label, and the third learning model is updated based on the cross entropy loss function. Embodiment 5

[0073] Fig.15The block diagram of the natural resource measurement system of the present invention is used to implement the natural resource measurement method based on remote sensing images. The natural resource measurement system includes: a camera unit, a distance measurement unit, a sensor unit, an image processing unit, a model generation unit, a resource identification unit, a first analysis unit, a second analysis unit and a data generation unit. Among them, the camera unit is configured to shoot a remote sensing image of a target area. The distance measurement unit is configured to collect an elevation image of the target area. The sensor unit is configured to detect the position parameters and posture parameters of the camera unit. The image processing unit is configured to generate a sample image and a corrected remote sensing image. The model generation unit is configured to train a first learning model, a second learning model and a third learning model according to multiple groups of sample images. The resource identification unit is configured to identify the building image, the first resource and the second resource of the corrected remote sensing image according to the first learning model, the second learning model and the third learning model. The first analysis unit is configured to estimate the volume data of the first resource. The second analysis unit is configured to estimate the volume data of the second resource.

[0074] Referring to Figure 16, the model generation unit also includes an input module, an external database, a processor, a first model memory, a second model memory, a third model memory, and an output module. The input module collects sample images, the external database stores resource feature tags, the processor is used for model training, and the first model memory, the second model memory, and the third model memory respectively update and store the first learning model, the second learning model, and the third learning model. The output unit is used to output the trained first learning model, the second learning model, and the third learning model.

[0075] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A natural resource measurement method based on remote sensing images, characterized in that: The following steps are involved: Step 1: The sample area has at least one or more of a building image, a first resource, and a second resource. Sample images of the sample area are obtained, and a first learning model, a second learning model, and a third learning model are trained according to multiple groups of sample images. The first resource is a river, and the second resource is a forest. Step 2: The camera unit collects multiple sets of remote sensing images of the target area, the ranging unit collects multiple sets of elevation images of the target area, the sensor unit detects the position parameters and attitude parameters of the camera unit, adjusts the remote sensing image based on the attitude parameters and the position parameters, and obtains the corrected remote sensing image; Step 3: Identify the building image of the corrected remote sensing image according to the first learning model, and identify the first resource of the corrected remote sensing image according to the second learning model; Step 4: Generate a center line of the first resource, fit the center line into a spline curve, extract at least one first sampling point of the spline curve, identify and correct the width data of the first sampling point in the remote sensing image and the elevation data of the first sampling point in the elevation image; Step 5: generating depth data of the first sampling point by combining the width data of the first sampling point in multiple corrected remote sensing images and the elevation data in multiple elevation images, and estimating volume data of the first resource by the depth data of the multiple first sampling points; Step 6: After deleting the building image and the first resource in the corrected remote sensing image, an intermediate image is generated, a second resource of the intermediate image is identified according to the third learning model, at least one boundary point of the second resource is extracted, shadow data of the boundary point is found, and height data of the boundary point is predicted according to the shadow data of the boundary point; Step 7: extract at least one second sampling point of the second resource, predict the height data of the second sampling point according to the height data of the plurality of boundary points, and estimate the volume data of the second resource according to the height data and elevation data of the plurality of second sampling points, wherein: In step 2, a rotation matrix is ​​constructed according to the posture parameters of the camera unit, a collinear equation function is generated according to the rotation matrix and the position parameters of the camera unit, an initial ground coordinate corresponding to the remote sensing image coordinate is preset, actual ground coordinates are iteratively generated according to the collinear equation function, the corrected remote sensing image coordinates are calculated according to the actual ground coordinates, the pixel points of the remote sensing image are mapped to the corrected remote sensing image according to the corrected remote sensing image coordinates, and a coordinate correction ratio is generated according to the remote sensing image coordinates and the corrected remote sensing image coordinates. In step 4, a smooth curve is fitted with the endpoints of the center line as control points. If the vertical distance d between at least one center line coordinate point on the center line and the smooth curve is greater than the distance threshold d', the maximum vertical distance d is set. max The corresponding center line coordinate point is used as a new control point and the smooth curve is refitted until the vertical distance between all the center line coordinate points on the center line and the smooth curve is less than or equal to the distance threshold d'. The finally generated smooth curve is used as a spline curve, and the first sampling threshold is adjusted according to the coordinate correction ratio. Any end point of the spline curve is used as the first sampling point, and the next curve coordinate point is searched along the spline curve with the first sampling point as the starting point. The curvature of the curve coordinate point and the plane distance and elevation difference between the curve coordinate point and the previous first sampling point are extracted, and the cross-sectional change value of the curve coordinate point is calculated. If the cross-sectional change value is greater than the first sampling threshold, the curve coordinate point is used as the first sampling point, and the next curve coordinate point is continued to be searched along the spline curve until the other end point of the spline curve is reached. In step 5, the cross-sectional data of the first sampling point is calculated according to the width data and the depth data of the first sampling point, and the cross-sectional function of the first resource is estimated according to the cross-sectional data of the plurality of first sampling points. The integral of the cross-sectional function in the length direction of the spline curve is the volume data of the first resource. In step 7, the second sampling threshold is adjusted based on the coordinate correction ratio, multiple groups of contour lines of the second resource are generated, the surface gradient of any contour line coordinate point on the contour line is calculated, the contour line coordinate point with the largest surface gradient is taken as the second sampling point, the next contour line coordinate point is searched along the contour line with the second sampling point as the starting point, the surface gradient of the contour line coordinate point and the plane distance between the contour line coordinate point and the previous second sampling point are extracted, the surface change value of the contour line coordinate point is calculated, if the surface change value is greater than the second sampling threshold, the contour line coordinate point is taken as the second sampling point, the next contour line coordinate point is searched along the contour line until all the contour line coordinate points are found, the elevation data of the second sampling point is found, the height data of any second sampling point (X, Y) of the second resource is determined by the least squares method, the height function H (X, Y) of the second resource is generated according to the height data and elevation data of multiple second sampling points, and the volume data of the second resource is obtained. , X max is the maximum horizontal coordinate point of the second resource in the corrected remote sensing image, X min is the minimum horizontal coordinate point of the second resource in the corrected remote sensing image, Y max is the maximum ordinate point of the second resource in the corrected remote sensing image, Y min It is the minimum vertical coordinate point of the second resource in the corrected remote sensing image.

2. The natural resource measurement method based on remote sensing images according to claim 1, characterized in that: In step 6, the pixel grayscale of the intermediate image is calculated, the shadow interval is extracted based on the shadow threshold, and a boundary point is selected between the shadow interval and the second resource dividing line. The shadow data is the shadow width corresponding to the boundary point, and the azimuth angle ε of the boundary point is obtained. The altitude angle θ and azimuth angle α of the camera unit at the time of remote sensing image shooting and the altitude angle ω and azimuth angle β of the sun are obtained. The altitude data of the boundary point is predicted based on the shadow width, ε, θ, α, β and ω.

3. A measurement system for implementing the natural resource measurement method based on remote sensing images as claimed in claim 1, characterized in that: include: A camera unit configured to capture a remote sensing image of a target area; a range-finding unit configured to collect an elevation image of a target area; A sensing unit configured to detect position parameters and posture parameters of the camera unit; an image processing unit configured to generate a sample image and a corrected remote sensing image; A model generation unit is configured to train a first learning model, a second learning model and a third learning model according to multiple groups of sample images, A resource identification unit configured to identify the building image of the corrected remote sensing image, the first resource, and the second resource according to the first learning model, the second learning model, and the third learning model; a first analysis unit configured to estimate volume data of a first resource; The second analysis unit is configured to estimate volume data of a second resource.

Images