A Tidal Flat Terrain Monitoring Method Based on Few-Shot Learning

Abstract

The present invention discloses a tidal flat terrain monitoring method based on few-shot learning. This method utilizes the non-orthogonal tidal flat flooding image data collected by the coastal monitoring system; with the support of a small number of labeled samples, through introducing the large-scale basic model SAM for reinforcement learning, it realizes the rapid and accurate extraction of the tidal flat waterline; uses the collinearity equation to convert the non-orthogonal waterline into an orthogonal waterline; combined with the real-time tidal level information provided by the water level monitoring station, this technology can construct a high-precision tidal flat terrain on the time scale of the tidal cycle. The advantages of the present invention are: it provides an effective solution strategy for the tidal flat terrain monitoring field with scarce labeled data, and further supports the establishment of a fully automated and dynamic monitoring system for the tidal flat terrain.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing measurement of tidal flat topography, and in particular to a tidal flat topography monitoring method based on few-shot learning. Background Art

[0002] Traditional methods for tidal flat topography mapping mainly rely on classical surveying techniques and aerial image interpretation. Although modern techniques such as real-time kinematic (RTK) and terrestrial laser scanning (TLS) show their advantages in high-frequency dynamic measurement of small areas, these methods are usually accompanied by high labor costs and limited measurement ranges.

[0003] In wide-area tidal flat topography mapping, the waterline method is widely used due to its ability to capture topographic features that change over time. Although satellite imagery has an advantage in terms of coverage, due to limitations in satellite overpass cycles and weather conditions, this method is more suitable for long-term monitoring, and its measurement accuracy is limited by the image resolution.

[0004] The rapid change characteristics of tidal flat topography make it difficult for existing satellite and unmanned aerial vehicle monitoring means to achieve real-time observation, and these methods also require a large amount of human input, restricting high-frequency monitoring of tidal flat topography changes. In contrast, coastal monitoring systems, with their automated image acquisition capabilities, high spatio-temporal resolution, and wide monitoring range, have become an effective monitoring alternative.

[0005] Image segmentation, as a key technical step in distinguishing water bodies from other objects, is the basis for effective research and monitoring. Although the progress of artificial intelligence and deep learning has promoted the automation of image segmentation, the high computational cost of deep learning methods in training and parameter adjustment, combined with the color and geomorphic complexity of tidal flats, makes it difficult to obtain representative annotated datasets, and the sample quantity and accuracy of the annotated datasets determine the accuracy of artificial intelligence and deep learning. Therefore, traditional measurement methods still have significant limitations. Summary of the Invention

[0006] The purpose of the present invention is to provide a tidal flat topography monitoring method based on few-shot learning according to the deficiencies of the above-mentioned prior art, and to provide an innovative solution for the automatic recognition of waterlines by supporting few-shot learning.

[0007] The purpose of the present invention is achieved by the following technical solutions:

[0008] A tidal flat topography monitoring method based on few-shot learning, characterized in that the method comprises the following specific steps:

[0009] S1. Obtain non-orthogonal tidal flat flooding image data through a coastal monitoring system;

[0010] S2. With the support of a small number of labeled samples, by introducing the large-scale basic model SAM for reinforcement learning, the rapid and accurate extraction of the tidal flat waterline is achieved;

[0011] S3. Using the collinearity equation, convert the non-orthogonal waterline into an orthogonal waterline;

[0012] S4. Combining the real-time tidal level information provided by the water level monitoring station, construct a high-precision tidal flat topography on the time scale of the tidal cycle.

[0013] The specific steps of the said step S1 are as follows:

[0014] (1) Use the coastal monitoring system to continuously take tidal flat images. The time interval is set according to the monitoring range. The smaller the time interval, the higher the terrain accuracy, but the requirement for remote image transmission is also higher;

[0015] (2) Take the two water level values at the beginning and complete inundation of the tidal flat image as the water level thresholds respectively, and screen the tidal flat inundation images located between the two threshold water levels.

[0016] The specific steps of the said step S2 are as follows:

[0017] (1) Use the image encoder and image embedder of the SAM model to extract features from the tidal flat inundation images, which provides the necessary image feature information for the subsequent steps;

[0018] (2) Through the prompt generation module, generate prompt points from the labeled and unlabeled samples. For the labeled samples, further expand the small samples through data augmentation operations, including rotating the original image by 90 degrees, 180 degrees, and 270 degrees, mirroring along the vertical and horizontal axes, and adjusting the intensity of the image through gamma correction. These operations generate a series of variant images, thereby enhancing the generalization ability of the model and improving the accuracy of feature recognition. For the unlabeled samples, use superpixel segmentation for various types of pixels, and separate the pixels containing water bodies from other types of landscapes by setting the comprehensive weights of color space similarity and planar distance. That is, further classify each pixel into two clusters using the Gaussian mixture algorithm: the water body cluster and the non-water body cluster. The Gaussian mixture algorithm is as follows:

[0019] (1)

[0020] (2)

[0021] where p(x) represents the probability of pixel x, C(i) represents the clustering component, w(i) represents the component weight of the i-th cluster, and μ(i) and σ(i) represent the mean and standard deviation of the i-th component respectively.

[0022] Next, a hint mask is generated based on the pixel characteristics of the tidal flat water body, and hint points are extracted from the mask.

[0023] (3) Using the segmentation module, based on the trained hint points, the tidal flat flooding image is segmented into water body and non-water body to extract the accurate water edge line.

[0024] The specific steps of step S3 are as follows:

[0025] (1) Lens distortion correction: Using the Matlab camera calibration toolbox, input the photos of the checkerboard taken at different angles, and calculate the internal parameters of the camera.

[0026] (2) Image calibration is used to generate an orthographic projection image: Image calibration follows the principle of linear transformation from three-dimensional world coordinates (x, y, z) to two-dimensional image coordinates (u, v) in photogrammetry. This principle defines the basic relationship between image coordinates and real-world coordinates through collinearity equations.

[0027] As follows:

[0028]

[0029]

[0030] where (u0, v0) is the image center, (x c , y c , z c ) is the position of the camera station, f is the focal length of the camera, λ u , λ v are the horizontal and vertical scale factors, and m i,j represents the rotation and translation coefficient;

[0031] Solving the above equation is quite complex, and there are a large number of unknown parameters in the equation. Therefore, the direct linear transformation (DLT) equation is used to solve it. This equation defines the linear relationship between image and world coordinates. The DLT equation is as follows:

[0032]

[0033]

[0034] where the 11 unknown L 1-11 coefficients are calculated through external calibration points combined with corresponding image coordinate points. When calculating, the physical characteristic parameters of the camera need to be input: image size (m, n), focal length f, image center coordinates (u0, v0), the position of the camera in real-world coordinates (x c , y c , z c ) and three Euler angles (φ, σ, τ). After calculating L1-11 After the coefficients, the inverse operations of formulas (3) and (4) will be applied to convert the image coordinates into the actual world coordinates. After fixing the camera in the field, take the field images containing the RTK calibration points, extract the pixel coordinates corresponding to the calibration points, and calculate the external parameters of the camera according to the above formulas;

[0035] (3) Using the internal and external parameters of the camera, orthogonally project the water edge line, that is, convert the coordinates of the water edge line set from pixel coordinates to geographical coordinates.

[0036] (4) Iteratively calculate the number of intersection points of each water edge line with the other water edge lines, and delete the water edge line with the largest number of intersection points until there are no intersecting water edge lines.

[0037] The specific steps of step 4 are as follows:

[0038] (1) Calibrate the water level sensor with RTK, and combine the real-time water level data to discretize the water edge line into elevation point clouds at the same moment;

[0039] (2) Divide the monitoring area grid according to the required resolution, and interpolate the elevation point clouds into the grid by the natural neighbor interpolation method to form elevation raster data, that is, obtain the tidal flat digital terrain elevation.

[0040] The advantages of the present invention are: it provides an effective solution strategy for the tidal flat terrain monitoring field with scarce labeled data, and further supports the establishment of a fully automated and dynamic monitoring system for tidal flat terrain. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is the flow chart of the present invention;

[0042] Figure 2 is the flow chart of the few-shot learning strategy in the present invention;

[0043] Figure 3 is the image segmentation effect diagram of the few-shot learning strategy in the present invention;

[0044] Figure 4 is the effect diagram of the internal parameter calibration result of the camera in the present invention;

[0045] Figure 5 is the effect diagram of the internal and external parameter calibration results of the camera in the present invention;

[0046] Figure 6 is the detailed flow chart of the tidal flat terrain construction in the present invention;

[0047] Figure 7 is the elevation point cloud diagram in the present invention;

[0048] Figure 8 is the tidal flat topographic map in the embodiment provided by the present invention;

[0049] Figure 9 This is the verification accuracy graph in the embodiments provided by the present invention. Specific implementation manners

[0050] The features of the present invention and other related features are further described in detail below through embodiments with reference to the accompanying drawings for the understanding of those skilled in the same industry:

[0051] Embodiment: As Figures 1 to 9 shown, the tidal flat terrain monitoring method based on few-shot learning in this embodiment includes the following steps:

[0052] Step S1: According to the tidal flat to be monitored, a 4G network camera is set at the spatial position (3360622.14, 570868.14, 9.09). Its photographing period is from 7:00 to 17:00 every day, and the tidal flat images are recorded at intervals of 10 minutes. Combining with the real-time water level data, the water level threshold [0.25m, 3.50m] is set to automatically screen the tidal flat flooding images.

[0053] Step S2: With the support of 300 labeled samples, through the use of the large-scale basic model SAM for reinforcement learning, the rapid and accurate extraction of the tidal flat water line is realized, as Figure 2 shown in the process. The result of water body extraction is as Figure 3 shown, and the result is very accurate.

[0054] Before converting the water line into an orthographic projection, it is necessary to first calculate the internal and external parameters of the camera. The camera internal parameter calibration result is as Figure 4 shown. The physical characteristic parameters of the camera adopted in this embodiment are as follows: image size (3840, 2160), image center point coordinates (1884.31, 1086.09), camera focal length (2691.56, 2680.53), the spatial position of the camera (3360622.14, 570868.14, 9.09), and the camera Euler angles (azimuth angle: 9.87°, pitch angle: 70.09°, roll angle: 1.00°). Substituting the parameters into the formula, the camera internal and external parameter calibration result is as Figure 5 shown. Then, using the internal and external parameters, the water line set is uniformly converted into an orthographic projection, that is, from pixel coordinates to geographic coordinates; and the intersecting water lines are deleted to obtain a non-intersecting water line set.

[0055] Step S4: Construct the tidal flat terrain according to the Figure 6 shown process. Combining the collected real-time water level (the water level has been calibrated by GPS-RTK), the water lines at the corresponding moments are uniformly converted into elevation point clouds, as Figure 7 shown. Then the monitoring area is divided into grids with a resolution of 0.5m, and the elevation point clouds are interpolated into a tidal flat DEM using the natural neighbor interpolation method, asFigure 8 as shown

[0056] Through the comprehensive application of a coastal monitoring system, the Segment Anything model, and a few-shot learning strategy, a method for tidal flat terrain monitoring based on few-shot learning was developed. The average absolute error of the obtained terrain was approximately 0.080 m, and the errors of 101 measured points are as Figure 9 shown, and this measurement accuracy basically meets the requirements

[0057] In summary, the measurement results of the present invention are accurate and are suitable for scenarios with a large range and rapid changes in tidal flat terrain

[0058] Although the above embodiments have detailed the concept and implementation of the object of the present invention with reference to the accompanying drawings, those of ordinary skill in the art can recognize that various improvements and transformations can still be made to the present invention without departing from the scope defined by the claims. Therefore, they are not elaborated here one by one

Claims

1. A tidal flat terrain monitoring method based on few-shot learning, characterized in that: It includes the following steps: S1. Set up a coastal monitoring system and a water level monitoring station in the monitored tidal flat, and obtain non-orthogonal tidal flat flooding image data of the tidal flat through the coastal monitoring system; S2. With the support of a small number of labeled samples, use the large-scale basic model SAM for reinforcement learning to extract the water edge line of the tidal flat in the non-orthogonal tidal flat flooding image; S3. Use the collinearity equation to convert the extracted non-orthogonal water edge line into an orthogonal water edge line; S4. Combine the real-time tide level information provided by the water level monitoring station and the converted orthogonal water edge line to construct the tidal flat topography on the time scale of the tidal cycle; Step S1 includes: using the coastal monitoring system to continuously capture images of the tidal flat at certain time intervals, taking the two water level values at the start of flooding and at full flooding in the images of the tidal flat as threshold water levels respectively, and screening the tidal flat flooding images located between the two threshold water levels; Step S2 includes: using the image encoder and image embedder of the SAM model to extract features from the non-orthogonal tidal flat flooding image; Using the prompt generation module to generate prompt points from labeled samples and unlabeled samples; For the labeled samples, augment the small samples through data augmentation operations, including rotating the original image by 90 degrees, 180 degrees, and 270 degrees, mirroring along the vertical and horizontal axes, and adjusting the intensity of the image through gamma correction to generate a series of variant images; For the unlabeled samples, segment various pixels using superpixels, distinguish the pixels containing water bodies from other types of landscapes by setting the comprehensive weights of color space similarity and planar distance, then generate a prompt mask according to the pixel characteristics of the tidal flat water bodies, and extract prompt points from it; Using the segmentation module, based on the trained prompt points, segment the tidal flat flooding image into water bodies and non-water bodies to extract the accurate water edge line.

2. The method for monitoring tidal flat topography based on few-shot learning according to claim 1, wherein: Use the Gaussian mixture algorithm to classify each pixel in the unlabeled samples into water body clusters and non-water body clusters.

3. The method for monitoring tidal flat topography based on few-shot learning according to claim 1, characterized in that: Step S3 includes: Lens distortion correction: Use the Matlab camera calibration toolbox, input square grid photos taken at different angles, and calculate the internal parameters of the camera of the coastal monitoring system; Image calibration: Solve the collinear linear equation using the direct linear transformation equation to convert the image coordinates into actual world coordinates; after the camera of the coastal monitoring system is fixed, take field images containing RTK calibration points, extract the pixel coordinates corresponding to the calibration points, and calculate the external parameters of the camera; Using the internal and external parameters of the camera, convert the coordinates of the water edge line set from pixel coordinates to geographic coordinates to achieve the orthographic projection of the water edge line; Iteratively calculate the number of intersection points of each water edge line with the remaining water edge lines, and delete the water edge line with the largest number of intersection points until there are no intersecting water edge lines.

4. A method for monitoring tidal flat topography based on few-shot learning according to claim 1, characterized in that: Step S4 includes: Calibrate the water level sensor of the water level monitoring station using RTK, and combine the real-time tide level data to discretize the water edge line at the same moment into elevation point clouds; Divide the monitoring area grid according to the required resolution, and interpolate the elevation point clouds into the grid through the natural neighbor interpolation method to form elevation raster data, that is, obtain the digital terrain elevation of the tidal flat.

Images