Coastal vegetation coverage inversion method based on image recognition

Abstract

The present application discloses a coastal vegetation coverage inversion method based on image recognition, relates to the field of data encryption technology, comprising collecting original remote sensing images of a coastal area, and synchronously recording tidal phase data at the time of collection, preprocessing the original remote sensing images to generate multispectral orthographic image maps and three-dimensional point cloud data; inputting the multispectral orthographic image maps, three-dimensional point cloud data and tidal phase data into a generative adversarial network to generate an extended training sample set of spectral features and vegetation coverage true value labels corresponding to tidal conditions; extracting spectral features, vegetation index features and three-dimensional structure features from the extended training sample set and measured data obtained through ground field investigation to construct an initial multidimensional feature training data set. The present application trains a physical guidance coastal vegetation coverage physical guidance semantic segmentation model, and generates a high-precision vegetation coverage map through inference and precision verification of the coastal vegetation coverage physical guidance semantic segmentation model.

Description

Technical Field

[0001] This invention relates to the field of data encryption technology, and in particular to a method for inverting coastal vegetation cover based on image recognition. Background Technology

[0002] With the rapid development of remote sensing technology, vegetation cover inversion technology based on Earth observation methods such as multispectral, hyperspectral and lidar has become an important support for ecological environment monitoring and research. Among the many technical routes, data-driven methods based on deep learning have shown great potential. They usually rely on large-scale, high-quality training sample sets, and by constructing complex convolutional neural networks or semantic segmentation models, they can directly learn the spectral, texture and spatial context features of vegetation from remote sensing images and output pixel-level vegetation cover prediction maps.

[0003] However, when applied to the specific and dynamically changing coastal habitat, the aforementioned existing technologies face a key technical challenge: the tidal cycle of flooding and exposing tidal flats causes drastic and nonlinear dynamic changes in surface moisture content, sediment reflectance characteristics, and the spectral response of vegetation canopy. The learning process of existing deep learning models heavily relies on the assumption that the training data and prediction data are in the same or similar environmental conditions, making it difficult to obtain fully labeled training samples that cover all tidal states. This results in a significant decrease in the generalization ability of the trained model when faced with prediction images that differ greatly from the tidal conditions of the training samples, leading to fluctuations in inversion accuracy. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an image recognition-based method for inverting coastal vegetation cover, which solves the core problems of insufficient model generalization ability and unstable inversion accuracy caused by dynamic tidal changes in existing technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a coastal vegetation cover inversion method based on image recognition, which includes: acquiring original remote sensing images of the coastal area and simultaneously recording tidal phase data at the acquisition time; preprocessing the original remote sensing images to generate multispectral orthophoto maps and three-dimensional point cloud data.

[0008] Multispectral orthophotos, 3D point cloud data, and tidal phase data are input into a generative adversarial network to generate an extended training sample set of spectral features and vegetation cover ground truth labels corresponding to tidal conditions.

[0009] Spectral features, vegetation index features, and three-dimensional structural features were extracted from the extended training sample set and measured data obtained through ground field surveys to construct an initial multidimensional feature training dataset.

[0010] Based on the tidal phase data and historical observation statistics, an empirical relationship function between tidal phase data and soil moisture is established. The equivalent surface moisture thickness of pixels in the multispectral orthophoto map is calculated. The equivalent surface moisture thickness is used to perform pixel-by-pixel physical correction of the spectral features and vegetation index features in the initial multidimensional feature training dataset, and the corrected multidimensional feature training dataset is generated.

[0011] Based on 3D point cloud data and solar elevation angle during imaging, the geometric consistency loss of vegetation projection is calculated. The corrected multidimensional features are used to train the dataset to train the physical-guided semantic segmentation model of coastal vegetation cover.

[0012] The multidimensional feature data obtained after feature extraction and physical correction of the area to be inverted is input into the trained coastal vegetation cover physical-guided semantic segmentation model to obtain a preliminary vegetation cover inversion map of the area to be inverted. The accuracy of the preliminary vegetation cover inversion map is verified, and the final coastal vegetation cover product is generated.

[0013] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the method includes: acquiring original remote sensing images of the coastal area and simultaneously recording tidal phase data at the acquisition time; preprocessing the original remote sensing images to generate multispectral orthophoto maps and three-dimensional point cloud data, including the following steps:

[0014] By using a drone equipped with a multispectral camera in the coastal area, raw remote sensing images are acquired, the exposure time of each frame of raw remote sensing image is recorded, and the tidal phase data corresponding to the acquisition time of each frame of raw remote sensing image is obtained by looking up the tide table.

[0015] Radiometric calibration is performed on the original remote sensing image to convert the pixel values ​​of the original remote sensing image into surface reflectance data. Atmospheric correction is then performed on the surface reflectance data to obtain the surface reflectance data.

[0016] By combining surface reflectance data with positioning and attitude data for geometric correction, geometric distortion is eliminated, and a georeferenced image is generated. The georeferenced image is then orthorectified to generate a multispectral orthophoto map with uniform planar accuracy and multispectral information.

[0017] Motion reconstruction of structure is performed on original remote sensing images with high overlap to generate sparse point clouds describing three-dimensional spatial locations, and then dense matching is performed to generate three-dimensional point cloud data.

[0018] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the method includes the following steps: inputting multispectral orthophotos, three-dimensional point cloud data, and tidal phase data into a generative adversarial network to generate extended training samples with spectral features and ground truth labels of vegetation cover corresponding to tidal conditions.

[0019] Multispectral orthophotos, 3D point cloud data, and tidal phase data are input into the generator of the generative adversarial network.

[0020] The generator of the generative adversarial network generates synthetic multispectral reflectance data based on multispectral orthophotos, 3D point cloud data, and tidal phase data.

[0021] The generator of the generative adversarial network synchronously generates and synthesizes vegetation cover true labels corresponding to multispectral reflectance data based on multispectral orthophoto maps, 3D point cloud data and tidal phase data.

[0022] The synthesized multispectral reflectance data were paired with the true values ​​of vegetation cover to form a synthetic sample;

[0023] The discriminator of the generative adversarial network (GAN) distinguishes between real and synthetic samples. Based on the feedback from the discriminator, the parameters of the generator and discriminator of the GAN are updated. The generator of the trained GAN outputs extended training samples with spectral features corresponding to tidal conditions and ground truth labels of vegetation cover.

[0024] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the following steps are included: extracting spectral features, vegetation index features, and three-dimensional structural features from the extended training sample set and measured data obtained through ground field surveys to construct an initial multidimensional feature training dataset:

[0025] The nonlinear relationship between the combination of near-infrared and red light band reflectance and vegetation cover was analyzed using the support vector machine regression method to form vegetation index characteristics.

[0026] The extracted spectral features, vegetation index features, and three-dimensional structural features are standardized and then aligned and merged according to the samples.

[0027] The merged features are associated with the expanded training sample set and the ground truth labels of vegetation cover from the measured data obtained through ground field surveys to form the initial multidimensional feature training dataset.

[0028] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the method includes the following steps: calculating the equivalent surface moisture thickness of pixels in a multispectral orthophoto image based on the tide-soil moisture empirical relationship function established by tidal phase data and historical observation statistics.

[0029] An empirical relationship function between tide and soil moisture was established by regression analysis of historical tidal phase data and synchronous field-measured soil volumetric water content data.

[0030] The tidal phase data corresponding to each pixel in the multispectral orthophoto image is input into the tidal-soil moisture empirical relationship function for calculation, and the predicted soil moisture content value of each pixel is output based on the input tidal phase data.

[0031] The predicted soil moisture content values ​​are converted into the equivalent surface moisture thickness.

[0032] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the method includes the following steps: pixel-by-pixel physical correction of the spectral features and vegetation index features in the initial multidimensional feature training dataset is performed using the equivalent surface water thickness to generate the corrected multidimensional feature training dataset.

[0033] The water-pair spectrum was obtained by consulting publicly available spectral libraries. Based on the equivalent surface water thickness and the water-pair spectrum attenuation coefficient, the spectral attenuation factor of each pixel was obtained.

[0034] The spectral features and vegetation index features in the initial multidimensional feature training dataset are simultaneously corrected using a spectral attenuation factor to obtain the corrected spectral features and vegetation index features.

[0035] The corrected spectral features and vegetation index features are merged with the three-dimensional structural features in the initial multidimensional feature training dataset;

[0036] The merged corrected spectral features, corrected vegetation index features, and three-dimensional structural features are re-associated with the ground truth labels of vegetation cover in the initial multidimensional feature training dataset.

[0037] The re-associated feature-label pairs are combined to generate a corrected multidimensional feature training dataset.

[0038] Combine all the re-associated feature-label pairs to generate a corrected multidimensional feature training dataset.

[0039] As a preferred embodiment of the coastal vegetation cover inversion method based on image recognition described in this invention, the following steps are included: Calculating the vegetation projection geometric consistency loss based on three-dimensional point cloud data and the solar elevation angle during imaging:

[0040] The direction of the light source is obtained based on the solar altitude angle and solar azimuth angle at the time of imaging;

[0041] Based on the three-dimensional point cloud data of the light source direction and potential vegetation area, the theoretical vegetation projection range is obtained through light projection simulation.

[0042] The vegetation cover predicted by the physical-guided semantic segmentation model of coastal vegetation cover. Figure 2 Values ​​are converted to obtain a predicted vegetation distribution map;

[0043] By spatially overlaying and comparing the theoretical vegetation projection range with the predicted vegetation distribution map, the difference between the theoretical vegetation projection range and the predicted vegetation distribution map is calculated, and the vegetation projection geometric consistency loss is obtained.

[0044] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the method includes the following steps: training a coastal vegetation cover physical-guided semantic segmentation model using a corrected multidimensional feature training dataset:

[0045] Input the feature vectors from the corrected multidimensional feature training dataset into the coastal vegetation cover physical-guided semantic segmentation model;

[0046] The coastal vegetation cover physical-guided semantic segmentation model performs forward propagation on the input feature vector and outputs a predicted vegetation cover map.

[0047] Data item loss based on the difference between the ground truth labels of vegetation cover in the vegetation cover map and the corrected multidimensional feature training dataset;

[0048] The total loss of the coastal vegetation cover physical-guided semantic segmentation model is obtained by weighted summing of the data item loss and the vegetation projection geometric consistency loss.

[0049] The network parameters of the coastal vegetation cover physical-guided semantic segmentation model are updated using the backpropagation algorithm with the total loss, resulting in the trained coastal vegetation cover physical-guided semantic segmentation model.

[0050] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the method includes the following steps: inputting the multidimensional feature data obtained after feature extraction and physical correction of the area to be inverted into the trained coastal vegetation cover physical-guided semantic segmentation model to obtain a preliminary vegetation cover inversion map of the area to be inverted.

[0051] The multidimensional feature data obtained after feature extraction and physical correction of the region to be inverted is input into the trained coastal vegetation cover physical-guided semantic segmentation model. The trained coastal vegetation cover physical-guided semantic segmentation model performs forward propagation on the multidimensional feature data obtained after feature extraction and physical correction of the input region to be inverted.

[0052] The trained coastal vegetation cover physical-guided semantic segmentation model outputs the vegetation cover probability value for each pixel in the area to be inverted.

[0053] The vegetation cover probability values ​​of the pixels are combined to form a preliminary vegetation cover inversion map of the area to be inverted.

[0054] As a preferred embodiment of the image recognition-based coastal vegetation cover inversion method of the present invention, the following steps are included: verifying the accuracy of the preliminary vegetation cover inversion map to generate the final coastal vegetation cover product:

[0055] A ground-based field survey was conducted in areas that did not participate in the training of the coastal vegetation cover physical-guided semantic segmentation model. Independent validation sample sets were obtained by using the quadrat method and positioning instruments.

[0056] Based on the independent validation sample set, the predicted vegetation cover value is extracted from the preliminary vegetation cover inversion map;

[0057] The extracted predicted vegetation cover values ​​are compared with the actual vegetation cover measurements in the independent validation sample set. Linear regression analysis is used to calculate the accuracy evaluation index between the predicted vegetation cover values ​​and the actual vegetation cover measurements.

[0058] Based on the accuracy evaluation index, the threshold values ​​for the evaluation index of the preliminary vegetation cover inversion map are set.

[0059] The preliminary vegetation cover inversion map that meets the evaluation index threshold of the preliminary vegetation cover inversion map is mapped and refined by adding information such as geographic coordinates, legend and scale to obtain the refined preliminary vegetation cover inversion map, thus forming the final coastal vegetation cover product.

[0060] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the coastal vegetation cover inversion method based on image recognition as described in the first aspect of the present invention.

[0061] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the coastal vegetation cover inversion method based on image recognition as described in the first aspect of the present invention.

[0062] The beneficial effects of this invention are as follows: By collecting original remote sensing images of coastal areas and simultaneously recording tidal phase data, multispectral orthophoto maps and three-dimensional point cloud data are generated after preprocessing. Generative adversarial networks are used to synthesize an extended training sample set corresponding to tidal conditions. Multidimensional features are extracted from the sample set to construct a training dataset. The equivalent surface moisture thickness is calculated through the tide-soil moisture empirical relationship function. The spectral and vegetation index features are physically corrected. The geometric consistency loss of vegetation projection is calculated by combining three-dimensional point cloud and solar altitude angle. A physical-guided semantic segmentation model for coastal vegetation cover is trained. A high-precision vegetation cover map is generated through reasoning and accuracy verification of the physical-guided semantic segmentation model for coastal vegetation cover. Attached Figure Description

[0063] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of an image recognition-based method for inverting coastal vegetation cover.

[0065] Figure 2 A schematic diagram illustrating the generation of training samples for adversarial networks.

[0066] Figure 3 This is a flowchart of the data preprocessing process.

[0067] Figure 4 This is a schematic diagram of vegetation cover inversion. Detailed Implementation

[0068] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0069] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0070] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0071] Reference Figures 1-4 This is one embodiment of the present invention, which provides a coastal vegetation cover inversion method based on image recognition, including the following steps:

[0072] S1. Collect raw remote sensing images of the coastal area and simultaneously record tidal phase data at the time of collection. Preprocess the raw remote sensing images to generate multispectral orthophoto maps and three-dimensional point cloud data.

[0073] S1.1. Using a drone equipped with a multispectral camera, acquire raw remote sensing images in the coastal area, record the exposure time of each frame of raw remote sensing image, and obtain the tidal phase data corresponding to the acquisition time of each frame of raw remote sensing image by querying the tide table.

[0074] Furthermore, a drone equipped with a multispectral camera flies along a flight path over the coastal area. During flight, the multispectral camera continuously images the ground surface to obtain raw remote sensing images composed of multiple bands. The high-precision global positioning system and inertial measurement unit on the drone record the precise latitude and longitude coordinates and attitude angle data at the moment of exposure of each frame of raw remote sensing image. Simultaneously, the exposure time of each frame of raw remote sensing image is recorded, and the exposure time is matched and queried with the tide table. The tide table contains data on the predicted tide height changes over time. By matching, the tidal phase data that strictly corresponds to the acquisition time of each frame of raw remote sensing image is obtained.

[0075] S1.2. Radiometric calibration is performed on the original remote sensing image. The pixel values ​​of the original remote sensing image are converted into surface reflectance data. Atmospheric correction is performed on the surface reflectance data to obtain the surface reflectance data.

[0076] Furthermore, radiometric calibration is performed on the original remote sensing image. The radiometric calibration process uses calibration coefficient files provided by the multispectral camera manufacturer to convert the pixel digital quantization values ​​of the original remote sensing image into physically meaningful apparent reflectance data. Atmospheric correction is then performed on the apparent reflectance data. Atmospheric correction adopts a method based on the radiative transfer model, inputting atmospheric parameters such as aerosol optical thickness and water vapor content during imaging to eliminate the influence of atmospheric scattering and absorption on the surface reflection signal, thereby obtaining surface reflectance data that is closer to the true physical properties of the surface.

[0077] S1.3 Combine surface reflectance data with positioning and attitude data for geometric correction to eliminate geometric distortion and generate georeferenced images. Perform orthorectification on the georeferenced images to generate multispectral orthophoto maps with uniform planar accuracy and multispectral information.

[0078] Furthermore, atmospherically corrected surface reflectance data is combined with positioning and attitude data recorded by the UAV's GPS and inertial measurement unit for geometric correction. The geometric correction process involves constructing a rigorous sensor imaging geometric model and using ground control points or direct georeferenced techniques to correct image geometric distortions caused by terrain undulations and sensor attitude changes, generating a geographic coordinate reference image. Then, orthorectification is performed on the geographic reference image. Orthorectification uses a digital elevation model to perform differential correction on the image to eliminate projection differences, generating a multispectral orthorectified image map.

[0079] S1.4. Motion reconstruction of structure is performed using original remote sensing images with high overlap to generate sparse point clouds describing three-dimensional spatial locations, and dense matching is then performed to generate three-dimensional point cloud data.

[0080] Furthermore, motion reconstruction processing is performed on the original remote sensing images, where both forward and lateral overlap are achieved. This motion reconstruction processing uses algorithms such as feature point extraction and matching, and incremental reconstruction to calculate the external orientation elements of each original remote sensing image and generate sparse 3D point clouds. Based on the sparse point clouds, a dense matching algorithm is applied to calculate the 3D coordinates of each pixel, thereby generating high-density, high-precision 3D point cloud data.

[0081] S2. Input the multispectral orthophoto map, 3D point cloud data and tidal phase data into the generative adversarial network to generate extended training samples with spectral features and vegetation cover ground truth labels corresponding to tidal conditions.

[0082] S2.1 Input the multispectral orthophoto map, 3D point cloud data and tidal phase data into the generator of the generative adversarial network.

[0083] Furthermore, the multi-band reflectance values ​​of each pixel in the multispectral orthophoto image, the three-dimensional structural feature vectors such as elevation and slope corresponding to each pixel extracted from the three-dimensional point cloud data, and the tidal phase data corresponding to each pixel are concatenated into a multi-dimensional conditional vector; the concatenated multi-dimensional conditional vectors are then batch-input into the input layer of the generator of the generative adversarial network.

[0084] S2.2 The generator of the generative adversarial network generates synthesized multispectral reflectance data based on multispectral orthophotos, 3D point cloud data and tidal phase data.

[0085] Furthermore, through a series of deconvolutional and upsampling layers for forward propagation, the nonlinear activation function in the network transforms the features, generating synthetic multispectral reflectance data in the output layer that corresponds to the input conditional vector pixel by pixel in space, maintaining the same dimensionality as the real multispectral orthophoto map.

[0086] S2.3 The generator of the generative adversarial network synchronously generates and synthesizes the vegetation cover true value labels corresponding to the multispectral reflectance data based on the multispectral orthophoto map, 3D point cloud data and tidal phase data.

[0087] Furthermore, a parallel branch structure within the network utilizes the same multidimensional conditional vector and intermediate feature representations to generate a ground truth vegetation cover map that is spatially fully registered with the synthesized multispectral reflectance data through an additional output head.

[0088] S2.4 Pair the synthesized multispectral reflectance data with the true vegetation cover labels to form a synthetic sample.

[0089] Furthermore, strict pairing is performed according to pixel location; each pairing unit contains a pixel with a synthetic multispectral reflectance vector and its corresponding ground truth label value of vegetation cover, and all pairing units of pixels are combined to form a complete synthetic sample.

[0090] S2.5 The discriminator of the generative adversarial network (GAN) distinguishes between real and synthetic samples. Based on the feedback of the discriminator, the parameters of the generator and the discriminator of the GAN are updated. The generator of the trained GAN outputs extended training samples with spectral features corresponding to tidal conditions and ground truth labels of vegetation cover.

[0091] Furthermore, the discriminator judges the authenticity of the input samples and outputs the discrimination probability, along with the discriminator's discrimination loss and the generator's generation loss. The backpropagation algorithm is used to simultaneously update the network weight parameters of the generator and discriminator in the generative adversarial network (GAN) based on the loss values. After multiple iterations of adversarial training, the generator of the GAN produces high-quality synthetic samples that are difficult for the discriminator to recognize. Finally, the generator of the stably trained GAN is used to generate an expanded training sample set in batches, containing spectral features and ground truth labels of vegetation cover corresponding to tidal conditions.

[0092] S3. Extract spectral features, vegetation index features, and three-dimensional structural features from the extended training sample set and the measured data obtained through ground field surveys to construct the initial multidimensional feature training dataset.

[0093] S3.1. Use the support vector machine regression method to analyze the nonlinear relationship between the combination of near-infrared and red light band reflectance and vegetation coverage, and form vegetation index characteristics.

[0094] Furthermore, the near-infrared and red band reflectance values ​​of each pixel are extracted from the multispectral orthophoto portion of the expanded training sample set and the measured data obtained through field surveys. These extracted near-infrared and red band reflectance values ​​are combined to form an input feature vector. The input feature vector and the corresponding ground truth vegetation cover label are then input into a support vector machine (SVM) regression model for training. The trained SVM regression model learns the complex nonlinear mapping relationship between the near-infrared and red band reflectance combination and vegetation cover. The trained SVM regression model is then used to predict the near-infrared and red band reflectance combinations of all pixels in the expanded training sample set and the measured data obtained through field surveys. The continuous value of the predicted output is the newly constructed vegetation index feature.

[0095] S3.2. The extracted spectral features, vegetation index features and three-dimensional structural features are standardized and then aligned and merged according to the samples.

[0096] Furthermore, the spectral features extracted from the extended training sample set and the measured data obtained through ground field surveys, the vegetation index features formed using the support vector machine regression method, and the three-dimensional structural features were standardized. The standardization process adopted the Z-score method, which involves subtracting the mean of each type of feature data and dividing it by its standard deviation, so that the mean of each type of feature data after processing is 0 and the standard deviation is 1. The standardized spectral features, standardized vegetation index features, and standardized three-dimensional structural features were strictly aligned according to the sample number. The aligned standardized spectral features, standardized vegetation index features, and standardized three-dimensional structural features were then combined and merged along the feature dimension to form a unified multi-dimensional feature vector for each sample.

[0097] S3.3. Associate the merged features with the expanded training sample set and the ground truth labels of vegetation cover in the measured data obtained through ground field surveys to form the initial multidimensional feature training dataset.

[0098] Furthermore, the multidimensional feature vectors of each merged sample are associated one-to-one with the corresponding ground truth labels of vegetation cover in the expanded training sample set and the measured data obtained through field surveys. The association operation ensures that each multidimensional feature vector has a correct ground truth label value of vegetation cover as a supervision signal. All the associated multidimensional feature vectors and ground truth labels of vegetation cover are combined to form an initial multidimensional feature training dataset with a complete structure and standardized format.

[0099] S4. Based on the tidal phase data and historical observation statistics, establish the tidal-soil moisture empirical relationship function to calculate the equivalent surface moisture thickness of pixels in the multispectral orthophoto image.

[0100] S4.1. An empirical relationship function between tide and soil moisture was established by regression analysis of historical tidal phase data and synchronous field-measured soil volumetric water content data.

[0101] Furthermore, historical tidal phase data and soil volumetric water content data measured simultaneously at the same time and location were collected to form sample pairs. Nonlinear regression analysis was used to fit the sample pairs. The goal of nonlinear regression analysis is to establish a mathematical function to describe the mapping relationship between tidal phase data and soil volumetric water content data. The parameters in the function were optimized by minimizing the error between the predicted and measured values, and finally the specific mathematical form and parameter values ​​of the tidal-soil moisture empirical relationship function were determined.

[0102] S4.2 Input the tidal phase data corresponding to each pixel in the multispectral orthophoto image into the tidal-soil moisture empirical relationship function for calculation, and output the predicted soil moisture content value of each pixel based on the input tidal phase data.

[0103] The expression for soil moisture content is:

[0104] ;

[0105] in, For the first The predicted soil moisture content value for each pixel. For the first The tidal phase data corresponding to each pixel, The initial saturation coefficient, This is the permeation attenuation rate coefficient. To dominate the tidal cycle, For the background moisture constant term, For cell index;

[0106] Furthermore, the tidal phase data corresponding to each pixel in the multispectral orthophoto image is used as input and substituted into the established tidal-soil moisture empirical relationship function. The tidal-soil moisture empirical relationship function calculates the input tidal phase data according to its internal mathematical expression. The calculation process of the function involves exponential and sine terms. The exponential term simulates the nonlinear effect of water infiltration and retention, and the sine term captures the influence of the periodic fluctuation of tides. After the function performs the calculation, it outputs a continuous predicted soil moisture content value for each pixel.

[0107] S4.3 Convert the predicted soil moisture content values ​​into the equivalent surface moisture thickness.

[0108] Furthermore, the predicted soil moisture content value output by the tide-soil moisture empirical relationship function is multiplied by an equivalent layer thickness parameter to convert the water content per unit volume into an equivalent water layer depth with thickness dimensions, thus forming the surface equivalent water thickness for each pixel.

[0109] S5. Using the equivalent surface water thickness, perform pixel-by-pixel physical correction on the spectral features and vegetation index features in the initial multidimensional feature training dataset to generate the corrected multidimensional feature training dataset.

[0110] S5.1 Obtain the water-pair spectrum by consulting publicly available spectral libraries, and obtain the spectral attenuation factor for each pixel based on the equivalent surface water thickness and the water-pair spectrum attenuation coefficient.

[0111] Furthermore, the absorption coefficients of liquid water in key bands such as near-infrared and red light are queried from publicly available spectral databases as the attenuation coefficients of water on the spectrum; the obtained attenuation coefficients of water on the spectrum are multiplied by the equivalent surface water thickness of each pixel; the product is used as the exponential part of an exponential function for calculation; and the negative exponent of the natural constant is calculated to obtain the spectral attenuation factor of each pixel.

[0112] S5.2. Use the spectral attenuation factor to synchronously correct the spectral features and vegetation index features in the initial multidimensional feature training dataset to obtain the corrected spectral features and vegetation index features.

[0113] Furthermore, the spectral feature values ​​in the initial multidimensional feature training dataset are divided by the spectral attenuation factor of the corresponding pixel to obtain the corrected spectral features through reverse compensation; the vegetation index feature values ​​in the initial multidimensional feature training dataset are regarded as intermediate results obtained from the original spectral features, and the corrected vegetation index features are obtained again by using the support vector machine regression method based on the corrected spectral feature values.

[0114] S5.3. Merge the corrected spectral features, the corrected vegetation index features, and the three-dimensional structural features in the initial multidimensional feature training dataset.

[0115] Furthermore, the corrected spectral features, corrected vegetation index features, and three-dimensional structural features in the initial multidimensional feature training dataset are aligned with the same sample order and pixel position; the aligned corrected spectral features, corrected vegetation index features, and three-dimensional structural features are then combined and merged along the feature dimension to form a new merged feature vector.

[0116] S5.4 Reassociate the merged corrected spectral features, corrected vegetation index features, and three-dimensional structural features with the ground truth labels of vegetation cover in the initial multidimensional feature training dataset.

[0117] Furthermore, the sample identifiers corresponding to the merged corrected spectral features, corrected vegetation index features, and three-dimensional structural features are matched with the ground truth labels of vegetation cover stored in the initial multidimensional feature training dataset; a one-to-one correspondence is established between the merged feature vectors and the original ground truth labels of vegetation cover.

[0118] S5.5. Combine the re-associated feature-label pairs to generate a corrected multidimensional feature training dataset;

[0119] Furthermore, the corrected spectral features and corrected vegetation index features obtained after physical correction are merged with the three-dimensional structural features in the initial multidimensional feature training dataset. The merged feature vectors are then associated one-to-one with the corresponding ground truth labels of vegetation cover in the initial multidimensional feature training dataset to form reassociated feature-label pairs. All reassociated feature-label pairs are collected and integrated to finally generate a corrected multidimensional feature training dataset with a complete structure and corrected for the influence of tidal moisture.

[0120] S5.6 Combine all the re-associated feature-label pairs to generate the corrected multidimensional feature training dataset.

[0121] Furthermore, each re-associated feature-label pair, namely the feature vector formed by merging the corrected spectral features, the corrected vegetation index features, and the three-dimensional structural features, and the strictly corresponding ground truth label of vegetation cover, is treated as a complete data unit. All data units are systematically collected and integrated to ultimately generate a corrected multidimensional feature training dataset with a unified structure and standardized format.

[0122] S6. Calculate the vegetation projection geometric consistency loss based on 3D point cloud data and solar elevation angle during imaging.

[0123] S6.1. Obtain the direction of the light source based on the solar altitude angle and solar azimuth angle during imaging.

[0124] Furthermore, based on the solar altitude angle and solar azimuth angle during imaging; the solar altitude angle is defined as the angle between the sun's rays and the horizontal plane, and the solar azimuth angle is defined as the angle between the projection of the sun's rays onto the horizontal plane and the due north direction; the solar altitude angle and solar azimuth angle are converted into a unit direction vector by trigonometric functions, pointing from the Earth's surface to the solar light source, and the unit direction vector is the direction of the light source.

[0125] S6.2. Based on the three-dimensional point cloud data of the light source direction and the potential vegetation area, the theoretical vegetation projection range is obtained through light projection simulation.

[0126] Furthermore, based on the direction of the light source and the potential vegetation areas identified from the 3D point cloud data, which consist of a set of 3D point cloud data points with elevations significantly higher than the surrounding flat terrain, the theoretical vegetation projection range is calculated through ray projection simulation. The ray projection simulation starts from the 3D point cloud data vertex of each potential vegetation area and emits light rays in the opposite direction to the light source. The first point where each ray intersects with the digital terrain surface constructed from the 3D point cloud data is traced. All intersections of the ray rays and the terrain are projected onto a two-dimensional horizontal plane to generate the theoretical vegetation projection range.

[0127] S6.3, Using the physical guidance semantic segmentation model to predict vegetation cover in coastal areas. Figure 2 Values ​​are converted to obtain the predicted vegetation distribution map.

[0128] Furthermore, the vegetation cover map predicted by the coastal vegetation cover physical-guided semantic segmentation model is binarized. A threshold is set for the binarization process. For example, pixels with a vegetation cover value greater than 0.5 are assigned a value of 1, indicating that the area is predicted to be vegetation, and pixels with a vegetation cover value less than or equal to 0.5 are assigned a value of 0, indicating that the area is predicted to be non-vegetated. The binary image obtained after the assignment operation is the predicted vegetation distribution map.

[0129] S6.4. Spatially overlay and compare the theoretical vegetation projection range with the predicted vegetation distribution map, calculate the difference measure between the theoretical vegetation projection range and the predicted vegetation distribution map, and obtain the vegetation projection geometric consistency loss.

[0130] The expression for vegetation projection geometric consistency loss is:

[0131] ;

[0132] in, For vegetation projection geometric consistency loss, This represents the theoretical vegetation projection range. To predict vegetation distribution maps.

[0133] Furthermore, the theoretical vegetation projection range is spatially overlaid and compared with the predicted vegetation distribution map; the number of pixels in the intersection region between the theoretical vegetation projection range and the predicted vegetation distribution map is calculated; the number of pixels in the union region between the theoretical vegetation projection range and the predicted vegetation distribution map is calculated; the number of intersection pixels is divided by the number of union pixels to obtain the cross-union ratio; the value minus the cross-union ratio is used as the difference measure, which is the vegetation projection geometric consistency loss.

[0134] S7. Using the corrected multidimensional feature training dataset, train the physical-guided semantic segmentation model of coastal vegetation cover.

[0135] S7.1 Input the feature vectors from the corrected multidimensional feature training dataset into the coastal vegetation cover physical-guided semantic segmentation model.

[0136] Furthermore, the feature vectors of each sample in the corrected multidimensional feature training dataset are used as input data and fed into the input layer of the coastal vegetation cover physical-guided semantic segmentation model in batches.

[0137] S7.2 The coastal vegetation cover physical-guided semantic segmentation model performs forward propagation on the input feature vector and outputs the predicted vegetation cover map.

[0138] Furthermore, the coastal vegetation cover physical-guided semantic segmentation model extracts multi-level features from the input feature vector through its internal encoder network, which consists of convolutional layers, pooling layers, and activation functions. The extracted features are then upsampled and spatial information restored through the decoder network, which consists of deconvolutional layers and skip connections. The output layer of the coastal vegetation cover physical-guided semantic segmentation model uses the sigmoid activation function to generate a predicted vegetation cover map of the same size as the input, where each pixel value represents the probability of the presence of vegetation.

[0139] S7.3, Data item loss based on the true vegetation cover labels in the training dataset after vegetation cover map and correction of multidimensional features.

[0140] Furthermore, based on the predicted vegetation cover map output by the coastal vegetation cover physical-guided semantic segmentation model and the true vegetation cover ground truth labels stored in the corrected multidimensional feature training dataset, the difference between the predicted vegetation cover map and the true vegetation cover ground truth labels is calculated pixel by pixel. The cross-entropy loss function is used as the data term loss. The cross-entropy of the predicted value of each pixel in the predicted vegetation cover map and the true vegetation cover ground truth label value is calculated, and the average of the cross-entropy loss values ​​of all pixels is taken to obtain the data term loss value representing the overall deviation between the entire predicted vegetation cover map and the true label.

[0141] S7.4. The data item loss and the vegetation projection geometric consistency loss are weighted and summed to obtain the total loss of the coastal vegetation cover physical guided semantic segmentation model.

[0142] Furthermore, the data item loss and the vegetation projection geometric consistency loss are weighted and summed. The weight coefficient of the data item loss is used to adjust the model's emphasis on the accuracy of fitting the training data, and the weight coefficient of the vegetation projection geometric consistency loss is used to adjust the degree of emphasis of the coastal vegetation cover physical-guided semantic segmentation model on the degree of adherence to physical laws. The weighted data item loss and the weighted vegetation projection geometric consistency loss are added together to obtain the total loss of the coastal vegetation cover physical-guided semantic segmentation model.

[0143] S7.5. Use the total loss to update the network parameters of the coastal vegetation cover physical-guided semantic segmentation model through the backpropagation algorithm to obtain the trained coastal vegetation cover physical-guided semantic segmentation model.

[0144] Furthermore, the gradient descent optimization algorithm is adopted to update the model's weight parameters and bias parameters based on the calculated gradient values; the forward propagation, back propagation and parameter update process are repeated until the total loss converges to a stable value, thus obtaining the trained coastal vegetation cover physical-guided semantic segmentation model.

[0145] S8. Input the multidimensional feature data obtained after feature extraction and physical correction of the area to be inverted into the trained coastal vegetation cover physical-guided semantic segmentation model to obtain the preliminary vegetation cover inversion map of the area to be inverted.

[0146] S8.1 Input the multidimensional feature data obtained after feature extraction and physical correction of the region to be inverted into the trained coastal vegetation cover physical-guided semantic segmentation model. The trained coastal vegetation cover physical-guided semantic segmentation model performs forward propagation on the multidimensional feature data obtained after feature extraction and physical correction of the input region to be inverted.

[0147] Furthermore, the multidimensional feature data obtained after feature extraction and physical correction of the region to be inverted is loaded into memory in batches; the trained coastal vegetation cover physical-guided semantic segmentation model receives the multidimensional feature data as input; the coastal vegetation cover physical-guided semantic segmentation model internally performs convolution operations, activation function transformations and feature map upsampling and downsampling operations in sequence according to the forward propagation path determined in the training phase, and processes the input multidimensional feature data layer by layer.

[0148] S8.2. The trained coastal vegetation cover physical-guided semantic segmentation model outputs the vegetation cover probability value of each pixel in the area to be inverted.

[0149] Furthermore, the final layer of the trained coastal vegetation cover physical-guided semantic segmentation model applies the Sigmoid activation function to the processed feature map; the Sigmoid activation function compresses the output value of each pixel to the range of 0 to 1; the output value is interpreted as the probability that the corresponding pixel has vegetation cover, i.e., the vegetation cover probability value; the coastal vegetation cover physical-guided semantic segmentation model sequentially outputs the vegetation cover probability values ​​of all pixels in the region to be inverted.

[0150] S8.3 Combine the vegetation cover probability values ​​of the pixels to form a preliminary vegetation cover inversion map of the area to be inverted.

[0151] Furthermore, the vegetation cover probability values ​​of all pixels are arranged according to their original spatial location in the area to be inverted; the arranged vegetation cover probability value matrix is ​​converted into a georeferenced raster image; the raster image is the preliminary vegetation cover inversion map of the area to be inverted, and the gray value of each pixel in the image represents the vegetation cover probability of the location.

[0152] S9. Verify the accuracy of the preliminary vegetation cover inversion map and generate the final coastal vegetation cover product.

[0153] S9.1. Conduct ground field surveys in areas that did not participate in the training of the coastal vegetation cover physical-guided semantic segmentation model, and obtain independent validation sample sets by using quadratic box method and positioning instrument measurement.

[0154] Furthermore, the quadrat method was used to conduct on-site measurements of vegetation cover within the selected sample plots. At the same time, high-precision positioning instruments were used to record the geographic coordinates of the center point of each quadrat. The ratio of the actual vegetation cover area measured in each quadrat to the quadrat area was taken as the true vegetation cover measurement value of the quadrat. The geographic coordinates of all quadrats and the corresponding true vegetation cover measurement values ​​were combined to form an independent validation sample set.

[0155] S9.2. Based on the independent validation sample set, extract the predicted vegetation cover value from the preliminary vegetation cover inversion map.

[0156] Furthermore, based on the geographic coordinates of each quadrat in the independent validation sample set, the corresponding pixel location is located on the preliminary vegetation cover inversion map; the vegetation cover probability value of the pixel location is read as the predicted vegetation cover value of the quadrat location; and the predicted vegetation cover values ​​corresponding to all quadrats are extracted to form a predicted value sequence.

[0157] S9.3. Compare the extracted predicted vegetation cover values ​​with the actual vegetation cover measurements in the independent validation sample set, and use linear regression analysis to calculate the accuracy evaluation index between the predicted vegetation cover values ​​and the actual vegetation cover measurements.

[0158] The expression for the accuracy evaluation index is:

[0159] ;

[0160] in, As an accuracy evaluation index, For independent validation of the sample set, For the first The predicted vegetation cover value for each sample. For the first The actual vegetation cover measurement value of each sample. For sample index;

[0161] Furthermore, the extracted predicted vegetation cover value sequence and the actual vegetation cover measurement value sequence in the independent validation sample set are used as input data for linear regression analysis. The linear regression analysis establishes a linear relationship model between the predicted vegetation cover value and the actual vegetation cover measurement value. The squared value of the coefficient of determination of the linear regression model is calculated, and the squared value of the coefficient is used as the accuracy evaluation index.

[0162] S9.4. Based on the accuracy evaluation index, set the threshold for the evaluation index of the preliminary vegetation cover inversion map.

[0163] Furthermore, based on actual application needs and minimum requirements for product accuracy, a pass threshold for accuracy evaluation indicators is pre-set; the accuracy evaluation indicators are then compared with the pass threshold.

[0164] S9.5. For the preliminary vegetation cover inversion map that meets the evaluation index threshold, map editing and finishing are carried out, and information such as geographic coordinates, legend and scale are added to obtain the preliminary vegetation cover inversion map after map editing and finishing, forming the final coastal vegetation cover product.

[0165] Furthermore, for the preliminary vegetation cover inversion map that meets the preset qualified threshold for accuracy evaluation indicators, map elements such as geographic coordinate system, legend, scale, compass, and necessary titles and explanatory text are added; the output format of the refined map is converted and optimized to generate a deliverable final coastal vegetation cover product that conforms to industry standards.

[0166] This embodiment also provides a computer device applicable to the coastal vegetation cover inversion method based on image recognition, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the coastal vegetation cover inversion method based on image recognition as proposed in the above embodiment.

[0167] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0168] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the coastal vegetation cover inversion method based on image recognition as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0169] In summary, this invention involves: acquiring original remote sensing images of coastal areas and simultaneously recording tidal phase data; generating multispectral orthophoto maps and 3D point cloud data through preprocessing; synthesizing an extended training sample set corresponding to tidal conditions using a generative adversarial network; extracting multidimensional features from the sample set to construct a training dataset; calculating the equivalent surface moisture thickness using the tide-soil moisture empirical relationship function; physically correcting spectral and vegetation index features; calculating the vegetation projection geometric consistency loss by combining 3D point cloud data and solar altitude angle; training a physical-guided semantic segmentation model for coastal vegetation cover; and generating a high-precision vegetation cover map through inference and accuracy verification using the physical-guided semantic segmentation model for coastal vegetation cover.

[0170] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for inverting coastal vegetation cover based on image recognition, characterized in that: This includes acquiring raw remote sensing images of the coastal area and simultaneously recording tidal phase data at the time of acquisition, preprocessing the raw remote sensing images, and generating multispectral orthophoto maps and three-dimensional point cloud data. Multispectral orthophotos, 3D point cloud data, and tidal phase data are input into a generative adversarial network to generate extended training samples with spectral features and ground truth labels for vegetation cover corresponding to tidal conditions. This includes the following steps: Multispectral orthophotos, 3D point cloud data, and tidal phase data are input into the generator of the generative adversarial network. The generator of the generative adversarial network generates synthetic multispectral reflectance data based on multispectral orthophotos, 3D point cloud data, and tidal phase data. The generator of the generative adversarial network synchronously generates and synthesizes vegetation cover true labels corresponding to multispectral reflectance data based on multispectral orthophoto maps, 3D point cloud data and tidal phase data. The synthesized multispectral reflectance data were paired with the true values ​​of vegetation cover to form a synthetic sample; The discriminator of the generative adversarial network (GAN) distinguishes between real and synthetic samples. Based on the feedback of the discriminator, the parameters of the generator and the discriminator of the GAN are updated. The generator of the trained GAN outputs extended training samples with spectral features corresponding to tidal conditions and ground truth labels of vegetation cover. Spectral features, vegetation index features, and three-dimensional structural features were extracted from the extended training sample set and measured data obtained through ground field surveys to construct an initial multidimensional feature training dataset. Based on the tidal phase data and historical observation statistics, an empirical relationship function between tidal phase data and soil moisture is established. The equivalent surface moisture thickness of pixels in the multispectral orthophoto map is calculated. The equivalent surface moisture thickness is used to perform pixel-by-pixel physical correction of the spectral features and vegetation index features in the initial multidimensional feature training dataset, and the corrected multidimensional feature training dataset is generated. Based on 3D point cloud data and solar elevation angle during imaging, the geometric consistency loss of vegetation projection is calculated. The corrected multidimensional features are used to train the dataset to train the physical-guided semantic segmentation model of coastal vegetation cover. The physical-guided semantic segmentation model of coastal vegetation cover includes an encoder network and a decoder network. The encoder network consists of convolutional layers, pooling layers, and activation functions. The decoder network consists of deconvolutional layers and skip connections. The output layer uses the sigmoid activation function. The data item loss is based on the loss between the predicted vegetation cover map and the ground truth vegetation cover label. The data item loss and the vegetation projection geometric consistency loss are weighted and summed to obtain the total loss of the coastal vegetation cover physical guided semantic segmentation model. The multidimensional feature data obtained after feature extraction and physical correction of the area to be inverted is input into the trained coastal vegetation cover physical-guided semantic segmentation model to obtain a preliminary vegetation cover inversion map of the area to be inverted. The accuracy of the preliminary vegetation cover inversion map is verified, and the final coastal vegetation cover map is generated.

2. The coastal vegetation cover inversion method based on image recognition as described in claim 1, characterized in that: Acquire raw remote sensing images of the coastal area and simultaneously record tidal phase data at the time of acquisition. Preprocess the raw remote sensing images to generate multispectral orthophoto maps and 3D point cloud data, including the following steps: By using a drone equipped with a multispectral camera in the coastal area, raw remote sensing images are acquired, the exposure time of each frame of raw remote sensing image is recorded, and the tidal phase data corresponding to the acquisition time of each frame of raw remote sensing image is obtained by looking up the tide table. Radiometric calibration is performed on the original remote sensing image to convert the pixel values ​​of the original remote sensing image into surface reflectance data. Atmospheric correction is then performed on the surface reflectance data to obtain the surface reflectance data. By combining surface reflectance data with positioning and attitude data for geometric correction, geometric distortion is eliminated, and a georeferenced image is generated. The georeferenced image is then orthorectified to generate a multispectral orthophoto map with uniform planar accuracy and multispectral information. Motion reconstruction of structure is performed on original remote sensing images with high overlap to generate sparse point clouds describing three-dimensional spatial locations, and then dense matching is performed to generate three-dimensional point cloud data.

3. The coastal vegetation cover inversion method based on image recognition as described in claim 2, characterized in that: Spectral features, vegetation index features, and three-dimensional structural features are extracted from the extended training sample set and measured data obtained through ground field surveys to construct an initial multidimensional feature training dataset, including the following steps: The nonlinear relationship between the combination of near-infrared and red light band reflectance and vegetation cover was analyzed using the support vector machine regression method to form vegetation index characteristics. The extracted spectral features, vegetation index features, and three-dimensional structural features are standardized and then aligned and merged according to the samples. The merged features are associated with the expanded training sample set and the ground truth labels of vegetation cover from the measured data obtained through ground field surveys to form the initial multidimensional feature training dataset.

4. The coastal vegetation cover inversion method based on image recognition as described in claim 1, characterized in that: Based on the tidal phase data and historical observation statistics, an empirical relationship function between tides and soil moisture is established to calculate the equivalent surface moisture thickness of pixels in a multispectral orthophoto image. This includes the following steps: An empirical relationship function between tide and soil moisture was established by regression analysis of historical tidal phase data and synchronous field-measured soil volumetric water content data. The tidal phase data corresponding to each pixel in the multispectral orthophoto image is input into the tidal-soil moisture empirical relationship function for calculation, and the predicted soil moisture content value of each pixel is output based on the input tidal phase data. The predicted soil moisture content values ​​are converted into the equivalent surface moisture thickness.

5. The coastal vegetation cover inversion method based on image recognition as described in claim 4, characterized in that: The spectral features and vegetation index features in the initial multidimensional feature training dataset are physically corrected pixel by pixel using the equivalent surface water thickness to generate the corrected multidimensional feature training dataset, including the following steps: The water-pair spectrum was obtained by consulting publicly available spectral libraries. Based on the equivalent surface water thickness and the water-pair spectrum attenuation coefficient, the spectral attenuation factor of each pixel was obtained. The spectral features and vegetation index features in the initial multidimensional feature training dataset are simultaneously corrected using a spectral attenuation factor to obtain the corrected spectral features and vegetation index features. The corrected spectral features and vegetation index features are merged with the three-dimensional structural features in the initial multidimensional feature training dataset; The merged corrected spectral features, corrected vegetation index features, and three-dimensional structural features are re-associated with the ground truth labels of vegetation cover in the initial multidimensional feature training dataset. The re-associated feature-label pairs are combined to generate a corrected multidimensional feature training dataset. Combine all the re-associated feature-label pairs to generate a corrected multidimensional feature training dataset.

6. The coastal vegetation cover inversion method based on image recognition as described in claim 5, characterized in that: Based on 3D point cloud data and the solar elevation angle during imaging, the geometric consistency loss of vegetation projection is calculated, including the following steps: The direction of the light source is obtained based on the solar altitude angle and solar azimuth angle at the time of imaging; Based on the three-dimensional point cloud data of the light source direction and potential vegetation area, the theoretical vegetation projection range is obtained through light projection simulation. Binarize the vegetation cover map predicted by the coastal vegetation cover physical-guided semantic segmentation model to obtain the predicted vegetation distribution map. By spatially overlaying and comparing the theoretical vegetation projection range with the predicted vegetation distribution map, the difference between the theoretical vegetation projection range and the predicted vegetation distribution map is calculated, and the vegetation projection geometric consistency loss is obtained.

7. The coastal vegetation cover inversion method based on image recognition as described in claim 6, characterized in that, The corrected multidimensional feature training dataset is used to train a physical-guided semantic segmentation model for coastal vegetation cover, including the following steps: Input the feature vectors from the corrected multidimensional feature training dataset into the coastal vegetation cover physical-guided semantic segmentation model; The coastal vegetation cover physical-guided semantic segmentation model performs forward propagation on the input feature vector and outputs a predicted vegetation cover map. Data item loss based on the difference between the ground truth labels of vegetation cover in the vegetation cover map and the corrected multidimensional feature training dataset; The total loss of the coastal vegetation cover physical-guided semantic segmentation model is obtained by weighted summing of the data item loss and the vegetation projection geometric consistency loss. The network parameters of the coastal vegetation cover physical-guided semantic segmentation model are updated using the backpropagation algorithm with the total loss, resulting in the trained coastal vegetation cover physical-guided semantic segmentation model.

8. The coastal vegetation cover inversion method based on image recognition as described in claim 7, characterized in that: The multidimensional feature data obtained after feature extraction and physical correction of the area to be inverted is input into the trained coastal vegetation cover physical-guided semantic segmentation model to obtain a preliminary vegetation cover inversion map of the area to be inverted, including the following steps: The multidimensional feature data obtained after feature extraction and physical correction of the region to be inverted is input into the trained coastal vegetation cover physical-guided semantic segmentation model. The trained coastal vegetation cover physical-guided semantic segmentation model performs forward propagation on the multidimensional feature data obtained after feature extraction and physical correction of the input region to be inverted. The trained coastal vegetation cover physical-guided semantic segmentation model outputs the vegetation cover probability value for each pixel in the area to be inverted. The vegetation cover probability values ​​of the pixels are combined to form a preliminary vegetation cover inversion map of the area to be inverted.

9. The coastal vegetation cover inversion method based on image recognition as described in claim 8, characterized in that: The accuracy of the preliminary vegetation cover inversion map is verified, and the final coastal vegetation cover map is generated, including the following steps: A ground-based field survey was conducted in areas that did not participate in the training of the coastal vegetation cover physical-guided semantic segmentation model. Independent validation sample sets were obtained by using the quadrat method and positioning instruments. Based on the independent validation sample set, the predicted vegetation cover value is extracted from the preliminary vegetation cover inversion map; The extracted predicted vegetation cover values ​​are compared with the actual vegetation cover measurements in the independent validation sample set. Linear regression analysis is used to calculate the accuracy evaluation index between the predicted vegetation cover values ​​and the actual vegetation cover measurements. Based on the accuracy evaluation index, the threshold values ​​for the evaluation index of the preliminary vegetation cover inversion map are set. The preliminary vegetation cover inversion map that meets the evaluation index threshold of the preliminary vegetation cover inversion map is mapped and refined by adding information such as geographic coordinates, legend and scale to obtain the refined preliminary vegetation cover inversion map, thus forming the final coastal vegetation cover map.

Images