A method for monitoring the vegetation coverage degree of an intertidal zone in ecological restoration of a shoreline

Abstract

The application discloses a kind of ecological restoration coast line intertidal zone vegetation coverage dynamic monitoring method, belong to the technical field of coastal zone ecology monitoring, solve intertidal zone vegetation coverage monitoring because of tidal boundary drift, vegetation type mixed and topography fluctuation leads to data distortion problem. Including synchronous acquisition satellite and unmanned aerial vehicle image;Intertidal zone range is dynamically demarcated based on real-time tidal data, generates boundary vector layer and calculates total area;After image and boundary are superimposed, input convolutional neural network model and output mangrove, salt marsh herb and mutual flower rice classification result;Classification result is optimized to morphology, and initial vertical projection area is calculated;According to slope distribution, sampling grid is laid out, and vegetation canopy projection area is measured to generate error correction factor optimization model parameter;Calibration projection area is calculated by reclassifying using optimization model;Finally, effective coverage is calculated. Mainly used for the accurate monitoring and acceptance certification of vegetation coverage in coastal zone ecological restoration engineering.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of coastal ecological monitoring, and particularly relates to a method for dynamically monitoring vegetation coverage in an intertidal zone of an ecological restoration shoreline. BACKGROUND

[0002] In the monitoring of coastal ecological restoration projects, the accurate assessment of vegetation coverage in the intertidal zone faces a number of technical challenges:

[0003] 1. Dynamic shift of the intertidal zone boundary. The range of the intertidal zone is affected by the periodic inundation of tides, and the land-side boundary shrinks towards the land during high tide and expands towards the sea during low tide. Traditional methods use fixed elevation lines or manually drawn boundaries, which are difficult to match the actual inundation range. In particular, in areas with artificial dikes or roads, the land-side boundary shrinks by more than 10% due to dike foot accumulation; in natural shorelines without structures, the boundary is strongly subjective due to the ambiguity of vegetation debris traces. The undulating topography of the tidal flat further causes the measured position of the mean low tide line to deviate from the theoretical value, exacerbating the jaggedness of the boundary.

[0004] 2. Classification confusion of vegetation types. The spectral characteristics of mangroves, salt marshes and invasive species Spartina alterniflora overlap significantly in conventional remote sensing images: the similarity of near-infrared reflectivity between mangroves and Spartina alterniflora is over 85%; the spectral variation of salt marshes during the withering period causes them to be easily misclassified as bare beaches during the withering period; and the high reflectivity of flower spikes in flowering mangroves is misclassified as salt marshes. Existing classification models rely on a single time image and lack adaptability to phenological changes, with a classification accuracy of less than 70% in mixed vegetation areas.

[0005] 3. Topographic undulation interference with sampling representativeness: the micro-topography of the intertidal zone is highly heterogeneous, with a slope of over 5° in the tidal creek area and a slope of less than 2° on the beach. Traditional uniform sampling grids (such as 100m x 100m) miss topographic abruptness features due to sparse sampling on steep slopes, and waste resources due to excessive sampling on gentle slopes. The projection area calculation deviation caused by slope is over 30%, and cannot be eliminated by post-statistical correction.

[0006] The above problems have complex causes: the tidal movement is uncontrollable, the vegetation phenology responds nonlinearly to environmental factors, and the topography is continuously modified by water dynamics. Existing technologies attempt to alleviate errors by increasing the frequency of manual reconnaissance or high-resolution images, but the cost is high and dynamic monitoring is difficult to achieve. These defects restrict the standardization of ecological restoration shoreline verification. SUMMARY

[0007] The present application aims to solve the following technical problems in the monitoring of vegetation coverage in the intertidal zone: the statistical range is distorted due to the shift of the tidal boundary; the coverage of native and invasive species is confused due to the mixing of vegetation types; and the sampling deviation is caused by topographic undulation, resulting in calculation errors of coverage.

[0008] When artificial dykes coexist with natural coastlines, traditional boundary demarcation methods cannot unify land-side boundary standards, leading to systematic deviation of intertidal zone range.

[0009] Conventional remote sensing classification ignores phenological changes and spectral overlap, resulting in a misclassification rate of over 30% for mangroves, salt marsh herbs, and Spartina alterniflora.

[0010] Image classification results have isolated noise points, internal holes, and jagged edges, which destroy the continuity of vegetation patches and deviate the area calculation from the actual distribution.

[0011] Uniform sampling cannot adapt to the sudden change in intertidal zone slope, and data is sparse in steep slope areas and redundant in gentle slope areas, lacking terrain representation.

[0012] Errors in mixed vegetation areas are transmitted across categories, overestimating the pollution of native vegetation statistics by Spartina alterniflora, and the coupling of correction factors exacerbates distortion.

[0013] The coverage monitoring results are inconsistent with the engineering acceptance standards, and cannot automatically generate spatial targeted repair schemes.

[0014] Different types of canopy projection areas cannot be physically separated in mixed vegetation quadrats, and the measured values are misallocated.

[0015] The superpixel boundary is blurred due to the reflection of water bodies in the intertidal zone, and the segmentation results are inconsistent with the actual vegetation distribution.

[0016] Visual interpretation of micro-morphological characteristics is prone to errors due to changes in field illumination and physiological variations of vegetation.

[0017] To solve the above problems, the present application provides a kind of ecological restoration coast intertidal zone vegetation coverage dynamic monitoring method, comprising: synchronously obtaining satellite image and unmanned aerial vehicle aerial image of target intertidal zone area, covering mangrove, salt marsh herb and Spartina alterniflora distribution area;Based on real-time tide table data, determine the tidal state at monitoring time;Intertidal zone range is dynamically demarcated: at high tide, the land-side boundary is the trace line of the high tide of the average high tide of many years or the outer edge line of vegetation growth;At low tide, the sea-side boundary is the average low tide line;Generate intertidal zone dynamic boundary vector layer, calculate its total area S 总; superimpose the image and the boundary layer, input the convolutional neural network model, and output the classification results of the mangrove, salt marsh herb, and Spartina alterniflora; perform morphological optimization on the classification results, and calculate the initial vertical projection areas of the mangrove, salt marsh herb, and Spartina alterniflora respectively; according to the slope distribution of the intertidal zone, the sampling grid is arranged densely in steep slope areas and sparsely in gentle slope areas; the vertical projection area of the vegetation canopy in the grid is measured in the field; the error correction factor is generated by comparing the initial projection area with the measured value; the convolutional neural network model parameters are iteratively optimized according to the error correction factor; the classification and morphological optimization are performed again using the optimized convolutional neural network model to obtain the calibrated vegetation vertical projection area; the effective coverage of the ecological function C = (the sum of the calibrated projection areas of the mangrove and the salt marsh herb) / S 总 ) x 100%.

[0018] Preferably, when there is an artificial dam or road in the step 2 of dynamically delineating the intertidal zone range of the present application, the land side boundary adopts the outer edge line of the dam top; when there is no artificial structure, the land side boundary adopts the line connecting the measured position points of the mean low water line.

[0019] Preferably, the step 3 of the present application specifically comprises: extracting near-infrared band reflectivity data and red light band reflectivity data from the satellite image; based on the near-infrared band reflectivity data and the red light band reflectivity data, calculating the normalized vegetation index band according to the formula: normalized vegetation index = (near-infrared band reflectivity - red light band reflectivity) / (near-infrared band reflectivity + red light band reflectivity); combining the normalized vegetation index band, the near-infrared band, and the red light band into three-channel input data; inputting the three-channel input data into the DeepLabV3+ convolutional neural network model; the model learns the vegetation features through training data, and the training data includes red mangrove flowering period images, salt marsh herb withering period images, and Spartina alterniflora invasion area images; the model outputs the classification probability map of the mangrove, salt marsh herb, and Spartina alterniflora; and the classification probability map is converted into the vegetation type classification result by threshold segmentation.

[0020] Preferably, the morphological optimization in the step 4 of the present application performs the following operations in sequence: performing an opening operation on the classification results of the mangrove, salt marsh herb, and Spartina alterniflora respectively, and the opening operation uses a 3x3 pixel rectangular structure element to eliminate isolated noise points; performing a closing operation on the classification results processed by the opening operation, and the closing operation uses a 5x5 pixel rectangular structure element to fill the internal holes of the vegetation patches; performing sliding window smoothing on the classification results processed by the closing operation, and the sliding window size is 7x7 pixels, and the pixels with a vegetation type proportion of more than 60% in the window are uniformly assigned to the type.

[0021] Preferably, the slope distribution acquisition in step 5 of the present application comprises: generating a digital elevation model through aerial images taken by a drone; generating a slope grid map by calculating slope values pixel by pixel based on the digital elevation model; marking the area with a slope value of 0-2° in the slope grid map as a gentle slope area; marking the area with a slope value of 2-5° in the slope grid map as a transition area; marking the area with a slope value > 5° in the slope grid map as a steep slope area; and outputting a spatial distribution map with slope partition markers for guiding the layout of the sampling grid.

[0022] Preferably, the error correction factor calculation in step 5 of the present application comprises: setting a 1m x 1m quadrat measured vegetation canopy vertical projection area within the sampling grid; extracting the initial vertical projection area calculated in step 4 at the same location and range; independently calculating the correction factors according to the vegetation types: mangrove error correction factor = mangrove measured vertical projection area / mangrove initial vertical projection area; salt marsh herb error correction factor = salt marsh herb measured vertical projection area / salt marsh herb initial vertical projection area; and Spartina alterniflora error correction factor = Spartina alterniflora measured vertical projection area / Spartina alterniflora initial vertical projection area; and independently feeding the three types of correction factors back to the convolutional neural network model in step 3 for parameter optimization.

[0023] Preferably, the present application further comprises step 7: acquiring the ecological function effective coverage C calculated in step 6; calling the biological beach vegetation coverage threshold in the specification; when C ≥ the biological beach vegetation coverage threshold, generating a standard certification report, which contains the coverage value and the passing certification conclusion; and when C < the biological beach vegetation coverage threshold, performing the following operations: extracting the area where the combined coverage of mangroves and salt marsh herbs is lower than the biological beach vegetation coverage threshold from the vegetation distribution map output in step 6; converting the low-coverage area boundary into a vector polygon layer; and generating a repair recommendation report, which contains the coverage value, the non-compliance conclusion, and the geographic location identification of the attached low-coverage area vector polygon.

[0024] Preferably, when setting the 1m x 1m quadrat measured vegetation canopy vertical projection area in step 5 of the present application: when there are mixed vegetation types within the quadrat, a high-definition near-ground photograph is taken to obtain a bird's eye view of the quadrat; the bird's eye view is divided into homogeneous vegetation patches through a superpixel segmentation algorithm; the vegetation types of the patches are manually labeled and the projection area proportion of each type is calculated; and the total measured area is allocated to mangroves, salt marsh herbs, and Spartina alterniflora according to the proportions.

[0025] Preferably, the superpixel segmentation algorithm of the present application performs the following sequential operations: performing a simple linear iterative clustering algorithm to generate an initial superpixel segmentation result; converting the initial superpixel segmentation result to HSV color space; performing edge detection on the hue component of the HSV color space to optimize superpixel boundaries; calculating spectral similarity of adjacent superpixels; merging adjacent superpixels with spectral similarity exceeding 95% to form homogeneous vegetation patches; outputting a vector boundary of the homogeneous vegetation patches containing patch ID and geographic coordinates.

[0026] Preferably, the artificial labeling process of the present application uses a mobile augmented reality assistant tool to perform the following sequential operations: loading a pre-stored typical vegetation micro-morphological feature library, the feature library containing a mangrove leaf vein fractal dimension dataset, a salt marsh herb stem internode length dataset, and a Spartina alterniflora inflorescence spikelet arrangement pattern dataset; capturing real-time high-definition images of the vegetation surface within the sample plot through the mobile camera; performing local feature extraction on the real-time images, extracting the mangrove leaf vein fractal dimension, the salt marsh herb stem internode length, and the Spartina alterniflora inflorescence spikelet arrangement pattern; matching the extracted features with the feature library: labeling as mangrove when the mangrove leaf vein fractal dimension matching error is ≤5%; labeling as salt marsh herb when the salt marsh herb stem internode length matching error is ≤3mm; labeling as Spartina alterniflora when the Spartina alterniflora inflorescence spikelet arrangement pattern matching coincidence degree is ≥90%; automatically associating the labeling results with the geographic coordinates of the sampling points, and returning the results to the central database through the mobile network to update the model training sample set.

[0027] The present application at least includes the following beneficial effects:

[0028] The present application realizes the consistency of coverage calculation and intertidal zone physical process by tidal dynamic boundary locking, vegetation classification calibration and terrain driven sampling, and meets the accuracy requirements of ecological restoration project acceptance. The differentiated boundary standard of artificial and natural shorelines eliminates the land side range deviation caused by dam siltation and trace ambiguity, and guarantees the objectivity of intertidal area statistics. Multi-temporal training and NDVI enhanced spectral separation suppresses the phenological misjudgment of flowering mangrove and withering herb, and improves the robustness of invasive species Spartina alterniflora identification. Cascade morphological operation eliminates classification noise, holes and edge sawtooth, generates continuous vegetation patches, and ensures that the projected area reflects the actual coverage state of the canopy. The slope partition driven sampling grid is adaptively laid out, the steep slope area is captured at high density, the gentle slope area is matched at low density, and the data representativeness is optimized. The correction factor is independently calculated for different types of vegetation, which blocks the error transmission of Spartina alterniflora to native vegetation, and realizes the targeted optimization of model parameters. The coverage result is automatically compared with the standard threshold, and the authentication conclusion and low coverage area vector positioning map are directly output, which connects the decision chain from monitoring to restoration. Near-ground photography and superpixel segmentation realize the non-destructive decoupling of vegetation area in mixed sample, which guarantees the accurate allocation of measured values according to physical proportion. HSV hue boundary optimization and spectral similarity merging suppress the interference of water body reflection, and reconstruct the homogeneous vegetation patches in accordance with the actual continuous distribution. The micro-morphological feature library matching overcomes the influence of environmental variation, improves the accuracy of field vegetation labeling, and supports high-reliability error correction.

[0029] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A mangrove picture of a mangrove restoration area;

[0031] Figure 2 A Spartina alterniflora picture of a mangrove restoration area;

[0032] Figure 3 A collected image of a mangrove restoration area; the left is a satellite image of the area, the picture in the upper right corner of the right is a UAV aerial image, and the picture in the lower right corner of the right is a downward-looking UAV aerial image;

[0033] Figure 4 A mudflat distribution map;

[0034] Figure 5 A flow framework diagram of the intertidal zone vegetation coverage dynamic monitoring method of the ecological restoration shoreline of the present application. DETAILED DESCRIPTION

[0035] The present application will be further described in detail below with reference to examples, so that those skilled in the art can implement it according to the description.

[0036] It should be understood that the terms such as "have", "contain" and "include" used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0037] As shown in the method for dynamically monitoring vegetation coverage in the intertidal zone of an ecological restoration coastline according to the present application, the method comprises the following steps: Figure 5

[0038] Step 1: synchronously acquire satellite images and unmanned aerial vehicle aerial images of the target intertidal zone region, covering mangrove, salt marsh herb and Spartina alterniflora distribution areas;

[0039] Step 2: based on real-time tide table data, determine the tidal level state at the monitoring time; dynamically determine the intertidal zone range: at high tide, the land side boundary is the multi-year average high tide mark line or the vegetation growth outer edge line; at low tide, the sea side boundary is the average low tide line; generate an intertidal zone dynamic boundary vector layer, and calculate the total area S 总 .

[0040] Step 3: superimpose the images of step 1 and the boundary layer of step 2, input a convolutional neural network model, and output the classification results of mangrove, salt marsh herb and Spartina alterniflora;

[0041] Step 4: morphologically optimize the classification results of step 3, and calculate the initial vertical projection areas of mangrove, salt marsh herb and Spartina alterniflora, respectively;

[0042] Step 5: according to the slope distribution of the intertidal zone, densely arrange the sampling grid in steep slope areas and sparsely arrange the sampling grid in gentle slope areas; measure the vertical projection area of the vegetation canopy in the grid in the field; compare the initial projection area of step 4 with the measured value to generate an error correction factor; and iteratively optimize the convolutional neural network model parameters of step 3 according to the error correction factor;

[0043] Step 6: use the optimized convolutional neural network model to re-execute steps 3-4 to obtain the calibrated vegetation vertical projection area; calculate the effective coverage of ecological function C = (the sum of the calibrated projection areas of mangrove and salt marsh herb / S 总 ) x 100%.

[0044] ​In the process of building the model, the convolutional neural network model adopts DeepLabV3+ architecture, the main network of which is Xception65, including an encoder (Encoder), an atrous spatial pyramid pooling (ASPP) module and a decoder (Decoder) module. The encoder gradually extracts features and reduces the spatial dimension through depthwise separable convolution and step convolution; the ASPP module captures multi-scale context information using different dilated convolutions; and the decoder fuses shallow feature maps to restore spatial details and upsample to the original resolution. The input layer is set to 256 pixels x 256 pixels x 3 channels, and the three-channel data are: normalized vegetation index (NDVI) band, near-infrared band (Band 8) and red band (Band 4) in turn. The output layer has 3 channels, respectively corresponding to the classification probabilities of mangrove, salt marsh herb and Spartina alterniflora, and is output through a Softmax activation function, ensuring that the sum of the probabilities of three categories for each pixel is 1.

[0045] The model training process is as follows: the training data is selected from historical multi-temporal satellite and unmanned aerial vehicle images, including samples of typical scenes such as mangrove flowering period, salt marsh herb growing and withering period (green and withered), and Spartina alterniflora invasion area, totaling 10,000 256x256 pixel slices. According to the ratio of 7:2:1, the training set, the validation set and the test set are randomly divided. During training, a weighted cross-entropy loss function is used to deal with the class imbalance problem, and the weight is inversely proportional to the frequency of each class in the training set. The optimizer uses Adam, the initial learning rate is set to 0.001, and the learning rate is reduced to 0.1 of the original every 10 training cycles (epochs). The batch size is set to 16, and the total training period is 100. To enhance the robustness of the model, data augmentation is used during training, including random horizontal flipping, vertical flipping, ±30° random rotation and ±20% brightness adjustment. All training is completed on a server equipped with an NVIDIA V100 GPU based on the TensorFlow 2.8 framework. The average intersection over union (mIoU) on the validation set is used as the main evaluation indicator of the model performance, and when the indicator no longer improves for 10 consecutive cycles, the training is terminated in advance and the optimal model weight is saved.

[0046] Specifically, for data collection and boundary demarcation

[0047] Satellite imagery was generated using 10-meter resolution Sentinel-2 data, while drone aerial photography employed 0.2-meter resolution DJI IPhantom 4 RTK imagery, covering mangrove, salt marsh herbaceous, and Spartina alterniflora distribution areas. Tidal data was sourced from the National Marine Information Center's real-time tide tables, with tide levels automatically acquired via API interface at the monitored times. At high tide, the landside boundary was selected as follows: if historical survey data existed, the multi-year average high tide trace line was used; otherwise, the outer edge of vegetation growth identified by the drone was used. At low tide, the seaside boundary was generated by connecting points (single-point spacing ≤ 20 meters) measured in-situ using GNSS-RTK. The boundary vector layer was generated in ArcGIS Pro, with a total area of ​​S... 总 Calculated automatically by software.

[0048] During operation, the drone flew along a preset route to take pictures at an altitude of 100 meters, with a forward overlap of 80% and a lateral overlap of 70%. The GNSS-RTK equipment was set up on the tidal flat and measured on foot along the land-water boundary at average low tide, with the point spacing automatically recorded as 20 meters. Tidal data was automatically called via an API through a Python script, synchronized with the image acquisition time.

[0049] Vegetation classification and optimization

[0050] Imagery and boundary layers were overlaid and cropped in QGIS to preserve data within the intertidal zone. The CNN model was built using TensorFlow 2.8 + DeepLabV3 + a pre-trained model (public code repository). Input data: 256×256 pixel slices in three channels (near-infrared / red / NDVI). Training data: included mangrove flowering period, salt marsh herbaceous vegetation growth and decay period, and Spartina alterniflora invasion area. The input data selected a combination of NDVI, near-infrared, and red bands because NDVI effectively enhances the distinction between vegetation and non-vegetation features such as bare beaches and water bodies; the near-infrared band is sensitive to chlorophyll content and cell structure, effectively reflecting vegetation biomass; and the red band helps identify vegetation growth and decay status. This combination of three bands fully utilizes the spectral differences between different phenological stages and vegetation types, significantly improving the spectral separation ability of mangroves, salt marsh herbs, and Spartina alterniflora in the convolutional neural network model. This is an optimized input setting for monitoring complex intertidal vegetation cover scenarios.

[0051] When optimizing sequences morphologically, the opening operation uses a 3×3 rectangular structuring element to eliminate single-pixel noise. The closing operation uses a 5×5 rectangular structuring element with a padding of ≤25 pixels. 2 Holes (square pixels, same below). A smooth sliding window uses a 7×7 window; reclassification occurs when the dominant vegetation within the window accounts for ≥60%.

[0052] NDVI calculation, near-infrared band (Band 8), red band (Band 4) reflectance into the formula. Classification threshold: mangrove probability ≥ 0.7, salt marsh herb ≥ 0.6, mutual grass ≥ 0.8.

[0053] Error correction and coverage calculation

[0054] For slope zoning: Digital Elevation Model (DEM) is generated by Pix4Dmapper from UAV images. Slope calculation is through ArcGIS Slope tool, neighborhood analysis window 3x3 pixels. Zoning threshold: 0-2° gentle slope, 2-5° transition, >5° steep slope.

[0055] Sampling grid layout: gentle slope area grid 200m x 200m, steep slope area 50m x 50m, transition area 100m x 100m. Each grid center is set to 1m x 1m quadrat (PVC frame demarcation).

[0056] The actual projection area uses a camera focal length of 35mm, 1.5m from the ground, and a vertical view of the sample plan view. The area is calculated by manually outlining the crown layer projection profile using ImageJ software.

[0057] Error correction factor calculation: mangrove correction factor = actual mangrove projection area / initial mangrove projection area (same location). Salt marsh herb correction factor and mutual grass correction factor are independently calculated. Among them, the convolutional neural network model parameters according to the error correction factor iterative optimization step 3 are specifically realized by fine-tuning: load the trained DeepLabV3+ model weight, freeze the parameters of the encoder and ASPP module, and only update the parameters of the decoder part. In the optimization process, the loss function is adjusted to the weighted cross-entropy loss of each class, and the weight is dynamically scaled by the corresponding error correction factor (Km, Ks, Ki), that is, the mangrove loss term is multiplied by Km, the salt marsh herb loss term is multiplied by Ks, and the mutual grass loss term is multiplied by Ki. Fine-tune using the Adam optimizer, learning rate set to 0.0001, batch size 8, training period 20. Through this type of weighted fine-tuning strategy, the model can specifically correct the classification bias on specific vegetation types and achieve parameter optimization.

[0058] Coverage calculation: calibrated projection area = optimized model recalculated projection area x corresponding correction factor. Ecological function effective coverage C = (∑mangrove + salt marsh herb calibration area) / S 总 x 100%.

[0059] In the existing vegetation coverage monitoring methods such as HY / T 254-2018 Beach Quality Evaluation and Classification, a fixed elevation boundary (such as the mean high water spring line) is used, ignoring the dynamic changes of tides; the maximum likelihood method is used for vegetation classification, only inputting RGB images; a 100m x 100m sampling grid is uniformly laid out; coverage = vegetation projection area / sample area x 100%. Among them, the fixed boundary leads to the omission of low-tide beach, the misjudgment of phenological change in single-phase classification, the neglect of terrain variation in uniform sampling, and the isolation of Spartina alterniflora interference. The present embodiment adopts a tide-driven dynamic boundary, multi-phase training and NDVI enhancement, slope-driven adaptive grid, type correction factor independent feedback, through dynamic boundary demarcation, multi-phase vegetation classification and terrain-driven sampling, to realize the consistency of intertidal zone vegetation coverage monitoring results and field ecological distribution, and to meet the acceptance accuracy requirements of coastal ecological restoration projects.

[0060] Further, in another embodiment, when there is an artificial dam or road in step 2 of dynamically demarcating the intertidal zone range, the land side boundary uses the outer edge line of the dam top; when there is no artificial structure, the land side boundary uses the trace line of the distribution of coastal vegetation debris; and the sea side boundary uses the line generated by connecting the measured position points of the mean low tide line.

[0061] Specifically, when there is an artificial dam or road in the intertidal zone area, the land side boundary uses the outer edge line of the dam top as the demarcation reference. In specific implementation, the outer edge line of the dam top can be generated by collecting coordinate points on site with GNSS-RTK measurement equipment. The coordinate point collection interval can be selected as 10 meters, 20 meters or 30 meters, and the preferred interval is 20 meters to ensure the smoothness of the boundary. During measurement, the RTK mobile station is fixed on the top of the measurement rod, and the equipment automatically records the point coordinates (planar accuracy ± 2 cm) by continuously walking along the outer edge of the dam top. The collected point data is imported into a geographic information system software, and a vector boundary line is generated by the "point to line" tool. The vector line is used as the land side dynamic boundary of the intertidal zone, and is updated synchronously with the tide data.

[0062] When there is no artificial structure in the natural coast, the land side boundary is demarcated using the trace line of the distribution of coastal vegetation debris. In implementation, a high-definition image of the coastline (resolution ≤ 5 cm) is obtained by aerial photography using a drone equipped with a visible light lens. The aerial photography height is set to 80 meters, 100 meters or 120 meters, and the preferred height is 100 meters to balance the coverage range and detail resolution. After the image is spliced by Agisoft Metashape software, the vegetation debris distribution area is extracted by visual interpretation or threshold segmentation, with color characteristics: RGB values 170-190, 150-170, 120-140. The outer edge line of the debris is drawn by the ROI tool of ENVI software, and is exported as a vector layer. This layer is superimposed with real-time tide data, and is used as the land side boundary at high tide.

[0063] The low tide line is measured by connecting the points along the coastline. The measurement is carried out at low tide, using a GNSS-RTK device (such as Leica GS18) along the water-land boundary. The distance between points can be selected as 15 meters, 20 meters or 25 meters, preferably 20 meters to match the complexity of the terrain. The device is set up on a tripod, preferably at a height of 1.5 meters to reduce multipath errors, and the single-point collection time is ≥ 30 seconds. The collected points are connected by the "point line" tool in QGIS software to generate a continuous vector line, which is corrected by Gauss-Kruger projection (coordinate system CGCS2000) and used as the dynamic boundary of the sea side.

[0064] As in the prior art such as "HY / T 254-2018 Beach Quality Evaluation and Classification", the fixed elevation line such as the mean high tide line is used to demarcate the boundary, without distinguishing between artificial and natural shorelines. The present embodiment uses differentiated demarcation rules: the outer edge line of the dike top is used for artificial shorelines to avoid boundary shrinkage caused by dike foot siltation; the vegetation debris trace line is used for natural shorelines to solve the subjectivity of visual sketching; and the low tide line point line is used for the sea side to eliminate the deviation between the theoretical mean low tide line and the actual terrain. The above dynamic boundary demarcation method reduces the statistical error of the intertidal zone range from ≥ 10% in the prior art to ≤ 3%, without increasing the frequency of artificial reconnaissance.

[0065] Further, in another embodiment, step 3 specifically comprises: extracting near-infrared band reflectance data and red light band reflectance data from satellite images; based on the near-infrared band reflectance data and the red light band reflectance data, calculating the normalized vegetation index band according to the formula: normalized vegetation index = (near-infrared band reflectance - red light band reflectance) / (near-infrared band reflectance + red light band reflectance); combining the normalized vegetation index band, the near-infrared band and the red light band into three-channel input data; inputting the three-channel input data into a DeepLabV3+ convolutional neural network model; the model learns vegetation features through training data, and the training data includes mangrove flowering image, salt marsh herbaceous senescence image and Spartina alterniflora invasion area image; the model outputs classification probability maps of mangrove, salt marsh herbaceous and Spartina alterniflora; and threshold segmentation is used to convert the classification probability maps into vegetation type classification results.

[0066] Specifically, the reflectance data of the near-infrared band (wavelength 842 nm ± 10 nm) and the red light band (wavelength 665 nm ± 10 nm) are extracted from the satellite image. In the implementation, Sentinel-2 L2A level data (spatial resolution 10 meters) is selected, and the Band 8 (near-infrared) reflectance value Band 8_ref and the Band 4 (red light) reflectance value Band 4_ref are extracted through the band operation tool of ENVI software. The reflectance value range is limited to the typical vegetation interval of 0.1-0.3. The normalized difference vegetation index (NDVI) is calculated according to the formula: NDVI = (Band8_ref - Band4_ref) / (Band8_ref + Band4_ref). After calculation, the NDVI band (value range -1.0 to 1.0) is generated, and the NDVI value of the vegetation coverage area is >0.3.

[0067] The NDVI band, the near-infrared band (Band 8), and the red light band (Band 4) are combined into three-channel input data. The implementation steps are as follows: read the three-band data in the Python environment using the GDAL library; stack according to the channel order (NDVI→Band8→Band4) through the stack() function of NumPy; crop the stacked data to 256×256 pixel slices with an overlap rate of 30%. Build a DeepLabV3+ convolutional neural network model: load the pre-trained weights based on the TensorFlow framework (the backbone network is Xception65), and set the input layer to 256×256×3. The training data includes three typical scenarios: mangrove flowering period, salt marsh herbaceous growth period, and Spartina alterniflora invasion area.

[0068] The model outputs the classification probability map of mangrove, salt marsh herb, and Spartina alterniflora (resolution 10 meters / pixel). The implementation process is as follows: use tf.keras to load the trained DeepLabV3+ model; input three-channel slices and output three-class probability values (value range 0.0-1.0) for each pixel; convert the probability map through threshold segmentation: mark as mangrove when the mangrove probability is ≥0.7; mark as salt marsh herb when the salt marsh herb probability is ≥0.6; mark as Spartina alterniflora when the Spartina alterniflora probability is ≥0.8; implement threshold segmentation in ArcGIS Pro by using the raster calculator tool (Con() function) to generate a vegetation type classification raster map.

[0069] The existing technology of "HY / T 254-2018" uses RGB images and maximum likelihood method for classification, which has two major defects: insufficient spectral information, only visible light bands (R / G / B) are used, which cannot distinguish between the near-infrared high reflectance of mangrove and the near-infrared reflectance of Spartina alterniflora; poor phenology adaptability, single-phase training leads to misjudgment of flowering mangrove as salt marsh herb (spike high reflectance similar to herb characteristics).

[0070] This implementation method enhances vegetation characteristics through NDVI, improving the spectral separation between vegetation and bare beach; NDVI > 0.3 indicates effective vegetation. Training data covering key phenological stages such as flowering and withering periods are used to suppress phenological misclassification. Differentiated thresholds are set for different vegetation types: 0.7 for mangroves, 0.6 for herbaceous plants, and 0.8 for Spartina alterniflora, reducing the misidentification rate of invasive species. This method can improve the classification accuracy of mixed vegetation areas to over 90%, and reduce the misclassification rate of Spartina alterniflora to < 5%.

[0071] Furthermore, in another implementation, the morphological optimization in step 4 performs the following operations in sequence: opening operations are performed on the classification results of mangroves, salt marsh herbs, and Spartina alterniflora, respectively, and the opening operation uses a 3×3 pixel rectangular structural element to eliminate isolated noise; closing operations are performed on the classification results after the opening operation, and the closing operation uses a 5×5 pixel rectangular structural element to fill the internal holes of the vegetation patches; sliding window smoothing is performed on the classification results after the closing operation, and the sliding window size is 7×7 pixels, and pixels in the window with a vegetation type accounting for more than 60% are uniformly assigned the type.

[0072] Specifically, the opening operation was performed on the classification results of mangroves, salt marsh herbs, and Spartina alterniflora to eliminate isolated noise. Implementation steps: In ENVI software, a morphological tool such as the Morphological Operations module was called; the Opening operation was selected, and the structuring element was set to a 3×3 pixel rectangle with a fixed size; the three types of vegetation raster data were processed sequentially: mangrove classification raster (input file mangrove.tif); salt marsh herb classification raster (saltmarsh.tif); and Spartina alterniflora classification raster (spartina.tif); the opening operation eliminated areas ≤1 pixel. 2 Isolated pixels (square pixels, the same below) are like scattered bare beach misclassifications in salt marsh grass.

[0073] Perform a closing operation on the classification results after the opening operation to fill the internal pores of the vegetation patches. Implementation steps: Select the Closing operation in the same ENVI module; set the structuring element to a 5×5 pixel rectangle with a fixed size; process the three types of vegetation data sequentially: fill the mangrove canopy interior ≤25 pixels. 2 Small holes (square pixels, the same below), such as vegetation gaps caused by bird nests; repairing holes formed by tidal erosion in salt marsh herbaceous patches; closing internal gaps caused by image shadows in dense areas of Spartina alterniflora.

[0074] The classification results after the closing operation are executed with a sliding window smoothing to optimize the jagged edges of the patches. Implementation steps: In ArcGIS Pro, use the focal statistics tool such as Focal Statistics; set a rectangular neighborhood window of 7*7 pixels with a total of 49 pixels; calculate the proportion of the main vegetation type in the window: if the proportion of a certain type is greater than or equal to 60%, reclassify the center pixel of the window as the type, such as mangrove pixels in the mangrove window being greater than or equal to 30; if the proportion of all types is less than 60%, keep the original classification of the center pixel; output the smoothed vegetation classification raster to eliminate the pepper and salt noise of mixed pixels in the transition area.

[0075] There are three defects in using only a single morphological filter such as median filtering: poor parameter universality, fixed size structure element cannot simultaneously process noise (small size required) and holes (large size required); edge jaggedness remains, no patch continuity optimization step is designed, and the vegetation boundary in the steep slope area is ladder-shaped; over-smoothing, uniform filtering leads to the deletion of small vegetation patches such as <10 pixels 2 (squared pixels, the same below) are deleted. The present embodiment solves the above problems through cascading operations: open and close operations are separated, 3*3 open operation accurately eliminates noise and retains ≥4 pixel 2 (squared pixels, the same below) small patches, 5*5 closing operation focuses on filling holes and retains the original boundary shape; adaptive smoothing, 7*7 window combined with 60% proportion threshold, only modifies classification ambiguous pixels such as mangrove-salt marsh herb boundary, avoids the loss of details caused by overall smoothing. This method makes the vegetation patch continuity index (PCI) improve to more than 0.92, and the area calculation error is less than or equal to 3%.

[0076] Further, in another embodiment, the slope distribution acquisition in step 5 includes: generating a digital elevation model through unmanned aerial vehicle aerial image; calculating the slope value of each pixel based on the digital elevation model to generate a slope raster map; marking the area with a slope value of 0-2° in the slope raster map as a gentle slope area; marking the area with a slope value of 2-5° in the slope raster map as a transition area; marking the area with a slope value >5° in the slope raster map as a steep slope area; outputting a spatial distribution map with slope partition markers for guiding the layout of the sampling grid.

[0077] Specifically, a digital elevation model (DEM) is generated by aerial images of the unmanned aerial vehicle. The implementation steps are as follows: use DJI unmanned aerial vehicle to collect images, set the flight height to 80 meters, 100 meters or 120 meters, preferably 100 meters, the heading overlap rate is ≥80%, and the lateral overlap rate is ≥70%; import the image into three-dimensional reconstruction software such as Pix4Dmapper, set the ground control point (GCP) coordinates (planar accuracy ±3cm, elevation accuracy ±5cm); perform dense point cloud reconstruction, point density ≥50 points / square meter, generate initial DEM (resolution 0.1m); optimize DEM by Poisson surface reconstruction algorithm, eliminate vegetation canopy interference, output bare land DEM, file format GeoTIFF.

[0078] Based on DEM, the slope value is calculated pixel by pixel and the slope zoning map is generated. The implementation steps are as follows: in ArcGIS Pro, call the slope tool Slope, set the neighborhood analysis window to 3x3 pixels, fixed size; the calculation rule adopts the second-order difference algorithm such as Zevenbergen&Thorne method, output the slope grid, value range 0°-90°, accuracy 0.1°; the slope zoning threshold is:

[0079] 0°-2.0° is gentle slope area, marked as green; 2.1°-5.0° is transition area, marked as yellow; 5.0° is steep slope area, marked as red; use the reclassification tool such as Reclassify to convert the slope grid to the zoning grid, output the spatial distribution map with color label (EPSG: 4326 coordinate system).

[0080] The slope zoning map guides the adaptive layout of the sampling grid. The implementation rules are as follows: the steep slope area >5.0° is laid out with a 50m x 50m grid sampling point, such as the edge of the tidal ditch; the transition area 2.1°-5.0° is laid out with a 100m x 100m grid, such as the edge of the tidal channel; the gentle slope area 0°-2.0° is laid out with a 200m x 200m grid, such as flat mudflat; load the slope zoning map in QGIS, generate a vector sampling grid layer through the "create grid" tool such as Create Grid, and the grid center point is the actual sampling location.

[0081] The uniform grid sampling method has the following defects: (1) poor terrain adaptability, overlarge grid in steep slope area > 5° leads to missing micro-topographic mutations such as tidal ditch boundaries, and over-dense grid in gentle slope area leads to resource waste; (2) projection area deviation, when the slope > 5°, the vertical projection area calculation error is large. The present embodiment can solve the above problems by the following methods: (1) high-precision reconstruction of DEM, 0.1-meter resolution bare land DEM accurately reflects micro-topography; (2) dynamic zoning of slope, 0°-2° for gentle slope, 2.1°-5° for transition, > 5° for steep slope, three threshold values match the intertidal zone geomorphic features; (3) grid density grading, the grid density in steep slope area is multiple times of that in gentle slope area, and the terrain mutations are accurately captured. The present method can significantly reduce the projection area calculation deviation and improve the sampling efficiency.

[0082] Further, in another embodiment, the error correction factor calculation in step 5 includes: setting a 1-meter x 1-meter quadrat measured vegetation canopy vertical projection area in the sampling grid; extracting the initial vertical projection area calculated in step 4 at the same location and range; independently calculating the correction factor according to the vegetation type: mangrove error correction factor = mangrove measured vertical projection area / mangrove initial vertical projection area; salt marsh herb error correction factor = salt marsh herb measured vertical projection area / salt marsh herb initial vertical projection area; Spartina alterniflora error correction factor = Spartina alterniflora measured vertical projection area / Spartina alterniflora initial vertical projection area; independently feeding the three types of correction factors to the convolutional neural network model in step 3 for parameter optimization.

[0083] Specifically, a 1-meter x 1-meter quadrat measured vegetation canopy vertical projection area is set in the sampling grid. The implementation steps are as follows: using a PVC frame with an outer dimension of 1.0 meter x 1.0 meter and a wall thickness of 5 millimeters to fix the quadrat range; using a single-lens reflex camera to vertically set at a height of 1.5 meters above the quadrat with a focal length of 35 mm; taking a high-definition overhead view with a resolution of 7952 x 5304 pixels and saving it as a RAW format; manually outlining the vegetation canopy contour using ImageJ software, calculating the projection area with an accuracy of ± 0.01 square meters; extracting the projection area at the same location from the initial vegetation classification map generated in step 4, which can use the GIS spatial query tool Extract by Mask. Independently calculating the error correction factor according to the vegetation type, the mangrove correction factor is: wherein A m_field is the measured mangrove area, A m_init is the initial projection area of the mangrove at the same location; the salt marsh herb correction factor is: wherein A s_field is the measured salt marsh herb area, A m_init is the initial projection area of the salt marsh herb at the same location; the Spartina alterniflora correction factor is: wherein A i_field is the measured Spartina alterniflora area, A m_initis the initial projected area of Spartina alterniflora, and the correction factor of all sample plots of the same vegetation type is taken as the arithmetic mean. The correction factor is fed back to the convolutional neural network model to optimize the parameters, and the original DeepLabV3+ model is loaded in the TensorFlow framework. According to the vegetation type, the output layer is separated, the mangrove output layer weight Wm, the salt marsh herb output layer weight Ws, and the Spartina alterniflora output layer weight Wi are scaled by the correction factor, and the loss function is scaled by the correction factor. The mangrove loss is , the herb loss is , the Spartina alterniflora loss is , and the total loss is Back propagation is used with the Adam optimizer and a learning rate of 0.001 to update the weights.

[0084] The overall correction factor , is the sum of the measured areas of all vegetation types, is the sum of the initial projected areas of all vegetation types, and there are serious defects: the overestimation error of Spartina alterniflora is transmitted to the mangrove / herb, resulting in the contamination of the area of the original vegetation; the uniform adjustment of the model weight cannot repair the specific vegetation classification layer.

[0085] The present embodiment blocks the transmission of Spartina alterniflora error to the mangrove / herb by calculating independently by type, and uses the loss function weighting to dynamically scale the loss of each category according to the correction factor, so that the high error category obtains a larger gradient update. The output layer separation optimization makes the mangrove weight Wm only affected by Km, and the Spartina alterniflora weight Wi only adjusted by Ki. This method makes the calculation error of the mangrove area controlled within ±3%, and the misjudgment rate of Spartina alterniflora significantly reduced.

[0086] Further, in another embodiment, it further includes step 7: obtaining the ecological function effective coverage C calculated in step 6; calling the existing specifications such as “DB33T 2368-2021 Coastal Restoration Repair Evaluation Technical Specification” and other biological beach vegetation coverage thresholds 30% (or other set values such as 20% or 40% and the like); when C≥30%, a standard authentication report is generated, which contains the coverage value and the passing authentication conclusion; when C<30%, the following operations are performed: extracting the area where the coverage of mangrove and salt marsh herb is less than 30% from the vegetation distribution map output in step 6; converting the low coverage area boundary to a vector polygon layer; generating a repair suggestion report, which contains the coverage value, the non-compliance conclusion, and the geographic location identification of the low coverage area vector polygon.

[0087] Specifically, the ecological function effective coverage C (unit: %) calculated in step 6 is compared with the biological beach vegetation coverage threshold value specified in the existing specification, such as 30%. Implementation steps: call the specification text in the Python script, and extract the threshold value using the PDF parsing library PyPDF2; logical judgment: if C≥30%, execute the standard certification report generation process; if C<30%, execute the repair recommendation report generation process.

[0088] When C≥30%, a standard certification report is generated, and the report content includes: title "Tidal Zone Vegetation Ecological Function Compliance Certification Report"; main text: coverage value C (retaining one decimal place, such as "32.7%"), monitoring time, and coordinate range; conclusion page: "After monitoring, the ecological function effective coverage of the target area is C%, which meets the specification requirements".

[0089] When C<30%, a repair recommendation report is generated, and when extracting low coverage areas, load the vegetation distribution map output in step 6 in ArcGIS Pro; use the raster calculator Con("mangrove + saltmarsh"<30, 1, 0) to generate a binary mask, with 1 representing low coverage areas; output a vector polygon layer through a raster-to-face tool such as Raster to Polygon, with the field name Low_Coverage.

[0090] The existing technology such as "HY / T 254-2018 Beach Quality Evaluation and Classification" uses manual judgment of coverage and relies on on-site evaluation reports, without an automatic threshold comparison mechanism; there is a disconnection between the report and the spatial data, and only the coverage percentage is described in the conclusion, without associating the spatial location of low coverage areas. The present embodiment automatically calls the specification threshold, directly reads the industry specification threshold of 30% through the PDF parsing library, eliminating manual table lookup errors; the logical judgment is integrated into the processing flow, enabling automatic discrimination of coverage compliance. Spatial targeted repair, with low coverage area vector polygons accurate to 10-meter grid units. This method significantly improves the efficiency of developing repair plans, with a positioning error of low coverage areas ≤5 meters.

[0091] Further, in another embodiment, when setting the 1-meter x 1-meter sample plot to measure the vertical projection area of the vegetation canopy in step 5: when there are mixed vegetation types within the sample plot, use high-definition near-ground photography to obtain a bird's eye view of the sample plot; divide the bird's eye view into homogeneous vegetation patches through a superpixel segmentation algorithm; manually label the vegetation type of each patch and calculate the projection area proportion of each type; allocate the total measured area to mangroves, salt marsh herbs, and Spartina alterniflora according to the proportion.

[0092] Specifically, when mixed vegetation types exist within a 1 m x 1 m plot, a high-definition near-ground camera is used to obtain an overhead view of the plot. The camera has a resolution of 7952 x 5304 and is equipped with a 35mm fixed-focus lens. It is vertically fixed above the plot by a tripod, with the center of the lens being 1.5 meters above the ground, and an error of ±0.05 meters. The light intensity requirement is ≥50000 Lux, and a supplementary light is needed on cloudy days. During image acquisition, RAW format original images are taken, EXIF geographic coordinates are retained, the focus mode is set to manual, the aperture is f / 8.0, the ISO is 100, and the shutter speed is 1 / 500 seconds. Three images are taken for each plot, at the center point, northeast 30°, and northwest 30°, respectively, for three-dimensional correction.

[0093] The overhead view is divided into homogeneous vegetation patches by a superpixel segmentation algorithm. In the initial segmentation, the SLIC (Simple Linear Iterative Clustering) algorithm is executed in Python using the skimage library, with the parameters set to cluster number = 100, compactness factor = 20, and maximum iteration number = 10. The initial superpixel segmentation map is output, with each superpixel ≈100 cm 2 . In the boundary optimization, the HSV color space is converted, and the hue component H is extracted, with a value range of 0-360°. The Canny edge detection is used, with a low threshold of 50 and a high threshold of 150, to optimize the superpixel boundary and merge adjacent superpixels with a spectral similarity >95% (Euclidean distance <5 in Lab space). The vector output generates a homogeneous vegetation patch vector boundary (GeoJSON format), with an attribute table containing patch ID, area (m 2 ), and Lab mean (L / a / b).

[0094] The patch vegetation type is manually labeled and the measured total area is assigned. The labeling rule is to load the patch vector into a tablet computer, such as an iPad Pro + Apple Pencil, and visually determine the texture characteristics of the patch: mangrove leaves are leathery and alternate; salt marsh herbs are linear leaf veins of grasses; and Spartina alterniflora has dense small spikelets in a conical inflorescence. The area allocation calculation is set as the measured total area of the plot A total =1.0m 2 ; the area ratio of each patch type is calculated: the mangrove ratio , the salt marsh herb ratio , and the Spartina alterniflora ratio ; and the measured area is assigned: the measured area of mangrove , the measured area of salt marsh herb , and the measured area of Spartina alterniflora .

[0095] The prior art such as LY / T 1812-2021 adopts manual estimation and distribution, visually judges the proportion of each type of vegetation in the mixed sample plot, has large errors, adopts destructive sampling, cuts and weighs the vegetation by type, destroys the ecological structure of the sample plot, and has no spatial positioning, loses patch position information, and cannot be associated with remote sensing pixels. The present embodiment adopts high-definition photography to retain the in-situ state of the vegetation, while the prior art needs physical cutting, and the super-pixel segmentation precision reaches 1 cm 2 which is higher than the prior art. The physical proportion precision distribution makes the patch area proportion calculation based on actual geometric segmentation, and the area distribution error is ≤3%. The spatial position binds the patch vector boundary with geographic coordinates (such as E119°15'23", N25°02'41"), which can be directly matched with remote sensing classification pixels, and supports error correction factor spatial tracing. This method improves the projection area distribution accuracy of the mixed sample plot to 97%, and shortens the single sample plot processing time.

[0096] Further, in another embodiment, the super-pixel segmentation algorithm performs the following sequential operations: performing a simple linear iterative clustering algorithm to generate an initial super-pixel segmentation result; converting the initial super-pixel segmentation result to an HSV color space; performing edge detection optimization on the hue component of the HSV color space to optimize the super-pixel boundary; calculating the spectral similarity of adjacent super-pixels; merging adjacent super-pixels with a spectral similarity exceeding 95% to form homogeneous vegetation patches; and outputting a homogeneous vegetation patch vector boundary containing patch ID and geographic coordinates.

[0097] Specifically, a simple linear iterative clustering (SLIC) algorithm is performed to generate an initial super-pixel segmentation result, and the parameters are set as

[0098] The input image is a sample plot overhead view (RGB format, resolution 7952x5304); the cluster number is set to 80, 100 or 120, preferably 100; the compactness factor (compactness) is selected as 15, 20 or 25, preferably 20; and the maximum number of iterations is fixed at 10 times. The algorithm is executed using the Python library skimage.segmentation.slic() function; the image is converted from the RGB space to the CIELAB color space, retaining the lightness L, color a / b channels; an initial super-pixel label map is generated, and the area of each super-pixel is approximately 1% of the sample plot area. The output is optimized by morphological closing operation (structure element 3x3 pixels) to fill small cracks, and saved as a GeoTIFF raster (spatial reference WGS84 coordinate system).

[0099] In HSV color space, the superpixel boundary is optimized. The initial segmentation map is converted to HSV space by calling the cvtColor() function of OpenCV. The hue component H is extracted, with a value range of 0-180°. Canny edge detection is performed on the H component, with a low threshold of 50 and a high threshold of 150. The detection results are superimposed on the initial superpixel boundary. Short edge segments with a length of ≤5 pixels are removed. Based on the optimized edge point set, Delaunay triangulation is used to reconstruct the continuous boundary, and the superpixel vector surface with topological relationship (Shapefile format) is output.

[0100] Merge spectrally similar superpixels and generate homogeneous patches. When calculating spectral similarity, the mean color difference of adjacent superpixels is calculated in CIELAB space , where L is the lightness (Lightness) color, L=0 is pure black, L=100 is pure white, and the value range is 0-100; a is the red-green axis (a-axis), +a is the red color, -a is the green color, and the value range is -128-127; b is the yellow-blue axis (b-axis), +b is the yellow color, -b is the blue color, and the value range is -128-127; for example, assuming that the color values of two vegetation regions are as follows: region 1: L1=60, a1=-20, b1=30 (dark green); region 2: L2=62, a2=-18, b2=28 (slightly lighter green); then the color difference ΔE = √[(60-62) 2 + (-20+18) 2 + (30-28) 2 ]= √[4 + 4 + 4] = √12 ≈ 3.46, since the color difference 3.46<5, it belongs to homogeneous vegetation, and can be merged into the same patch. The merging threshold ΔE≤5 is set, which is equivalent to the color difference that cannot be distinguished by the human eye. The patch merging rule is to traverse all adjacent superpixel pairs, and if ΔE≤5 and the area ratio ≤1.5, merge them into a homogeneous patch, and iterate until there are no superpixel pairs that meet the conditions. Vector output: generate the final homogeneous vegetation patch vector boundary, and the attribute table contains patch ID (unique code), center point coordinates (including latitude and longitude, precision 0.0001°), and spectral mean (L / a / b value, precision 0.1).

[0101] If the intertidal water body reflection causes the boundary to be blurred when directly using the RGB color space in the RGB space segmentation, the reflection area DE≥20, the fixed merging threshold, the uniform color difference threshold DE≤10, will over-merge the heterogeneous patches, and the output raster label map cannot support GIS spatial analysis without topological reconstruction. The embodiment optimizes the HSV anti-reflection, the hue component (H) is not sensitive to the water body mirror reflection, and the H value variation of the reflection area is ≤5°; Canny edge detection focuses on the real vegetation boundary and eliminates the reflection artifact. Precise control of spectral merging, strict color difference threshold DE≤5, ensures that only homogeneous vegetation is merged, and the area ratio limit ≤1.5 avoids unreasonable fusion of large patches. When topological vector output, Delaunay triangulation reconstruction ensures the geometric closure of the patch, and the attribute table contains spectral characteristics, supporting remote sensing data linkage analysis. The patch boundary error is ≤2 pixels (0.4 cm) from the actual coincidence, the spectral consistency of homogeneous patches is 98%, and the processing time of a single sample is short.

[0102] Further, in another embodiment, the human annotation process uses a mobile end augmented reality auxiliary tool to perform the following sequential operations: load a pre-stored typical vegetation micro-morphological feature library, the feature library includes a mangrove leaf vein fractal dimension dataset, a salt marsh herb stem internode length dataset, and a Spartina alterniflora inflorescence spikelet arrangement pattern dataset; real-time capture high-definition images of the vegetation surface in the sample area through the mobile end camera; perform local feature extraction on the real-time images, extract the mangrove leaf vein fractal dimension, the salt marsh herb stem internode length, and the Spartina alterniflora inflorescence spikelet arrangement pattern; match the extracted features with the feature library: when the mangrove leaf vein fractal dimension matching error is ≤5%, it is labeled as mangrove; when the salt marsh herb stem internode length matching error is ≤3 mm, it is labeled as salt marsh herb; when the matching coincidence degree of the Spartina alterniflora inflorescence spikelet arrangement pattern is ≥90%, it is labeled as Spartina alterniflora; the annotation result is automatically associated with the geographic coordinates of the sampling point, and is returned to the central database through the mobile network to update the model training sample set.

[0103] Specifically, the pre-stored typical vegetation micro-morphological feature library includes three types of data sets:

[0104] The mangrove leaf feature library collects mature Kandelia candel leaf samples, and obtains vein images through a 2400 dpi scanner; calculates the vein fractal dimension, with a value range of 1.25-1.35 (box counting method, box size 2-20 pixels); stores it as a floating point data set (mean 1.30, standard deviation 0.03).

[0105] The salt marsh herb stem feature library measures the internode length of Halocnemum strobilaceum stem, with a vernier caliper accuracy of ±0.1 mm; the length value range is 5.0-8.0 mm, stored according to normal distribution, with a mean of 6.5 mm and a standard deviation of 0.8 mm.

[0106] The inflorescence feature library microphotography collects the arrangement pattern of spikelets, generates a binary template image, and labels the geometric feature of the spiral angle of 120°±5°.

[0107] The field operation is performed by a mobile end augmented reality tool, and the image acquisition uses a tablet computer. A built-in LiDAR sensor camera is used to capture images at a vertical distance of 20-30 cm from the surface of the vegetation. High-definition images are automatically captured, with a resolution of 1024×768 pixels and a depth of field control of ≤5 cm.

[0108] Then, feature extraction is performed. The mangrove frame selects a single leaf blade area (≥2 cm 2 ), and the fractal dimension of the leaf vein (box counting method) is calculated. The marsh herb identifies the node position of the stem, measures the distance between adjacent nodes (pixel calibration scale: 1 mm:8.5 pixels), and divides the inflorescence area of the Spartina alterniflora. The arrangement pattern of the spikelets is matched (coverage calculation).

[0109] Finally, type determination is performed. For the mangrove, the error between the measured fractal dimension and the average value of the feature library is ≤5%, for example, the measured value is 1.28, the library value is 1.32, and the error is 3.0%. For the marsh herb, the deviation between the measured value and the library value of the internode length is ≤3 mm, for example, the measured value is 7.2 mm, the library value is 6.5 mm, and the deviation is 0.7 mm. For the Spartina alterniflora, the matching degree of the spikelet arrangement pattern is ≥90%, for example, the area matching degree is 92%.

[0110] The labeled results are bound in space and the database is updated. The geographical coordinates are associated with the device GNSS module to obtain the sampling point coordinates, such as WGS84 coordinates, with an accuracy of ±0.5 m. Attribute records are generated, including the type of vegetation, the characteristic value (such as the fractal dimension of the mangrove 1.28), the longitude and latitude (such as E119°15'23", N25°02'41"). The central database updates the data to the PostgreSQL database through the 4G / 5G network, and dynamically expands the training sample set: the mangrove adds a fractal dimension of 1.28 samples, and the marsh herb adds an internode length of 7.2 mm samples.

[0111] Currently, artificial visual determination of vegetation type usually has a misjudgment rate of ≥15%, paper record coordinates have a spatial error of usually ≥10 m, and the sample set update period is ≥1 year, which lags behind ecological changes. The micro-morphological quantification of the present embodiment is anti-interference, the fractal dimension resists leaf size variation, and the leaf dimension difference is ≤1% for 10-50 cm 2 leaf size variation. The stem internode length excludes the interference of withered leaves, and the spiral structure of the spikelet maintains the stability during the flowering period. Real-time data is returned from the field, the model is updated within 2 minutes, the coordinates are matched with the remote sensing pixels, and the position error is ≤0.5 m. Therefore, the classification accuracy of the mangrove during the flowering period can be significantly improved, the reliability of the identification of the marsh herb during the withered and yellow period is high, and the timeliness of the Spartina alterniflora invasion monitoring is improved to the hour level. Embodiment

[0112] As shown in Figure 1 and Figure 2 , a certain bay mangrove restoration area is selected as the biological beach verification object, and the vegetation types in this area include mangrove and Spartina alterniflora, etc. Figure 1 Mangrove, Figure 2 Spartina alterniflora.

[0113] When the boundary is drawn, the land side boundary is drawn according to the scenario without artificial structures, and the trace line of the coastal vegetation debris distribution is used. The sea side boundary is generated by connecting the measured points of the average low tide line, the point distance is 20 meters, GNSS-RTK measurement, the accuracy is ±0.5 meters. When the image is obtained, the satellite image uses Sentinel-2 data, the resolution is 10 meters, and it covers the mangrove and Spartina alterniflora area; the unmanned aerial vehicle image uses DJI Phantom 4 RTK aerial photography, the resolution is 0.1 meters, the flight height is 100 meters, and the overlap rate is 80%. As shown in Figure 3 , the left side is the satellite image of the area, the right upper corner picture on the right side is the unmanned aerial vehicle aerial image, and the right lower corner picture on the right side is the downward aerial image of the unmanned aerial vehicle.

[0114] When the vegetation classification and area calculation are performed, the input data of the classification: near-infrared band (Band 8), red light band (Band 4), and NDVI band; the model uses DeepLabV3+ convolutional neural network, and the output is the classification probability of mangrove and Spartina alterniflora; the classification threshold is mangrove ≥0.7 and Spartina alterniflora ≥0.8. Morphological optimization is performed, 3×3 pixel rectangle is used for opening operation to eliminate noise points; 5×5 pixel rectangle is used for closing operation to fill holes; 7×7 window is used for sliding window smoothing, and the proportion of main vegetation ≥60% is reclassified. Vertical projection area calculation is performed, the initial area of mangrove A m_init , and the initial area of Spartina alterniflora A i_init .

[0115] Field verification and coverage calculation are performed. The sampling grid is laid out according to the slope zoning: 200m×200m grid is laid out in the gentle slope area 0°-2°, and 50m×50m grid is laid out in the steep slope area >5°. Mixed quadrat processing is adopted for the mixed area of mangrove and Spartina alterniflora, 1m×1m quadrat high-definition aerial photography is adopted, superpixel segmentation is adopted to divide homogeneous patches, SLIC algorithm is adopted, and the cluster number is 100; patch type is manually labeled, and the measured total area is allocated according to the area proportion. When the vegetation coverage is calculated, the total coverage , where A m_cal , A i_cal are the calibrated projection areas; the single coverage of mangrove is . As shown in Figure 4The shown certain gulf beach distribution map, due to the underwater bank slope water depth is less than 2 meters in the gulf, the slope is particularly gentle, therefore the intertidal zone is wide, the planted mangrove is about 87 hectares, the total area of the gulf is 21.29 square kilometers, the area of the restored mangrove accounts for about 5% of the intertidal zone area.The gulf beach distribution has a large area of spartina alterniflora, the high beach outside the mangrove is almost occupied by spartina alterniflora, the total area proportion of the mangrove and spartina alterniflora in the intertidal zone is calculated, and the vegetation coverage of the gulf intertidal zone is calculated.

[0116] The threshold comparison is carried out, the vegetation coverage C of the gulf intertidal zone is greater than 30%, it is determined to reach the standard, the single coverage C of the mangrove is less than 30%, and it is marked that the original vegetation needs to be supplemented. total m The authentication conclusion is that the total coverage reaches the standard, the restoration suggestion is marked the low coverage area of the mangrove, and it is suggested that the native species kandelia candel seedlings are supplemented.The gulf beach area is wide, through mangrove restoration, the mangrove seedlings are introduced into the beach where the mangrove seedlings can grow, the mangrove wetland ecological system is reconstructed, a stable mangrove community and ecological system are formed, the community structure of benthic organisms is continuously optimized, and similar ecological functions to the original plants are provided.

[0117] Although the embodiments of the present application have been disclosed as above, it is not limited to the application listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present application. For those skilled in the art, other modifications can be easily realized.​

Claims

1. A method for dynamically monitoring vegetation coverage in the intertidal zone of an ecologically restored shoreline, characterized in that, The method comprises the following steps: Step 1, synchronously acquiring satellite images and unmanned aerial vehicle aerial images of a target intertidal zone region, covering mangrove, salt marsh herb and Spartina alterniflora distribution areas; Step 2, based on real-time tidal table data, determine the tidal level state at the monitoring time; dynamically delineate the intertidal zone range: when the high tide level, the land side boundary is the trace line of the mean spring tide or the outer edge line of the vegetation growth; when the low tide level, the sea side boundary is the mean low tide line; generate the dynamic boundary vector layer of the intertidal zone, and calculate the total area S 总 ; Step 3, superimposing the images of step 1 and the boundary layer of step 2, inputting a convolutional neural network model, and outputting classification results of the mangrove, salt marsh herb and Spartina alterniflora; Step 4, performing morphological optimization on the classification results of step 3, and calculating initial vertical projection areas of the mangrove, salt marsh herb and Spartina alterniflora respectively; Step 5, according to the slope distribution of the intertidal zone, the sampling grid is arranged densely in steep slope areas and sparsely in gentle slope areas; the vertical projection area of the vegetation canopy in the grid is measured in the field; the error correction factor is generated by comparing the initial vertical projection area of step 4 with the measured value; and the convolutional neural network model parameters of step 3 are iteratively optimized according to the error correction factor; Step 6: Re-execute steps 3-4 using the optimized convolutional neural network model to obtain the calibrated vertical projection area of vegetation; calculate the effective coverage of ecological function C = (the sum of the calibrated projection areas of mangrove and salt marsh herb / S 总 ) x 100%. Step 3 specifically comprises: extracting near-infrared reflectance data and red light reflectance data from the satellite images; based on the near-infrared reflectance data and the red light reflectance data, calculating the normalized vegetation index band according to the formula: normalized vegetation index = (near-infrared reflectance - red light reflectance) / (near-infrared reflectance + red light reflectance); combining the normalized vegetation index band, the near-infrared band and the red light band into three-channel input data; inputting the three-channel input data into a DeepLabV3+ convolutional neural network model; the model learns vegetation features through training data, and the training data includes red mangrove flowering image, salt marsh herb withering image and Spartina alterniflora invasion area image; the model outputs classification probability maps of mangrove, salt marsh herb and Spartina alterniflora; the classification probability map is converted into vegetation type classification results by threshold segmentation; In step 4, the morphological optimization performs the following operations in sequence: performing an opening operation on the classification results of the mangrove, salt marsh herb and Spartina alterniflora respectively, and the opening operation uses a 3*3 pixel rectangular structure element to eliminate isolated noise points; performing a closing operation on the classification results processed by the opening operation, and the closing operation uses a 5*5 pixel rectangular structure element to fill holes in the vegetation patch; performing sliding window smoothing on the classification results processed by the closing operation, and the sliding window size is 7*7 pixels, and the pixels in the window with a vegetation type proportion of more than 60% are uniformly assigned to the type; In step 5, the slope distribution acquisition comprises: generating a digital elevation model through unmanned aerial vehicle aerial images; generating a slope grid map by calculating the slope value of each pixel based on the digital elevation model; marking the area with a slope value of 0-2° in the slope grid map as a gentle slope area; marking the area with a slope value of 2-5° in the slope grid map as a transition area; marking the area with a slope value >5° in the slope grid map as a steep slope area; and outputting a spatial distribution map with slope partition markers for guiding the arrangement of the sampling grid.

2. The method of claim 1, wherein, In step 2, when there are artificial dikes or roads, the land side boundary adopts the outer edge line of the dike top; when there are no artificial structures, the land side boundary adopts the coastline vegetation debris distribution trace line; and the sea side boundary adopts the line connecting the measured position points of the mean low tide line.

3. The method of claim 1, wherein, Step 5, error correction factor calculation, includes: setting up 1m × 1m quadrats within the sampling grid and measuring the vertical projection area of ​​the vegetation canopy; extracting the initial vertical projection area calculated in Step 4 from the same location and range; and independently calculating the correction factor according to vegetation type. Mangrove error correction factor = measured vertical projection area of ​​mangrove / initial vertical projection area of ​​mangrove; Error correction factor for salt marsh herbaceous plants = Measured vertical projected area of ​​salt marsh herbaceous plants / Initial vertical projected area of ​​salt marsh herbaceous plants; Spartina alterniflora error correction factor = measured vertical projected area of ​​Spartina alterniflora / initial vertical projected area of ​​Spartina alterniflora; The three types of correction factors are independently fed back to the convolutional neural network model in step 3 for parameter optimization.

4. The method of claim 3, wherein, The process also includes step 7: obtaining the effective ecological function coverage C calculated in step 6; calling the biota vegetation coverage threshold in the standard; generating a standard compliance certification report when C ≥ biota vegetation coverage threshold, the report includes coverage values ​​and a certification conclusion; and performing the following operations when C < biota vegetation coverage threshold: extracting areas from the vegetation distribution map output in step 6 where the combined coverage of mangroves and salt marsh herbs is lower than the biota vegetation coverage threshold; converting the low-coverage area boundaries into a vector polygon layer; and generating a restoration recommendation report, which includes coverage values, a non-compliance conclusion, and a geographic location identifier for the low-coverage area vector polygon.

5. The method of claim 3, wherein, When setting the measured vertical projection area of ​​the vegetation canopy in a 1m×1m quadrat in step 5: when there are mixed vegetation types in the quadrat, high-definition near-ground photography is used to obtain the top view of the quadrat; the top view is divided into homogeneous vegetation patches by superpixel segmentation algorithm; The vegetation types of the patches were manually labeled and the proportion of the projected area of ​​each type was calculated; the total measured area was allocated to mangroves, salt marsh herbs and Spartina alterniflora according to the proportion.

6. The method of claim 5, wherein, The superpixel segmentation algorithm performs the following sequential operations: 1) Execute a simple linear iterative clustering algorithm to generate initial superpixel segmentation results; 2) Convert the initial superpixel segmentation results to the HSV color space; Edge detection is performed on the hue components of the HSV color space to optimize superpixel boundaries; the spectral similarity of adjacent superpixels is calculated; adjacent superpixels with a spectral similarity of more than 95% are merged to form homogeneous vegetation patches; and the vector boundaries of homogeneous vegetation patches containing patch IDs and geographic coordinates are output.

7. The method of claim 6, wherein, The manual annotation process uses mobile augmented reality assistance tools to perform the following sequential operations: Load the pre-stored typical vegetation micromorphological feature library, which includes a mangrove leaf vein fractal dimension dataset, a salt marsh herb stem internode length dataset, and an Spartina alterniflora inflorescence spikelet arrangement pattern dataset. High-definition images of vegetation surfaces within sample plots are captured in real time using a mobile camera. Local feature extraction was performed on real-time images to extract the fractal dimension of mangrove leaf veins, the internode length of salt marsh herbaceous stems, and the spikelet arrangement pattern of Spartina alterniflora inflorescence. The extracted features were matched with the feature library for similarity: when the fractal dimension matching error of the leaf veins of mangroves was ≤5%, they were labeled as mangroves; when the internode length matching error of the stems of salt marsh herbs was ≤3 mm, they were labeled as salt marsh herbs; when the overlap of the spikelet arrangement pattern of Spartina alterniflora was ≥90%, they were labeled as Spartina alterniflora. The annotation results are automatically associated with the geographical coordinates of the sampling points and transmitted back to the central database via mobile network to update the model training sample set.

Images