A mangrove vegetation carbon sink monitoring and measurement method based on drone and AI technology

Abstract

The present invention relates to a mangrove vegetation carbon sink monitoring and measurement method based on drone and AI technology. A mangrove identification model based on the U‑Net improved model of the PyTorch deep learning framework is constructed through drone AI technology to realize intelligent identification and distinction of various types of mangrove vegetation and their regional area and quantity, with an accuracy rate of more than 90%. Through regular drone aerial photography monitoring, combined with field-measured growth parameters and variable parameters derived from drone aerial photography data, an optimal fitting model of mangrove tree crown width, aboveground biomass, and unit leaf area based on drone aerial survey is constructed to accurately calculate the carbon storage and carbon fixation of mangrove regional vegetation, and comprehensively evaluate the carbon sink capacity of mangroves in the monitoring area. This application realizes intelligent, digital, and comprehensive monitoring and evaluation of the carbon sink capacity of mangrove wetland vegetation.

Description

Technical Field

[0001] The present invention relates to the technical field of ecological environment protection, and in particular to a mangrove vegetation carbon sink monitoring and measurement method based on drone and AI technology. Background Art

[0002] Mangroves grow mostly in tropical and subtropical coastal intertidal flats, and have the functions of purifying water quality, maintaining biodiversity, preventing wind and consolidating embankments, promoting siltation and protecting beaches. However, in recent years, due to the invasion of alien species and unreasonable development activities such as deforestation for farmland and reclamation of land from the sea, the mangrove growth area has shrunk and mangrove resources have been severely damaged. This not only affects its ecological service function, but also directly weakens its carbon sequestration capacity. Therefore, using advanced, efficient and rapid measurement methods to obtain information on the growth area, distribution area and community space of mangroves is crucial for effectively monitoring the carbon sequestration function of mangrove vegetation, and is also of great significance for protecting and planning the growth of mangroves and maintaining ecological security.

[0003] Common monitoring methods for mangrove vegetation include: field surveys and satellite remote sensing monitoring. However, since mangroves generally grow in inaccessible muddy tidal flats with poor accessibility, traditional field survey methods are time-consuming and inefficient, and it is difficult to quickly obtain large-scale survey data. There are many uncertainties, and the probability of personnel making mistakes in long-term work is also high. Unlike traditional monitoring methods, remote sensing technology has the characteristics of fast monitoring speed, wide range, and strong effectiveness, and is widely used. Among them, satellite remote sensing technology has high camera efficiency, large coverage area, and time continuity, but its resolution and accuracy are relatively low, and it is easily affected by atmospheric effects, which limits its application in high-precision monitoring of mangrove communities.

[0004] In recent years, the rise of drone aerial photography technology has provided a new means for monitoring and management of mangrove communities, especially for high-precision monitoring of small-scale areas (such as protected areas, cities and counties, and patch scales). Drone technology has the characteristics of easy portability, low cost, short data update cycle, and high spatial resolution. It can also carry various types of sensors, such as hyperspectral and laser radar, which can derive a large number of spectral, texture and structural features. In particular, airborne laser radar (Light Detection And Ranging, LiDAR) technology can obtain three-dimensional structural information of different tree species in complex scenes and generate canopy height models (Canopy Height Model, CHM). However, due to the high cost of LiDAR data acquisition, it has not yet been widely used in tree species research. At the same time, airborne visible light band optical cameras have the advantages of low cost, low requirements for take-off sites and weather, and greater operational flexibility, but there are still problems such as large amount of image data, high requirements for computer performance for image stitching and processing, and the accuracy of data post-processing combined with model analysis needs to be further improved and the workload is large.

[0005] Therefore, there is an urgent need to provide a mangrove vegetation carbon sink monitoring and measurement method based on drones and AI technology to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings and defects of the prior art and provide a mangrove vegetation carbon sink monitoring and measurement method based on drones and AI technology, combining the high-precision monitoring capability of drones and the intelligent analysis capability of AI technology to achieve rapid and accurate acquisition of mangrove growth area, distribution area and community spatial information, providing scientific basis and technical support for the protection, restoration and carbon sink management of mangroves.

[0007] A mangrove vegetation carbon sink monitoring and measurement method based on drone and AI technology includes the following steps:

[0008] Step 1: Collect training samples using a multispectral camera on a drone;

[0009] Step 2: Build a mangrove recognition model based on the U-Net improved model of the PyTorch deep learning framework;

[0010] Step 3: Collect orthophoto images and point cloud distribution images of mangroves through drone aerial survey, and use the mangrove recognition model trained in step 2 to identify the distribution area of ​​mangroves and the distribution area and quantity of various types of mangroves in the monitoring area;

[0011] Step 4: Obtain the tree height and crown width data of mangroves based on the UAV LiDAR. Combined with the area and number of each species calculated in Step 3, build a model of aboveground biomass, tree height and crown width of different types of mangrove vegetation, and calculate the aboveground biomass and carbon storage based on the UAV LiDAR.

[0012] Step 5: Obtain the tree height and crown width data of mangroves based on the UAV LiDAR. Combined with the area and number of each species calculated in Step 3, construct models of leaf area, tree height and crown width of different types of mangroves. Calculate the daily vegetation carbon sequestration based on the UAV LiDAR according to the measured daily vegetation carbon sequestration efficiency per unit leaf area of ​​mangroves.

[0013] Step 6: Calculate the carbon storage value CC of mangroves in the monitoring area based on the aboveground biomass and carbon storage method of UAV laser radar calculated in step 4 a According to the calculation in step 5, the daily vegetation carbon sequestration method based on UAV lidar is used to calculate the carbon sequestration value CC of mangroves in the monitoring area. b , when the formula is satisfied When, through The optimized carbon sink value is obtained and used to dynamically evaluate the changes in the carbon sink capacity of mangroves in the monitored area.

[0014] As a preferred technical solution of the present invention, the step 1 of collecting training samples using a multispectral camera carried by an unmanned aerial vehicle is specifically as follows: different mangrove areas are selected as training samples according to the number of species and forest age, and the training samples are divided into single species, two species, and multiple species according to species composition, and the training ratios are set to 10%, 20%, and 70%, respectively. Among the samples of multiple species, the training ratios are set to 20%, 20%, and 60%, respectively, according to the forest age of 1-3 years, the forest age of 3-5 years, and the forest age of more than 5 years;

[0015] 1500-5000 mangrove training samples were collected, including 15-25 typical mangrove species and 100-200 samples of each.

[0016] As a preferred technical solution of the present invention, the specific method of constructing a mangrove identification model of the U-Net improved model based on the PyTorch deep learning framework in step 2 is:

[0017] Step 2.1, obtain the RGB image data of the training samples collected by the drone, and use CVAT to intelligently label the mangrove species in the training samples;

[0018] Step 2.2, perform image enhancement on the sample data, including flipping, rotation, scaling, brightness and noise operations;

[0019] Step 2.3, extracting sample data features, including band data fusion, vegetation index feature calculation and texture feature extraction;

[0020] Step 2.4: Build an improved U-Net model based on the PyTorch deep learning framework, including the model architecture design of the input layer, encoder, bottleneck layer, decoder, and output layer;

[0021] Step 2.5: Train the built model, including data set division, parameter setting, training iteration, and training verification;

[0022] Step 2.6, test and optimize the trained model: analyze whether the model results are ideal. When the model result accuracy is ≥90%, F1-score ≥0.8 and IoU ≥0.5, the model results are ideal; when the model results are not ideal, further optimize the model parameters, which include data enhancement, model architecture improvement, loss function optimization, optimizer and learning rate adjustment, and training process optimization;

[0023] Step 2.7. Save the model.

[0024] As a preferred technical solution of the present invention, the step 3 of collecting orthophoto images and point cloud distribution images of mangroves through drone aerial survey is specifically as follows: drones are used for aerial survey in the monitoring area at a predetermined period, and aerial surveys are conducted during low tide, wind speed below level five, clear weather, and suitable sunlight incidence angles; the drone aerial survey route and parameter settings include a flight speed of 10m / s, a flight altitude of 60m, a "W"-shaped flight route, a heading overlap of 80%, a lateral overlap of 75%, and a pitch angle of 45° for the inclined route, and orthophoto images and point cloud distribution images of mangroves are regularly collected at preset time intervals.

[0025] As a preferred technical solution of the present invention, the specific method of using the mangrove identification model trained in step 2 in step 3 to identify the mangrove distribution area and the distribution area and quantity of various types of mangroves in the monitoring area is:

[0026] Step 3.1, import the model;

[0027] Step 3.2: Recognize each image collected by the drone aerial survey, including image preprocessing, feature extraction and prediction operations;

[0028] Step 3.3, merging and mosaicking the recognition results of multiple images, including image stitching and result fusion operations;

[0029] Step 3.4: Output the model identification results, showing the distribution area and quantity of each type of mangrove in the monitoring area.

[0030] As a preferred technical solution of the present invention, in step 4, the tree height and crown width data of mangroves are obtained based on the UAV laser radar, and the area and quantity of each species calculated in step 3 are combined to construct a model of plant mass, tree height and crown width of different types of mangroves. The specific method for calculating the aboveground biomass and carbon storage based on the UAV laser radar is as follows:

[0031] Step 4.1. Use the UAV-mounted laser radar to obtain 3D point cloud data and digital surface model (DSM) data, use the local maximum algorithm to count trees and use the regional growing algorithm to segment individual mangrove trees, and estimate the tree height (H) and crown width (CW) growth data of individual trees;

[0032] Step 4.2, obtain the mangrove tree height (H) and diameter at breast height (DBH) growth data through ground sample plot measurement, and use the biomass allometric growth equation to calculate the aboveground biomass (AGB) data parameters;

[0033] Step 4.3: Use linear regression model and random forest model to construct the best fitting model based on the relationship between aboveground biomass and tree height and crown width of different mangrove species;

[0034] Step 4.4: Identify the mangrove distribution area and the distribution area and quantity data of various types of mangroves in the monitoring area according to the mangrove identification model, calculate and summarize the total aboveground biomass of vegetation using the best fitting model, and obtain the carbon storage CS of the entire mangrove area by applying the carbon storage replacement coefficients of different types of mangroves. The formula is as follows:

[0035] CS = a×

[0036] I is the mangrove species; N is the corresponding mangrove tree; n is the nth tree; a is the carbon stock replacement coefficient; CS is the carbon stock in the mangrove monitoring area.

[0037] As a preferred technical solution of the present invention, in step 4.3, the linear regression model formula is as follows:

[0038]

[0039] β0 is the intercept; β1 and β2 are the regression coefficients of tree height and crown area, indicating their influence on biomass; ϵ is the error term, indicating the part that the model cannot explain.

[0040] The random forest model formula is as follows:

[0041]

[0042] T is the number of trees; Treet(X) is the prediction of the t-th tree for the input feature X.

[0043] As a preferred technical solution of the present invention, in step 5, the tree height and crown width data of mangroves are obtained based on the UAV laser radar, and the leaf area, tree height and crown width models of different types of mangroves are constructed in combination with the area and quantity of each species calculated in step 3. According to the measured daily carbon sequestration efficiency of mangrove vegetation per unit leaf area, the specific method for calculating the daily vegetation carbon sequestration based on the UAV laser radar is as follows:

[0044] Step 5.1. Use the UAV-mounted laser radar to obtain 3D point cloud data and digital surface model (DSM) data, use the local maximum algorithm to count trees and use the regional growing algorithm to segment individual mangrove trees, and estimate the tree height (H) and crown width (CW) growth data of individual trees;

[0045] Step 5.2, obtain the mangrove tree height (H), crown width (CW) and leaf area growth data through ground sample plot measurement, and use the linear regression model to construct the best fitting model based on the relationship between leaf area and tree height and crown width of different types of mangroves;

[0046] Step 5.3: Use photosynthetic meters to measure and calculate the daily carbon sequestration parameters per unit leaf area of ​​different types of mangroves;

[0047] Step 5.4: Identify the mangrove distribution area and the distribution area and quantity data of various types of mangroves in the monitoring area according to the mangrove identification model, calculate the total leaf area of ​​different types of mangroves, and obtain the carbon sequestration CC of the entire mangrove area through the daily vegetation carbon sequestration parameter per unit leaf area. The formula is as follows:

[0048]

[0049]

[0050]

[0051] I is the mangrove type; N is the corresponding number of mangroves; Ain is the leaf area of ​​the nth mangrove under the i-th type of mangrove; bi is the daily carbon sequestration per unit leaf area of ​​the i-th type of mangrove; GTLin is the carbon sequestration of the nth mangrove under the i-th type of mangrove; CC is the carbon sequestration in the mangrove monitoring area.

[0052] As a preferred technical solution of the present invention, the carbon sink measurement optimization method in step 6 is specifically as follows:

[0053] Step 6.1: Conduct drone aerial surveys in the mangrove monitoring area at a predetermined period. According to the method for calculating aboveground biomass and carbon storage based on drone lidar in step 4, calculate the carbon storage value CS1 of the mangroves in the monitoring area at that time, obtain the carbon storage value CS0 of the mangroves in the monitoring area last time, and subtract the carbon storage value CS0 of the mangroves in the monitoring area last time from the carbon storage value CS1 of the mangroves in the monitoring area at that time to obtain the carbon storage value CC of the mangroves in the monitoring area based on aboveground biomass. a ;

[0054] Step 6.2: According to the method for calculating the daily vegetation carbon sequestration based on UAV lidar in step 5, conduct field measurements of different types of mangroves within a predetermined period, calculate their daily unit leaf area carbon sequestration parameters 3-12 times, calculate their average values, and obtain the mean values ​​of the daily unit leaf area carbon sequestration parameters of different types of mangroves. Then calculate the carbon sequestration value CC of mangroves in the monitoring area within the predetermined period. b ;

[0055] Step 6.3: When the formula is satisfied When, through The carbon sink optimization value is obtained to dynamically evaluate the changes in the carbon sink capacity of mangroves in the monitoring area; When When adjusting the fitting model or parameters of tree height and crown width, aboveground biomass and leaf area of ​​mangroves, .

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The present invention uses UAV AI technology to construct a mangrove identification model based on the U-Net improved model of the PyTorch deep learning framework, and realizes intelligent identification and differentiation of various types of mangrove vegetation and their regional area and quantity, with an accuracy rate of more than 90%; through regular UAV aerial monitoring, combined with field-measured growth parameters and variable parameters derived from UAV aerial data, the best fitting model of mangrove tree crown width, aboveground biomass, and unit leaf area based on UAV aerial survey is constructed, and the carbon storage and carbon fixation of mangrove regional vegetation are accurately calculated, and the carbon sink capacity of mangroves in the monitoring area is comprehensively evaluated. This application realizes intelligent, digital, and comprehensive monitoring and evaluation of the carbon sink capacity of mangrove wetland vegetation. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of the steps of the present invention.

[0059] Figure 2 A schematic diagram of the mangrove identification model flow chart of the present invention for constructing a U-Net improved model based on the PyTorch deep learning framework.

[0060] Figure 3 This is a schematic diagram of the flow of the model of the present invention identifying the distribution areas of various types of mangrove vegetation within the monitoring area.

[0061] Figure 4 This is a schematic diagram of the overall process of calculating carbon reserves and carbon fixation in the present invention.

[0062] Figure 5 Schematic diagram of the method for calculating carbon storage based on UAV lidar and aboveground biomass in the present invention.

[0063] Figure 6 Schematic diagram of the method for calculating vegetation carbon sequestration based on UAV laser radar and unit leaf area in the present invention.

[0064] Figure 7 Schematic diagram of the process flow of the carbon sink measurement optimization method of the present invention. DETAILED DESCRIPTION

[0065] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention. The specific implementation process of the present invention is as follows:

[0066] like Figures 1 to 7 As shown, a mangrove vegetation carbon sink monitoring and measurement method based on drone and AI technology includes the following steps:

[0067] Step 1: Collect training samples using a multispectral camera on a drone;

[0068] Step 2: Build a mangrove recognition model based on the U-Net improved model of the PyTorch deep learning framework;

[0069] Step 3: Collect orthophoto images and point cloud distribution images of mangroves through drone aerial survey, and use the mangrove recognition model trained in step 2 to identify the distribution area of ​​mangroves and the distribution area and quantity of various types of mangroves in the monitoring area;

[0070] Step 4: Obtain the tree height and crown width data of mangroves based on the UAV LiDAR. Combined with the area and number of each species calculated in Step 3, build a model of aboveground biomass, tree height and crown width of different types of mangrove vegetation, and calculate the aboveground biomass and carbon storage based on the UAV LiDAR.

[0071] Step 5: Obtain the tree height and crown width data of mangroves based on the UAV LiDAR. Combined with the area and number of each species calculated in Step 3, construct models of leaf area, tree height and crown width of different types of mangroves. Calculate the daily vegetation carbon sequestration based on the UAV LiDAR according to the measured daily vegetation carbon sequestration efficiency per unit leaf area of ​​mangroves.

[0072] Step 6: Calculate the carbon storage value CC of mangroves in the monitoring area based on the aboveground biomass and carbon storage method of UAV laser radar calculated in step 4 a According to the calculation in step 5, the daily vegetation carbon sequestration method based on UAV lidar is used to calculate the carbon sequestration value CC of mangroves in the monitoring area. b , when the formula is satisfied When, through The optimized carbon sink value is obtained and used to dynamically evaluate the changes in the carbon sink capacity of mangroves in the monitored area.

[0073] Reference Figures 2 to 3 , the above step 1 uses a multispectral camera on a drone to collect training samples. Specifically, different mangrove areas are selected as training samples according to the number of species and forest age. The training samples are divided into single species, two species, and multiple species according to species composition, and the training ratios are set to 10%, 20%, and 70% respectively. In the samples of multiple species, the training ratios are set to 20%, 20%, and 60% respectively according to the forest age of 1-3 years, 3-5 years, and >5 years. It should be noted that the sample plots must have the characteristics of clear boundaries of each system element, stable plant growth, and a canopy density between 40-70% to improve the efficiency of model training and the wide adaptability of the model.

[0074] In the first embodiment of the present invention, 1500 mangrove training samples are collected, including 15 typical mangrove species and 100 samples of each.

[0075] According to the collected training samples, a mangrove recognition model based on the U-Net improved model of the PyTorch deep learning framework is constructed, including:

[0076] Step 2.1: Obtain the RGB image data of the training samples collected by the drone, and use the CVAT tool that supports intelligent interaction to intelligently label the mangrove species in the training samples.

[0077] Step 2.2, perform image enhancement on sample data, including flipping, rotation, scaling, brightness and noise operations. Specifically, after normalizing the pixel values ​​of each band of 1500 samples with a pixel size of 2560 × 2560 to the range of [0, 1], perform flipping processing, which includes horizontal flipping and vertical flipping operations, and then perform random rotation processing on the image with a random rotation angle between ±10 degrees, and then perform random scaling processing with a scaling ratio between 0.9 times and 1.1 times, and brightness adjustment between ±10%. Finally, perform noise processing, which includes processing and no processing operations; after the above-mentioned geometric transformation and spectral enhancement processing, 9000 equivalent samples are finally obtained. After expanding the training sample size, it can better adapt to changes in illumination, changes in viewing angle and noise interference, increase the diversity of mangrove samples, and reduce the cost of drone aerial surveys.

[0078] Step 2.3, extract sample data features, including band data fusion, vegetation index feature calculation and texture feature extraction, to obtain 5-band image data, 4 vegetation index feature data and 4 texture feature data. Specifically:

[0079] Band data fusion: The five-band (red R, green G, blue B, infrared RE, and near-infrared NIR band) data obtained by the drone are spliced ​​to form a multi-channel input.

[0080] Vegetation index feature calculation: Extract and calculate vegetation index feature data based on Python calculation, including but not limited to Normalized Difference Vegetation Index NDVI, Normalized Red Edge Index NDRE, Modified Soil Adjusted Vegetation Index MSAVI, and Ratio Vegetation Index RVI.

[0081] Texture feature extraction: Based on Python, use GLCM (gray-level co-occurrence matrix), LBP (local binary pattern) and other methods to extract texture features and calculate texture feature data, including but not limited to contrast, homogeneity, entropy, and correlation.

[0082] Step 2.4: Build an improved U-Net model based on the PyTorch deep learning framework, including the U-Net model architecture. The U-Net model architecture includes an input layer, an encoder, a bottleneck layer, a decoder, and an output layer.

[0083] Input layer: The input data shape is (2560, 2560, 13), including 5-band image data, 4 vegetation index feature data and 4 texture feature data, where 2560x2560 is the length and width of the image, and 13 is the number of channels.

[0084] Encoder: It includes convolution blocks, each of which contains two 3x3 convolution layers followed by ReLU activation function and BatchNormalization; the convolution layer uses 3x3 convolution kernel to extract features; ReLU (Rectified Linear Unit) activation function is used to introduce nonlinearity; each convolution block is followed by a 2x2 maximum pooling layer to gradually reduce the spatial dimension.

[0085] Bottleneck layer: A convolutional block consisting of two 3x3 convolutional layers with a total of 512 filters.

[0086] Decoder: The upsampling layer uses transposed convolution (2x2) to gradually restore the spatial dimension; the jump connection first filters the encoder output through the "attention mechanism module" to highlight important features, suppress noise or background information, and splices the feature map of the encoder corresponding layer with the decoder layer feature map to retain spatial information and improve segmentation (classification) accuracy; each decoding layer contains two 3x3 convolution layers, and the number of filters is the same as the encoder corresponding layer.

[0087] Output layer: The convolution layer uses 1x1 convolution to map the feature map to the number of mangrove species. The Softmax activation function is used for multi-class classification. The Softmax activation function converts the output into a probability distribution to ensure that the sum of the probabilities of all categories is 1.

[0088] In specific applications, the 3x3 convolution kernel of the encoder is used to extract feature variables, and the ReLU (Rectified Linear Unit) activation function is used to introduce nonlinear factors to map the input to between 0 and 1; the 512 filters of the bottleneck layer and two 3x3 convolution layers are used to further abstract and compress the features in order to restore higher-resolution feature maps in the decoding stage; the jump connection is used to first filter the encoder output through the "attention mechanism module" to highlight important features and suppress noise or background information, and the feature map of the corresponding layer of the encoder is spliced ​​with the feature map of the decoder layer to retain spatial information and improve segmentation (classification) accuracy; the 1x1 convolution is used to map the feature map to the number of mangrove categories, and then the Softmax activation function is used to convert the output into a probability distribution to ensure that the sum of all outputs is 1. .

[0089] z is the output vector; i represents the i-th element in the input vector; j is an index used to represent all elements in the input vector; the denominator represents the sum of the exponential values ​​of all input elements to ensure that the sum of the probability distribution of the final output is 1.

[0090] Step 2.5: Train the built model, including:

[0091] Dataset division: randomly set 70% of the samples as training set, 20% as validation set, and 10% as test set. That is, the training set is 6300 samples (70%), the validation set is 1800 samples (20%), and the test set is 900 samples (10%), for a total of 9000 labeled samples.

[0092] Parameter setting: Batch size is set to 100, and 100 samples are used for parameter update in each training iteration; the number of training epochs is set to 50, and each epoch model will perform a complete training process on the entire training dataset and update the model parameters through multiple training iterations.

[0093] Training iteration: Randomly extract 100 samples (i.e. a batch) from the training set, input these samples into the model, perform forward propagation, and calculate the output; use the cross entropy loss function to evaluate the gap between the model output and the true label; based on the calculated loss value, calculate the gradient through the back propagation algorithm and use the optimizer to update the model parameters; record the loss value of each batch for subsequent analysis.

[0094] The above cross-entropy loss function (Cross-Entropy Loss) has the following formula:

[0095]

[0096] Where yi is the true label and xi is the predicted probability.

[0097] The above optimizer uses the Adam optimizer. The learning rate is set to 0.001, which determines the step size setting for each parameter update; β1 (beta1): the decay rate of the first-order moment estimate, set to 0.9, used to control the influence of the past gradient on the current gradient; β2 (beta2): the decay rate of the second-order moment estimate, set to 0.999, to control the influence of the past square gradient on the current square gradient; ε (epsilon): a small constant, set to 1e-7, used to prevent zero division errors.

[0098] Training and validation: After each epoch, the validation set is used to evaluate the performance of the model. The specific steps are to input the samples in the validation set into the model for forward propagation; calculate the loss and performance indicators of the validation set, and evaluate the performance of the model through performance indicators, including but not limited to accuracy, F1-score, and IoU.

[0099] Accuracy: In most cases, over 90% is considered a good performance, and the formula is as follows:

[0100]

[0101] Where TPn (True Positives for class) is the number of samples correctly predicted by the model as class n; FPn (False Positives for class) is the number of samples incorrectly predicted by the model as class n; FNn (False Negatives for class) is the number of samples incorrectly predicted by the model as classes other than class n; n is the total number of classes (in mangrove species identification, n is the number of mangrove species).

[0102] F1-score: refers to the harmonic mean of precision and recall. Generally, a value above 0.8 is considered good. In the case of data imbalance, F1-score can better reflect model performance than accuracy. The formula is as follows:

[0103]

[0104] Among them, Precision is the ratio of correctly predicted positive samples to all predicted positive samples, and Recall is the ratio of correctly predicted positive samples to all true positive samples.

[0105] IoU: An indicator that measures the degree of overlap between the predicted area and the true area. Generally, a value above 0.5 is considered good. The formula is as follows:

[0106]

[0107] Finally, validation results are recorded to monitor the model's training progress and evaluate the model's performance.

[0108] Step 2.6, model testing and optimization: Analyze whether the model results are ideal, and use the test set to finally evaluate the model's generalization ability, usually used after training and verification. Perform visual analysis on the test set to intuitively show the comparison between the model's segmentation effect and the true label. When the model result accuracy is ≥90%, F1-score ≥0.8, and IoU ≥0.5, the model result is ideal; when the model result is not ideal, further optimize the model parameters, and the optimization of the model parameters includes:

[0109] Data enhancement: By transforming and expanding the training data, the diversity and richness of the data can be increased, thereby improving the generalization ability of the model.

[0110] Model architecture improvement: Increase the depth (increase the number of convolution blocks) or width (increase the number of filters in each convolution block) of U-Net to improve the expressiveness of the model; or try using other activation functions, such as Leaky ReLU or Swish, to improve the learning ability of the model.

[0111] Loss function optimization: including but not limited to weighted loss function and combined loss function; weighted loss function means that if there is a category imbalance, a weighted cross entropy loss function can be used to give a higher weight to the minority class to improve the model's recognition ability for these categories. The combined loss function is a combination of cross entropy loss and Dice loss (a loss function commonly used in image segmentation tasks), which can optimize both classification accuracy and segmentation effect.

[0112] Optimizer and learning rate scheduling: Choose a suitable optimizer. In addition to Adam, you can also try RMSProp or AdamW (Adam with weight decay) to improve convergence speed and model performance. Use learning rate decay strategies (such as ReduceLROnPlateau) or learning rate warm-up strategies to dynamically adjust the learning rate to improve training results.

[0113] Optimize the training process: adjust the batch size according to the GPU memory and model complexity. A larger batch size can speed up training. Or use early stopping to monitor the model performance on the validation set. If there is no improvement within a few epochs, stop training early to prevent overfitting.

[0114] Step 2.7, Save the model: After the model training is completed, save the trained model as a file for subsequent use. The saved model file should include the structure and parameters of the model, usually stored in .pth or .pt format.

[0115] Through the above steps, based on UAV multispectral remote sensing images and combined with feature fusion, the mangrove identification model constructed by the deep learning neural network U-Net model can efficiently process multi-band image data, vegetation index features and texture feature data, and realize the rapid and accurate acquisition of mangrove growth area, distribution area and community spatial information, with an accuracy rate of ≥90%.

[0116] The specific steps of step 2 of collecting orthophoto images and point cloud distribution images of mangroves through drone aerial survey are as follows: drones are used for aerial survey in the monitoring area at a predetermined period of every quarter, and aerial survey is carried out during a time period of low tide, wind speed below level 5, clear weather, and suitable angle of sunlight incidence (10:00 to 16:00 during the day); the drone aerial survey route and parameter settings include: flight speed of 10m / s, flight altitude of 60m, "W"-shaped flight route, heading overlap of 80%, lateral overlap of 75%, and a pitch angle of 45° for the inclined route, and orthophoto images and point cloud distribution images of mangroves are collected regularly at preset time intervals.

[0117] Based on the mangrove identification model trained in steps 2.1 to 2.7 above, the specific steps for identifying the mangrove distribution area and the distribution area and quantity of various types of mangroves in the monitoring area in step 3 include:

[0118] Step 3.1. Import the model: First load the saved model, import the model into the working environment by specifying the model file path, and set it to evaluation mode to ensure the performance of the model during inference.

[0119] Step 3.2: Recognize each image collected by the drone aerial survey, including:

[0120] Image preprocessing: Perform the same preprocessing as during training on each image, including normalization and other processing, to ensure that the format of the input data is consistent with that during training.

[0121] Feature extraction and prediction operation: Input the preprocessed image into the model, perform forward propagation, and obtain the prediction result; the model will output the category probability of each pixel, and the user can determine the final category of each pixel based on the probability value.

[0122] Step 3.3: Merge and mosaic the recognition results of multiple images (images taken by drones usually cover a large area), including:

[0123] Image stitching: stitch adjacent images together to form a large image. Stitching is done based on the overlapping areas when shooting to ensure seamless connection.

[0124] Result fusion: For overlapping areas, a voting mechanism or weighted average method is used to determine the final category.

[0125] Step 3.4: Output the model recognition results as a visual image or report to show the distribution, area and quantity of each type of mangrove in the monitoring area. Users can also generate a map containing the recognition results, marking the distribution areas of different mangrove species and the area and quantity of each type of mangrove.

[0126] Reference Figure 5 In step 4, the tree height and crown width data of mangroves are obtained based on the UAV laser radar. Combined with the area and quantity of each species calculated in step 3, the aboveground biomass, tree height and crown width models of different types of mangrove vegetation are constructed. The specific method for calculating the aboveground biomass and carbon storage based on the UAV laser radar is as follows:

[0127] Step 4.1: Use the UAV equipped with LiDAR to obtain 3D point cloud data and digital surface model (DSM) data, use the local maximum algorithm to count trees and use the regional growing algorithm to segment individual mangrove trees, and calculate the tree height (H) and crown width (CW) growth data of individual trees; Specifically, the tree height survey step includes obtaining 3D point cloud data, LiDAR emitting laser pulses to the ground (each pulse may generate multiple echo signals, the first echo usually comes from the top of the canopy, and the last echo usually comes from the ground or lower vegetation), canopy height model (CHM) = digital surface model (DSM) - digital terrain model (DTM). The crown width survey step includes emitting laser pulses, receiving multiple echo signals, recording the position and intensity of each echo point, distinguishing between canopy echoes and ground echoes, calculating the weight of each echo point by counting the number of echo points of each pulse, identifying and accumulating the weight of the canopy echo points, and then dividing by the total number of pulses to obtain the effective canopy coverage.

[0128] Step 4.2: Obtain the growth data of mangrove tree height (H) and diameter at breast height (DBH) through ground sample plot measurements, and use the biomass allometric growth equation to calculate the aboveground biomass (AGB) data parameters; specifically, measure the key parameters related to mangrove growth on the spot, set up sampling points and sample plots, and use a laser meter to measure tree height; measure the DBH at 1.3 m above the ground, and use the allometric growth equation (based on the main mangrove species in South China) to evaluate the aboveground biomass.

[0129] Step 4.3: Use linear regression model and random forest model respectively to construct the best fitting model based on the relationship between aboveground biomass and tree height and crown width of different types of mangroves. The performance of the model can be evaluated by using indicators such as mean square error (MSE) and determination coefficient (R²) to obtain the best fitting model.

[0130] It is assumed that there is a linear relationship between biomass and tree height and crown area, and the linear regression model formula is:

[0131]

[0132] β0 is the intercept; β1 and β2 are the regression coefficients of tree height and crown area, indicating their influence on biomass; ϵ is the error term, indicating the part that the model cannot explain.

[0133] The random forest model improves the accuracy of the model by building multiple decision trees and combining their prediction results. The prediction of each tree can be expressed as: AGB = predict(x), the specific formula is as follows:

[0134] AGB

[0135] T is the number of trees; Treet(X) is the prediction of the t-th tree for the input feature X.

[0136] Step 4.4: Identify the mangrove distribution area and the distribution area and quantity data of various types of mangroves in the monitoring area according to the mangrove identification model, calculate and summarize the total aboveground biomass of vegetation using the best fitting model, and obtain the carbon storage CS of the entire mangrove area by applying the carbon storage replacement coefficients of different types of mangroves. The formula is as follows:

[0137] CS = a×

[0138] I is the mangrove species; N is the corresponding mangrove tree; n is the nth tree; a is the carbon stock replacement coefficient; CS is the carbon stock in the mangrove monitoring area.

[0139] Reference Figure 6In step 5, the tree height and crown width data of mangroves are obtained based on the UAV laser radar. Combined with the area and quantity of each species calculated in step 3, the leaf area, tree height and crown width models of different types of mangroves are constructed. According to the measured daily unit leaf area carbon sequestration efficiency of mangroves, the specific method for calculating the daily vegetation carbon sequestration based on the UAV laser radar is as follows:

[0140] Step 5.1: Use UAV laser radar to obtain 3D point cloud data and digital surface model (DSM) data; use the local maximum algorithm to count trees and the regional growing algorithm to segment individual mangrove trees, and calculate the tree height (H) and crown width (CW) growth data of individual trees. The calculation steps are the same as step 4.1.

[0141] Step 5.2: Obtain the growth data of mangrove tree height (H), crown width (CW) and leaf area through ground sample plots, and use the linear regression model to construct the best fitting model based on the relationship between leaf area and tree height and crown width of different types of mangroves. The formula is as follows:

[0142]

[0143] β0 is the intercept; β1 and β2 are the regression coefficients of tree height and crown area, indicating their influence on biomass; ϵ is the error term, indicating the part that the model cannot explain.

[0144] Step 5.3: Use a portable photosynthetic meter to measure and calculate the daily carbon sequestration parameters per unit leaf area of ​​different mangrove species.

[0145] Step 5.4: Identify the mangrove distribution area and the distribution area and quantity data of various types of mangroves in the monitoring area according to the mangrove identification model, calculate the total leaf area of ​​different types of mangroves, and obtain the carbon sequestration CC of the entire mangrove area through the daily vegetation carbon sequestration parameter per unit leaf area. The formula is as follows:

[0146]

[0147]

[0148]

[0149] I is the mangrove type; N is the corresponding number of mangrove trees; n is the nth tree; Hin is the height of the nth mangrove under the i-th mangrove; Win is the crown width of the nth mangrove under the i-th mangrove; Ain is the leaf area of ​​the nth mangrove under the i-th mangrove; bi is the daily carbon sequestration per unit leaf area of ​​the i-th mangrove; GTLin is the carbon sequestration of the nth mangrove under the i-th mangrove; CC is the carbon sequestration in the mangrove monitoring area.

[0150] Reference Figure 7 The carbon sink measurement optimization method in step 6 is specifically as follows:

[0151] Step 6.1: Carry out drone aerial survey in the mangrove monitoring area at a predetermined period of each quarter (90 days). According to the method of calculating aboveground biomass and carbon storage based on drone lidar in step 4, calculate the carbon storage value CS1 of the mangroves in the monitoring area at that time, obtain the carbon storage value CS0 of the mangroves in the monitoring area last time, and subtract the carbon storage value CS0 of the mangroves in the monitoring area last time from the carbon storage value CS1 of the mangroves in the monitoring area at that time to obtain the carbon storage value CC of the mangroves in the monitoring area based on aboveground biomass. a .

[0152] Step 6.2: According to the method for calculating the daily vegetation carbon sequestration based on UAV lidar in step 5, conduct field measurements of different types of mangroves within a predetermined period, calculate their daily unit leaf area carbon sequestration parameters 3-12 times, and calculate the average value to obtain the mean value of the daily unit leaf area carbon sequestration parameters of different types of mangroves. Then calculate the carbon sequestration value CC of mangroves in the monitoring area within the predetermined period of each quarter (90 days). b .

[0153] Step 6.3: When the formula is satisfied When the carbon storage value of mangroves in the monitoring area calculated in step 6.1 is CC a The carbon sequestration value CC of mangroves in the monitoring area calculated in step 6.2 b After adding and averaging, the optimal carbon sink value is obtained, which is used to dynamically evaluate the changes in the carbon sink capacity of mangroves in the monitoring area. When When adjusting the tree height and crown width, the aboveground biomass and leaf area of ​​mangroves, the fitting model or parameters are adjusted so that .

[0154] The above-mentioned embodiments only express the implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A mangrove vegetation carbon sink monitoring and measurement method based on drone and AI technology, characterized in that: The following steps are involved: Step 1: Collect training samples using a multispectral camera on a drone; Step 2: Build a mangrove recognition model based on the U-Net improved model of the PyTorch deep learning framework; Step 3: Collect orthophoto images and point cloud distribution images of mangroves through drone aerial survey, and use the mangrove recognition model trained in step 2 to identify the distribution area of ​​mangroves and the distribution area and quantity of various types of mangroves in the monitoring area; Step 4: Obtain the tree height and crown width data of mangroves based on the UAV LiDAR. Combined with the area and number of each species calculated in Step 3, build a model of aboveground biomass, tree height and crown width of different types of mangrove vegetation, and calculate the aboveground biomass and carbon storage based on the UAV LiDAR. Step 5: Obtain the tree height and crown width data of mangroves based on the UAV LiDAR. Combined with the area and number of each species calculated in Step 3, construct models of leaf area, tree height and crown width of different types of mangroves. Calculate the daily vegetation carbon sequestration based on the UAV LiDAR according to the measured daily vegetation carbon sequestration efficiency per unit leaf area of ​​mangroves. Step 6: Calculate the carbon storage value CC of mangroves in the monitoring area based on the aboveground biomass and carbon storage method of UAV laser radar calculated in step 4 a According to the calculation in step 5, the daily vegetation carbon sequestration method based on UAV lidar is used to calculate the carbon sequestration value CC of mangroves in the monitoring area. b , when the formula is satisfied When, through The optimized carbon sink value is obtained to dynamically evaluate the changes in the carbon sink capacity of mangroves in the monitoring area; The specific method for calculating aboveground biomass and carbon storage based on UAV lidar in step 4 is: Step 4.1: Use the UAV-mounted laser radar to obtain 3D point cloud data and digital surface model (DSM) data, use the local maximum algorithm to count trees and use the regional growing algorithm to segment individual mangrove trees, and estimate the tree height H and crown width CW growth data of individual trees; Step 4.2, obtain the mangrove tree height H and breast diameter DBH growth data through ground sample plot measurement, and use the biomass allometric growth equation to calculate the aboveground biomass AGB data parameters; Step 4.3: Use linear regression model and random forest model to construct the best fitting model based on the relationship between aboveground biomass and tree height and crown width of different mangrove species; Step 4.4: Identify the mangrove distribution area and the distribution area and quantity data of various types of mangroves in the monitoring area according to the mangrove identification model, calculate and summarize the total aboveground biomass of vegetation using the best fitting model, and obtain the carbon storage CS of the entire mangrove area by applying the carbon storage replacement coefficients of different types of mangroves. The formula is as follows: , I is the mangrove species; N is the corresponding mangrove tree; n is the nth tree; a is the carbon storage replacement coefficient; CS is the carbon storage in the mangrove monitoring area; AGB is in is the total aboveground biomass of the nth mangrove tree under the ith mangrove forest type.

2. According to claim 1, a method for monitoring and measuring carbon sinks in mangrove vegetation based on drone and AI technology is characterized in that: The step 1 of collecting training samples using a multispectral camera carried by an unmanned aerial vehicle is specifically as follows: different mangrove areas are selected as training samples according to the number of species and forest age, and the training samples are divided into single species, two species, and multiple species according to species composition, and the training ratios are set to 10%, 20%, and 70%, respectively. Among the samples of multiple species, the training ratios are set to 20%, 20%, and 60%, respectively, according to the forest age of 1-3 years, the forest age of 3-5 years, and the forest age of more than 5 years; 1500-5000 mangrove training samples were collected, including 15-25 typical mangrove species and 100-200 samples of each.

3. According to claim 1, a method for monitoring and measuring carbon sinks in mangrove vegetation based on drone and AI technology is characterized in that: The specific method for constructing the mangrove identification model of the U-Net improved model based on the PyTorch deep learning framework in step 2 is: Step 2.1, obtain the RGB image data of the training samples collected by the drone, and use CVAT to intelligently label the mangrove species in the training samples; Step 2.2, perform image enhancement on the sample data, including flipping, rotation, scaling, brightness and noise operations; Step 2.3, extracting sample data features, including band data fusion, vegetation index feature calculation and texture feature extraction; Step 2.4: Build an improved U-Net model based on the PyTorch deep learning framework, including the model architecture design of the input layer, encoder, bottleneck layer, decoder, and output layer; Step 2.5: Train the built model, including data set division, parameter setting, training iteration, and training verification; Step 2.6, test and optimize the trained model: analyze whether the model results are ideal. When the model result accuracy is ≥90%, F1-score ≥0.8 and IoU ≥0.5, the model results are ideal; when the model results are not ideal, further optimize the model parameters, which include data enhancement, model architecture improvement, loss function optimization, optimizer and learning rate adjustment, and training process optimization; Step 2.

7. Save the model.

4. According to claim 1, a method for monitoring and measuring carbon sinks in mangrove vegetation based on drone and AI technology is characterized in that: The specific steps of collecting orthophoto images and point cloud distribution images of mangroves by drone aerial survey in step 3 are as follows: drones are used for aerial survey in the monitoring area at a predetermined period, and aerial surveys are conducted during low tide, wind speed below level 5, clear weather, and suitable sunlight incidence angles; the drone aerial survey route and parameter settings include a flight speed of 10m / s, a flight altitude of 60m, a "W"-shaped flight route, a heading overlap of 80%, a lateral overlap of 75%, and a pitch angle of 45° for the inclined route, and orthophoto images and point cloud distribution images of mangroves are collected regularly at preset time intervals.

5. According to claim 1, a method for monitoring and measuring carbon sinks in mangrove vegetation based on drone and AI technology is characterized in that: In step 3, the mangrove identification model trained in step 2 is used to identify the distribution area of ​​mangroves and the distribution area and quantity of various types of mangroves in the monitoring area. The specific method is: Step 3.1, import the model; Step 3.2: Recognize each image collected by the drone aerial survey, including image preprocessing, feature extraction and prediction operations; Step 3.3, merging and mosaicking the recognition results of multiple images, including image stitching and result fusion operations; Step 3.4: Output the model identification results, showing the distribution area, area and quantity of each type of mangrove in the monitoring area.

6. According to claim 1, a method for monitoring and measuring carbon sinks in mangrove vegetation based on drone and AI technology is characterized in that: In step 4.3, the linear regression model formula is as follows: ， β0 is the intercept; H is the tree height; CW is crown width; β1 and β2 are the regression coefficients of tree height and crown area, indicating their influence on biomass; is the error term, which represents the part that the model cannot explain; The random forest model formula is as follows: ， T is the number of trees; t (X) is the prediction of the t-th tree for the input feature X.

7. According to claim 1, a method for monitoring and measuring carbon sinks in mangrove vegetation based on drone and AI technology is characterized in that: In step 5, the tree height and crown width data of mangroves are obtained based on the UAV laser radar, and the leaf area, tree height and crown width models of different types of mangroves are constructed in combination with the area and quantity of each species calculated in step 3. According to the measured daily carbon sequestration efficiency of mangrove vegetation per unit leaf area, the specific method for calculating the daily vegetation carbon sequestration based on the UAV laser radar is as follows: Step 5.1: Use the UAV-mounted laser radar to obtain 3D point cloud data and digital surface model (DSM) data, use the local maximum algorithm to count trees and use the regional growing algorithm to segment individual mangrove trees, and estimate the tree height H and crown width CW growth data of individual trees; Step 5.2, obtain the mangrove tree height H, crown width CW and leaf area growth data through ground sample plot measurement, and use the linear regression model to construct the best fitting model based on the relationship between leaf area and tree height and crown width of different types of mangroves; Step 5.3: Use photosynthetic meters to measure and calculate the daily carbon sequestration parameters per unit leaf area of ​​different mangrove species. Step 5.4: Identify the mangrove distribution area and the distribution area and quantity data of various types of mangroves in the monitoring area according to the mangrove identification model, calculate the total leaf area of ​​different types of mangroves, and obtain the carbon sequestration CC of the entire mangrove area through the daily vegetation carbon sequestration parameter per unit leaf area. The formula is as follows: ， I is the type of mangrove; N is the number of corresponding mangrove trees; Ain is the leaf area of ​​the nth mangrove under the i-th type of mangrove; bi is the daily carbon fixation per unit leaf area of ​​the i-th type of mangrove; GTL in is the carbon sequestration of the nth mangrove under the ith mangrove; CC is the carbon sequestration in the mangrove monitoring area.

8. According to claim 1, a method for monitoring and measuring carbon sinks in mangrove vegetation based on drone and AI technology is characterized in that: The carbon sink measurement optimization method in step 6 is specifically as follows: Step 6.1: Conduct drone aerial surveys in the mangrove monitoring area at a predetermined period. According to the method for calculating aboveground biomass and carbon storage based on drone lidar in step 4, calculate the carbon storage value CS1 of the mangroves in the monitoring area at that time, obtain the carbon storage value CS0 of the mangroves in the monitoring area last time, and subtract the carbon storage value CS0 of the mangroves in the monitoring area last time from the carbon storage value CS1 of the mangroves in the monitoring area at that time to obtain the carbon storage value CC of the mangroves in the monitoring area based on aboveground biomass. a ; Step 6.2: According to the method for calculating the daily vegetation carbon sequestration based on UAV lidar in step 5, conduct field measurements of different types of mangroves within a predetermined period, calculate their daily unit leaf area carbon sequestration parameters 3-12 times, calculate their average values, and obtain the mean values ​​of the daily unit leaf area carbon sequestration parameters of different types of mangroves. Then calculate the carbon sequestration value CC of mangroves in the monitoring area within the predetermined period. b ; Step 6.3: When the formula is satisfied When, through The carbon sink optimization value is obtained to dynamically evaluate the changes in the carbon sink capacity of mangroves in the monitoring area; When When adjusting the fitting model or parameters of tree height and crown width, aboveground biomass and leaf area of ​​mangroves, .

Images