A UAV-based fine classification method for dense mangroves in complex terrain

Abstract

The present invention discloses a method for fine classification of dense mangroves in complex landforms by using unmanned aerial vehicles, which relates to the field of image processing technology and comprises the following steps: S1, collecting unmanned aerial vehicle data of mangrove ecological areas; S2, preprocessing the unmanned aerial vehicle data to obtain multispectral raw data, panchromatic raw data and lidar raw data; S3, dividing the panchromatic raw data into sample sets; S4, processing the multispectral raw data, the lidar raw data and the output results of the sample label acquisition unit to obtain classification features; S5, processing the classification features to obtain an improved mangrove fine classification feature group; S6, processing the improved mangrove fine classification feature group. In the present invention, the texture features are further improved by changing the dimension reduction method of the image, and the multispectral features, lidar features and vegetation index are combined for target recognition at the same time, thereby improving the accuracy of fine classification of dense mangroves in complex landforms.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for finely classifying dense mangroves with complex landforms using an unmanned aerial vehicle. Background Art

[0002] Mangroves are woody plant communities that grow in the intertidal zones of tropical and subtropical coasts and are periodically submerged by seawater. They are important wetland ecosystems on the coast and play an important role in purifying coastal water quality, maintaining ecological balance and biodiversity, nutrient cycling, and carbon fixation and storage. At the same time, mangroves have become an important part of "blue carbon" with their significant carbon sink capacity. However, due to the differences in carbon sink capacity among different tree species, fine classification of mangroves has become our primary task to accurately assess their carbon sink capacity, but the time and labor costs of manual measurement are too high; extracting mangroves through high-resolution satellite remote sensing is expensive and has poor fine classification accuracy. The existing mangrove fine classification method collects drone images of the target mangrove area, extracts features from them first, and then performs fine classification through a classifier to determine the specific interspecies distribution and area of ​​mangroves in the target area. However, in the existing mangrove fine classification method, the commonly used single image features are lacking in accuracy. Although the mixed features can improve the ability to extract objects to a certain extent, the anti-interference ability of complex landforms is still lacking, which leads to low precision of mangrove fine classification. Summary of the invention

[0003] In order to solve the above problems, the present invention proposes a method for fine classification of dense mangroves in complex terrain using an unmanned aerial vehicle.

[0004] The technical solution of the present invention is: a method for fine classification of dense mangroves with complex landforms by drone comprises the following steps:

[0005] S1. Collect drone data of mangrove ecological areas, including multispectral drone data, panchromatic drone data and lidar drone data;

[0006] S2, preprocessing the UAV data using a data preprocessing unit to obtain multispectral raw data, panchromatic raw data, and lidar raw data;

[0007] S3, using a sample label acquisition unit to divide the full-color original data into sample sets;

[0008] S4, using a feature extraction unit to process the multispectral raw data, the laser radar raw data, and the output result of the sample label acquisition unit to obtain classification features;

[0009] S5, using a feature screening unit to process the classification features to obtain an improved mangrove fine classification feature group;

[0010] S6. Use the ablative classification units to process the improved mangrove fine classification feature group and complete the classification.

[0011] Further, the feature extraction unit includes a label conversion layer, an improved texture feature extraction module, a vegetation index calculation layer, a lidar feature extraction module, a first adder A1, a second adder A2 and a third adder A3;

[0012] The input end of the label conversion layer serves as the input end of the feature extraction unit and is connected to the output end of the sample label acquisition unit; the first output end of the label conversion layer is connected to the first input end of the improved texture feature extraction module; the second output end of the label conversion layer is respectively connected to the first input end of the first adder, the first input end of the second adder and the first input end of the third adder; the second input end of the improved texture feature extraction module is connected to the first output end of the data preprocessing unit; the output end of the improved texture feature extraction module is connected to the second input end of the first adder A1; the input end of the vegetation index calculation layer is connected to the second output end of the data preprocessing unit; the output end of the vegetation index calculation layer is connected to the second input end of the second adder A2; the input end of the lidar feature extraction module is connected to the third output end of the data preprocessing unit; the output end of the lidar feature extraction module is connected to the second input end of the third adder A3; the output end of the first adder A1, the output end of the second adder A2 and the output end of the third adder A3 are all used as the output end of the feature extraction unit.

[0013] The beneficial effect of the above further scheme is as follows: in the present invention, the multispectral data and lidar data of the drone are processed to obtain three kinds of feature data, which are then combined with the label data output by the label conversion layer to form independent feature groups in each adder. Each feature group has multiple single features, thereby providing a large amount of material for subsequent feature fusion.

[0014] Further, the improved texture feature extraction module includes a data dimension reduction layer, a first data transcoding layer, a first xgboost model training layer, a first feature importance calculation layer, a first feature separation layer, a gray level co-occurrence matrix construction layer, a first normalization layer and a Haralick feature calculation layer;

[0015] The input end of the data dimension reduction layer serves as the first input end of the improved texture feature extraction module; the first input end of the first data transcoding layer serves as the second input end of the improved texture feature extraction module; the output end of the data dimension reduction layer is connected to the second input end of the first data transcoding layer; the output end of the first data transcoding layer, the first xgboost model training layer, the first feature importance calculation layer, the first feature separation layer, the gray level co-occurrence matrix construction layer, the first normalization layer and the input end of the Haralick feature calculation layer are connected in sequence; the output end of the Haralick feature calculation layer serves as the output end of the improved texture feature extraction module.

[0016] The beneficial effect of the above further scheme is: the data dimension reduction layer in the present invention processes the multispectral features to improve the dimension reduction method to achieve feature concentration and reduce the interference of noise on feature extraction. The data after dimension reduction and the sample features are converted into a data structure dedicated to the xgboost model by the data transcoding layer, and the first important feature is calculated by the first xgboost model to construct the gray level co-occurrence matrix. After the matrix is ​​normalized, the Haralick feature group with strong anti-interference ability can be calculated.

[0017] Furthermore, the lidar feature extraction module includes a ground point extraction layer, an elevation calculation layer, a second normalization layer, and an intensity calculation layer;

[0018] The input end of the ground point extraction layer and the input end of the second normalization layer are both used as the input end of the lidar feature extraction module; the output end of the ground point extraction layer is connected to the input end of the elevation calculation layer; the output end of the second normalization layer is connected to the input end of the intensity calculation layer; the output end of the elevation calculation layer and the output end of the intensity calculation layer are both used as the output end of the lidar feature extraction module.

[0019] The beneficial effects of the above further scheme are as follows: in the present invention, the ground extraction layer first extracts the ground information of the lidar data and then calculates the corresponding elevation features through the elevation calculation layer; the second normalization layer processes the lidar data and obtains the corresponding intensity features through the intensity calculation layer. The two features together constitute the lidar feature group, which can be used to counteract the interference information with the height difference between the mangroves and the complex landforms.

[0020] Further, the feature screening unit includes a second data transcoding layer, a second xgboost model training layer, a third xgboost model training layer, a fourth xgboost model training layer, a second feature importance calculation layer, a second feature separation layer and a multiplier U1;

[0021] The input end of the second data transcoding layer serves as the input end of the feature screening unit; the first output end, the second output end and the third output end of the second data transcoding layer are respectively connected to the input end of the second xgboost model training layer, the input end of the third xgboost model training layer and the input end of the fourth xgboost model training layer in one-to-one correspondence; the output end of the second xgboost model training layer, the output end of the third xgboost model training layer and the output end of the fourth xgboost model training layer are respectively connected to the first input end, the second input end and the third input end of the second feature importance calculation layer in one-to-one correspondence; the first output end, the second output end and the third output end of the second feature importance calculation layer are respectively connected to the first input end, the second input end and the third input end of the second feature separation layer in one-to-one correspondence; the first output end, the second output end and the third output end of the second feature separation layer are respectively connected to the first input end, the second input end and the third input end of the multiplier U1 in one-to-one correspondence; the output end of the multiplier U1 serves as the output end of the feature screening unit.

[0022] The beneficial effect of the above further scheme is as follows: in the present invention, three groups of feature groups are converted from the data transcoding layer into a data structure dedicated to the xgboost model, and the feature importance is calculated by the second, third, and fourth xgboost models respectively, and then the features with a similarity higher than 70% are eliminated, and the remaining features are separated and then enter the multiplier U1 together with the multispectral features to form a feature group for fine classification of mangroves. The fused feature group can show a strong classification ability in the fine classification of mangroves and the ability to resist interference in complex landforms.

[0023] Furthermore, the expression of the data dimension reduction layer is:

[0024] ;

[0025] ;

[0026] U=[p1,p2,…,pk] ;

[0027] In the formula, , represents the noise covariance matrix, represents the final signal estimate, Indicates that the main components that can represent the signal are selected from the eigenvectors of the noise covariance matrix. feature vectors, Represents the signal portion of the original input data mapped to the feature vector, represents the transpose operation, represents the original input data, represents the eigenvector of the first largest eigenvalue in the noise covariance matrix, represents the eigenvector of the second largest eigenvalue in the noise covariance matrix, represents the noise covariance matrix The eigenvector with the largest eigenvalue, Indicates the amount of data, represents the identity matrix, represents a Gaussian filter.

[0028] The beneficial effect of the above further scheme is: the present invention separates various environmental noises such as shadows, water ripples and other interference signals in the original multispectral data, and reduces the feature dimension by removing the noise. Therefore, the present invention utilizes the characteristics of data dimensionality reduction to improve the image signal-to-noise ratio, enhances the characteristic distinction between mangroves and non-mangroves, and then improves the anti-interference and generalization capabilities of the classification model, thereby achieving enhanced fine classification of mangroves.

[0029] Furthermore, the expression of the first feature importance calculation layer is:

[0030] ;

[0031] T i = 1 2 [ G L 2 H L +μ + G R 2 H R +μ + ( G L + G R ) 2 H L + H R +μ ]-γ ;

[0032] In the formula, represents the feature importance, represents the feature gain, Indicates the corresponding features The number of times the tree was used for splits, Indicates the corresponding features The number of times the tree was used for splits, Indicates the corresponding features The number of samples to split child nodes at in the tree, Indicates the corresponding features The number of samples to split child nodes at in the tree, represents the gradient statistics of the left subtree, represents the gradient statistics of the right subtree, represents the second-order gradient statistics of the left subtree, represents the second-order gradient statistics of the right subtree, represents the L2 regularization coefficient, Represents the complexity penalty coefficient of the tree.

[0033] The beneficial effect of the above further scheme is: the present invention uses the number of influences of each feature on the samples at the split node in the xgboost model as the feature importance, and the average gain after feature splitting as the feature importance, and both use the proportion of the number of times each feature is used as a splitting criterion in all trees to all instances as the weight, thereby further increasing its reliability as the basis for feature importance.

[0034] The beneficial effects of the present invention are as follows: the present invention obtains the original multispectral image after preprocessing the collected multispectral UAV data, obtains the original high-resolution panchromatic image after preprocessing the collected panchromatic UAV data, and obtains the original lidar point cloud data after preprocessing the collected laser radar UAV data; the original multispectral image and the original lidar data are processed by a feature extraction unit to obtain four types of original features, namely, improved texture features, vegetation index, lidar features, and multispectral features; the four types of original features are processed by a feature screening unit to obtain a fine classification feature group; finally, the target complex landform dense mangroves are finely classified by the fine classification feature group to obtain a high-precision classification result. In the present invention, the texture features are further improved by changing the dimensionality reduction method of the image, and the multispectral features, lidar features, and vegetation index are combined for target recognition at the same time, thereby improving the accuracy of fine classification of complex landform dense mangroves. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Flow chart of the method for fine classification of dense mangroves in complex terrain using drones;

[0036] Figure 2 It is a structural diagram of a feature extraction unit;

[0037] Figure 3 A schematic diagram of the structure of the improved texture feature extraction module;

[0038] Figure 4 This is a structural diagram of the lidar feature extraction module;

[0039] Figure 5 It is a structural schematic diagram of a feature screening unit;

[0040] Figure 6 A comparison chart of classification accuracy of a fusion feature group ablation experiment based on the improved texture feature of the present invention;

[0041] Figure 7 A comparison chart of classification accuracy between a fusion feature group based on the improved texture features of the present invention and a mainstream texture feature fusion feature group in a control experiment;

[0042] Figure 8This is a comparison chart of the classification accuracy of the improved mangrove fine classification feature group of the present invention and the mainstream single feature control experiment. DETAILED DESCRIPTION

[0043] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0044] like Figure 1 As shown, the present invention provides a method for fine classification of dense mangroves in complex terrain using an unmanned aerial vehicle, comprising the following steps:

[0045] S1. Collect drone data of mangrove ecological areas, including multispectral drone data, panchromatic drone data and lidar drone data;

[0046] S2, preprocessing the UAV data using a data preprocessing unit to obtain multispectral raw data, panchromatic raw data, and lidar raw data;

[0047] S3, using a sample label acquisition unit to divide the full-color original data into sample sets;

[0048] S4, using a feature extraction unit to process the multispectral raw data, the laser radar raw data, and the output result of the sample label acquisition unit to obtain classification features;

[0049] S5, using a feature screening unit to process the classification features to obtain an improved mangrove fine classification feature group;

[0050] S6. Use the ablative classification units to process the improved mangrove fine classification feature group and complete the classification.

[0051] In the embodiment of the present invention, Figure 2 As shown, the feature extraction unit includes a label conversion layer, an improved texture feature extraction module, a vegetation index calculation layer, a lidar feature extraction module, a first adder A1, a second adder A2 and a third adder A3;

[0052] The input end of the label conversion layer serves as the input end of the feature extraction unit and is connected to the output end of the sample label acquisition unit; the first output end of the label conversion layer is connected to the first input end of the improved texture feature extraction module; the second output end of the label conversion layer is respectively connected to the first input end of the first adder, the first input end of the second adder and the first input end of the third adder; the second input end of the improved texture feature extraction module is connected to the first output end of the data preprocessing unit; the output end of the improved texture feature extraction module is connected to the second input end of the first adder A1; the input end of the vegetation index calculation layer is connected to the second output end of the data preprocessing unit; the output end of the vegetation index calculation layer is connected to the second input end of the second adder A2; the input end of the lidar feature extraction module is connected to the third output end of the data preprocessing unit; the output end of the lidar feature extraction module is connected to the second input end of the third adder A3; the output end of the first adder A1, the output end of the second adder A2 and the output end of the third adder A3 are all used as the output end of the feature extraction unit.

[0053] In the present invention, three kinds of feature data are obtained by processing the multispectral data and lidar data of the drone, and then combined with the label data output by the label conversion layer to form independent feature groups in each adder. Each feature group has multiple single features, thereby providing a large amount of material for subsequent feature fusion.

[0054] In the embodiment of the present invention, Figure 3 As shown, the improved texture feature extraction module includes a data dimension reduction layer, a first data transcoding layer, a first xgboost model training layer, a first feature importance calculation layer, a first feature separation layer, a gray level co-occurrence matrix construction layer, a first normalization layer and a Haralick feature calculation layer;

[0055] The input end of the data dimension reduction layer serves as the first input end of the improved texture feature extraction module; the first input end of the first data transcoding layer serves as the second input end of the improved texture feature extraction module; the output end of the data dimension reduction layer is connected to the second input end of the first data transcoding layer; the output end of the first data transcoding layer, the first xgboost model training layer, the first feature importance calculation layer, the first feature separation layer, the gray level co-occurrence matrix construction layer, the first normalization layer and the input end of the Haralick feature calculation layer are connected in sequence; the output end of the Haralick feature calculation layer serves as the output end of the improved texture feature extraction module.

[0056] In the present invention, the data dimension reduction layer processes the multispectral features to improve the dimension reduction method to achieve feature concentration and reduce the interference of noise on feature extraction. The data after dimension reduction and the sample features are converted into a data structure dedicated to the xgboost model by the data transcoding layer. The first important feature is calculated by the first xgboost model to construct a gray level co-occurrence matrix. After the matrix is ​​normalized, the Haralick feature group with strong anti-interference ability can be calculated.

[0057] In the embodiment of the present invention, Figure 4 As shown, the lidar feature extraction module includes a ground point extraction layer, an elevation calculation layer, a second normalization layer, and an intensity calculation layer;

[0058] The input end of the ground point extraction layer and the input end of the second normalization layer are both used as the input end of the lidar feature extraction module; the output end of the ground point extraction layer is connected to the input end of the elevation calculation layer; the output end of the second normalization layer is connected to the input end of the intensity calculation layer; the output end of the elevation calculation layer and the output end of the intensity calculation layer are both used as the output end of the lidar feature extraction module.

[0059] In the present invention, the ground extraction layer first extracts the ground information of the lidar data and then calculates the corresponding elevation features through the elevation calculation layer; the second normalization layer processes the lidar data and obtains the corresponding intensity features through the intensity calculation layer. The two features together constitute the lidar feature group, which can be used to counteract the interference information with the height difference between the mangroves and the complex landforms.

[0060] In the embodiment of the present invention, Figure 5 As shown, the feature screening unit includes a second data transcoding layer, a second xgboost model training layer, a third xgboost model training layer, a fourth xgboost model training layer, a second feature importance calculation layer, a second feature separation layer and a multiplier U1;

[0061] The input end of the second data transcoding layer serves as the input end of the feature screening unit; the first output end, the second output end and the third output end of the second data transcoding layer are respectively connected to the input end of the second xgboost model training layer, the input end of the third xgboost model training layer and the input end of the fourth xgboost model training layer in one-to-one correspondence; the output end of the second xgboost model training layer, the output end of the third xgboost model training layer and the output end of the fourth xgboost model training layer are respectively connected to the first input end, the second input end and the third input end of the second feature importance calculation layer in one-to-one correspondence; the first output end, the second output end and the third output end of the second feature importance calculation layer are respectively connected to the first input end, the second input end and the third input end of the second feature separation layer in one-to-one correspondence; the first output end, the second output end and the third output end of the second feature separation layer are respectively connected to the first input end, the second input end and the third input end of the multiplier U1 in one-to-one correspondence; the output end of the multiplier U1 serves as the output end of the feature screening unit.

[0062] In the present invention, three groups of feature groups are converted from the data transcoding layer to a data structure dedicated to the xgboost model, and the feature importance is calculated by the second, third, and fourth xgboost models respectively, and then the features with a similarity higher than 70% are eliminated, and the remaining features are separated and combined with the multispectral features into a feature group for fine classification of mangroves by multiplier U1. The fused feature group can show a strong classification ability in the fine classification of mangroves and an anti-interference ability to cope with complex landforms.

[0063] In the present invention, by fine-tuning different xgboost models in a targeted manner, a reliable model environment is provided for different feature screening, thereby ensuring that each set of features has its own feature specificity. In the debugging model, a sub-dataset is randomly extracted from the data set to ensure that the model has high versatility and avoid screening out singular value features that are only valid locally.

[0064] In the embodiment of the present invention, the first xgboost model, the second xgboost model, the third xgboost model and the fourth xgboost model are all subjected to multiple rounds of model fine-tuning before model training, and the specific steps are as follows:

[0065] A1. Randomly extract 10 sub-datasets from the feature dataset for model adjustment, use the first fine-tuning model as the substrate while retaining the seed, and perform multiple rounds of model training by changing the feature dataset;

[0066] A2. Confirm the results of each round of model training, remove the best and worst results, and if the remaining results have a difference of more than 10% with the first fine-tuning model results and account for more than or equal to one-third, the model is considered not to be generalizable and the seed is retained for a second round of fine-tuning; otherwise, the model is considered to be generalizable and can be used to calculate feature importance.

[0067] In the present invention, by fine-tuning different xgboost models in a targeted manner, a reliable model environment is provided for different feature screening, thereby ensuring that each set of features has its own feature specificity. In the debugging model, a sub-dataset is randomly extracted from the data set to ensure that the model has high versatility and avoid screening out singular value features that are only valid locally.

[0068] In the embodiment of the present invention, the expression of the data dimension reduction layer is:

[0069] ;

[0070] ;

[0071] U=[p1,p2,…,pk] ;

[0072] In the formula, , represents the noise covariance matrix, represents the final signal estimate, Indicates that the main components that can represent the signal are selected from the eigenvectors of the noise covariance matrix. feature vectors, Represents the signal portion of the original input data mapped to the feature vector, represents the transpose operation, represents the original input data, represents the eigenvector of the first largest eigenvalue in the noise covariance matrix, represents the eigenvector of the second largest eigenvalue in the noise covariance matrix, represents the noise covariance matrix The eigenvector of the largest eigenvalue, Indicates the amount of data, represents the identity matrix, represents a Gaussian filter.

[0073] The present invention separates various environmental noises such as shadows, water ripples and other interference signals in the original multispectral data, and reduces the feature dimension by removing the noise. Therefore, the present invention utilizes the characteristics of data dimensionality reduction to improve the image signal-to-noise ratio, enhances the characteristic distinction between mangroves and non-mangroves, and further improves the anti-interference and generalization capabilities of the classification model, thereby achieving enhanced fine classification of mangroves.

[0074] In this embodiment of the present invention, the expression of the first feature importance calculation layer is:

[0075] ;

[0076] T i = 1 2 [ G L 2 H L +μ + G R 2 H R +μ + ( G L + G R ) 2 H L + H R +μ ]-γ ;

[0077] In the formula, represents the feature importance, represents the feature gain, Indicates the corresponding features The number of times the tree was used for splits, Indicates the corresponding features The number of times the tree was used for splits, Indicates the corresponding features The number of samples to split child nodes at in the tree, Indicates the corresponding features The number of samples to split child nodes at in the tree, represents the gradient statistics of the left subtree, represents the gradient statistics of the right subtree, represents the second-order gradient statistics of the left subtree, represents the second-order gradient statistics of the right subtree, represents the L2 regularization coefficient, Represents the complexity penalty coefficient of the tree. Here we take 0. Here we take 0.1.

[0078] The present invention uses the number of samples affected by each feature in the split node in the xgboost model and the average gain after feature splitting as feature importance, and uses the proportion of the number of times each feature is used as a splitting criterion in all trees to all instances as weight, further increasing its reliability as a basis for feature importance. The expression of the first feature importance calculation layer is the same as the expression of the second feature importance calculation layer.

[0079] In the embodiment of the present invention, all mangrove fine classification models are subjected to multiple rounds of decision tree number adjustment before model training, and the specific steps are as follows:

[0080] B1. Set the step size n (n=50, 25, 10...1) within the decision tree threshold range (min: 1, max: 500), input it into the model one by one, and use the training set of the feature data set adjusted with the number of decision trees to train, and obtain multiple training models;

[0081] B2. Use the validation set of the feature data set used to adjust the number of decision trees to perform accuracy verification and screening on each round of models trained in B1, gradually narrow the screening interval, and finally select the model with the best performance, and use the number of its decision trees as the result;

[0082] B3. Use the test set of the feature data set to adjust the number of decision trees to perform performance test on the model to be tested. If the effect is poor, return to B2 and use the ranking result.

[0083] In the present invention, targeted adjustments are made to different random forest model decision trees to provide an optimal environment for the classification of different feature groups, thereby maximizing the role of mixed features in the fine classification of mangroves while ensuring that the classification results of each feature group are the optimal results of the feature group, thereby confirming the excellent performance of the fusion features.

[0084] like Figure 6 As shown, an ablation experiment is performed by removing different features and removing the feature while removing the improved texture feature composition of the present invention; the evaluation indicators are overall accuracy (used to measure the overall accuracy of the classification, with a value range of 0 to 1), Kappa coefficient (used to measure the consistency between the predicted results and the actual observation results, with a value range of -1 to 1), allocation inconsistency (used to measure the proportion of samples that are not correctly assigned to their true categories, with a value range of 0 to 1), and quantity inconsistency (used to measure the proportion of samples that are not accurately predicted for each category, with a value range of 0 to 1).

[0085] like Figure 7 As shown in the figure, the classification feature group based on the improved texture feature fusion of the present invention is compared with the classification feature group based on other mainstream texture feature fusions in the fine classification of mangroves; the evaluation indicators are overall accuracy (used to measure the overall accuracy of the classification, with a value range of 0 to 1), Kappa coefficient (used to measure the consistency between the predicted results and the actual observation results, with a value range of -1 to 1), allocation inconsistency (used to measure the proportion of samples that are not correctly assigned to their true categories, with a value range of 0 to 1), and quantity inconsistency (used to measure the proportion of samples that are not accurately predicted for each category, with a value range of 0 to 1).

[0086] like Figure 8As shown in the figure, the improved mangrove fine classification feature group of the present invention is compared with the mainstream single feature in the mangrove fine classification; the evaluation indicators are overall accuracy (used to measure the overall accuracy of the classification, with a value range of 0 to 1), Kappa coefficient (used to measure the consistency between the predicted results and the actual observation results, with a value range of -1 to 1), allocation inconsistency (used to measure the proportion of samples that failed to be correctly assigned to their true categories, with a value range of 0 to 1), and quantity inconsistency (used to measure the proportion of samples that failed to accurately predict the number of samples in each category, with a value range of 0 to 1).

[0087] The present invention collects drone images of Gaoqiao mangrove ecological zone in Zhanjiang City, Guangdong Province, pre-processes the collected multispectral drone data to obtain original multispectral images, pre-processes the collected panchromatic drone data to obtain original high-resolution panchromatic images, and pre-processes the collected laser radar drone data to obtain original lidar point cloud data. Sample labels are collected through high-resolution panchromatic images, and are divided into four data sets according to their uses, namely, feature data sets for feature screening, feature data sets for model signs, feature data sets for decision tree number adjustment, and feature data sets for fine classification, and each data set is divided into a training set, a test set, and a validation set according to a ratio of 7:2:1. The original multispectral image and the original lidar data are processed by a feature extraction unit to obtain four types of original features, namely, improved texture features, vegetation index, lidar features, and multispectral features; the four types of original features are processed by a feature screening unit to obtain a fine classification feature group. Finally, the target complex landform dense mangroves are finely classified by the fine classification feature group to obtain a high-precision classification result. The present invention further improves the texture characteristics by changing the dimensionality reduction method of the image, and simultaneously combines multispectral features, lidar features and vegetation index for target recognition, thereby improving the accuracy of fine classification of dense mangroves in complex landforms.

[0088] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for fine classification of dense mangroves in complex terrain by drone, characterized in that: The following steps are involved: S1. Collect drone data of mangrove ecological areas, including multispectral drone data, panchromatic drone data and lidar drone data; S2, preprocessing the UAV data using a data preprocessing unit to obtain multispectral raw data, panchromatic raw data, and lidar raw data; S3, using a sample label acquisition unit to divide the full-color original data into sample sets; S4, using a feature extraction unit to process the multispectral raw data, the laser radar raw data, and the output result of the sample label acquisition unit to obtain classification features; S5, using a feature screening unit to process the classification features to obtain an improved mangrove fine classification feature group; S6. Use the ablative classification unit to process the improved mangrove fine classification feature group and complete the classification; The feature extraction unit includes a label conversion layer, an improved texture feature extraction module, a vegetation index calculation layer, a lidar feature extraction module, a first adder A1, a second adder A2 and a third adder A3; The input end of the label conversion layer is used as the input end of the feature extraction unit and is connected to the output end of the sample label acquisition unit; the first output end of the label conversion layer is connected to the first input end of the improved texture feature extraction module; the second output end of the label conversion layer is respectively connected to the first input end of the first adder, the first input end of the second adder and the first input end of the third adder; the second input end of the improved texture feature extraction module is connected to the first output end of the data preprocessing unit; the output end of the improved texture feature extraction module is connected to the second input end of the first adder A1; the input end of the vegetation index calculation layer is connected to the second output end of the data preprocessing unit; the output end of the vegetation index calculation layer is connected to the second input end of the second adder A2; the input end of the lidar feature extraction module is connected to the third output end of the data preprocessing unit; the output end of the lidar feature extraction module is connected to the second input end of the third adder A3; the output end of the first adder A1, the output end of the second adder A2 and the output end of the third adder A3 are all used as the output end of the feature extraction unit; The improved texture feature extraction module includes a data dimension reduction layer, a first data transcoding layer, a first xgboost model training layer, a first feature importance calculation layer, a first feature separation layer, a gray level co-occurrence matrix construction layer, a first normalization layer and a Haralick feature calculation layer; The input end of the data dimension reduction layer serves as the first input end of the improved texture feature extraction module; the first input end of the first data transcoding layer serves as the second input end of the improved texture feature extraction module; the output end of the data dimension reduction layer is connected to the second input end of the first data transcoding layer; the output end of the first data transcoding layer, the first xgboost model training layer, the first feature importance calculation layer, the first feature separation layer, the gray level co-occurrence matrix construction layer, the first normalization layer and the input end of the Haralick feature calculation layer are connected in sequence; the output end of the Haralick feature calculation layer serves as the output end of the improved texture feature extraction module; The lidar feature extraction module includes a ground point extraction layer, an elevation calculation layer, a second normalization layer and an intensity calculation layer; The input end of the ground point extraction layer and the input end of the second normalization layer are both used as the input end of the lidar feature extraction module; the output end of the ground point extraction layer is connected to the input end of the elevation calculation layer; the output end of the second normalization layer is connected to the input end of the intensity calculation layer; the output end of the elevation calculation layer and the output end of the intensity calculation layer are both used as the output end of the lidar feature extraction module; The feature screening unit includes a second data transcoding layer, a second xgboost model training layer, a third xgboost model training layer, a fourth xgboost model training layer, a second feature importance calculation layer, a second feature separation layer and a multiplier U1; The input end of the second data transcoding layer serves as the input end of the feature screening unit; the first output end, the second output end and the third output end of the second data transcoding layer are respectively connected to the input end of the second xgboost model training layer, the input end of the third xgboost model training layer and the input end of the fourth xgboost model training layer in one-to-one correspondence; the output end of the second xgboost model training layer, the output end of the third xgboost model training layer and the output end of the fourth xgboost model training layer are respectively connected to the first input end, the second input end and the third input end of the second feature importance calculation layer in one-to-one correspondence; the first output end, the second output end and the third output end of the second feature importance calculation layer are respectively connected to the first input end, the second input end and the third input end of the second feature separation layer in one-to-one correspondence; the first output end, the second output end and the third output end of the second feature separation layer are respectively connected to the first input end, the second input end and the third input end of the multiplier U1 in one-to-one correspondence; the output end of the multiplier U1 serves as the output end of the feature screening unit.

2. The method for fine classification of dense mangroves in complex landforms by drone according to claim 1 is characterized in that: The expression of the data dimension reduction layer is: ； ； ； ； In the formula, represents the noise covariance matrix, represents the final signal estimate, Indicates that the main components that can represent the signal are selected from the eigenvectors of the noise covariance matrix. feature vectors, Represents the signal portion of the original input data mapped to the feature vector, represents the transpose operation, represents the original input data, represents the eigenvector of the first largest eigenvalue in the noise covariance matrix, represents the eigenvector of the second largest eigenvalue in the noise covariance matrix, represents the noise covariance matrix The eigenvector of the largest eigenvalue, Indicates the amount of data, represents the identity matrix, represents a Gaussian filter.

3. The method for fine classification of dense mangroves in complex landforms by drone according to claim 1 is characterized in that: The expression of the first feature importance calculation layer is: ； ； In the formula, represents the feature importance, represents the feature gain, Indicates the corresponding features The number of times the tree was used for splits, Indicates the corresponding features The number of times the tree was used for splits, Indicates the corresponding features The number of samples to split child nodes at in the tree, Indicates the corresponding features The number of samples to split child nodes at in the tree, represents the gradient statistics of the left subtree, represents the gradient statistics of the right subtree, represents the second-order gradient statistics of the left subtree, represents the second-order gradient statistics of the right subtree, represents the L2 regularization coefficient, represents the complexity penalty coefficient of the tree, Indicates the number of nodes in the tree.

Images