Earth surface plant classification statistical method based on unmanned aerial vehicle aerial image

Abstract

The invention relates to the technical field of classification statistics, in particular to a surface plant classification statistics method based on unmanned aerial vehicle aerial images, which comprises the following steps of: defining latitude and longitude boundaries and a total area of a target region, and setting unmanned aerial vehicle flight parameters, plant classification categories and statistical indexes; selecting a multispectral unmanned aerial vehicle and a camera which are adaptive to each other, completing camera interior orientation element calibration and radiation correction, and obtaining core parameters; according to the invention, classification is carried out by integrating spectrum, texture and form features and combining a machine learning model, so that the precision of plant classification is remarkably improved, and subjective errors are reduced; by matching ground verification and a model correction mechanism, the reliability of a classification result is further guaranteed, meanwhile, multi-dimensional information of distribution, number and coverage area of plants can be comprehensively obtained, data integrity is high and traceable, the degree of standardization of an operation process is high, and the method is suitable for ground surface plant classification statistics requirements under different scales and different terrain conditions. And the application scene is wider.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of classification statistics, and particularly to a surface plant classification and statistics method based on unmanned aerial vehicle aerial images. BACKGROUND

[0002] Surface plant classification is a basic work that divides different plants into several categories with clear distinguishing standards according to inherent properties such as morphological characteristics, growth habits and ecological properties, and combines the growth and distribution state of the plants on the ground, so as to clearly define the exclusive characteristics and distribution range of each plant, and lay a classification foundation for subsequent plant resource investigation, ecological research and other work.

[0003] Surface plant classification is an important prerequisite for ecological environment monitoring, natural resource management, biodiversity protection and regional planning and construction, and through classification, the composition structure, distribution law and growth condition of surface plants can be accurately mastered, which provides a scientific basis for analyzing ecological system stability, evaluating vegetation restoration effect, formulating biodiversity protection strategies and reasonably planning land use methods, and also provides data support for practical applications such as agricultural production optimization, forestry resource cultivation and ecological disaster prevention.

[0004] However, generally, the traditional surface plant classification method mainly includes manual field investigation method, sampling method, line transect method and early remote sensing monitoring method, and these methods have obvious limitations. The manual field investigation method and the sampling method and the line transect method need a large amount of manpower and material resources, have a long investigation period, are difficult to apply to large-scale and complex terrain areas, have low data acquisition efficiency, are greatly affected by the subjective judgment of the investigators, and are prone to classification deviation. The early remote sensing monitoring method has problems of insufficient spatial resolution and limited classification accuracy, is difficult to accurately distinguish closely related species or low and hidden plant types, cannot obtain detailed plant growth information, has high professional requirements for data interpretation and cannot guarantee the consistency of the results.

[0005] In summary, a surface plant classification and statistics method based on unmanned aerial vehicle aerial images is needed to solve the above problems. SUMMARY

[0006] The present application relates to the technical field of classification statistics, and particularly to a surface plant classification and statistics method based on unmanned aerial vehicle aerial images.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0008] The present application provides a surface plant classification and statistics method based on unmanned aerial vehicle aerial images, comprising the following steps:

[0009] S1. Define the target area latitude and longitude boundary and total area, set the unmanned aerial vehicle flight parameters, plant classification categories and statistical indicators;

[0010] S2. Select a suitable multi-spectral unmanned aerial vehicle and camera, complete the camera internal orientation element calibration and radiation correction, and obtain the core parameters;

[0011] S3. Fly according to the planned route, collect multi-spectral images with the set parameters, monitor the clarity in real time and supplement the shooting, and form a complete data set;

[0012] S4. Distortion correction, splicing and radiation correction are performed on the original images to generate a target area orthographic image with a resolution that meets the standard;

[0013] S5. Extract the vegetation area image by calculating the NDVI and setting the threshold, and count the total number of vegetation pixels;

[0014] S6. Extract the spectrum, texture and morphology of the vegetation, and integrate them to form a 14-dimensional feature vector for each vegetation pixel;

[0015] S7. Divide the sample set to train the random forest model, and after verification, apply it to all feature vectors to generate a preliminary classification image;

[0016] S8. Set up ground verification sample plots to collect field data, compare and correct the classification model, and generate a precision classification image with a Kappa coefficient that meets the standard;

[0017] S9. Based on the precision classification image, count the number of vegetation pixels, total coverage area and number of plants that meet the threshold for each type of plant;

[0018] S10. According to the total coverage area of each type of plant obtained in S9 and the total area of the survey area in S1, calculate the coverage of each type of plant, and simultaneously summarize the plant number and total coverage area data in S9, combine the region boundary information in S1 and the classification distribution image in S8, and generate a ground plant classification statistical report.

[0019] Preferably, the implementation steps of step S1 are:

[0020] S1.1. Clearly define the target area of the ground plant classification statistics, import the latitude and longitude boundary data of the target area through GIS software, specifically 116.32°-116.35° east longitude and 39.92°-39.95° north latitude, complete the spatial definition of the survey area;

[0021] S1.2. Based on the defined spatial range, determine the total area of the survey area, and set the core flight parameters of the unmanned aerial vehicle;

[0022] S1.3. The explicit classification and statistical core target, the classification target covers four types of plants, i.e. arbor, shrub, herb and liana, and the statistical index is determined as the number, coverage area and coverage of each type of plant.

[0023] Preferably, the implementation steps of step S2 are:

[0024] S2.1. According to the ground resolution requirement set in S1.2, a quadcopter unmanned aerial vehicle carrying a 6-channel multispectral camera is selected to ensure that the payload performance matches the investigation accuracy requirement;

[0025] S2.2. Confirm the core parameters of the multispectral camera;

[0026] S2.3. Carry out camera interior orientation element calibration, obtain the principal point coordinate pixel and focal length through a professional calibration board, and complete the basic data acquisition of distortion correction;

[0027] S2.4. Perform camera radiation correction, determine the correction coefficient of each waveband through calibration experiment, and provide basis for subsequent image radiation correction.

[0028] Preferably, the implementation steps of step S3 are:

[0029] S3.1. Based on the latitude and longitude boundary delineated in S1.1 and the flight height set in S1.2, a grid-shaped full-coverage route is designed using the unmanned aerial vehicle ground station software to ensure that there is no missing investigation area;

[0030] S3.2. Set the flight and image acquisition auxiliary parameters, image acquisition time interval and route inflection point hovering time to ensure the continuity of image acquisition;

[0031] S3.3. Start the real-time positioning function of the unmanned aerial vehicle, rely on the GPS system to realize positioning with an accuracy of ±1m, execute multispectral image acquisition according to the planned route, and record the information of the acquisition process synchronously;

[0032] S3.4. Real-time monitor the clarity of the acquired images, if there are blurred frames, trigger the automatic re-shooting mechanism, finally obtain a complete data set containing 3200 original images, which meets the subsequent processing requirements.

[0033] Preferably, the implementation steps of step S4 are:

[0034] S4.1. Based on the camera interior orientation elements obtained in S2.3, perform distortion correction on the 3200 original multispectral images collected in S3.4 to eliminate the influence of lens optical distortion;

[0035] S4.2. According to the heading 80% and lateral 70% overlap rate set in S1.2, use a stitching algorithm based on feature point matching, set the feature point detection threshold to 0.03, and stitch the distortion-corrected images into complete area images.

[0036] S4.3. Apply the radiation correction coefficient determined in S2.4 to the complete image after splicing to perform radiation correction, and remove the interference caused by illumination changes and atmospheric scattering;

[0037] S4.4. Integrate the correction and splicing results to generate a target area orthographic image with a resolution of 5 cm and a size of 6000 pixels x 5000 pixels, thereby providing basic data for subsequent vegetation extraction.

[0038] Preferably, the implementation steps of step S5 are as follows:

[0039] S5.1. Based on the 5 cm resolution orthographic image generated in S4.4, extract the normalized difference vegetation index of each band reflectance data;

[0040] S5.2. Set the NDVI threshold value to 0.3, and clearly define the pixels with NDVI ≥ 0.3 as vegetation pixels and the pixels with NDVI < 0.3 as non-vegetation pixels to establish a vegetation identification standard;

[0041] S5.3. Perform binary processing on the orthographic image according to the set NDVI threshold value to separate the vegetation and non-vegetation areas, and extract a complete vegetation coverage image consistent with the survey area boundary defined in S1.1;

[0042] S5.4. Count the total number of vegetation pixels in the vegetation coverage image to finally determine the total number of vegetation pixels and provide a data volume reference for subsequent feature extraction.

[0043] Preferably, the implementation steps of step S6 are as follows:

[0044] S6.1. For the vegetation coverage image obtained in S5.3, extract the reflectance mean value features of the 6 multispectral bands, with the value ranges of the respective bands being 0.12-0.25 for the blue band, 0.20-0.35 for the green band, 0.10-0.22 for the red band, 0.30-0.45 for the red edge band, 0.40-0.65 for the near-infrared band, and 0.25-0.40 for the long-wave infrared band;

[0045] S6.2. Calculate the contrast, entropy, correlation, and uniformity four texture features under a 3x3 window using the gray level co-occurrence matrix method, with the value ranges of the respective indexes being 0-200 for the contrast, 0-4 for the entropy, -1-1 for the correlation, and 0-1 for the uniformity;

[0046] S6.3. Extract the crown area, perimeter, circularity, and aspect ratio four morphological features, set the crown area threshold value to 0.1 m2, the perimeter threshold value to 1.2 m, the circularity to 0.3-1.0, and the aspect ratio to 1.0-3.5, and clearly define the morphological determination standard;

[0047] S6.4. Integrate the three types of features of spectrum, texture, and morphology to form a 14-dimensional feature vector for each vegetation pixel, and the total number of feature vectors is consistent with the total number of vegetation pixels counted in S5.4.

[0048] Preferably, the implementation steps of step S7 are:

[0049] S7.1. Randomly divide the sample set from the feature vectors generated in S6.4, and select 20% as the model training set and the remaining 80% as the model test set;

[0050] S7.2. Configure the random forest classification model parameters, set the number of decision trees to 100, the maximum tree depth to 15 layers, the minimum number of samples for node splitting to 5, and the out-of-bag data error threshold to 0.05, and establish the model basic framework;

[0051] S7.3. Take the four types of plants, i.e., trees, shrubs, herbs, and vines, as the classification labels, train the model using the training set, and verify the classification accuracy of the model through the test set, and stop training when the accuracy reaches 93% or above;

[0052] S7.4. Apply the classification model that has met the training requirements to all feature vectors to determine the type of each vegetation pixel one by one and form the preliminary classification results;

[0053] S7.5. Based on the type determination results of the vegetation pixels, generate a preliminary classification image containing the distribution information of the four types of plants, and keep the image resolution at 5 cm as set in S4.4 to ensure consistent spatial accuracy.

[0054] Preferably, the implementation steps of step S8 are:

[0055] S8.1. Based on the preliminary classification image in S7.5 and the survey area delineated in S1.1, design 60 1m x 1m ground verification quadrats using systematic sampling method, randomly generate the center longitude and latitude of the quadrats through GIS software, and ensure that the quadrats cover the distribution areas of all types of plants;

[0056] S8.2. Go to the field to conduct verification investigation, record the actual type, quantity, and coverage of plants in each quadrat, and form a complete field verification data set;

[0057] S8.3. Compare the field verification data set with the classification results of the corresponding quadrats in the preliminary classification image, calculate the confusion matrix and obtain the Kappa coefficient to evaluate the accuracy of the classification results;

[0058] S8.4. If the Kappa coefficient is lower than 0.88, adjust the random forest model parameters set in S7.2 based on the field data of the validation sample plots, increase the number of decision trees to 120 and adjust the minimum sample size for node splitting to 3, and re-classify the feature vectors in S6.4;

[0059] S8.5. Repeat the validation and adjustment process until the Kappa coefficient reaches 0.88 or above, and finally generate an accurate plant type distribution image to ensure the reliability of the classification results.

[0060] Preferably, the implementation step of step S9 is:

[0061] S9.1. Based on the plant type distribution image corrected in S8.5, count the number of vegetation pixels corresponding to each type of plant for trees, shrubs, herbs and lianas respectively;

[0062] S9.2. Calculate the total coverage area of each type of plant by multiplying the number of vegetation pixels of each type of plant by the area of a single pixel based on the 5cm resolution determined in S4.4;

[0063] S9.3. According to the 0.1m2 crown area threshold set in S6.3, count the number of plants with a single crown area greater than or equal to 0.1m2 in each type of plant, and exclude scattered vegetation with a crown area less than 0.1m2;

[0064] S9.4. Integrate the total coverage area and plant number statistics of each type of plant to form a preliminary statistical result, ensuring that the data corresponds to the classification results in S8.5, and laying a foundation for subsequent coverage calculation.

[0065] Compared with the prior art, the present application has the advantages that: the method of the present application compensates for the shortcomings of traditional methods through high-resolution unmanned aerial vehicles and multispectral images, has the characteristics of high investigation efficiency and wide coverage, can quickly obtain complete vegetation image data of the target area, and avoids the cumbersome process of manual investigation; by integrating spectral, texture and morphological features and combining machine learning models for classification, the accuracy of plant classification is significantly improved, and subjective errors are reduced; with the ground verification and model correction mechanism, the reliability of the classification results is further ensured, and at the same time, multi-dimensional information such as the distribution, number and coverage area of plants can be obtained comprehensively, the data integrity is strong and traceable, the standardization degree of the operation process is high, and the method is suitable for plant classification and statistical requirements under different scales and different terrain conditions, and has a wider application scenario. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A flowchart of a surface plant classification and statistical method based on unmanned aerial vehicle aerial images is shown;

[0067] Figure 2The present application shows a surface plant classification and statistics method based on unmanned aerial vehicle aerial images. DETAILED DESCRIPTION

[0068] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0069] Embodiment 1, please refer to Figures 1-2 The present application provides a surface plant classification and statistics method based on unmanned aerial vehicle aerial images, comprising the following steps:

[0070] S1. Defining the longitude and latitude boundary and total area of the target area, setting the unmanned aerial vehicle flight parameters, plant classification categories and statistics indicators;

[0071] S2. Selecting a suitable multispectral unmanned aerial vehicle and camera, completing camera interior orientation element calibration and radiation correction, and obtaining core parameters;

[0072] S3. Flying according to the planned route, collecting multispectral images with the set parameters, monitoring the clarity in real time and supplementing the shooting, and forming a complete data set;

[0073] S4. Distortion correction, splicing and radiation correction are performed on the original images to generate a target area orthographic image map with up-to-standard resolution;

[0074] S5. Extracting the vegetation area image by calculating the NDVI and setting the threshold, and counting the total number of vegetation pixels;

[0075] S6. Extracting the spectrum, texture and morphology of the vegetation, and integrating to form a 14-dimensional feature vector of each vegetation pixel;

[0076] S7. Dividing the sample set to train the random forest model, verifying that the standard is met, and applying it to all feature vectors to generate a preliminary classification image;

[0077] S8. Setting up ground verification quadrats to collect field data, comparing and correcting the classification model, and generating a precision classification image with a Kappa coefficient up to standard;

[0078] S9. Based on the precision classification image, counting the number of vegetation pixels, total coverage area and number of plants meeting the threshold of each type of plant;

[0079] S10. Calculate the coverage of each type of plant according to the total coverage area of each type of plant obtained in S9 and the total area of the survey region in S1, and meanwhile, aggregate the plant quantity and total coverage area data in S9, combine the region boundary information in S1 and the classification distribution image in S8, and generate a ground plant classification statistics report.

[0080] In the embodiment, it is also necessary to explain that the implementation steps of step S1 are as follows:

[0081] S1.1. Define the target region of the ground plant classification statistics, import the longitude and latitude boundary data of the target region through the GIS software, specifically, 116.32°-116.35° east longitude and 39.92°-39.95° north latitude, and complete the spatial definition of the survey range;

[0082] S1.2. Determine the total area of the survey region based on the defined spatial range, and set the core flight parameters of the unmanned aerial vehicle;

[0083] S1.3. Define the core target of classification and statistics, the classification target covers four types of plants, namely, trees, shrubs, herbs and lianas, and the statistical indicators are determined as the quantity, coverage area and coverage of each type of plant.

[0084] In the embodiment, it is also necessary to explain that the implementation steps of step S2 are as follows:

[0085] S2.1. According to the ground resolution requirement set in S1.2, a quadcopter unmanned aerial vehicle carrying a 6-channel multispectral camera is selected to ensure that the payload performance matches the survey accuracy requirement;

[0086] S2.2. Confirm the core parameters of the multispectral camera;

[0087] S2.3. Carry out camera interior orientation element calibration, obtain the principal point coordinate pixel and focal length through a professional calibration board, and complete the distortion correction basic data acquisition;

[0088] S2.4. Perform camera radiation correction, determine the correction coefficients of each waveband through calibration experiments, and provide a basis for subsequent image radiation correction.

[0089] In the embodiment, it is also necessary to explain that the implementation steps of step S3 are as follows:

[0090] S3.1. Based on the longitude and latitude boundary defined in S1.1 and the flight height set in S1.2, design a grid-shaped full-coverage route using the unmanned aerial vehicle ground station software to ensure that there is no missed survey region;

[0091] S3.2. Set the flight and image acquisition auxiliary parameters, image acquisition time interval and route inflection point hovering time to ensure the continuity of image acquisition;

[0092] S3.3. Start the real-time positioning function of the UAV, rely on the GPS system to achieve a positioning accuracy of ±1 m, and execute multispectral image acquisition according to the planned flight route, and record the information of the acquisition process synchronously;

[0093] S3.4. Real-time monitor the clarity of the acquired images, and if a blurred frame appears, trigger an automatic retake mechanism, and finally obtain a complete data set containing 3200 original images, which meets the subsequent processing requirements.

[0094] In this embodiment, it should also be noted that the implementation steps of step S4 are:

[0095] S4.1. Based on the camera interior orientation elements obtained in S2.3, perform distortion correction on the 3200 original multispectral images acquired in S3.4 to eliminate the influence of lens optical distortion;

[0096] S4.2. According to the 80% forward and 70% lateral overlap rate set in S1.2, use a stitching algorithm based on feature point matching, set the feature point detection threshold to 0.03, and stitch the distortion-corrected images into complete area images;

[0097] S4.3. Apply the radiation correction coefficient determined in S2.4 to perform radiation correction on the stitched complete images to remove the interference caused by changes in light and atmospheric scattering;

[0098] S4.4. Integrate the correction and stitching results to generate a target area orthophoto map with a resolution of 5 cm and a size of 6000 pixels x 5000 pixels, which provides basic data for subsequent vegetation extraction.

[0099] In this embodiment, it should also be noted that the implementation steps of step S5 are:

[0100] S5.1. Based on the 5 cm resolution orthophoto map generated in S4.4, extract each band reflectance data normalized vegetation index;

[0101] S5.2. Set the NDVI decision threshold to 0.3, and clearly define that NDVI≥0.3 is a vegetation pixel and NDVI<0.3 is a non-vegetation pixel, and establish a vegetation identification standard;

[0102] S5.3. Perform binaryzation processing on the orthophoto map according to the set NDVI threshold, separate the vegetation and non-vegetation areas, and extract a complete vegetation coverage image consistent with the survey area boundary defined in S1.1;

[0103] S5.4. Count the total number of vegetation pixels in the vegetation coverage image, and finally determine the total number of vegetation pixels, which provides a data volume reference for subsequent feature extraction.

[0104] In this embodiment, it should also be noted that the implementation steps of step S6 are:

[0105] S6.1. Extract the mean reflectance features of 6 multispectral bands from the vegetation cover image obtained in S5.3, and the value ranges of each band are 0.12-0.25 for blue band, 0.20-0.35 for green band, 0.10-0.22 for red band, 0.30-0.45 for red edge band, 0.40-0.65 for near-infrared band, and 0.25-0.40 for long-wave infrared band;

[0106] S6.2. Calculate the contrast, entropy, correlation, and uniformity of 4 texture features under a 3x3 window using the gray level co-occurrence matrix method, and the value ranges of each index are 0-200 for contrast, 0-4 for entropy, -1-1 for correlation, and 0-1 for uniformity;

[0107] S6.3. Extract 4 morphological features of vegetation crown area, perimeter, circularity, and aspect ratio, set the crown area threshold to 0.1 m2, the perimeter threshold to 1.2 m, the circularity to 0.3-1.0, and the aspect ratio to 1.0-3.5, and define the morphological determination criteria;

[0108] S6.4. Integrate the spectral, texture, and morphological features to form a 14-dimensional feature vector for each vegetation pixel, and the total number of feature vectors is consistent with the total number of vegetation pixels counted in S5.4.

[0109] In this embodiment, it is also necessary to explain that the implementation steps of step S7 are:

[0110] S7.1. Randomly divide the sample set from the feature vectors generated in S6.4, select 20% as the model training set, and the remaining 80% as the model test set;

[0111] S7.2. Configure the random forest classification model parameters, set the number of decision trees to 100, the maximum tree depth to 15 layers, the minimum number of node splitting samples to 5, and the out-of-bag data error threshold to 0.05, and establish the model basic framework;

[0112] S7.3. Take the 4 types of plants, i.e., trees, shrubs, herbs, and vines, defined in S1.3 as the classification labels, use the training set to train the model, and at the same time, verify the model classification accuracy through the test set, and stop training when the accuracy reaches 93% or above;

[0113] S7.4. Apply the classification model that has met the training requirements to all feature vectors to complete the plant type determination of each vegetation pixel one by one, and form the preliminary classification results;

[0114] S7.5. Based on the vegetation pixel type determination results, generate a preliminary classification image containing the distribution information of the 4 types of plants, and the image resolution remains 5 cm as set in S4.4 to ensure consistent spatial accuracy.

[0115] In this embodiment, it also needs to be explained that the implementation steps of step S8 are:

[0116] S8.1. Based on the preliminary classification image of S7.5 and the survey area delineated in S1.1, 60 1m x 1m ground verification sample plots are designed using systematic sampling method, and the sample plot center longitude and latitude are randomly generated through GIS software to ensure that the sample plots cover the distribution areas of various plants;

[0117] S8.2. Go to the field to carry out verification investigation, record the actual type, quantity and coverage of plants in each sample plot, and form a complete field verification data set;

[0118] S8.3. Compare the field verification data set with the classification results of the corresponding sample plots in the preliminary classification image, calculate the confusion matrix and obtain the Kappa coefficient to evaluate the accuracy of the classification results;

[0119] S8.4. If the Kappa coefficient is less than 0.88, adjust the random forest model parameters set in S7.2 based on the field data of the verification sample plots, increase the number of decision trees to 120, and adjust the minimum sample number for node splitting to 3, and reclassify the feature vectors in S6.4;

[0120] S8.5. Repeat the verification and adjustment process until the Kappa coefficient reaches 0.88 or more, and finally generate accurate plant type distribution images to ensure the reliability of the classification results.

[0121] In this embodiment, it also needs to be explained that the implementation steps of step S9 are:

[0122] S9.1. Based on the plant type distribution image corrected in S8.5, the number of vegetation pixels corresponding to each type of plant is counted according to the four types of trees, shrubs, herbs and lianas;

[0123] S9.2. Combined with the 5cm resolution determined in S4.4, the total coverage area of each type of plant is calculated by multiplying the number of vegetation pixels of each type of plant by the area of a single pixel;

[0124] S9.3. According to the 0.1m2 crown area threshold set in S6.3, the number of plants with a single crown area greater than or equal to 0.1m2 is counted for each type of plant, and the scattered vegetation with a crown area less than 0.1m2 is excluded;

[0125] S9.4. Integrate the total coverage area and plant number statistics of each type of plant to form a preliminary statistical result, ensuring that the data corresponds to the classification results of S8.5 one by one, and laying a foundation for subsequent coverage calculation.

[0126] Embodiment 2, please refer to Figure 1In practical applications, the present application proposes a surface plant classification and statistics method based on unmanned aerial vehicle aerial images, specifically including the following steps:

[0127] S1. Investigation area definition and core parameter planning:

[0128] S1.1. Clearly define the target area of surface plant classification and statistics, import the latitude and longitude boundary data of the target area through GIS software, specifically 116.32°-116.35° east longitude and 39.92°-39.95° north latitude, complete the spatial definition of the investigation range;

[0129] S1.2. Based on the defined spatial range, determine the total area of the investigation area as 15000㎡, and set the core flight parameters of the unmanned aerial vehicle, including flight height 120 meters, ground resolution 5cm, heading overlap rate 80%, and lateral overlap rate 70%;

[0130] S1.3. Clearly define the classification and statistics core target, the classification target covers four types of plants: trees, shrubs, herbs, and vines, and the statistics index is determined as the number, coverage area, and coverage of each type of plant;

[0131] By clearly defining the investigation area boundary, core flight parameters and classification and statistics target, a clear range and unified standard are set for the entire surface plant classification and statistics work, laying a standardized and implementable foundation, avoiding problems such as ambiguous range and parameter conflicts in subsequent processes;

[0132] S2. Multi-spectral unmanned aerial vehicle and load selection and calibration:

[0133] S2.1. According to the 5cm ground resolution requirement set in S1.2, a quadcopter unmanned aerial vehicle equipped with a 6-channel multi-spectral camera is selected to ensure that the load performance matches the investigation accuracy requirement;

[0134] S2.2. Confirm the core parameters of the multi-spectral camera, the camera focal length is 16mm, and the center wavelengths of each channel are set as 450nm (blue band), 550nm (green band), 650nm (red band), 730nm (red edge band), 850nm (near-infrared band), and 900nm (long-wave infrared band);

[0135] S2.3. Carry out camera interior orientation element calibration, obtain the principal point coordinates x0=1024 pixels, y0=768 pixels and focal length f=16mm through a professional calibration board, and complete the distortion correction basic data acquisition;

[0136] S2.4. Perform camera radiation correction, determine the correction coefficients K1-K6 of each band as 1.02, 1.05, 1.03, 1.01, 1.04, and 1.06 through calibration experiments, which provides a basis for subsequent image radiation correction;

[0137] Through the precise selection of the multi-spectral unmanned aerial vehicle and camera, and the completion of the interior orientation elements and radiation correction, the performance of the image acquisition equipment is ensured to match the investigation accuracy requirements, the influence of the equipment itself error on the data is eliminated, and a high-quality and reliable original data source is provided for subsequent image processing and classification analysis;

[0138] S3. Route planning and multi-spectral image acquisition:

[0139] S3.1. Based on the latitude and longitude boundary drawn in S1.1 and the flight height of 120 meters set in S1.2, a grid-shaped full-coverage route is designed using the unmanned aerial vehicle ground station software, ensuring that there is no missing investigation area;

[0140] S3.2. Set the flight and image acquisition auxiliary parameters, the flight speed is 8 m / s, the image acquisition time interval is 0.5 seconds, and the route inflection point hovering time is 2 seconds, which ensures the continuity of image acquisition;

[0141] S3.3. Start the real-time positioning function of the unmanned aerial vehicle, rely on the GPS system to realize the positioning with an accuracy of ±1 m, and execute the multi-spectral image acquisition according to the planned route, and record the information of the acquisition process synchronously;

[0142] S3.4. Real-time monitoring of the clarity of the acquired images, if there are blurred frames, the automatic re-shooting mechanism is triggered, and finally a complete data set containing 3200 original images is obtained, which meets the subsequent processing requirements;

[0143] Through scientific planning of full-coverage grid route, setting reasonable flight and acquisition parameters, combining real-time positioning and image quality monitoring mechanism, efficient, non-missing, high-quality acquisition of multi-spectral images of the target area is realized, forming a complete original data set with sufficient quantity and clarity up to standard, providing sufficient data support for subsequent image stitching and preprocessing;

[0144] S4. Original image preprocessing:

[0145] S4.1. Based on the camera interior orientation elements (x0 = 1024 pixels, y0 = 768 pixels, f = 16 mm) obtained in S2.3, the 3200 original multi-spectral images collected in S3.4 are subjected to distortion correction to eliminate the influence of lens optical distortion;

[0146] S4.2. According to the heading 80% and lateral 70% overlap rate set in S1.2, a stitching algorithm based on feature point matching is used, the feature point detection threshold is set to 0.03, and the distortion-corrected images are stitched into complete area images;

[0147] S4.3. Apply the radiation correction coefficients (K1-K6) determined in S2.4 to the stitched complete images for radiation correction to remove the interference caused by light changes and atmospheric scattering;

[0148] S4.4. Integrating the correction and stitching results to generate a target area orthographic image with a resolution of 5 cm and a size of 6000 pixels x 5000 pixels, to provide basic data for subsequent vegetation extraction;

[0149] By performing a series of preprocessing operations such as distortion correction, stitching, and radiation correction on the original image, the effects of external factors such as lens distortion, light difference, and atmospheric interference are effectively eliminated, and the discrete original image is integrated into a spatially complete, resolution-compliant, and data-accurate orthographic image, providing standardized and high-quality basic image data for subsequent vegetation area extraction and feature analysis;

[0150] S5. Vegetation area extraction:

[0151] S5.1. Based on the 5 cm resolution orthographic image generated in S4.4, extract the reflectance data of each band, and calculate the normalized vegetation index according to the formula NDVI = (near-infrared band reflectance - red band reflectance) / (near-infrared band reflectance + red band reflectance);

[0152] S5.2. Set the NDVI decision threshold to 0.3, and clearly define NDVI >= 0.3 as vegetation pixels and NDVI < 0.3 as non-vegetation pixels (such as bare land, buildings, etc.), to establish a vegetation identification standard;

[0153] S5.3. Perform binary processing on the orthographic image according to the set NDVI threshold to separate vegetation and non-vegetation areas, and extract a complete vegetation coverage image consistent with the 15000 m2 survey area boundary defined in S1.1;

[0154] S5.4. Count the total number of vegetation pixels in the vegetation coverage image, and finally determine the total number of vegetation pixels as 2860000, to provide a data volume reference for subsequent feature extraction;

[0155] By calculating NDVI and setting a reasonable decision threshold, the precise separation of vegetation and non-vegetation areas is achieved, and a complete vegetation coverage image consistent with the survey area boundary is successfully extracted, while the total number of vegetation pixels is clearly defined, providing a data volume reference for subsequent plant classification feature extraction and reducing the interference of non-vegetation areas on the classification process;

[0156] S6. Plant classification feature extraction:

[0157] S6.1. Extract the average reflectance of 6 multispectral bands from the vegetation cover image obtained in S5.3, with the value ranges of each band being 0.12-0.25 for blue band, 0.20-0.35 for green band, 0.10-0.22 for red band, 0.30-0.45 for red edge band, 0.40-0.65 for near-infrared band, and 0.25-0.40 for long-wave infrared band;

[0158] S6.2. Calculate the contrast, entropy, correlation, and uniformity of 4 texture features using the gray level co-occurrence matrix method under a 3x3 window, with the value ranges of each index being 0-200 for contrast, 0-4 for entropy, -1-1 for correlation, and 0-1 for uniformity;

[0159] S6.3. Extract 4 morphological features of vegetation crown area, perimeter, circularity, and aspect ratio, set the crown area threshold to 0.1 m2, the perimeter threshold to 1.2 m, the circularity to 0.3-1.0, and the aspect ratio to 1.0-3.5, and define the morphological determination criteria;

[0160] S6.4. Integrate the spectral, texture, and morphological features to form a 14-dimensional feature vector for each vegetation pixel, with the total number of feature vectors being consistent with the total number of vegetation pixels calculated in S5.4, both being 2860000;

[0161] By systematically extracting the spectral, texture, and morphological core differentiating features of plants and integrating them into high-dimensional feature vectors, the essential attribute information of different plant types is comprehensively captured, a high-dimensional feature space is constructed to support accurate classification, and sufficient, comprehensive, and effective feature basis is provided for subsequent classification model training, improving the discrimination and accuracy of model classification;

[0162] S7. Classification model training and vegetation type determination:

[0163] S7.1. Randomly divide the sample set from the 2860000 feature vectors generated in S6.4, select 20% (572000) as the model training set, and the remaining 80% (2288000) as the model test set;

[0164] S7.2. Configure the random forest classification model parameters, set the number of decision trees to 100, the maximum tree depth to 15 layers, the minimum number of node splitting samples to 5, and the out-of-bag data error threshold to 0.05, and establish the model basic framework;

[0165] S7.3. Use the 4 plant types of trees, shrubs, herbs, and vines defined in S1.3 as classification labels, train the model using the training set, and verify the model classification accuracy using the test set, stop training when the accuracy reaches 93% or above;

[0166] S7.4. Apply the trained classification model to all 2860000 feature vectors to determine the plant type of each vegetation pixel one by one, forming a preliminary classification result;

[0167] S7.5. Based on the vegetation pixel type determination result, generate a preliminary classification image containing 4 plant distribution information, the image resolution remains 5cm set in S4.4, ensuring consistent spatial accuracy;

[0168] By reasonably dividing the sample set, optimizing the configuration of random forest model parameters and carrying out training and verification, a high-precision classification model suitable for target plant classification is established, the type determination of all vegetation pixels is successfully completed, a preliminary classification image with consistent 5cm spatial resolution is generated, and the effective conversion from high-dimensional feature data to vegetation type spatial distribution result is realized, providing a foundation for subsequent verification and correction;

[0169] S8. Ground verification and classification result correction:

[0170] S8.1. Based on the preliminary classification image of S7.5 and the 15000㎡ survey area delineated in S1.1, 60 1m×1m ground verification quadrats are designed using systematic sampling method, the center longitude and latitude of the quadrat is randomly generated by GIS software to ensure that the quadrat covers all types of plant distribution areas;

[0171] S8.2. Go to the field to carry out verification investigation, record the actual type, quantity and coverage of plants in each quadrat, and form a complete field verification data set;

[0172] S8.3. Compare the field verification data set with the classification results of the corresponding quadrat in the preliminary classification image, calculate the confusion matrix and obtain the Kappa coefficient to evaluate the accuracy of the classification results;

[0173] S8.4. If the Kappa coefficient is less than 0.88, adjust the random forest model parameters set in S7.2 based on the field data of the verification quadrat, increase the number of decision trees to 120 and adjust the minimum sample number for node splitting to 3, and reclassify the feature vectors in S6.4;

[0174] S8.5. Repeat the verification and adjustment process until the Kappa coefficient reaches more than 0.88, finally generate a precise plant type distribution image to ensure the reliability of the classification results;

[0175] Through setting up a comprehensive ground verification sample plot to obtain real field data, establishing an iterative mechanism of "preliminary classification-ground verification-model optimization-reclassification", ensuring that the classification result Kappa coefficient reaches a high precision standard of 0.88 or above, effectively correcting the bias in preliminary classification, significantly improving the accuracy, reliability and credibility of plant classification results, and making the classification image truly and accurately reflect the actual distribution of ground plants;

[0176] S9. Plant quantity statistics and crown area accounting:

[0177] S9.1. Based on the plant type distribution image corrected in S8.5, the number of vegetation pixels corresponding to each type of plant is counted respectively according to trees, shrubs, herbs and lianas, ensuring that the statistical objects are consistent with the classification results;

[0178] S9.2. Combined with the 5cm resolution (single pixel area 0.0025㎡) determined in S4.4, the total coverage area of each type of plant is calculated by multiplying the number of vegetation pixels of each type of plant by the area of a single pixel;

[0179] S9.3. According to the 0.1㎡ crown area threshold set in S6.3, the number of plants with a single crown area of ≥0.1㎡ is counted in each type of plant, and the scattered vegetation with a crown area of less than 0.1㎡ is excluded (regarded as clumped individuals);

[0180] S9.4. Integrating the total coverage area and plant quantity statistics data of each type of plant, the preliminary statistical results are formed, ensuring that the data correspond one-to-one with the classification results in S8.5, laying a foundation for subsequent canopy coverage calculation;

[0181] Based on the accurate plant type distribution image, through the statistics of vegetation pixel number, the accounting of total coverage area and the screening of plant number meeting the crown area threshold, the accurate quantification of key quantity indicators of each type of plant is realized, forming structured and traceable preliminary statistical data, and the data strictly correspond to the classification results, providing accurate and standardized core data support for subsequent canopy coverage calculation and final result integration;

[0182] S10. Canopy coverage calculation and statistical result integration output:

[0183] S10.1. Based on the total coverage area of each type of plant obtained in S9.4 and the total area of the survey area determined in S1.2, the canopy coverage of each type of plant is calculated according to the formula: canopy coverage = total coverage area of a certain type of plant / total area of the survey area × 100%;

[0184] S10.2. The plant quantity, total coverage area data in S9.4 and the canopy coverage data in S10.1 are summarized, and the regional boundary information in S1.1 and the classification distribution image in S8.5 are associated to form a complete statistical data set;

[0185] S10.3. Based on the complete statistical data set, a ground plant classification statistical report is made, the report including a distribution diagram of 4 types of plants, a quantity statistical table, a coverage area comparison graph and a coverage ratio pie chart, and the result visualization degree is improved;

[0186] S10.4. The statistical result is output in a standard format, the statistical data is saved in a CSV format, the image data is saved in a GeoTIFF format, and finally, the complete result meeting the ground plant classification statistical requirement is output, and all the data can be traced back to the core parameters and processing results in the previous steps;

[0187] Through calculating the coverage of each type of plant, comprehensively summarizing multi-dimensional statistical data and spatial image information, making a visual classification statistical report and outputting in a standard format, the complete classification statistical result of standardization, visualization and traceability is formed, which not only meets the actual application requirement of ground plant classification statistics, but also guarantees the standardization and reusability of data, and provides direct support for subsequent ecological monitoring, resource management and the like.

[0188] In actual application, the specific algorithm process of the random forest classification of the application is as follows:

[0189] 1. Sample initialization: based on the 2860000 14-dimensional vegetation feature vectors generated in S6.4, the training set (572000 vectors) and the test set (2288000 vectors) are randomly divided in a ratio of 2:8 according to S7.1, each vector is associated with the clear 4-class labels of arbor, shrub, herb and liana in S1.3, and is recorded as a class set:

[0190] Y={y1,y2,y3,y4};

[0191] 2. Decision tree construction: for the training set, N samples are randomly extracted by bootstrap resampling method (N=572000, consistent with the size of the training set), and m features (m=4, determined based on feature dimension and classification efficiency optimization) are randomly selected from the 14-dimensional features as the splitting candidate features of each decision tree;

[0192] 3. Node splitting rule: the internal nodes of each decision tree are split by the purity index (Gini coefficient) minimization principle until the number of node samples is less than the set minimum sample number (5) or the sample classes are completely consistent, and T=100 independent decision trees (initial parameters) are generated;

[0193] 4. Model verification and optimization: the classification result of each decision tree is verified by using the test set, the classification accuracy of a single tree is calculated, and the preliminary classification result of the random forest is obtained through the integrated voting mechanism, if the overall accuracy does not reach 93%, the number of decision trees, the number of feature sampling and other parameters are adjusted;

[0194] 5). Final classification decision: apply the optimized model to all 2860000 feature vectors, determine the final class of each vegetation pixel through the ensemble voting mechanism, and generate the preliminary classification image (maintain the set 5cm resolution).

[0195] It should be further pointed out that the decision tree node splitting purity index (Gini coefficient) is:

[0196]

[0197] In the formula, D is the sample set of the current node, p k is the proportion of the kth plant (corresponding to arbor, shrub, herb, liana) in the sample set D; the Gini coefficient takes the value range [0, 1], and the smaller the value, the more concentrated the node sample class, and the higher the splitting purity, so as to determine the optimal splitting feature and splitting threshold.

[0198] It should be further pointed out that the random forest ensemble voting mechanism is:

[0199]

[0200] In the formula, f(x) is the final classification result of the random forest for the input feature vector x (14-dimensional feature vector from S6.4); T is the total number of decision trees (initially 100, and after correction 120); h t (x) is the classification result of the tth decision tree for x; I(·) is an indicator function, which takes the value 1 when the condition in the parentheses is true, and 0 otherwise; the formula indicates that by counting the voting results of all decision trees for x, the class with the most votes is selected as the final classification label, realizing the fault tolerance and high accuracy of ensemble learning.

[0201] It should be further pointed out that the classification accuracy calculation is:

[0202]

[0203] In the formula, Acc is the model classification accuracy, M is the number of test set samples (2288000); x i is the ith feature vector in the test set, y i is its true class label (from the field verification data in S8.2); the formula is used to verify the model training effect, to ensure that the accuracy is above 93%, and to ensure the classification reliability.

[0204] In summary, the present application adaptively matches the multi-dimensional features (14-dimensional spectrum + texture + morphological features) of ground plants, optimizes the feature sampling number (m=4) and the number of decision trees (100-120), and solves the problem of insufficient classification discrimination of single features;

[0205] Form a closed loop with the previous feature extraction (S6) and the later ground verification (S8), when the Kappa coefficient is lower than 0.88, the model parameters (the number of decision trees, the minimum sample size) are dynamically adjusted based on the field data, and the iteration optimization of the classification result is realized;

[0206] Combined with the high resolution characteristics (5cm) of the unmanned aerial vehicle image, the pixel-level features are accurately associated with the plant types, the limitations of the traditional remote sensing classification "spectrum confusion" are broken through, and the classification accuracy of the close relative species and low vegetation is improved.

[0207] Through the above steps, the method of the present application compensates for the shortcomings of the traditional method through the high resolution of the unmanned aerial vehicle and the multispectral image, has the characteristics of high investigation efficiency and wide coverage, can quickly obtain complete vegetation image data of the target area, avoids the cumbersome process of manual investigation; by integrating the three types of features of spectrum, texture and morphology and combining the machine learning model for classification, the accuracy of plant classification is significantly improved, and the subjective error is reduced; with the ground verification and model correction mechanism, the reliability of the classification result is further guaranteed, at the same time, the multi-dimensional information of the distribution, quantity and coverage area of the plants can be comprehensively obtained, the data integrity is strong and traceable, the standardization degree of the operation process is high, and the method is suitable for the classification and statistics of the surface plants under different scales and different terrain conditions, and the application scene is more extensive.

[0208] Although the embodiments of the present application have been shown and described, it can be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery, characterized in that, Includes the following steps: S1. Define the latitude and longitude boundaries and total area of ​​the target region, and set the UAV flight parameters, plant classification categories and statistical indicators; S2. Select and match the multispectral UAV and camera, complete the orientation element calibration and radiometric correction within the camera, and obtain the core parameters; S3. Fly along the planned route, collect multispectral images with set parameters, monitor the clarity in real time and take additional photos to form a complete dataset; S4. Perform distortion correction, stitching and radiometric correction on the original image to generate an orthophoto map of the target area with the required resolution; S5. By calculating NDVI and setting a threshold, extract the image of the vegetation area and count the total number of vegetation pixels; S6. Extract the three types of features of vegetation: spectral, texture, and morphology, and integrate them to form a 14-dimensional feature vector for each vegetation pixel; S7. Divide the sample set to train the random forest model, and after verification, apply it to all feature vectors to generate preliminary classification images; S8. Set up ground verification plots to collect field data, compare and correct the classification model, and generate accurate classification images with Kappa coefficient meeting the standard; S9. Based on accurate classification images, count the number of vegetation pixels, total coverage area, and number of plants that meet the threshold for each type of plant. S10. Based on the total coverage area of ​​various plants obtained in S9 and the total area of ​​the surveyed area in S1, calculate the coverage of various plants. At the same time, summarize the plant number and total coverage area data in S9, and combine the regional boundary information in S1 and the classification distribution image in S8 to generate a statistical report on the classification of surface plants.

2. The method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 1, characterized in that, The implementation steps of step S1 are as follows: S1.

1. Define the target area for the classification and statistics of surface plants, and import the latitude and longitude boundary data of the target area through GIS software, specifically 116.32°-116.35° east longitude and 39.92°-39.95° north latitude, to complete the spatial definition of the survey area; S1.

2. Based on the defined spatial range, determine the total area of ​​the survey area and set the core flight parameters of the UAV; S1.

3. Define the core objectives of classification and statistics. The classification objectives cover four types of plants: trees, shrubs, herbs, and vines. The statistical indicators are the quantity, coverage area, and coverage of each type of plant.

3. The method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 2, characterized in that, The implementation steps of step S2 are as follows: S2.

1. Based on the ground resolution requirements set in S1.2, a quadcopter UAV equipped with a 6-channel multispectral camera is selected to ensure that the payload performance matches the survey accuracy requirements. S2.

2. Confirm the core parameters of the multispectral camera; S2.

3. Conduct in-camera orientation element calibration, obtain principal point coordinate pixels and focal length through a professional calibration board, and complete the basic data collection for distortion correction; S2.

4. Perform camera radiometric correction and determine the correction coefficients for each band through calibration experiments to provide a basis for subsequent image radiometric correction.

4. The method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 3, characterized in that, The implementation steps of step S3 are as follows: S3.

1. Based on the latitude and longitude boundaries defined in S1.1 and the flight altitude set in S1.2, a grid-like full-coverage flight route is designed using UAV ground station software to ensure that no survey area is missed; S3.

2. Set auxiliary parameters for flight and image acquisition, image acquisition time interval, and hovering time at flight path inflection points to ensure the continuity of image acquisition; S3.

3. Enable the real-time positioning function of the UAV, rely on the GPS system to achieve positioning accuracy of ±1m, perform multispectral image acquisition according to the planned route, and record the acquisition process information simultaneously; S3.

4. Real-time monitoring of the clarity of the acquired images. If a blurry frame appears, an automatic re-shooting mechanism is triggered. Finally, a complete dataset containing 3200 original images is obtained to meet the needs of subsequent processing.

5. A method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 4, characterized in that, The implementation steps of step S4 are as follows: S4.

1. Based on the camera interior orientation elements obtained in S2.3, perform distortion correction on the 3200 original multispectral images acquired in S3.4 to eliminate the influence of lens optical distortion; S4.

2. Based on the 80% overlap rate in the heading direction and 70% overlap rate in the lateral direction set in S1.2, a stitching algorithm based on feature point matching is adopted, and the feature point detection threshold is set to 0.03 to stitch the distortion-corrected images into a complete regional image. S4.

3. Apply the radiometric correction coefficients determined in S2.4 to perform radiometric correction on the stitched complete image to remove interference caused by changes in illumination and atmospheric scattering; S4.

4. Integrate the correction and stitching results to generate an orthophoto map of the target area with a resolution of 5cm and a size of 6000 pixels × 5000 pixels, providing basic data for subsequent vegetation extraction.

6. The method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 5, characterized in that, The implementation steps of step S5 are as follows: S5.

1. Based on the 5cm resolution orthophoto map generated by S4.4, extract the normalized vegetation index of reflectance data for each band; S5.

2. Set the NDVI determination threshold to 0.3, and define pixels with NDVI ≥ 0.3 as vegetation pixels and pixels with NDVI < 0.3 as non-vegetation pixels, thus establishing a vegetation identification standard. S5.

3. Binarize the orthophoto map according to the set NDVI threshold, separate the vegetation and non-vegetation areas, and extract the complete vegetation cover image that is consistent with the boundary of the survey area defined in S1.

1. S5.

4. Count the total number of vegetation pixels in the vegetation cover image to determine the total number of vegetation pixels and provide a data volume reference for subsequent feature extraction.

7. A method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 6, characterized in that, The implementation steps of step S6 are as follows: S6.

1. Based on the vegetation cover image obtained in S5.3, extract the mean reflectance features of 6 multispectral bands. The value ranges for each band are as follows: blue band 0.12-0.25, green band 0.20-0.35, red band 0.10-0.22, red edge band 0.30-0.45, near-infrared band 0.40-0.65, and long-wave infrared band 0.25-0.

40. S6.

2. The gray-level co-occurrence matrix method is used to calculate four texture features: contrast, entropy, correlation, and uniformity in a 3×3 window. The value range of each index is 0-200 for contrast, 0-4 for entropy, -1-1 for correlation, and 0-1 for uniformity. S6.

3. Extract four morphological features of vegetation canopy area, perimeter, roundness, and aspect ratio, and set thresholds of 0.1㎡ for canopy area, 1.2m for perimeter, 0.3-1.0 for roundness, and 1.0-3.5 for aspect ratio to clarify the morphological judgment criteria; S6.

4. Integrate the three types of features: spectrum, texture, and morphology, to form a 14-dimensional feature vector for each vegetation pixel. The total number of feature vectors is consistent with the total number of vegetation pixels counted in S5.

4.

8. A method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 7, characterized in that, The implementation steps of step S7 are as follows: S7.

1. Randomly divide the sample set from the feature vectors generated in S6.4, select 20% as the model training set, and the remaining 80% as the model test set; S7.

2. Configure the parameters of the random forest classification model, set the number of decision trees to 100, the maximum tree depth to 15 layers, the minimum number of samples for node splitting to 5, and the out-of-bag error threshold to 0.05, and establish the basic framework of the model; S7.

3. Using the four plant categories of trees, shrubs, herbs and vines specified in S1.3 as classification labels, the model is trained using the training set, and the classification accuracy of the model is verified using the test set. Training is stopped when the accuracy reaches 93% or higher. S7.

4. Apply the trained classification model to all feature vectors to determine the plant type of each vegetation pixel one by one, and form a preliminary classification result; S7.

5. Based on the vegetation pixel type determination results, generate a preliminary classification image containing distribution information of 4 types of plants. The image resolution remains at 5cm as set in S4.4 to ensure consistent spatial accuracy.

9. A method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 8, characterized in that, The implementation steps of step S8 are as follows: S8.

1. Based on the preliminary classification imagery of S7.5 and the survey area delineated in S1.1, 60 1m×1m ground verification quadrats were designed using systematic sampling. The latitude and longitude of the quadrats center were randomly generated using GIS software to ensure that the quadrats covered the distribution areas of various plants. S8.

2. Conduct field verification surveys, record the actual type, quantity and coverage of plants in each quadrat, and form a complete field verification dataset; S8.

3. Compare the classification results of the field validation dataset with the corresponding quadrat in the preliminary classification image, calculate the confusion matrix and obtain the Kappa coefficient, and evaluate the accuracy of the classification results; S8.

4. If the Kappa coefficient is lower than 0.88, adjust the random forest model parameters set in S7.2 based on the field data of the validation sample plots, increase the number of decision trees to 120, adjust the minimum number of samples for node splitting to 3, and reclassify the feature vectors in S6.

4. S8.

5. Repeat the verification and adjustment process until the Kappa coefficient reaches 0.88 or higher, and finally generate accurate plant type distribution images to ensure the reliability of classification results.

10. A method for classifying and statistically analyzing surface vegetation based on UAV aerial imagery according to claim 9, characterized in that, The implementation steps of step S9 are as follows: S9.

1. Based on the plant type distribution image corrected in S8.5, count the number of vegetation pixels corresponding to each of the four types of plants: trees, shrubs, herbs, and vines. S9.

2. Based on the 5cm resolution determined in S4.4, the total coverage area of ​​various plants is calculated by multiplying the number of vegetation pixels of each type of plant by the area of ​​a single pixel. S9.

3. Based on the 0.1㎡ crown area threshold set in S6.3, count the number of plants with a single crown area ≥ 0.1㎡ in each type of plant, and exclude scattered vegetation with a crown area less than 0.1㎡. S9.

4. Integrate the total coverage area and number of plants of various types to form preliminary statistical results, ensuring that the data corresponds one-to-one with the classification results in S8.5, and laying the foundation for subsequent coverage calculations.

Images