A method for identifying grassland rodent burrows based on UAV aerial photography and machine learning

Abstract

A grassland rodent burrow identification method based on UAV aerial photography and machine learning is proposed. First, a UAV is used to photograph the survey area to obtain high- and low-altitude images. Then, the high-altitude images are pre-processed using the coordinates of ground control points to obtain the corresponding DOM (Domain Image) images. Next, both the DOM and low-altitude images undergo color space transformation and multi-scale segmentation to obtain corresponding high- and low-altitude polygon images. The low-altitude polygon images are then divided into training and validation samples. The algorithm classifier is repeatedly trained and validated using these samples until the required accuracy is achieved, resulting in a mature algorithm classifier. Finally, the high-altitude polygon images are fed into the algorithm classifier to output the vector data of the rodent burrow polygons. This design not only offers good accuracy and low time consumption but also boasts high recognition speed and strong timeliness, making it suitable for precise dynamic monitoring and prediction of rodent distribution.

Description

Technical Field

[0001] This invention relates to a method for identifying mouse holes, belonging to the field of grassland protection, and particularly to a grassland mouse hole identification method based on drone aerial photography and machine learning. Background Technology

[0002] In recent decades, overgrazing and irrational development have led to frequent outbreaks of grassland rodent infestations. The digging and grazing activities of these rodents damage grassland vegetation, causing significant soil erosion and resulting in large areas of secondary bare land and barren "black soil beaches."

[0003] Furthermore, rodents consume large quantities of high-quality forage, competing with livestock for food, exacerbating the conflict between livestock and forage, reducing the carrying capacity of grasslands, and causing great harm to livestock production. This seriously affects the healthy development of grassland animal husbandry and threatens the safety of the ecosystem, even leading to system dysfunction or loss. Therefore, making full use of modern technology to conduct large-scale monitoring of grassland rodent pests, making scientific monitoring, accurate judgment and assessment of rodent pest conditions, and establishing an effective dynamic monitoring system for grassland rodent pests are directly related to the coordinated development of animal husbandry and the safety, construction and utilization of grassland ecosystems.

[0004] Currently, the biggest problem in controlling grassland rodent pests is the lack of precision and sustainability in control methods. If the number and distribution of rodent burrows within a large area of ​​grassland could be accurately and quickly obtained, more targeted control measures could be taken, leading to more effective rodent control.

[0005] Traditional rodent monitoring methods mainly rely on manual surveys, such as fixed-point observation, blocking burrows to detect thieves, and day / night observations to investigate rodent populations. However, these methods are often time-consuming and labor-intensive, and have drawbacks such as high survey costs, small survey areas, and poor representativeness. Only by making timely and accurate predictions of grassland rodent infestations can we scientifically formulate integrated prevention and control plans, implement timely control measures, and achieve the goal of economically and effectively controlling rodent infestations.

[0006] The information disclosed in this background section is intended only to enhance understanding of the overall background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings and problems of poor accuracy and long time consumption in the existing technology, and to provide a grassland rodent burrow identification method based on UAV aerial photography and machine learning with better accuracy and less time consumption.

[0008] To achieve the above objectives, the technical solution of the present invention is: a grassland rodent burrow identification method based on UAV aerial photography and machine learning, wherein the rodent burrow identification method includes the following steps:

[0009] Step 1: First, set up multiple ground control points in the survey area, and then use drones to take high-altitude and low-altitude aerial photos of the survey area to obtain high-altitude and low-altitude images. The high-altitude images include the entire area of ​​the survey area, and the low-altitude images include a part of the survey area. The low-altitude images show any one or any combination of mouse holes, bare ground, and grassland.

[0010] The second step is to first use the coordinates of the ground control points to perform preprocessing on the high-altitude image to obtain the corresponding DOM image. Then, color space transformation and multi-scale segmentation are performed on both the DOM image and the low-altitude image to obtain the corresponding high-altitude polygon image and low-altitude polygon image.

[0011] The third step is to first divide the low-altitude polygonal image into training images and validation images. Then, feature extraction is performed on each training image to obtain the corresponding feature subset. The training images and their corresponding feature subsets constitute a set of training data. All the training data constitute training samples. At the same time, the validation images are used as validation samples. Then, the algorithm classifier is repeatedly trained and validated using the training samples and validation samples until the accuracy meets the requirements, thereby obtaining a mature algorithm classifier.

[0012] Step 4: First, input the high-altitude polygonal image into a well-trained algorithm classifier, then run the algorithm classifier to output the vector data of the polygons of the mouse holes. Then, identify the mouse holes based on the vector data of the polygons. This identification includes obtaining the number of mouse holes, the area of ​​each mouse hole, and the spatial distribution characteristics of the mouse holes in the survey area.

[0013] In the first step, the low-altitude image is taken at an altitude of 20-40 meters.

[0014] In the second step, the color space transformation refers to converting the RGB color space into the HSV color space;

[0015] The aforementioned multi-scale image segmentation refers to: first loading the image obtained after color space transformation in the eCognition Developer software, then selecting multi-scale segmentation, and setting scale parameters, image band weights, and object homogeneity standards to obtain a polygonal image.

[0016] The scale parameter is 25;

[0017] The image band weight refers to the band weight ratio of chroma, brightness, and saturation being 1:2:1;

[0018] The homogeneity criteria of the objects are: color factor 0.7, shape factor 0.3, and within the shape factor, smoothness factor 0.5 and compactness factor 0.5.

[0019] In the third step, dividing the low-altitude polygonal image into training images and verification images means:

[0020] First, visual interpretation is performed on the low-altitude polygonal images to identify the objects in the images as mouse holes, bare ground, and grassland. Then, the identified low-altitude polygonal images are divided into training images and verification images, and the objects displayed in the training images and verification images include the three categories of mouse holes, bare ground, and grassland.

[0021] In the third step, performing feature extraction on each training image to obtain a corresponding feature subset means:

[0022] The eCognition software is used to sequentially extract, select, and optimize features from the training image to obtain a feature subset corresponding to the training image; the feature subset includes spectral features, geometric features, texture features, and exponential features.

[0023] The spectral features include the mean, standard deviation, ratio, maximum difference (MaxDif), and brightness value of the three visible light bands, totaling 11 features.

[0024] The geometric features include five features: area, perimeter (BL), aspect ratio (LW), shape index (SI), and density.

[0025] The texture features include six features: contrast (GLCMCon), correlation (GLCMCor), homogeneity (GLCMHom), entropy (GLCMEnt), energy (GLCMAsm), and mean (GLCMMean).

[0026] The index features are calculated through spectral features, including the improved soil-adjusted vegetation index V-MSAVI and the visible light three-band differential vegetation index VDVI, totaling two features.

[0027] The feature subset includes 11 features: Mean_blue, Mean_green, Mean_red, MaxDif, Ratio_green, Brightness, Area, GLCMMean, GLCMHom, V-MSAVI, and VDVI.

[0028] In the third step, the algorithm classifier refers to: using the SVM algorithm to build an SVM classifier, and using the radial basis function neural network kernel function as the kernel function of the SVM algorithm. During the training process, the radial basis function neural network kernel function is used to calculate the distance between training samples.

[0029] The accuracy requirement means that the accuracy is greater than or equal to 90%.

[0030] In the second step, the preprocessing of the aerial image using the coordinates of ground control points to obtain the corresponding DOM image refers to:

[0031] First, the uniformly lit and colored aerial images, along with the coordinates of all ground control points, are imported into the Inpho image processing platform. Then, aerial triangulation is performed to obtain DSM images. Next, digital surface model filtering is performed to obtain DEM images. Finally, radiometric correction, differential correction, and mosaicking are performed to obtain DOM images.

[0032] During aerial triangulation, the coordinate information of ground control points is used as a reference to correct the aerial images.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. In this invention, a grassland mouse burrow identification method based on UAV aerial photography and machine learning, firstly, high-altitude and low-altitude images are captured by a UAV. The low-altitude images allow for visual identification of three types of terrain features: mouse burrows, bare ground, and grassland. Then, using the high-altitude and low-altitude images as source material, high-altitude polygon images and low-altitude polygon images are obtained respectively. Next, feature extraction is performed based on the low-altitude polygon images to obtain training and acceptance samples. The algorithm classifier is then repeatedly trained and validated using these training and acceptance samples until the accuracy meets the requirements, thus obtaining a mature algorithm classifier. Finally, the high-altitude polygon images are substituted into the mature algorithm classifier to output the polygonal vector data of the mouse burrows. The advantages of this design include... First, this design, by combining high-altitude and low-altitude imagery, achieves high resolution (approximately 1 cm), accurately distinguishing between low vegetation and small rodent burrows, facilitating the identification of burrows, bare ground, and grassland with excellent precision. Second, the vector data of the polygonal shapes of the rodent burrows acquired by this design can not only determine the number of burrows but also calculate the area of ​​each burrow and obtain the spatial distribution characteristics of the burrows in the survey area, such as sparseness or density, resulting in good identification performance and wide applicability. Third, the entire operation of this design relies on drones and computer operation, saving significant manpower and resources, reducing costs, and providing fast identification speed and high timeliness, which is beneficial for precise dynamic monitoring and forecasting of rodent distribution. Therefore, this invention is not only highly accurate but also time-efficient, and has significant meaning and role in the integrated management of grassland rodent pests.

[0035] 2. In the grassland mouse burrow identification method based on UAV aerial photography and machine learning of this invention, when extracting features from each training image to obtain a corresponding feature subset, the feature subset is limited to include spectral features, geometric features, texture features, and exponential features. This balances accuracy and information redundancy, ensuring relatively accurate and fast identification results. The effect is even better when the feature subset includes Mean_blue, Mean_green, Mean_red, MaxDif, Ratio_green, Brightness, Area, GLCMMMean, GLCMMHom, V-MSAVI, and VDVI. Therefore, this invention has high accuracy and fast identification speed.

[0036] 3. In the grassland rodent burrow identification method based on UAV aerial photography and machine learning of this invention, multiple ground control points are set up in the survey area before aerial imagery is generated. Then, the coordinates of the ground control points are used to correct the aerial imagery to improve its accuracy, thereby ensuring the accuracy of the final rodent burrow identification result. Therefore, this invention has high precision. Attached Figure Description

[0037] Figure 1 This is a flowchart of the aerial image acquisition process in this invention.

[0038] Figure 2 This is a flowchart illustrating the construction and application of the algorithm classifier in this invention.

[0039] Figure 3 This is a schematic diagram of the DOM image corresponding to the high-altitude image in this invention.

[0040] Figure 4 This is a schematic diagram of the low-altitude polygonal image in this invention.

[0041] Figure 5 This is a schematic diagram of the mouse hole identification results in this invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] See Figure 1 — Figure 5 A method for identifying grassland rodent burrows based on drone aerial photography and machine learning, the method comprising the following steps:

[0044] Step 1: First, set up multiple ground control points in the survey area, and then use drones to take high-altitude and low-altitude aerial photos of the survey area to obtain high-altitude and low-altitude images. The high-altitude images include the entire area of ​​the survey area, and the low-altitude images include a part of the survey area. The low-altitude images show any one or any combination of mouse holes, bare ground, and grassland.

[0045] The second step is to first use the coordinates of the ground control points to perform preprocessing on the high-altitude image to obtain the corresponding DOM image. Then, color space transformation and multi-scale segmentation are performed on both the DOM image and the low-altitude image to obtain the corresponding high-altitude polygon image and low-altitude polygon image.

[0046] The third step is to first divide the low-altitude polygonal image into training images and validation images. Then, feature extraction is performed on each training image to obtain the corresponding feature subset. The training images and their corresponding feature subsets constitute a set of training data. All the training data constitute training samples. At the same time, the validation images are used as validation samples. Then, the algorithm classifier is repeatedly trained and validated using the training samples and validation samples until the accuracy meets the requirements, thereby obtaining a mature algorithm classifier.

[0047] Step 4: First, input the high-altitude polygonal image into a well-trained algorithm classifier, then run the algorithm classifier to output the vector data of the polygons of the mouse holes. Then, identify the mouse holes based on the vector data of the polygons. This identification includes obtaining the number of mouse holes, the area of ​​each mouse hole, and the spatial distribution characteristics of the mouse holes in the survey area.

[0048] In the first step, the low-altitude image is taken at an altitude of 20-40 meters.

[0049] In the second step, the color space transformation refers to converting the RGB color space into the HSV color space;

[0050] The aforementioned multi-scale image segmentation refers to: first loading the image obtained after color space transformation in the eCognition Developer software, then selecting multi-scale segmentation, and setting scale parameters, image band weights, and object homogeneity standards to obtain a polygonal image.

[0051] The scale parameter is 25;

[0052] The image band weight refers to the band weight ratio of chroma, brightness, and saturation being 1:2:1;

[0053] The homogeneity criteria of the objects are: color factor 0.7, shape factor 0.3, and within the shape factor, smoothness factor 0.5 and compactness factor 0.5.

[0054] In the third step, dividing the low-altitude polygonal image into training images and verification images means:

[0055] First, visual interpretation is performed on the low-altitude polygonal images to identify the objects in the images as mouse holes, bare ground, and grassland. Then, the identified low-altitude polygonal images are divided into training images and verification images, and the objects displayed in the training images and verification images include the three categories of mouse holes, bare ground, and grassland.

[0056] In the third step, performing feature extraction on each training image to obtain a corresponding feature subset means:

[0057] The eCognition software is used to sequentially extract, select, and optimize features from the training image to obtain a feature subset corresponding to the training image; the feature subset includes spectral features, geometric features, texture features, and exponential features.

[0058] The spectral features include the mean, standard deviation, ratio, maximum difference (MaxDif), and brightness value of the three visible light bands, totaling 11 features.

[0059] The geometric features include five features: area, perimeter (BL), aspect ratio (LW), shape index (SI), and density.

[0060] The texture features include six features: contrast (GLCMCon), correlation (GLCMCor), homogeneity (GLCMHom), entropy (GLCMEnt), energy (GLCMAsm), and mean (GLCMMean).

[0061] The index features are calculated through spectral features, including the improved soil-adjusted vegetation index V-MSAVI and the visible light three-band differential vegetation index VDVI, totaling two features.

[0062] The feature subset includes 11 features: Mean_blue, Mean_green, Mean_red, MaxDif, Ratio_green, Brightness, Area, GLCMMean, GLCMHom, V-MSAVI, and VDVI.

[0063] In the third step, the algorithm classifier refers to: using the SVM algorithm to build an SVM classifier, and using the radial basis function neural network kernel function as the kernel function of the SVM algorithm. During the training process, the radial basis function neural network kernel function is used to calculate the distance between training samples.

[0064] The accuracy requirement means that the accuracy is greater than or equal to 90%.

[0065] In the second step, the preprocessing of the aerial image using the coordinates of ground control points to obtain the corresponding DOM image refers to:

[0066] First, the uniformly lit and colored aerial images, along with the coordinates of all ground control points, are imported into the Inpho image processing platform. Then, aerial triangulation is performed to obtain DSM images. Next, digital surface model filtering is performed to obtain DEM images. Finally, radiometric correction, differential correction, and mosaicking are performed to obtain DOM images.

[0067] During aerial triangulation, the coordinate information of ground control points is used as a reference to correct the aerial images.

[0068] The principle of this invention is explained as follows:

[0069] In this invention, the meanings of DSM, DEM, and DOM are as follows:

[0070] DSM stands for Digital Surface Model, abbreviated as DSM, and is often referred to as a digital surface model.

[0071] DEM stands for Digital Elevation Model.

[0072] DOM stands for Digital Orthophoto Map, abbreviated as DOM, and is often referred to as a digital orthophoto map.

[0073] Visual interpretation in this invention refers to direct observation by the human eye, without the need for instruments.

[0074] In order for the drone to automatically take pictures of the survey area, the present invention requires setting parameters such as the drone's flight trajectory and altitude before flight. In this way, the drone can fly autonomously according to the pre-set route and altitude. After the aerial photography, the flight quality needs to be checked, that is, the quality of the image data needs to be checked, such as exposure, clarity, and whether the photos can be opened. If necessary, the pre-set route and altitude need to be adjusted and the aerial photography needs to be repeated.

[0075] Example 1:

[0076] See Figure 1 — Figure 5 A method for identifying grassland rodent burrows based on drone aerial photography and machine learning, the method comprising the following steps:

[0077] Step 1: First, set up multiple ground control points in the survey area, and then use drones to take high-altitude and low-altitude aerial photos of the survey area to obtain high-altitude and low-altitude images. The high-altitude images include the entire area of ​​the survey area, and the low-altitude images include a part of the survey area. The low-altitude images show any one or any combination of mouse holes, bare ground, and grassland.

[0078] Step 2: First, use the coordinates of ground control points to perform preprocessing on the aerial imagery to obtain the corresponding DOM imagery (e.g., ...). Figure 2 (as shown in the figure), then color space transformation and multi-scale segmentation are performed on the DOM image and the low-altitude image in sequence to obtain the corresponding high-altitude polygon image and low-altitude polygon image.

[0079] Step 3: First, extract the low-altitude polygonal image (such as...) Figure 3As shown, the images are divided into training images and validation images. Then, feature extraction is performed on each training image to obtain the corresponding feature subset. The training images and their corresponding feature subsets constitute a set of training data. All the training data constitute training samples. At the same time, the validation images are used as validation samples. Then, the algorithm classifier is repeatedly trained and validated using the training samples and validation samples until the accuracy meets the requirements, thereby obtaining a mature algorithm classifier.

[0080] Step 4: First, input the aerial polygon image into a well-trained algorithm classifier, then run the algorithm classifier to output vector data of the mouse hole polygons (e.g., vector data format of a shapefile). Then, identify the mouse hole based on the polygon vector data (e.g., ...). Figure 4 As shown in the figure, this is a scaled-down schematic diagram. The small black dots in the figure are mouse holes. When a small black dot is selected and zoomed in, the vector data of that mouse hole can be obtained. This identification includes obtaining the number of mouse holes, the area of ​​each mouse hole, and the spatial distribution characteristics of mouse holes in the survey area.

[0081] Example 2:

[0082] The basic content is the same as in Example 1, except that:

[0083] Because the features of rodent burrows, grasslands, and bare ground need to be visually identified from low-altitude imagery, the flight altitude for taking low-altitude images cannot be too high, generally between 20 and 40 meters, with 30 meters being the preferred altitude. This process involves manually collecting UAV imagery during flight. This data collection is not conducted in a specific corner of the survey area, but rather evenly distributed across the entire region, thus representing the characteristics of these three types of terrain features (rodent burrows, grasslands, and bare ground) throughout the area.

[0084] Example 3:

[0085] The basic content is the same as in Example 1, except that:

[0086] When setting up multiple ground control points in a survey area, control points must be set at the four corners of the area, and multiple control points also need to be set in the center of the area to ensure that the control points are evenly distributed. In addition, the control points should be clearly visible so that they can be clearly captured when the UAV takes high-altitude images. The coordinates of the ground control points can be accurately measured using instruments such as RTK, with high precision.

[0087] Example 4:

[0088] The basic content is the same as in Example 1, except that:

[0089] The color space transformation refers to converting the RGB color space to the HSV color space to increase the difference in grayscale values ​​between the mouse hole and the background, thereby improving the effect of multi-scale segmentation of the image and ultimately improving the accuracy of classification.

[0090] The reason for choosing HSV is that the hue represented by H can essentially determine a certain color, and combined with saturation and brightness information, it can be judged whether it exceeds a certain threshold. RGB, on the other hand, consists of three components, requiring the determination of the contribution ratio of each component. Therefore, the HSV space has a wider recognition range, a greater difference in grayscale values ​​between the mouse hole and the background, and can more effectively distinguish between the mouse hole and the background. The ultimate goal is to automatically extract the outline of the mouse hole from a DOM image; therefore, the greater the difference between the mouse hole and the background, the easier it is to extract, which is reflected in the grayscale values ​​of the bands in the image. Therefore, we transform the RGB color space into HSV.

[0091] The preferred implementation process is through the cv2.cvtColor function in the Python OpenCV library.

[0092] Example 5:

[0093] The basic content is the same as in Example 1, except that:

[0094] In the third step, performing feature extraction on each training image to obtain a corresponding feature subset means:

[0095] The training image is sequentially subjected to feature extraction, feature selection, and feature optimization using eCognition software (preferably using feature optimization tools provided by eCognition, such as Wrapper and Filter), to obtain a feature subset corresponding to the training image; the feature subset includes spectral features, geometric features, texture features, and exponential features.

[0096] The spectral features include the mean, standard deviation, ratio, maximum difference (MaxDif), and brightness value of the three visible light bands, totaling 11 features. Among them, the mean, standard deviation, and ratio of the three visible light bands are 9 features.

[0097] The geometric features include five features: area, perimeter (BL), aspect ratio (LW), shape index (SI), and density.

[0098] The texture features include six features: contrast (GLCMCon), correlation (GLCMCor), homogeneity (GLCMHom), entropy (GLCMEnt), energy (GLCMAsm), and mean (GLCMMean).

[0099] The index features are calculated through spectral features, including the improved soil-adjusted vegetation index V-MSAVI and the visible light three-band differential vegetation index VDVI, totaling two features.

[0100] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A grassland burrow identification method based on unmanned aerial vehicle aerial photography and machine learning, characterized in that: The rat hole recognition method comprises the following steps: First step: first, set multiple ground control points in the survey area, then use a UAV to take high-altitude aerial photography and low-altitude aerial photography on the survey area to obtain high-altitude images and low-altitude images, wherein the high-altitude images include the entire area of the survey area, and the low-altitude images include part of the area of the survey area, and any one or any combination of rat holes, bare land and grassland is displayed on the low-altitude images; Second step: first, use the coordinates of the ground control points to perform preliminary processing on the high-altitude images to obtain DOM images corresponding to the high-altitude images, then sequentially perform color space conversion and image multi-scale segmentation on the DOM images and the low-altitude images to obtain corresponding high-altitude polygon images and low-altitude polygon images; Third step: first, divide the low-altitude polygon images into training images and verification images, then perform feature extraction on each training image to obtain a corresponding feature subset, the training image and the corresponding feature subset thereof form a group of training data, all the training data constitute a training sample, at the same time, the verification images are taken as a verification sample, then the training sample and the verification sample are used to repeatedly train and verify the algorithm classifier until the accuracy meets the requirements, thereby obtaining a mature algorithm classifier; Fourth step: first, input the high-altitude polygon images into the mature algorithm classifier, then run the algorithm classifier to output vector data of the polygon of the rat hole, and then identify the rat hole according to the vector data of the polygon, which includes obtaining the number of rat holes, the area of each rat hole and the spatial distribution characteristics of the rat holes in the survey area.

2. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 1, characterized in that: In the first step, the shooting altitude of the low-altitude images is 20-40 meters.

3. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 1 or 2, characterized in that: In the second step, the color space conversion refers to converting the RGB color space into the HSV color space; The image multi-scale segmentation refers to: first, load the images obtained after the color space conversion on the eCognition Developer software, then select multi-scale segmentation and set the scale parameter, image band weight and object homogeneity standard to obtain the polygon images.

4. The grassland rat hole recognition method based on UAV aerial photography and machine learning according to claim 3, characterized in that: The scale parameter is 25; The image band weight refers to the band weight ratio of the chroma, luminance and saturation, which is 1:2:1; The object homogeneity standard refers to: the color factor is 0.7, the shape factor is 0.3, and in the shape factor, the smoothness factor is 0.5 and the compactness factor is 0.

5.

5. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 3, characterized in that: In the third step, the division of the low-altitude polygon images into training images and verification images refers to: First, visually interpret the low-altitude polygon images to identify the image objects as three types of low-altitude polygon images of rat holes, bare land and grassland, respectively, then divide the identified low-altitude polygon images into training images and verification images, and the image objects displayed in the training images and the verification images include the three types of rat holes, bare land and grassland.

6. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 5, characterized in that: In the third step, the feature extraction on each training image to obtain a corresponding feature subset refers to: The eCognition software is used for sequentially performing feature extraction, feature selection and feature optimization on the training image to obtain a feature subset corresponding to the training image; the feature subset includes spectral features, geometric features, texture features and index features; the spectral features include average values (Mean), standard deviations (Std), ratios (ratio) of three visible light bands, maximum difference values (MaxDif) and brightness values (Brightness), and totally 11 features; the geometric features include area (Area), perimeter (BL), length-width ratio (LW), shape index (SI) and density (Density), and totally 5 features; the texture features include contrast (GLCMCon), correlation (GLCMCor), homogeneity (GLCMHom), entropy (GLCMEnt), energy (GLCMAsm) and average value (GLCMMean), and totally 6 features; the index features are calculated based on the spectral features and include improved soil-adjusted vegetation index V-MSAVI and visible light three-band difference vegetation index VDVI, and totally 2 features.

7. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 6, characterized in that: The feature subset includes 11 features, namely Mean_blue, Mean_green, Mean_red, MaxDif, Ratio_green, Brightness, Area, GLCMMean, GLCMHom, V-MSAVI and VDVI.

8. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 1 or 2, characterized in that: In the third step, the algorithm classifier refers to that an SVM classifier is established by using an SVM algorithm, a radial basis neural network kernel function is used as the kernel function of the SVM algorithm, and the radial basis neural network kernel function is used to calculate the distance between training samples in the training process. The accuracy meeting the requirements refers to that the accuracy is greater than or equal to 90%.

9. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 1 or 2, characterized in that: In the second step, the use of the coordinates of the ground control points for the pre-processing of the high-altitude image to obtain the DOM image corresponding to the high-altitude image refers to: firstly, the high-altitude image with uniform light and color and the coordinates of all ground control points are imported into an Inpho image processing platform, then aerial triangulation is performed to obtain a DSM image, then a digital surface model filtering operation is performed to obtain a DEM image, then radiation correction, differential correction and tiling operations are performed to obtain a DOM image.

10. The grassland burrow identification method based on UAV aerial photography and machine learning according to claim 9, characterized in that: In the process of performing the aerial triangulation operation, the high-altitude image is corrected based on the coordinate information of the ground control points.

Images