Unmanned aerial vehicle remote sensing-based method for classifying and identifying withered and dead pine tree removal progress information

Abstract

The invention relates to an unmanned aerial vehicle remote sensing-based withered and dead pine tree removal progress information classification and identification method. And based on the unmanned aerial vehicle remote sensing image data in the burst period and the final period of removal of the pine dead wood in the monitoring area, the removal progress information of four kinds of pine dead wood including newly-added dead wood, uncut dead wood, cut uncleaned wood and cut and cleaned wood is obtained. The method comprises the following steps: firstly, acquiring digital orthoimages DOMT1 and DOMT2 and digital surface images DSMT1 and DSMT2 of an unmanned aerial vehicle in a bursaphelenchus xylophilus epidemic outbreak period and a pine withered and dead wood removal final period in a dynamic observation area; establishing a DDPTnet network model and two stages of DOM data by using a YOLOv8n algorithm to obtain information of newly increased dead wood; obtaining felling area information through grid difference value calculation, and carrying out superposition analysis on the felling area information and dead wood point positions before felling to obtain non-felling dead wood point position information and felling dead wood point position information; and obtaining two types of withered and dead wood treatment information, namely cut and uncleaned wood and cut and cleaned wood, through a YOLOv5n classification network. The method can provide objective and effective technical support for pine wood nematode disease prevention and control work.

Description

Technical Field

[0001] The invention relates to the technical field of pine wood nematode disease prevention and control monitoring, and in particular to a method for classifying and identifying progress information of pine dead wood removal based on unmanned aerial vehicle remote sensing. Background Art

[0002] The main way to prevent and control the pine wood nematode epidemic is to identify dead trees in time when the epidemic breaks out, and then cut down the dead trees to prevent the epidemic from spreading further. Therefore, monitoring the progress of dead tree treatment is the top priority for preventing and controlling the pine wood nematode epidemic. During the dead tree treatment stage, there are four types of dead tree treatment progress: newly added, not cut down, cut down but not cleared, and cut down and cleared.

[0003] UAV remote sensing is a flexible and fast-response monitoring method, and the use of this remote sensing technology combined with automated interpretation can further speed up the census of dead trees in the forest. At the same time, the generalization of the target detection network is also stronger than the existing automated interpretation method in the dead tree detection scenario of UAV remote sensing, and it is an important means in the field of dead tree detection. Comprehensive domestic and foreign research has found that there is still a lack of methods for applying UAV remote sensing images and target recognition networks to the classification of dead tree removal progress information. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a method for classifying and identifying the progress information of pine dead wood removal based on unmanned aerial vehicle remote sensing, and to provide a technical support solution for the prevention and control of pine wood nematode disease by establishing a highly robust dead wood identification network through unmanned aerial vehicle remote sensing images and combined with a target recognition network.

[0005] To achieve the above object, the present invention adopts the following technical solution: a method for classifying and identifying pine dead wood removal progress information based on UAV remote sensing, comprising the following steps:

[0006] Step S1: Use a drone to obtain aerial image data of the dead pine trees in the target area during the outbreak period T1 and the final period of treatment T2, and obtain digital orthophoto maps DOM of the dead pine trees during the outbreak period and the final period of treatment after processing. T1 , DOM T2 and Digital Surface Map (DSM) T1 、DSM T2 ;

[0007] Step S2: Building a DDPTnet network model for identifying dead wood based on YOLOv8n;

[0008] Step S3: The DOM obtained in step S1 T1 The DDPTnet network model is used to obtain the dead tree location information before felling;

[0009] Step S4: Combine the dead wood point data before felling obtained in step S3 with the buffer mask DOM T2 The data eliminates the impact of dead trees before felling. The buffer distance is the average size of the pine tree crown in the user's measured area. The DDPTnet network model and the masked DOM are used to T2 The data obtained new dead tree location information;

[0010] Step S5: The DSM obtained in step S1 T1 、DSM T2 Calculate the grid difference to obtain the elevation change matrix, and use threshold segmentation to obtain the felled area;

[0011] Step S6: Overlay and analyze the felled area obtained in step S5 and the dead tree location data before felling obtained in step S3 to obtain felled dead tree location information and unfelled dead tree location information;

[0012] Step S7: The felled dead tree point data obtained in step S6 are identified using the remote sensing scene classification model to obtain felled and cleaned data point information and felled but uncleaned data point information.

[0013] In a preferred embodiment: in step S1, the heading overlap rate and lateral overlap rate of the acquired UAV remote sensing image are both better than 70% and the image resolution or ground sampling distance GSD is selected to be 0.1-0.15m; the flight altitude is determined in combination with the elevation of the highest altitude area in the area and the required GSD requirements, and the conversion formula is as follows:

[0014]

[0015] Among them, p is the pixel size, H is the flight altitude, and f is the focal length.

[0016] In a preferred embodiment: the specific method of step S2 is as follows:

[0017] Step S21: Using YOLOv8n as the baseline of DDPTnet, in the original YOLOv8 Backbone, only the features of the P3, P4 and P5 layers are input, and the features extracted from the P2 layer are integrated into the baseline; the Lion optimizer is added to shorten the training time;

[0018] Step S22: Based on YOLOv8n, an LSK module is added to expand the receptive field and strengthen the recognition of environmental features. The relevant background of the target environment is combined as a feature to add useful environmental weighted features to the input of interest from the perspective of the background.

[0019] Step S23: The loss function of DDPTnet adopts the loss function of YOLOv8n itself, which includes category loss and position loss. Among them, the category loss adopts the BCE loss calculation method, and the position loss is CIoU and DFL. The calculation formulas are as follows:

[0020] (1) Category loss function:

[0021]

[0022] where y n is the target value of the nth sample, is the predicted value of the nth sample, N is the number of categories;

[0023] (2) Position loss function:

[0024]

[0025] Among them, B and B gt Represent the predicted box and the real box area respectively, and IoU represents the overlap between the predicted box and the real box; b and b gt Represent the center points of the predicted box and the real box respectively, and ρ represents the Euclidean distance between the two center points, c represents the diagonal length of the minimum enclosing rectangle that can contain both the predicted box and the real box, α represents the weight function, v represents the consistency parameter of the aspect ratio, and w gt and h gt Represent the width and height of the target box, w and h represent the width and height of the prediction box respectively;

[0026] DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 )) (7)

[0027]

[0028]

[0029] Where y represents the actual value of the label, y i Indicates the label integral value, y i+1 Indicates the adjacent label integral value, S i represents the predicted probability of category i, S i+1 represents the predicted probability of category i+1.

[0030] In a preferred embodiment: In step S3, first, DOM T1The images were cropped into 512×512 blocks with 20% overlap in the horizontal and vertical directions, and then the blocks were identified one by one using DDPTnet. Secondly, the identified blocks were reassembled according to the number of rows and columns of the original image, and finally the target recognition frame was converted into vector point data with coordinate information, and the geometric center of the vector result was used as the dead wood point information.

[0031] In a preferred embodiment: the specific method of step S4 is as follows:

[0032] Step S41: Establishing a buffer zone and DOM based on the dead wood points before felling T2 The dead tree crown is masked to obtain the DOM after masking T2 Data; the buffer distance is the average size of the pine tree crown in the area measured by the user;

[0033] Step S42: The masked DOM obtained in step S41 is T2 Data was collected and the location information of newly added dead trees was obtained through DDPTnet network identification.

[0034] In a preferred embodiment: In step S5, firstly, the raster operation module in the GDAL library is used to perform the DSM T1 and DSM T2 Perform difference calculation, the specific formula is as follows:

[0035] DSM Δ =DSM T1 -DSM T2 (10)

[0036] Among them, DSM Δ Output grid size and DOM T1 Finally, the elevation of the pine felling pile is taken as the lowest height threshold, and the crown height is taken as the highest threshold to obtain the change elevation grid in the felled area;

[0037] Output raster size and DSM T1 The consistent elevation change area was finally determined by taking the trunk height of the pine tree as the lowest threshold and the crown height as the highest threshold to obtain the elevation change raster within the felled area.

[0038] In a preferred embodiment: in step S6, the dead wood points before felling in step 3 and the felling area obtained in step S5 are overlaid and analyzed using GIS software to obtain felled dead wood and unfelled dead wood.

[0039] In a preferred embodiment: the specific method of step S7 is as follows:

[0040] Step S71: Create a classification scene sample based on the dead tree canopy size and image spatial resolution characteristics; center on the dead tree felling point, and T2 The upper cropping generates a single square sample of 64×64 size, and the point geographic projection information is assigned to the corresponding sample using the affine transformation method to obtain the sample data;

[0041] Step S72: Use the YOLOv5n network that has been fine-tuned by the target detection network to perform classification; add the output part of the YOLOv5n network to the fully connected layer that integrates the features and the multi-classification softmax activation function and add the classification output layer to perform the classification task. The formula of the softmax activation function is as follows:

[0042]

[0043] Among them, Z i and Z j They represent the i-th and j-th elements of the input vector, respectively, e is the base of the natural logarithm, and K is the number of categories;

[0044] Step S73: Combine the dead wood felling identification results and use the acquired dead wood felling point information to cut the DOM T2 Image,During the classification process, vector surface data consistent with the cropped image range is constructed.,When the network determines the cleaning category of each image, it also creates a field and writes the category into the surface vector data corresponding to the image;

[0045] Step S74: Combined with affine transformation, each vector surface data has the same spatial coordinates as the image data. At this time, the vector surface data not only overlaps with the original felling point data, but also has its cleaning type. Through overlay analysis, two types of felling information, namely felled but not cleaned, and felled and cleaned, are obtained.

[0046] Compared with the prior art, the present invention has the following beneficial effects: The present invention solves the problem that the progress information of the treatment cannot be obtained in time during the outbreak period of dead wood and the end of the treatment. Through the combination of drone remote sensing and target recognition algorithm, a classification and recognition method of four types of treatment progress information, namely, newly added dead wood, uncut dead wood, cut but not cleaned, and cut and cleaned, can be obtained through two phases of drone remote sensing images. The recognition results can provide objective and effective decision-making support for the prevention and control of pine wood nematode disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic diagram of a method flow of a preferred embodiment of the present invention;

[0048] Figure 2 It is a schematic diagram of the extraction of felled areas in the preferred implementation area of ​​the present invention;

[0049] Figure 3 Schematic diagram of the DDPTnet model structure of the preferred embodiment of the present invention;

[0050] Figure 4 This is the dead wood identification process of the preferred embodiment of the present invention;

[0051] Figure 5 It is a thematic map of the control progress information of County B (B1) according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0053] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0054] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0055] like Figure 1-5 As shown, this embodiment provides a method for classifying and identifying progress information of pine dead wood removal based on UAV remote sensing, which is characterized by comprising the following steps:

[0056] Step S1: Use a drone to obtain aerial image data of the dead pine trees in the target area during the outbreak period (T1) and the end of the treatment period (T2), and obtain two digital orthophoto maps DOM after processing. T1 , DOM T2 and Digital Surface Map (DSM) T1 、DSM T2 .

[0057] Step S2: Building a DDPTnet network model for identifying dead wood based on YOLOv8n.

[0058] Step S3: The DOM obtained in step S1 T1 The DDPTnet network model is used to obtain the dead tree location information before felling.

[0059] Step S4: Combine the dead wood point data before felling obtained in step S3 with the buffer mask DOM T2Eliminate the impact of dead trees before felling. The buffer distance is the average size of the pine tree crown in the user's measured area. In this example, a circle with a buffer radius of 4m can cover the crown of the dead trees in the example. T2 Get the newly added dead tree location information.

[0060] Step S5: The DSM obtained in step S1 T1 、DSM T2 Calculate the grid difference to obtain the elevation change matrix, and use threshold segmentation to obtain the felled area.

[0061] Step S6: Overlay and analyze the felled area obtained in step S5 and the dead tree location data before felling obtained in step S3 to obtain felled dead tree location information and unfelled dead tree location information.

[0062] Step S7: The felled dead tree point data obtained in step S6 are identified using the remote sensing scene classification model to obtain felled and cleaned point data information and felled but not cleaned point data information.

[0063] In this embodiment, the specific method of step S1 is: the heading overlap rate and the lateral overlap rate of the acquired UAV remote sensing image are both better than 70% or above, and the image resolution is 8-12cm.

[0064] In this embodiment, the specific method of step S2 is as follows:

[0065] Step S21: Using YOLOv8n as the baseline of DDPTnet, in the original YOLOv8 Backbone, only the P3, P4 and P5 layer features are input, and the features extracted from the P2 layer are integrated into the baseline; the Lion optimizer is added to shorten the training time.

[0066] Step S22: Based on YOLOv8n, an LSK module is added to expand the receptive field and strengthen the recognition of environmental features. The relevant background of the target's surrounding environment is combined as a feature to add useful environmental weighted features to the input of interest from the perspective of the background.

[0067] Step S23: The loss function of DDPTnet adopts the loss function of YOLOv8n itself, which includes category loss and position loss. Among them, the category loss adopts the BCE loss calculation method, and the position loss is CIoU and DFL. The calculation formulas are as follows:

[0068] (1) Category loss function:

[0069]

[0070] where yn is the target value of the nth sample, is the predicted value of the nth sample, N is the number of categories;

[0071] (2) Position loss function:

[0072]

[0073] Among them, B and B gt Represent the predicted box and the real box area respectively, and IoU represents the overlap between the predicted box and the real box. b and b gt Represent the center points of the predicted box and the real box respectively, and ρ represents the Euclidean distance between the two center points, c represents the diagonal length of the minimum enclosing rectangle that can contain both the predicted box and the real box, α represents the weight function, v represents the consistency parameter of the aspect ratio, and w gt and h gt Represent the width and height of the target box, w and h represent the width and height of the prediction box respectively;

[0074] DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(yy i )log(S i+1 )) (7)

[0075]

[0076] Where y represents the actual value of the label, i Indicates the label integral value, y i+1 Indicates the adjacent label integral value, S i represents the predicted probability of category i, S i+1 Represents the predicted probability of category i+1. .

[0077] In this embodiment, the specific method of step S3 is: first, DOM T1 The images were cropped into 512×512 blocks with 20% overlap in the horizontal and vertical directions, and then the blocks were identified in turn using DDPTnet. Secondly, the identified blocks were reassembled according to the number of rows and columns of the original image, and finally the target recognition frame was converted into vector point data with coordinate information, and the geometric center of the vector result was used as the dead wood point information.

[0078] In this embodiment, the specific method of step S4 is as follows:

[0079] Step S41: Based on the dead wood points before felling, a circle with a buffer distance of 4m is established between the DOM T2The dead tree crown is masked to obtain the DOM after masking T2 data.

[0080] Step S42: The masked DOM obtained in step S41 is T2 Data, using DDPTnet to identify and obtain the newly added dead tree location information. In this example, taking County B (B1) as an example, the accuracy indicators of the newly added dead tree identification results are shown in Table 1:

[0081] Table 1. New dead wood identification results

[0082]

[0083] In this embodiment, the specific method of step S5 is: first, use the raster operation module in the GDAL library to calculate the DSM T1 and DSM T2 Perform difference calculation, the specific formula is as follows:

[0084] DSM Δ =DSM T1 -DSM T2 (10)

[0085] Among them, DSM Δ Output grid size and DOM T1 Finally, the elevation of the pine felling pile was used as the lowest height threshold and the crown height was used as the highest threshold to obtain the elevation change grid in the felled area.

[0086] In this example, the trunk height of dead trees is generally higher than 3m, and the crown height is lower than 30m, so the threshold is set at 3-30m.

[0087] In this embodiment, the specific method of step S6 is: the dead wood points before felling in step 3 and the felling area obtained in step S5 are superimposed and analyzed using GIS software to obtain felled dead wood and unfelled dead wood. In this example, the accuracy index related to the felled dead wood identification result is shown in Table 2:

[0088] Table 2 Identification results of felled dead trees

[0089]

[0090] In this embodiment, the specific method of step S7 is as follows:

[0091] Step S71: Create a classification scene sample based on the dead tree canopy size and image spatial resolution. T2A single square sample of 64×64 size is generated by cropping, and the point geographic projection information is assigned to the corresponding sample using the affine transformation method to obtain the sample data.

[0092] Step S72: Classify using the YOLOv5n network that has undergone fine-tuning of the target detection network learning strategy.

[0093] The output part of the YOLOv5n network is added to the fully connected layer that integrates the features and the multi-classification softmax activation function and the classified output layer to perform the classification task. The formula of the softmax activation function is as follows:

[0094]

[0095] Among them, Z i and Z j represents the i-th and j-th elements of the input vector, e is the base of the natural logarithm, and K is the number of categories. In this example, the accuracy of the YOLOv5n network after fine-tuning the learning strategy of the target detection network reaches 80.07%. The specific accuracy is shown in Table 3:

[0096] Table 3. Accuracy of identification results of felled dead wood cleaning types

[0097]

[0098] Step S73: Combine the dead wood felling identification results and use the acquired dead wood felling point information to cut the DOM T2 Image, during the classification process, vector surface data consistent with the cropped image range is constructed. When the network determines the cleaning category of each image, it also creates a field and writes the category into the surface vector data corresponding to the image.

[0099] Step S74: Combined with affine transformation, each vector surface data has the same spatial coordinates as the image data. At this time, the vector surface data not only overlaps with the original felling point data, but also has its cleaning type. Through overlay analysis, two types of felling and cleaning information are obtained: felled but not cleaned, and felled and cleaned.

[0100] In particular, this embodiment includes three detection areas, namely, three sub-areas A, B, and C, among which A is located in County A of a certain city, with an altitude of 20-380m in the detection area, B is located in County B of a certain city, with an altitude of 110-647m in the detection area, and C is located in Quangang District of Quanzhou City, with an altitude of 100-616m in the detection area. These three places are typical hilly areas with obvious terrain undulations. This embodiment will be described in detail with an example below. This embodiment selects the dynamic observation area B1 of County B as the sampling detection area for the progress of eradication. Among them, the detection area of ​​B1 is 1.63km2. DJI Phantom 4RTK and DJI Mavic 3E drones are used in County B. The drone parameters are shown in Table 4:

[0101] Table 4 Details of route planning for sampling inspection area for control progress in County B (B1)

[0102]

[0103] Figure 3 is the extraction of felled area in B1, where a is the regional DOM T1 ; b is the regional DOM T2 ; c is the felling area and DOM T2 Overlay graph; d is DOM T1 Detailed image, e is DOM T2 Detailed picture; f is the felled area and DOM T2 Overlay detail image; g is the DSM difference image (DSM T1 -DSM T2 ).

[0104] Figure 5 It is a special map of the progress of dead wood removal in the B1 area. It can intuitively analyze the trends of dead trees in the forest area and obtain progress information on the removal of four types of dead trees: uncut dead trees, newly added dead trees, cut down but not cleared, and cut down and cleared.

[0105] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing, characterized in that: The following steps are involved: Step S1: Use a drone to obtain aerial image data of the dead pine trees in the target area during the outbreak period T1 and the final period of treatment T2, and obtain a digital orthophoto map DOM of the dead pine trees during the outbreak period and the final period of treatment after processing. T1 , DOM T2 and Digital Surface Map (DSM) T1 、DSM T2 ; Step S2: Building a DDPTnet network model for identifying dead wood based on YOLOv8n; Step S3: The DOM obtained in step S1 T1 Use the DDPTnet network model to obtain dead tree location information before felling; Step S4: Combine the dead wood point data before felling obtained in step S3 with the buffer mask DOM T2 The data eliminates the impact of dead trees before felling. The buffer distance is the average size of the pine tree crown in the user's measured area. The DDPTnet network model and the masked DOM are used to T2 The data obtained new dead tree location information; Step S5: The DSM obtained in step S1 T1 、DSM T2 Calculate the grid difference to obtain the elevation change matrix, and use threshold segmentation to obtain the felled area; Step S6: Overlay and analyze the felled area obtained in step S5 and the dead tree location data before felling obtained in step S3 to obtain felled dead tree location information and unfelled dead tree location information; Step S7: The felled dead tree point data obtained in step S6 are identified using the remote sensing scene classification model to obtain felled and cleaned data point information and felled but uncleaned data point information.

2. The method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing according to claim 1 is characterized in that: In step S1, the heading overlap rate and lateral overlap rate of the acquired UAV remote sensing image are both better than 70% and the image resolution or ground sampling distance GSD is selected as 0.1-0.15m; the flight altitude is determined in combination with the elevation of the highest altitude area in the area and the required GSD requirements, and the conversion formula is as follows: Among them, p is the pixel size, H is the flight altitude, and f is the focal length.

3. The method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing according to claim 1 is characterized in that: The specific method of step S2 is as follows: Step S21: Using YOLOv8n as the baseline of DDPTnet, in the original YOLOv8 Backbone, only the features of the P3, P4 and P5 layers are input, and the features extracted from the P2 layer are integrated into the baseline; the Lion optimizer is added to shorten the training time; Step S22: Based on YOLOv8n, an LSK module is added to expand the receptive field and strengthen the recognition of environmental features. The relevant background of the target environment is combined as a feature to add useful environmental weighted features to the input of interest from the perspective of the background. Step S23: The loss function of DDPTnet adopts the loss function of YOLOv8n itself, which includes category loss and position loss. Among them, the category loss adopts the BCE loss calculation method, and the position loss is CIoU and DFL. The calculation formulas are as follows: (1) Category loss function: where y n is the target value of the nth sample, is the predicted value of the nth sample, N is the number of categories; (2) Position loss function: Among them, B and B gt Represent the predicted box and the real box area respectively, and IoU represents the overlap between the predicted box and the real box; b and b gt Represent the center points of the predicted box and the real box respectively, and ρ represents the Euclidean distance between the two center points, c represents the diagonal length of the minimum enclosing rectangle that can contain both the predicted box and the real box, α represents the weight function, v represents the consistency parameter of the aspect ratio, and w gt and h gt Represent the width and height of the target box, w and h represent the width and height of the prediction box respectively; DFL(S i ,S i+1 )=-((y i+1 -y)log(S i )+(y-y i )log(S i+1 )) (7) Where y represents the actual value of the label, i Indicates the label integral value, y i+1 Indicates the adjacent label integral value, S i represents the predicted probability of category i, S i+1 represents the predicted probability of category i+1.

4. The method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing according to claim 1 is characterized in that: In step S3, first, the DOM T1 The images were cropped into 512×512 blocks with 20% overlap in the horizontal and vertical directions, and then the blocks were identified one by one using DDPTnet. Secondly, the identified blocks were reassembled according to the number of rows and columns of the original image, and finally the target recognition frame was converted into vector point data with coordinate information, and the geometric center of the vector result was used as the dead wood point information.

5. The method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing according to claim 1 is characterized in that: The specific method of step S4 is as follows: Step S41: Establishing a buffer zone and DOM based on the dead wood points before felling T2 The dead tree crown is masked to obtain the DOM after masking T2 Data; the buffer distance is the average size of the pine tree crown in the area measured by the user; Step S42: The masked DOM obtained in step S41 is T2 Data was collected and the location information of newly added dead trees was obtained through DDPTnet network identification.

6. The method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing according to claim 1 is characterized in that: In step S5, the DSM is firstly processed using the raster operation module in the GDAL library. T1 and DSM T2 Perform difference calculation, the specific formula is as follows: DSM Δ =DSM T1 -DSM T2 (10) Among them, DSM Δ Output grid size and DOM T1 Finally, the elevation of the pine felling pile is taken as the lowest height threshold, and the crown height is taken as the highest threshold to obtain the change elevation grid in the felled area; Output raster size and DSM T1 The consistent elevation change area was finally determined by taking the trunk height of the pine tree as the lowest threshold and the crown height as the highest threshold to obtain the elevation change raster within the felled area.

7. The method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing according to claim 1 is characterized in that: In step S6, the dead wood points before felling in step 3 and the felling area obtained in step S5 are overlaid and analyzed using GIS software to obtain felled dead wood and unfelled dead wood.

8. The method for classifying and identifying the progress information of pine dead wood removal based on UAV remote sensing according to claim 1 is characterized by: The specific method of step S7 is as follows: Step S71: Create a classification scene sample based on the dead tree canopy size and image spatial resolution characteristics; center on the dead tree felling point, and T2 The upper cropping generates a single square sample of 64×64 size, and the point geographic projection information is assigned to the corresponding sample using the affine transformation method to obtain the sample data; Step S72: Use the YOLOv5n network that has been fine-tuned by the target detection network to perform classification; add the output part of the YOLOv5n network to the fully connected layer that integrates the features and the multi-classification softmax activation function and add the classification output layer to perform the classification task. The formula of the softmax activation function is as follows: Among them, Z i and Z j They represent the i-th and j-th elements of the input vector, respectively, e is the base of the natural logarithm, and K is the number of categories; Step S73: Combine the dead wood felling identification results and use the acquired dead wood felling point information to cut the DOM T2 Image,During the classification process, vector surface data consistent with the cropped image range is constructed.,When the network determines the cleaning category of each image, it also creates a field and writes the category into the surface vector data corresponding to the image; Step S74: Combined with affine transformation, each vector surface data has the same spatial coordinates as the image data. At this time, the vector surface data not only overlaps with the original felling point data, but also has its cleaning type. Through superposition analysis, two types of felling information, namely felled but not cleaned, and felled and cleaned, are obtained.

Images