Populus euphratica single tree crown segmentation method, device and equipment based on aerial image

Abstract

The present application relates to the intelligent mining direction of remote sensing data, and particularly relates to a poplar single tree crown segmentation method, device and equipment based on aerial image. The method comprises the following steps: acquiring a first multi-spectral unmanned aerial vehicle aerial image, and preprocessing the first multi-spectral unmanned aerial vehicle aerial image; inputting the first multi-spectral unmanned aerial vehicle aerial image into a crown top identification model based on a D-LinkNet network to obtain a poplar crown top seed point graph; using a watershed segmentation algorithm based on label control, taking the poplar crown top seed point graph and a first band of the first multi-spectral unmanned aerial vehicle aerial image as inputs to perform calculation, and segmenting a poplar single tree crown based on the first multi-spectral unmanned aerial vehicle aerial image; wherein the first band is an infrared band or a red edge band.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent mining of remote sensing data, in particular to fine cognition of Populus euphratica remote sensing information, and more particularly to a Populus euphratica single tree crown segmentation method, device and equipment based on aerial images. BACKGROUND

[0002] Populus euphratica has the characteristics of salt-tolerant, wind-sand-resistant, drought-resistant, etc., and can adapt to the extreme environment of desert area. It is the only forest tree species in oasis in arid desert area, and plays a very important role in maintaining ecological balance such as water conservation, biodiversity conservation and oasis development. Due to the influence of human activities such as logging, reclamation, reclamation and climate change, the area of Populus euphratica has decreased sharply, so Populus euphratica is listed as one of the third endangered plant species in China, and is identified as a priority for protection of forest genetic resources in the world's arid and semi-arid regions. Carrying out single tree crown segmentation of Populus euphratica can accurately identify single Populus euphratica, which can provide decision support for forestry management departments to carry out precise management and protection of Populus euphratica forest such as disease and pest control and irrigation. Since the single Populus euphratica crown is not large, generally not more than 20m 2 , the spatial resolution of the data is required to be high, so multi-spectral unmanned aerial image data is used for Populus euphratica single tree crown segmentation.

[0003] The single tree crown segmentation method based on high spatial resolution remote sensing (or unmanned aerial vehicle) image mainly includes edge detection method, watershed algorithm and pattern recognition image segmentation method. Among them, the optical image single tree crown segmentation method based on watershed algorithm is more commonly used. The watershed segmentation algorithm includes unmarked and marked control methods. The unmarked watershed segmentation algorithm has good application effect for natural images with simple scene, but for rich and complex remote sensing images, it will appear serious under-segmentation or over-segmentation phenomenon. The watershed algorithm based on marked control can avoid these problems to a certain extent under the condition that the marker symbol is accurate, so the key is how to automatically generate an accurate marker symbol. Usually, seed point automatic marking is carried out by using distance change or local maximum filtering method, etc. The accuracy of these methods depends on the size of the filtering window. If the window is too large, small trees will be missed, and if the window is too small, a single tree crown will be divided into multiple tree crowns, and the segmentation effect is poor in the area where trees are dense. And since the watershed algorithm is based on morphological principle for segmentation, it is difficult to distinguish the tree species in the area. The tree species is generally identified by using semantic segmentation or classification method. In the area where Populus euphratica grows sparsely, as long as Populus euphratica is identified and classified, the segmentation of Populus euphratica single tree crown can be realized. In fact, most of Populus euphratica grows densely, especially the Populus euphratica forest growing on the river bank. There is overlap between tree crowns. Using ordinary semantic segmentation method can only identify Populus euphratica forest, but it is difficult to identify single Populus euphratica. SUMMARY

[0004] The application comprehensively analyzes the characteristics of Populus euphratica in multispectral unmanned aerial vehicle aerial images, combines the advantages of a deep learning network model and a watershed segmentation method based on marker control, and proposes a watershed segmentation algorithm based on deep learning to generate marker seed points to realize the segmentation of Populus euphratica single tree crown in multispectral unmanned aerial vehicle aerial images. The application uses the neural network model of D-LinkNet to automatically generate Populus euphratica crown top seed points as markers for the watershed algorithm, which can solve the problems that the watershed algorithm cannot identify tree species and the segmentation effect of dense area tree crown is poor, thereby realizing accurate and automatic segmentation of Populus euphratica single tree crown. It is a practical Populus euphratica single tree crown segmentation method.

[0005] The embodiment of the application describes a Populus euphratica single tree crown segmentation method, device and equipment based on aerial images, which combines the advantages of a deep learning network model that can identify tree species and a watershed segmentation method based on marker control that has a continuous and closed boundary after segmentation and the advantages that the segmented regions are disjointed. The method can realize automatic and accurate segmentation of Populus euphratica single tree crown, and the specific steps are as follows:

[0006] (I) Automatic generation of Populus euphratica crown top seed points based on D-LinkNet network

[0007] Due to the rich and complex information of remote sensing images, it is difficult for conventional marker-controlled watershed segmentation methods to accurately mark the seed points of dense tree crowns, and it is also difficult to accurately identify Populus euphratica. However, the solar radiation intensity measured at the highest point of the tree crown top is the largest, and the overlapping of tree crowns in dense areas generally occurs at the edge of the tree crown, so the information of the tree crown top is not easily covered. In view of the above problems and the characteristics of the tree crown, the application proposes to automatically generate Populus euphratica tree crown top seed points by using a deep learning-based semantic segmentation method. Since the Populus euphratica tree crown top target is small, and the target of sparse and withered Populus euphratica is weak, a D-LinkNet deep learning model with strong recognition ability, large receptive field and fusion of multi-scale information characteristics is selected, so that it has good effect in extracting Populus euphratica tree crown top information. The automatic generation of Populus euphratica tree crown top seed points based on D-LinkNet network mainly includes two steps of construction of Populus euphratica tree crown top sample set and model training;

[0008] The marking of the sample at the top of the poplar canopy is different from that of the sample of the spatial distribution of poplar. The marking of the sample at the top of the poplar canopy has no exact boundary, but the marking should be performed with the top of the poplar canopy as the center, and each canopy top should be independent and cannot intersect or overlap. When the sample set of the top of the poplar canopy is constructed, the economic forest area planted by people, the shrub vegetation type area, the dense poplar forest area, the sparse poplar forest area, and the land surface type area of the dry poplar should be selected. In order to increase the diversity of the sample set and cover as many remote sensing image scenes as possible, the image expression form is randomly and complexly transformed by using the data augmentation method, the image radiation change caused by different conditions during the unmanned aerial vehicle flight is simulated, and the model is adapted to the image under various complex atmospheric conditions.

[0009] During the model training, the near-infrared and red edge bands are very sensitive in the aspect of vegetation identification. Therefore, in order to fully utilize the spectral information of the unmanned aerial vehicle image and improve the recognition accuracy of the poplar, the input data is a multispectral unmanned aerial vehicle image with a size of 1024*1024*5 and the corresponding poplar canopy top marking sample, the D-LinkNet network model is used for training, and the poplar canopy top seed point is automatically generated. The generated canopy top seed point also has the characteristics that the canopy top is the center, each seed point is independent, and there is no intersection and overlap.

[0010] (II) Poplar canopy segmentation based on marker-controlled watershed segmentation algorithm

[0011] The accuracy of the marker-controlled watershed segmentation algorithm depends on the accuracy of the marker. Therefore, the generated canopy seed point is used as the marker control symbol of the watershed segmentation algorithm, which can greatly improve the accuracy of the poplar single tree canopy segmentation. First, the seed point is dilated to make its range larger than the distribution range of the poplar. Then, the seed point image after the dilation operation is used to mask the near-infrared band or the red edge band. The image after the OSTU binarization is obtained, which can reflect the distribution of the poplar more completely. Then, the open operation is used to remove the small white noise points in the binary image, and the dilation operation is performed to make part of the background become the boundary of the poplar. The black region in the obtained image is the real background. Then, the poplar canopy seed point image is input. The white region in the image is the real foreground. Thus, it can be determined which is the poplar region and which is the background region. Then, the marker is created according to the poplar canopy seed point image. The watershed algorithm regards the region with the marker 0 as an uncertain region, the region with the marker 1 as a background region, and the positive integer greater than 1 as a poplar. Finally, the segmentation result is obtained by using the watershed algorithm.

[0012] DRAWINGS

[0013] Figure 1A flow chart of poplar single tree crown segmentation based on aerial image is provided by the embodiment of the present application.

[0014] Figure 2 A poplar crown top sample marking schematic diagram is provided by the embodiment of the present application.

[0015] Figure 3A A poplar single tree crown segmentation result schematic diagram is provided by the embodiment of the present application.

[0016] Figure 3B A poplar single tree crown segmentation result schematic diagram is provided by the embodiment of the present application.

[0017] Figure 3C A poplar single tree crown segmentation result schematic diagram is provided by the embodiment of the present application.

[0018] Figure 3D A poplar single tree crown segmentation result schematic diagram is provided by the embodiment of the present application.

[0019] Figure 4 A flow chart of poplar single tree crown segmentation method based on aerial image is provided by the embodiment of the present application.

[0020] Figure 5 A schematic diagram of poplar single tree crown segmentation device based on aerial image is provided by the embodiment of the present application.

[0021] Figure 6 A structural schematic diagram of a computing device is provided by the embodiment of the present application. DETAILED DESCRIPTION

[0022] Next, the scheme of the present application is introduced in combination with the drawings.

[0023] The semantic segmentation of deep learning based on high spatial resolution remote sensing (or unmanned aerial vehicle) image generally refers to the accurate segmentation according to the boundary of a specific object. Since the solar radiation intensity measured at the highest point of the crown top is the largest, the tree crown overlap in the dense area generally occurs at the edge of the tree crown, and the crown top information is not easy to be covered, therefore, the present application uses the semantic segmentation method of deep learning to extract the crown top information as the marked seed point of the watershed segmentation algorithm, so as to solve the problems that the ordinary watershed algorithm cannot identify the tree species type and the dense area segmentation effect is poor, and the problem that the ordinary semantic segmentation cannot perform single tree segmentation, and realize the accurate and automatic segmentation of the poplar single tree crown.

[0024] The application comprehensively analyzes the characteristics of Populus euphratica in multi-spectral unmanned aerial vehicle aerial image, and proposes a watershed segmentation algorithm based on deep learning to generate labeled seed points to realize the segmentation of Populus euphratica single tree crown in multi-spectral unmanned aerial vehicle aerial image. The application uses the neural network model of D-LinkNet to automatically extract the Populus euphratica crown top information as the seed point of the labeled watershed algorithm, which can solve the problem that the watershed algorithm cannot identify tree species and the segmentation effect in dense areas is poor, so as to realize the accurate and automatic segmentation of Populus euphratica single tree crown. The application relates to three parts of unmanned aerial vehicle multi-spectral image preprocessing, seed point generation and crown segmentation.

[0025] The embodiment of the application provides a Populus euphratica single tree crown segmentation method based on multi-spectral unmanned aerial vehicle aerial image. Figure 1 The method comprises the following specific steps:

[0026] (I) data processing

[0027] The collected multi-spectral unmanned aerial vehicle aerial image is processed by using PhotoScan software, and the following processes are required: aligning photos, creating dense point clouds, creating a triangulated irregular network (TIN) model, creating a digital elevation model (DEM), creating an orthomosaic map, etc. Finally, the required orthomosaic result is exported. In order to avoid memory overflow and other problems in the entire calculation process, the data is divided into blocks of 1024*1024 size, and a 10-pixel overlap area is set. The spatial resolution of the collected multi-spectral unmanned aerial vehicle aerial image is 8 cm, and it contains blue (455-485 nm), green (550-570 nm), red (658-678 nm), red edge (712-722), and near-infrared (820-860) bands.

[0028] 1024 in the block of 1024*1024 size represents a pixel.

[0029] (II) automatic generation of Populus euphratica crown top seed points based on D-LinkNet network

[0030] Since the Populus euphratica crown top target is small, and the sparse and dry Populus euphratica target is weak, the D-LinkNet deep learning model with strong recognition ability, large receiving domain and multi-scale information fusion characteristics is selected, so that it has good effect in Populus euphratica crown top information extraction. The automatic generation of Populus euphratica crown top seed points based on D-LinkNet network mainly includes two steps of construction of Populus euphratica crown top sample set and model training.

[0031] (1) Construction of Populus euphratica crown top sample set

[0032] The label of the poplar crown top sample is different from that of the huangpopulus spatial distribution sample. The label of the crown top sample does not have a definite boundary, but the huangpopulus crown top should be marked as the center when labeling, and each crown top should be independent and cannot intersect and overlap. Among them, the huangpopulus spatial distribution refers to the huangpopulus crown distribution. The labeling can be performed manually. Figure 2 The characteristics of huangpopulus crown top labeling are shown. Since the unmanned aerial vehicle aerial image shot has rich ground object types, contains artificial economic forest, shrub and other vegetation types, therefore when constructing the huangpopulus crown top sample set, the artificial economic forest area, different vegetation type area, dense huangpopulus forest area, sparse huangpopulus forest area and dry huangpopulus area and other ground object type areas should be selected. At the same time, in order to increase the diversity of the sample set and cover as many remote sensing image scenes as possible, the image expression form is randomly and complexly transformed by using the conventional data augmentation methods such as rotation, mirror image and the methods such as Gamma correction, log correction and histogram equalization, so as to simulate the image radiation changes caused by the unmanned aerial vehicle sensor under different atmospheric conditions, geographical position, terrain and ground objects, so that the model is adapted to the image under various complex atmospheric conditions.

[0033] (2) D-LinkNet network model

[0034] D-LinkNet network is a kind of deep learning network using Linknet with precoder as backbone network and using a hollow convolution layer in the center part, with the characteristics of high precision and fast running speed. D-LinkNet uses ResNet34 pre-trained on ImageNet dataset as its encoder, which is an effective method when the number of training images is limited. The decoder uses transpose convolution layer for upsampling, which restores the size of the feature map from 32*32 to 1024*1024. Because the input image resolution is very high, the poplar type is more complex, and the poplar occupies a very small part in the whole image, it is necessary to preserve detailed spatial information. Therefore, the hollow convolution layer is used in the center part, which can adjust the receptive field of the feature map without reducing the resolution of the feature map. The middle part of the hollow convolution is connected with a skip connection, stacked in a cascade mode, and makes use of the multi-resolution function, so as to achieve the purpose of increasing the receptive field and preserving spatial information. Because near-infrared and red edge bands are very sensitive in vegetation identification, in order to fully utilize the spectral information of unmanned aerial image and improve the accuracy of poplar identification, the input data is 1024*1024*5 multispectral unmanned aerial image and the corresponding poplar canopy top labeled sample, which is trained by using D-LinkNet network model, so as to automatically generate poplar canopy top seed points. The seed points generated based on D-LinkNet deep learning network are basically consistent with the characteristics of the labeled seed points, that is, each seed point is independent and does not intersect, and the tree canopy top is the center. 1024 in 1024*1024*5 represents pixels, and 5 represents waveband.

[0035] (Three) Poplar canopy segmentation based on marker-controlled watershed segmentation algorithm

[0036] The accuracy of the marker-controlled watershed segmentation algorithm depends on the accuracy of the marker, and the poplar canopy top information generated based on the deep learning network has high accuracy, which can be used as the marker control seed point graph of the watershed segmentation algorithm, so as to improve the accuracy of poplar single tree canopy segmentation. The specific steps are as follows:

[0037] Because the near-infrared band and the red edge band are sensitive to vegetation, the near-infrared band or the red edge band of the multispectral unmanned aerial vehicle aerial image is used as input. The usual method is to use OSTU binarization to binarize the input image, but the centimeter-level aerial image information is rich, and the ground object type is complex, so it is difficult to completely extract Populus euphratica by using OSTU binarization. Therefore, the seed point graph obtained is simply dilated to make its range greater than the distribution range of Populus euphratica, and then the seed point graph after dilatation is used to mask the near-infrared band or the red edge band, and then the image after masking is OSTU binarized to obtain a binary graph, and then the open operation is used to remove the small white noise points in the binary graph, and dilatation operation is performed, so that part of the background becomes the boundary of Populus euphratica, and the black area in the obtained image is definitely the real background; wherein the background refers to the non-Populus euphratica region;

[0038] Then input the Populus euphratica crown seed point graph, and the white area in the image is definitely the real foreground; the foreground refers to the region where Populus euphratica is located.

[0039] Thus it can be determined which is the Populus euphratica region and which is the background region, and then the Populus euphratica crown seed point graph is created to create a label;

[0040] The watershed algorithm regards the region marked as 0 as an uncertain region, the region marked as 1 as a background region, and a positive integer greater than 1 as Populus euphratica, and finally the watershed algorithm is used to obtain the segmentation result.

[0041] Figure 3A 、 Figure 3B 、 Figure 3C 、 Figure 3D The segmentation result of the Populus euphratica single tree crown is shown, and the watershed segmentation algorithm based on the Populus euphratica crown seed point label generated by deep learning has achieved good results, which not only avoids the over-segmentation problem brought by the label-free method, but also eliminates the under-segmentation phenomenon caused by inaccurate labeling such as distance transform, and can accurately extract the number of Populus euphratica. By observing the dense growth area of Populus euphratica ( Figure 3B 、 Figure 3C 、 Figure 3D ) and the sparse growth area of Populus euphratica ( Figure 3A ), it is found that the segmentation results are good.

[0042] The application combines a deep learning network model and a marker control based watershed segmentation method to segment poplar single tree crowns from multispectral unmanned aerial vehicle aerial images, and proposes a poplar single tree crown segmentation method based on multispectral unmanned aerial vehicle aerial images. The D-LinkNet deep learning model is used to automatically extract the crown top information as the marker seed point of the watershed segmentation algorithm, which can automatically segment the poplar single tree crown and has good effect, and can solve the problem that the ordinary watershed algorithm cannot identify the tree species and the segmentation effect in dense areas is poor, and is a practical poplar single tree crown segmentation method.

[0043] Based on the poplar single tree crown segmentation method based on multispectral unmanned aerial vehicle aerial images described above, the embodiment of the application further provides a poplar single tree crown segmentation method based on multispectral unmanned aerial vehicle aerial images, which is described in detail below. Figure 4 The method comprises the following steps.

[0044] 401, acquiring a first multispectral unmanned aerial vehicle aerial image, and preprocessing the first multispectral unmanned aerial vehicle aerial image;

[0045] 402, inputting the first multispectral unmanned aerial vehicle aerial image into a crown top identification model based on a D-LinkNet network to obtain a poplar crown top seed point graph;

[0046] 403, using a marker control based watershed segmentation algorithm, taking the poplar crown top seed point graph and a first band of the first multispectral unmanned aerial vehicle aerial image as inputs to segment poplar single tree crowns from the first multispectral unmanned aerial vehicle aerial image; wherein the first band is an infrared band or a red edge band.

[0047] In some embodiments, the preprocessing comprises aligning photos, creating a dense point cloud, creating a triangular irregular network (TIN) model, creating a digital elevation model, creating an orthomosaic, and exporting the required orthomosaic.

[0048] In one example of this embodiment, the preprocessing comprises dividing the first multispectral unmanned aerial vehicle aerial image data into blocks of 1024*1024 size, and setting an overlap area of 10 pixels.

[0049] In some embodiments, the method further comprises acquiring a second multispectral unmanned aerial vehicle aerial image, and preprocessing the second multispectral unmanned aerial vehicle aerial image; labeling the poplar crown top as the center in the second multispectral unmanned aerial vehicle aerial image to obtain training data; wherein the different poplar crown tops are not intersected or overlapped; using the training data and the second multispectral unmanned aerial vehicle aerial image to train the D-LinkNet network to obtain the crown top identification model.

[0050] In one example of the embodiment, the training of the D-LinkNet network using the training data and the second multi-spectral unmanned aerial vehicle aerial image includes training the D-LinkNet network using 1024*1024*5 data in the training data and the second multi-spectral unmanned aerial vehicle aerial image.

[0051] In another example of the embodiment, the labeling of the Populus euphratica crown top in the second multi-spectral unmanned aerial vehicle aerial image to obtain the training data includes using rotation, mirror, and random complex transformation of image forms such as Gamma correction, log correction, and histogram equalization to increase the diversity of the training data.

[0052] In some embodiments, the calculation using the marker control-based watershed segmentation algorithm with the Populus euphratica crown top seed point graph and the first band of the first multi-spectral unmanned aerial vehicle aerial image as inputs includes: performing dilation operation on the Populus euphratica crown top seed point graph so that the range of the Populus euphratica crown top seed point graph is greater than the distribution range of the Populus euphratica; then using the Populus euphratica crown top seed point graph after the dilation operation to mask the near-infrared band or the red edge band; performing Otsu binarization on the masked image to obtain a binary graph; then using opening operation to remove small white noise points in the binary graph, and performing dilation operation so that part of the background becomes the boundary of the Populus euphratica, and the black region in the obtained image is the real background and the white region is the real foreground; and creating a marker according to the Populus euphratica crown top seed point graph.

[0053] The method provided by the embodiment of the application combines a deep learning network model and a marker control-based watershed segmentation method to segment Populus euphratica single-tree crowns from multi-spectral unmanned aerial vehicle aerial images, proposes a Populus euphratica single-tree crown segmentation method based on multi-spectral unmanned aerial vehicle aerial images, uses a D-LinkNet deep learning model to automatically extract crown top information as a marker seed point of a watershed segmentation algorithm, can automatically segment Populus euphratica single-tree crowns, and has good effect, and can solve the problems that an ordinary watershed algorithm cannot identify tree species and has poor segmentation effect in dense areas, and is a practical Populus euphratica single-tree crown segmentation method.

[0054] The embodiment of the application further provides a Populus euphratica single-tree crown segmentation device based on multi-spectral unmanned aerial vehicle aerial images, as shown in Figure 5 The device includes:

[0055] The acquisition unit 510 is configured to acquire a first multi-spectral unmanned aerial vehicle aerial image and pre-process the first multi-spectral unmanned aerial vehicle aerial image.

[0056] The obtaining unit 520 is configured to input the first multi-spectral unmanned aerial vehicle aerial image into a D-LinkNet network-based tree crown top identification model to obtain a poplar tree crown top seed point graph.

[0057] The computing unit 530 is configured to calculate the poplar tree crown top seed point graph and a first band of the first multi-spectral unmanned aerial vehicle aerial image as inputs by using a marker-controlled watershed segmentation algorithm, to segment the poplar single tree crown from the first multi-spectral unmanned aerial vehicle aerial image; the first band is an infrared band or a red edge band.

[0058] Referring to Figure 6 , the embodiment of the present application provides a computing device 600, which includes a processor 610 and a memory 620. The memory 620 is configured to store computer instructions. The processor 610 is configured to execute the computer instructions stored in the memory 620, so that the computing device 600 can execute the method embodiment shown in the above Figure 4 .

[0059] It can be understood that the processor in the embodiment of the present application can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor can be a microprocessor or any conventional processor.

[0060] The embodiment of the present application further provides a computer storage medium, which includes computer software instructions, and the computer software instructions include a program for implementing the method shown in the above Figure 4 . The computer readable storage medium can be any available medium or a data storage device including one or more available media that can store data. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk) and the like.

[0061] The embodiment of the present application further provides a computer program product, which includes a program for implementing the method shown in the above Figure 4 . The computer program product is a software or a program product that can run on a processor or be stored in any available medium.

[0062] Finally, it should be noted that: the above examples are used to illustrate the technical solutions of the present application, but not limited to them; although the present application is described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the protection scope of the technical solutions of the embodiments of the present application.

Claims

1. A poplar single tree crown segmentation method based on aerial image, characterized in that, The method comprises the following steps: obtain first multi-spectral unmanned aerial vehicle aerial image, and pre-process the first multi-spectral unmanned aerial vehicle aerial image; input the first multi-spectral unmanned aerial vehicle aerial image into a crown top identification model based on a D-LinkNet network to obtain a poplar crown top seed point graph; use a marker-controlled watershed segmentation algorithm to calculate the poplar crown top seed point graph and a first band of the first multi-spectral unmanned aerial vehicle aerial image as input to identify poplar single tree crowns in the first multi-spectral unmanned aerial vehicle aerial image; wherein the first band is an infrared band or a red edge band; wherein the use of the marker-controlled watershed segmentation algorithm to calculate the poplar crown top seed point graph and the first band of the first multi-spectral unmanned aerial vehicle aerial image as input comprises: performing dilation operation on the poplar crown top seed point graph so that the range of the poplar crown top seed point graph is greater than the distribution range of the poplar; then using the poplar crown top seed point graph after the dilation operation to mask the near-infrared band or the red edge band; then performing Otsu binarization on the masked image to obtain a binary graph; then using opening operation to remove small white noise points in the binary graph, and performing dilation operation so that part of the background becomes the boundary of the poplar, and the black area in the obtained image is the real background and the white area is the real foreground; create a marker according to the poplar crown top seed point graph.

2. The method of claim 1, wherein, The pre-processing comprises aligning photos, creating a dense point cloud, creating a TIN model, creating a digital elevation model, creating an orthomosaic, and exporting the required orthomosaic.

3. The method of claim 2, wherein, The pre-processing comprises dividing the first multi-spectral unmanned aerial vehicle aerial image data into blocks of 1024*1024 size and setting an overlap area of 10 pixels.

4. The method of claim 1, wherein, The method further comprises: obtain second multi-spectral unmanned aerial vehicle aerial image, and pre-process the second multi-spectral unmanned aerial vehicle aerial image; label the poplar crown top as the center in the second multi-spectral unmanned aerial vehicle aerial image to obtain training data; wherein the different poplar crown tops are not intersected and not overlapped. use the training data and the second multi-spectral unmanned aerial vehicle aerial image to train the D-LinkNet network to obtain the crown top identification model.

5. The method of claim 4, wherein, The use of the training data and the second multi-spectral unmanned aerial vehicle aerial image to train the D-LinkNet network comprises: use the training data and 1024*1024*5 data in the second multi-spectral unmanned aerial vehicle aerial image to train the D-LinkNet network.

6. The method of claim 4, wherein, The labeling of the poplar crown top as the center in the second multi-spectral unmanned aerial vehicle aerial image to obtain training data comprises: use rotation, mirror image, and Gamma correction, log correction, and histogram equalization to randomly and complexly transform the image representation to increase the diversity of the training data.

7. A poplar single tree crown segmentation device based on aerial image, characterized in that, The method comprises the following steps: an obtaining unit is configured to obtain first multi-spectral unmanned aerial vehicle aerial image, and pre-process the first multi-spectral unmanned aerial vehicle aerial image; An obtaining unit is configured to input the first multi-spectral unmanned aerial vehicle aerial image into a crown top identification model based on a D-LinkNet network to obtain a Populus euphratica crown top seed point map; A calculating unit is configured to calculate the Populus euphratica crown top seed point map and a first band of the first multi-spectral unmanned aerial vehicle aerial image as inputs by using a marker-controlled watershed segmentation algorithm to identify Populus euphratica individual tree crowns in the first multi-spectral unmanned aerial vehicle aerial image, wherein the first band is an infrared band or a red edge band. The calculation by using the marker-controlled watershed segmentation algorithm and taking the Populus euphratica crown top seed point map and the first band of the first multi-spectral unmanned aerial vehicle aerial image as inputs includes: Performing dilation operation on the Populus euphratica crown top seed point map so that the range of the Populus euphratica crown top seed point map is greater than the distribution range of Populus euphratica; Then, performing mask operation on the near-infrared band or the red edge band by using the Populus euphratica crown top seed point map after the dilation operation; Then, performing maximum inter-class variance method OSTU binarization on the image after the mask operation to obtain a binary image; Then, removing small white noise points in the binary image by using opening operation and performing dilation operation so that part of the background becomes the boundary of Populus euphratica, and the black region in the obtained image is the real background and the white region is the real foreground; According to the Populus euphratica crown top seed point map, a marker is created.

8. A computing device, comprising: A computer program product including a computer readable medium having stored computer program instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, A computer program product including computer program instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-6.

Images