A method and system for monitoring regional vegetation cover

Abstract

The application discloses a kind of regional vegetation coverage monitoring method and system, method includes: obtaining the aerial picture in monitoring area, the aerial picture is BGR format;The aerial picture is converted into first gray image, the number of non-zero pixels in the first gray image is counted;The aerial picture is thresholded, and only the mask containing suitable color range part is created;The aerial picture is operated by mask, and the part matched with color range is selected in the aerial picture, and the selected area is confirmed;The selected area is converted into second gray image, and the number of non-zero pixels in the second gray image is counted;Vegetation coverage is calculated according to the number of non-zero pixels in the first gray image and the number of non-zero pixels in the second gray image.This method can effectively process aerial image data, improve vegetation coverage monitoring efficiency and accuracy, effectively save manpower.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vegetation coverage monitoring, and in particular to a regional vegetation coverage monitoring method and system. BACKGROUND

[0002] Coverage is an important quantitative indicator of vegetation growth conditions, and is an important basic parameter in global and regional climate numerical simulation, land-atmosphere interaction, hydrology and ecology research. Accurate acquisition of long time series data of vegetation coverage is of great significance to related field application research.

[0003] According to the spatial scale characteristics of the monitoring sample area, the vegetation coverage index is mainly obtained by large-scale remote sensing quantitative inversion and ground sample scale manual survey monitoring. At present, ground sample scale manual survey is still the main method to obtain vegetation coverage. The traditional visual method has large error due to human subjectivity and sampling uncertainty, the survey frequency is usually annual, and is limited by human resources. In the actual vegetation monitoring business work, how to improve the time frequency of monitoring data has become a bottleneck for the development of traditional monitoring.

[0004] In recent years, unmanned aerial vehicle monitoring has gradually become a new monitoring method. Studies have shown that small unmanned aerial vehicles have great potential in vegetation monitoring, grassland resource management and health evaluation, and can improve the accuracy of vegetation coverage monitoring. Unmanned aerial vehicles have the advantages of high flexibility and timeliness, and have certain advantages in assisting manual monitoring and surveying. Various regions are continuously exploring the construction of positioning observation stations to collect a large number of image data sets. In traditional image data analysis, manual visual estimation or software human-computer interaction is needed to extract vegetation coverage, but this method has the problems of low efficiency, strong subjectivity and high labor cost, and cannot meet the demand for efficient and accurate automatic extraction of vegetation coverage. Therefore, a fast and accurate image analysis algorithm is an important basis for the automation and intelligent development of observation stations, and has become a technical problem to be solved. SUMMARY

[0005] In view of the defects in the prior art, the present application provides a regional vegetation coverage monitoring method and system to solve the technical problems proposed in the background art.

[0006] In a first aspect, the present application provides a regional vegetation coverage monitoring method, comprising:

[0007] Obtaining an aerial photograph in a monitoring area, the aerial photograph being in BGR format;

[0008] Converting the aerial photograph into a first gray-scale image, and counting the number of non-zero pixels in the first gray-scale image;

[0009] Performing threshold processing on the aerial photograph to create a mask containing only a suitable color range portion;

[0010] The aerial picture is operated by a mask, a part matching a color range is selected in the aerial picture, and a selected region is confirmed;

[0011] The selected region is converted into a second gray image, and the number of non-zero pixels in the second gray image is counted;

[0012] The vegetation coverage is calculated according to the number of non-zero pixels in the first gray image and the number of non-zero pixels in the second gray image.

[0013] Further, the ratio of the number of non-zero pixels in the second gray image to the number of non-zero pixels in the first gray image is taken as the vegetation coverage.

[0014] Further, the threshold processing on the aerial picture comprises:

[0015] The threshold processing on the aerial picture is performed by using an HSV color model.

[0016] Further, the parameters of the HSV color model comprise:

[0017] H min is 35, H max is 99, S min is 43, S max is 255, V min is 43, V max is 255; wherein,

[0018] H min represents a hue lower limit, H max represents a hue upper limit, S min represents a saturation lower limit, S max represents a saturation upper limit, V min represents a lightness lower limit, V max represents a lightness upper limit.

[0019] Further, before the aerial picture is converted into a first gray image, the method further comprises:

[0020] The aerial picture is preprocessed to obtain a high-resolution aerial picture.

[0021] Further, the preprocessing of the aerial picture specifically comprises:

[0022] Features of the aerial picture are extracted to obtain a first feature map;

[0023] The first feature map is scaled to obtain a second feature map;

[0024] convolve the second feature map by using a depth separable convolution algorithm to obtain a third feature map;

[0025] input the aerial picture into a bidirectional pyramid structure to obtain a fourth feature map;

[0026] add the third feature map and the fourth feature map to obtain the high-resolution aerial picture.

[0027] Further, the extracting the features of the aerial picture specifically comprises:

[0028] convolve the aerial picture by using a plurality of hollow convolution algorithms with different hollow rates to obtain a plurality of hollow convolution results;

[0029] fuse all the hollow convolution results to obtain the first feature map.

[0030] Further, the scaling the first feature map specifically comprises:

[0031] scaling the first feature map by using an up-sampling process;

[0032] performing 1*1 convolution on the scaled first feature map based on a multi-layer perception mechanism composed of LN and MLP to obtain the second feature map.

[0033] Further, the convolving the second feature map by using a depth separable convolution algorithm specifically comprises:

[0034] convolve the second feature map by using a depth separable convolution algorithm with 3*3 and 7*7 to obtain two first depth convolution results;

[0035] convolve the second feature map by using a depth separable convolution algorithm with 5*5 to obtain a second depth convolution result;

[0036] calculate key vectors, query vectors and value vectors of the second depth convolution result under different hollow rates based on a multi-scale global attention mechanism;

[0037] calculate weights of the two first depth convolution results and the second depth convolution result respectively according to the key vectors, the query vectors and the value vectors;

[0038] perform 1*1 convolution on the three weights to obtain three shared weights respectively;

[0039] input the shared weights into three depth separable convolution algorithms respectively;

[0040] The first deep convolution result and the second deep convolution result are fused based on a multi-layer perception mechanism composed of a multi-scale global attention mechanism, LN and MLP, and are corrected through 1*1 convolution to obtain the third feature map.

[0041] In a second aspect, the application further provides a regional vegetation coverage monitoring system, comprising:

[0042] A picture acquisition module is configured to acquire aerial pictures in a monitoring area, wherein the aerial pictures are in BGR format.

[0043] A first statistical module is configured to convert the aerial pictures into a first gray-scale image and count the number of non-zero pixels in the first gray-scale image.

[0044] A mask creation module is configured to perform threshold processing on the aerial pictures to create a mask containing only a part of a suitable color range.

[0045] A region selection module is configured to select a part matching the color range in the aerial pictures through mask operation on the aerial pictures to confirm a selected region.

[0046] A second statistical module is configured to convert the selected region into a second gray-scale image and count the number of non-zero pixels in the second gray-scale image.

[0047] A calculation module is configured to calculate the vegetation coverage according to the number of non-zero pixels in the first gray-scale image and the number of non-zero pixels in the second gray-scale image.

[0048] The application has the following beneficial effects:

[0049] The regional vegetation coverage monitoring method and system provided by the application can improve the vegetation coverage monitoring efficiency and accuracy by identifying and processing aerial pictures to calculate the vegetation coverage in a region, and can avoid the problems of lack of efficient image processing algorithms and lack of general image processing tools. The method can effectively process aerial image data, improve the vegetation coverage monitoring efficiency and accuracy, and effectively save manpower. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the specific embodiments of the application or the technical solutions in the prior art, the drawings needed in the specific embodiments or the prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0051] Figure 1 A flowchart of a regional vegetation coverage monitoring method provided by the present embodiment is shown in the figure.

[0052] Figure 2 This is a schematic diagram of the structure of a regional vegetation coverage monitoring system provided in this embodiment. Detailed Implementation

[0053] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore merely examples and should not be used to limit the scope of protection of the present invention. It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art to which this invention pertains.

[0054] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0055] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0056] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrases "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0057] The monitoring area mentioned in this embodiment can be an administrative region such as a county or township, or a geographical area. The meaning of the area is not specifically limited in this embodiment.

[0058] like Figure 1 As shown, this application provides a flowchart of a method for monitoring regional vegetation cover. This method is used to identify the vegetation cover within a monitoring area by analyzing aerial images. The application first processes the aerial images of the area, using OpenCV to calculate the ratio of the target pixel to all non-black pixels in the image, and then uses this ratio as the vegetation cover of the image. The method may include the following steps:

[0059] Step S110: Obtain an aerial picture in the monitoring area, the aerial picture being in BGR format.

[0060] In this embodiment, the aerial picture in the monitoring area can be obtained by aerial photography of an airplane or a drone, and the aerial picture is in BGR format.

[0061] It can be understood that, in this embodiment, the resolution of the aerial picture in the monitoring area obtained by aerial photography of an airplane or a drone can be low, and therefore, in order to improve the accuracy of monitoring and identification of vegetation coverage, the low-resolution aerial picture needs to be preprocessed to obtain a high-resolution aerial picture, so that a higher-resolution aerial picture can be obtained without changing the original hardware devices of the system.

[0062] In this embodiment, the preprocessing of the low-resolution aerial picture specifically includes:

[0063] extracting features of the low-resolution aerial picture to obtain a first feature map;

[0064] scaling the first feature map to obtain a second feature map;

[0065] performing convolution on the second feature map by using a deep separable convolution algorithm to obtain a third feature map;

[0066] inputting the low-resolution aerial picture into a bidirectional pyramid structure to obtain a fourth feature map;

[0067] adding the third feature map and the fourth feature map to obtain a high-resolution aerial picture.

[0068] In this embodiment, when the method converts the low-resolution aerial picture, the features of the low-resolution aerial picture are extracted to obtain a first feature map; the first feature map is scaled to obtain a second feature map; the second feature map is convolved by using a deep separable convolution algorithm to obtain a third feature map; the low-resolution aerial picture is input into a bidirectional pyramid structure to obtain a fourth feature map; and the third feature map and the fourth feature map are added to obtain a high-resolution aerial picture.

[0069] Further, in some embodiments, the extraction of the features of the low-resolution aerial picture specifically includes:

[0070] performing convolution on the low-resolution aerial picture by using a plurality of hole convolution algorithms with different hole rates to obtain a plurality of hole convolution results;

[0071] fusing all the hole convolution results to obtain the first feature map.

[0072] In the embodiment, when extracting the features of the low-resolution aerial photograph, the method adopts three branches to convolve the low-resolution aerial photograph, which are 3*3 hollow convolution algorithm (hollow rate = 1) + BN + ReLU, 3*3 hollow convolution algorithm (hollow rate = 2) + BN + ReLU, and 3*3 hollow convolution algorithm (hollow rate = 3) + BN + ReLU, and then fuses the hollow convolution results of the three hollow convolution algorithms through an efficient channel attention (ECA) module to obtain a first feature map. In this way, when extracting the features of the low-resolution aerial photograph, the method captures the features of the low-resolution aerial photograph at different scales through different hollow rates, improves the processing capacity of the method for multi-scale objects, and uses hollow convolution algorithms with different hollow rates to effectively expand the receptive field and capture more extensive context information.

[0073] Further, in some embodiments, the scale of the first feature map is specifically scaled by:

[0074] scaling the scale of the first feature map by upsampling processing;

[0075] performing 1*1 convolution on the scaled first feature map based on a multi-layer perception mechanism composed of LN and MLP to obtain a second feature map.

[0076] In the embodiment, the method first scales the scale of the first feature map by upsampling processing, and then performs 1*1 convolution on the scaled first feature map based on a multi-layer perception mechanism composed of LN and MLP to obtain a second feature map, and uses the second feature map to construct a high-resolution aerial photograph.

[0077] Further, in some embodiments, the convolution of the second feature map by the depth separable convolution algorithm specifically includes:

[0078] convolving the second feature map by 3*3 and 7*7 depth separable convolution algorithms to obtain two first depth convolution results;

[0079] convolving the second feature map by a 5*5 depth separable convolution algorithm to obtain a second depth convolution result;

[0080] calculating the key vector, query vector and value vector of the second depth convolution result under different hollow rates based on the multi-scale global attention mechanism;

[0081] calculating the weights of the two first depth convolution results and the second depth convolution result according to the key vector, the query vector and the value vector, respectively;

[0082] performing 1*1 convolution on the three weights to obtain three shared weights, respectively;

[0083] The shared weights are respectively brought into the three depth separable convolution algorithms;

[0084] The multi-layer perception mechanism based on the multi-scale global attention mechanism, LN and MLP is used to fuse the first depth convolution result and the second depth convolution result, and the 1*1 convolution is used for correction to obtain the third feature map.

[0085] In the embodiment, the method is used for reconstructing high-resolution aerial pictures, and three branches are used for operation, which are 3*3 depth separable convolution algorithm, 5*5 depth separable convolution algorithm and 7*7 depth separable convolution algorithm. The method further increases the multi-scale global attention mechanism (MSDA) after the 5*5 depth separable convolution algorithm, and different vector (key vector, query vector and value vector) calculations are realized through different hole rates, and then the three branches are respectively fused into the 3*3 depth separable convolution algorithm and the 7*7 depth separable convolution algorithm to calculate the weights of the three branches. Then, the method uses the 1*1 convolution to share the weights of the three branches to obtain three shared weights, which are respectively input into the 3*3 depth separable convolution algorithm, the 5*5 depth separable convolution algorithm and the 7*7 depth separable convolution algorithm. Finally, a multi-layer perception mechanism of multi-scale global attention mechanism (MSDA) + LN layer + MLP is embedded, and the channel number is corrected through 1*1 convolution to obtain the third feature map.

[0086] Further, in some embodiments, the bidirectional pyramid structure comprises, from top to bottom, in order:

[0087] 3*3 hole convolution algorithm with a hole rate of 1, 3*3 hole convolution algorithm with a hole rate of 2, 3*3 hole convolution algorithm with a hole rate of 3, up-sampling processing, 3*3 hole convolution algorithm with a hole rate of 3, 3*3 hole convolution algorithm with a hole rate of 2, 3*3 hole convolution algorithm with a hole rate of 1.

[0088] In the embodiment, the method establishes a long-distance residual module between the low-resolution aerial picture and the high-resolution aerial picture, which is used for learning long-distance dependency. The long-distance residual module adopts a bidirectional pyramid structure. The bidirectional pyramid structure comprises, from top to bottom, in order: 3*3 hole convolution algorithm with a hole rate of 1, 3*3 hole convolution algorithm with a hole rate of 2, 3*3 hole convolution algorithm with a hole rate of 3, up-sampling processing, 3*3 hole convolution algorithm with a hole rate of 3, 3*3 hole convolution algorithm with a hole rate of 2, 3*3 hole convolution algorithm with a hole rate of 1.

[0089] Step S120: converting the aerial picture into a first gray image, and counting the number of non-zero pixels in the first gray image.

[0090] The calculation object of the vegetation coverage is land, and the purpose is to monitor the water and soil resources of the land. The vegetation coverage can prevent water and soil loss and avoid reducing the quality of the ecological environment, thereby maintaining the ecological balance of the land. In this embodiment, the influence of water area on land is not considered, and the water area is removed by converting the aerial picture into a first gray image and counting the number of non-zero pixels in the first gray image.

[0091] Step S130: performing threshold processing on the aerial picture to create a mask containing only a part of a suitable color range.

[0092] The land with good soil and water conservation usually appears green in the aerial picture. This is because such land is usually covered with a large amount of vegetation, which can reflect green light and appear green in the aerial picture. In order to better separate the land area in the image, threshold processing is performed on the aerial picture.

[0093] Specifically, the threshold processing on the aerial picture includes: performing threshold processing on the aerial picture by using an HSV color model. The HSV color model is a way of describing colors, in which a color is composed of three components: hue (H), saturation (S), and value (V). The basic color range (fuzzy range) of HSV is shown in Table 1. Each color has a basic hue, saturation, and value range, and these ranges are fuzzy because the specific definition of a color will vary depending on the application. For a certain color, any color falling within its basic hue range can be considered to be that color.

[0094] Table 1: Basic color range (fuzzy range) of HSV

[0095]

[0096] In this embodiment, when the threshold processing is performed on the aerial picture, green and cyan are included in the threshold range. Since green and cyan are continuous on the H channel, and the lower and upper limits of the S and V channels are the same, the HSV color model is set as: H min is 35, H max is 99, S min is 43, S max is 255, V min is 43, V max is 255; wherein H min represents the lower limit of hue, H max represents the upper limit of hue, S min represents the lower limit of saturation, S maxSaturation upper limit, V min Lightness lower limit, V max Lightness upper limit.

[0097] Step S140: Selecting a part matching the color range in the aerial picture by masking operation, and confirming a selected region.

[0098] Selecting a part of the aerial picture region in green and cyan color by masking operation, and confirming a selected region.

[0099] Step S150: Converting the selected region into a second gray-scale image, and counting the number of non-zero pixels in the second gray-scale image.

[0100] Step S160: Calculating the vegetation coverage according to the number of non-zero pixels in the first gray-scale image and the number of non-zero pixels in the second gray-scale image.

[0101] Specifically, the ratio of the number of non-zero pixels in the second gray-scale image to the number of non-zero pixels in the first gray-scale image is taken as the vegetation coverage.

[0102] In summary, the area vegetation coverage monitoring method provided by the embodiment can improve the vegetation coverage monitoring efficiency and accuracy by identifying and processing the aerial picture to calculate the vegetation coverage in the area, and can avoid the problems of lack of efficient image processing algorithm, lack of general image processing tool, and difficulty in realizing automatic monitoring. The method can effectively process aerial image data, improve the vegetation coverage monitoring efficiency and accuracy, and effectively save manpower.

[0103] In a second aspect, the embodiment also provides an area vegetation coverage monitoring system. As shown in the Figure 2 The system comprises:

[0104] A picture acquisition module 201 is configured to acquire an aerial picture in a monitoring area, and the aerial picture is in BGR format.

[0105] A first counting module 202 is configured to convert the aerial picture into a first gray-scale image, and count the number of non-zero pixels in the first gray-scale image.

[0106] A mask creation module 203 is configured to perform threshold processing on the aerial picture, and create a mask containing only a part of suitable color range.

[0107] A region selection module 204 is configured to select a part matching the color range in the aerial picture by masking operation, and confirm a selected region.

[0108] The second statistical module 205 is configured to convert the selected region into a second gray scale image and count the number of non-zero pixels in the second gray scale image.

[0109] The calculation module 206 is configured to calculate the vegetation coverage according to the number of non-zero pixels in the first gray scale image and the number of non-zero pixels in the second gray scale image.

[0110] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the description of the present application.

Claims

1. A method of monitoring regional vegetation cover, characterized by, The method comprises the following steps: acquiring a drone picture in a monitoring area, the drone picture being in BGR format; preprocessing the drone picture to obtain a high-resolution drone picture; converting the drone picture into a first gray-scale image and counting the number of non-zero pixels in the first gray-scale image; performing threshold processing on the drone picture to create a mask containing only a suitable color range part; selecting a part matching the color range in the drone picture through mask operation to confirm a selected area; converting the selected area into a second gray-scale image and counting the number of non-zero pixels in the second gray-scale image; calculating the vegetation coverage according to the number of non-zero pixels in the first gray-scale image and the number of non-zero pixels in the second gray-scale image; wherein the preprocessing of the drone picture specifically comprises: extracting features of the drone picture to obtain a first feature map; scaling the first feature map to obtain a second feature map; performing convolution on the second feature map by using a deep separable convolution algorithm to obtain a third feature map; inputting the drone picture into a bidirectional pyramid structure to obtain a fourth feature map; adding the third feature map and the fourth feature map to obtain the high-resolution drone picture; wherein the feature extraction of the drone picture specifically comprises: performing convolution on the drone picture by using a plurality of hollow convolution algorithms with different hollow rates to obtain a plurality of hollow convolution results; fusing all the hollow convolution results to obtain the first feature map.

2. A method of monitoring regional vegetation cover according to claim 1, wherein, The ratio of the number of non-zero pixels in the second gray-scale image to the number of non-zero pixels in the first gray-scale image is taken as the vegetation coverage.

3. A method of monitoring regional vegetation cover according to claim 1, wherein, The threshold processing on the drone picture comprises: performing threshold processing on the drone picture by using an HSV color model.

4. A method of monitoring regional vegetation cover according to claim 3, wherein, The parameters of the HSV color model comprise: H min is 35, H max is 99, S min is 43, S max is 255, V min is 43, V max is 255; wherein, H min H represents a hue lower limit max S represents a hue upper limit min S represents a saturation lower limit max V represents a saturation upper limit min V represents a lightness lower limit max V represents a lightness upper limit.

5. The method of claim 1, wherein, The scaling of the first feature map specifically comprises: scaling the first feature map by using up-sampling processing; performing 1*1 convolution on the scaled first feature map based on a multi-layer perception mechanism composed of LN and MLP to obtain the second feature map.

6. A method of monitoring regional vegetation cover according to claim 1, wherein, The convolution on the second feature map by using the deep separable convolution algorithm specifically comprises: performing convolution on the second feature map by using 3*3 and 7*7 deep separable convolution algorithms to obtain two first deep convolution results; performing convolution on the second feature map by using a 5*5 deep separable convolution algorithm to obtain a second deep convolution result; calculating key vectors, query vectors and value vectors of the second deep convolution result under different hollow rates based on a multi-scale global attention mechanism; calculating the weights of the two first deep convolution results and the second deep convolution result respectively according to the key vectors, the query vectors and the value vectors; performing 1*1 convolution on the three weights to obtain three shared weights respectively; inputting the shared weights into three deep separable convolution algorithms respectively; fusing the first deep convolution result and the second deep convolution result based on a multi-scale global attention mechanism and a multi-layer perception mechanism composed of LN and MLP, and correcting through 1*1 convolution to obtain the third feature map.

7. A regional vegetation cover monitoring system characterized by, The method comprises the following steps: The picture acquisition module is configured to acquire a picture of aerial photography in a monitoring area, and the picture of aerial photography is in BGR format. The first statistical module is configured to convert the picture of aerial photography into a first gray image and count a number of non-zero pixels in the first gray image. The mask creation module is configured to perform threshold processing on the picture of aerial photography to create a mask containing only a part of a suitable color range. The region selection module is configured to select a part matching the color range in the picture of aerial photography through mask operation, and confirm a selected region. The second statistical module is configured to convert the selected region into a second gray image and count a number of non-zero pixels in the second gray image. The calculation module is configured to calculate a vegetation coverage degree according to the number of non-zero pixels in the first gray image and the number of non-zero pixels in the second gray image. Before the picture of aerial photography is converted into the first gray image, the picture of aerial photography is preprocessed to obtain a high-resolution picture of aerial photography. The preprocessing of the picture of aerial photography specifically includes: extracting features of the picture of aerial photography to obtain a first feature map; scaling a scale of the first feature map to obtain a second feature map; performing convolution on the second feature map by using a deep separable convolution algorithm to obtain a third feature map; inputting the picture of aerial photography into a bidirectional pyramid structure to obtain a fourth feature map; adding the third feature map and the fourth feature map to obtain the high-resolution picture of aerial photography. The extraction of the features of the picture of aerial photography specifically includes: performing convolution on the picture of aerial photography by using a plurality of hollow convolution algorithms with different hollow rates to obtain a plurality of hollow convolution results; fusing all the hollow convolution results to obtain the first feature map.

Images