UAV-RGB-based urban wetland vegetation coverage rate automatic extraction method

Abstract

The invention discloses an urban wetland vegetation coverage rate automatic extraction method based on UAV-RGB, and belongs to the field of wetland ecological monitoring. The method comprises the following steps: acquiring an original image of a target area by using an unmanned aerial vehicle remote sensing technology, reconstructing the original image into a high-resolution orthoimage, establishing a secondary land utilization type, subdividing a vegetation type, delimiting a test area for visual interpretation, and performing field verification and precision correction on a classification result. The result is used as a sample to analyze spectrum and texture features of different land utilization and vegetation type images in a test area, and a multi-scale segmentation method is adopted to carry out image segmentation and parameter correction to obtain an optimal segmentation scale parameter combination. And automatically segmenting the image of the target region according to the optimal parameter combination, vectorizing the segmentation result after the precision reaches the standard, counting the vegetation coverage rate by adopting an area proportion method, and carrying out land utilization and vegetation type spatial mapping in combination with a 3S technology. According to the method, unmanned aerial vehicle near-earth remote sensing and computer technologies are combined, and the method is a more efficient and accurate urban wetland vegetation coverage rate automatic extraction method.

Description

Technical Field

[0001] The present invention belongs to the field of wetland ecological monitoring, and particularly relates to a method for automatically extracting the vegetation coverage rate of urban wetlands based on UAV-RGB. Background Art

[0002] Urban wetlands are an indispensable part of the urban ecosystem and have various ecological service functions. It can regulate the local climate and affect air humidity and temperature through water evaporation and vegetation transpiration; it can purify water quality, and wetland vegetation and soil can filter, adsorb and decompose pollutants; at the same time, it provides habitats, foraging and breeding places for many wild animals and plants, and plays an important role in maintaining the urban ecological environment, protecting biodiversity, and providing human cultural and recreational places. In the context of the rapid development of urbanization, urban wetlands are facing problems such as area reduction, pollution aggravation, and vegetation damage. Therefore, scientific monitoring and evaluation of urban wetlands are the key to protecting and rationally utilizing wetland resources. As an important indicator for measuring the ecological status of wetlands, the accurate acquisition of vegetation coverage rate is in urgent need.

[0003] Traditional ground measurement methods, such as the quadrat method and the transect method, although they can obtain relatively accurate local vegetation information, have obvious deficiencies. For large-area urban wetlands, a large amount of manpower, material resources and time are required. Moreover, in some wetland areas with complex terrain, inconvenient transportation or potential safety hazards, it is extremely difficult to conduct ground surveys, or even impossible to carry out. In addition, ground surveys can only obtain discrete point or line data, and it is difficult to comprehensively reflect the vegetation coverage of the entire wetland. Satellite remote sensing technology has certain advantages in large-scale vegetation monitoring, but it has limitations for urban wetlands with medium and small scales and complex landscapes. The spatial resolution of satellite images often cannot meet the requirements of high-precision monitoring of urban wetlands, resulting in confusion when distinguishing vegetation in wetlands from other ground objects (such as buildings, roads, small water bodies, etc. in the city), especially at the urban-wetland boundary, where the error is large. Moreover, satellite remote sensing data is affected by factors such as weather and clouds, and it is difficult to ensure the timeliness and quality of data acquisition.

[0004] Unmanned aerial vehicles (UAVs) have the advantages of low cost, flexible operation, and the ability to fly at low altitudes. They can flexibly plan flight routes and altitudes according to the shape, size, and terrain characteristics of the research area, enabling near-ground observation of urban wetlands and obtaining high-resolution image data. This customized data collection method can effectively avoid data missing problems caused by complex terrain or inaccessibility. As a commonly used image type for UAVs, RGB images contain rich color information. By analyzing RGB images, these color differences can be utilized to develop specialized algorithms for vegetation identification. Moreover, the processing of RGB images is relatively simple. Compared with other multi-spectral or hyper-spectral images, they have a smaller data volume and faster processing speed, which is more conducive to the development and application of automatic extraction algorithms for vegetation coverage, thus providing a new technical approach for the rapid and accurate measurement of urban wetland vegetation coverage. Summary of the Invention

[0005] The purpose of the present invention is to overcome the limitations of traditional measurement methods, reduce the cost input of standard plot surveys, ensure high-precision vegetation coverage data, provide a reliable basis for wetland ecological research, protection, and management, and provide a design method for extracting wetland vegetation coverage based on UAV images. The present invention mainly focuses on three key points of the design method for extracting wetland vegetation coverage based on UAV images - the accuracy of image land use type and vegetation identification, the degree of automation in the extraction process, and the efficiency of coverage calculation. It quickly obtains vegetation-related feature information in wetland UAV images, comprehensively and quantitatively analyzes the deviation between different processing links and accurate vegetation coverage results, and optimizes the extraction process with high accuracy, providing a new design method for extracting wetland vegetation coverage based on UAV images.

[0006] The specific technical solutions adopted by the present invention are as follows:

[0007] The present invention provides an automatic extraction method for urban wetland vegetation coverage based on UAV-RGB, specifically as follows:

[0008] S1. Apply online satellite remote sensing image data to preliminarily delimit the target range, predict aerial photography operation conditions, and use a UAV integrated with GNSS-RTK to collect high-resolution original image data of the target urban wetland at low altitude;

[0009] S2. Preprocess all the original image data collected in S1 and splice them into a high-resolution orthophoto map of the target urban wetland;

[0010] S3. Based on the orthophoto map obtained in S2, establish the first-level land use types according to the target urban wetlands, automatically extract the vegetation and further subdivide its vegetation types, and then delimit the test area within the target range by combining GIS technology and computer vision technology; conduct manual visual interpretation on the test area, finely vectorize the first-level land use types and vegetation types, and do not conduct any manual visual interpretation outside the test area.

[0011] S4. Based on the refined land use and vegetation classification results within the test area in S3, verify the refined land types and their ranges within the test area on the ground, conduct accuracy verification and correct the visual interpretation vector results within the test area in S3 to obtain highly accurate vectorized results of land use and vegetation types for standby accuracy evaluation at multiple segmentation scales.

[0012] S5. Based on the classification results corrected in S4, use the objects within the test area as the test sample set, and then analyze the spectral characteristics and texture characteristics of the sample set, and extract the same and different values between different spectral characteristics and texture characteristics.

[0013] S6. Based on the spectral and texture feature analysis results in S5, estimate the optimal segmentation scale, apply the multi-scale segmentation algorithm to automatically segment the image ranges of different land uses and refined vegetation types within the test area, overlay and verify the test results with the refined land use types and their ranges obtained in S4, and use the control variable method to optimize the parameter combinations at each scale until the segmentation results within the test area reach the highest accuracy, and record the optimal parameter combination.

[0014] S7. Based on the optimal parameter combination of multi-scale segmentation obtained in S6, further use the multi-scale segmentation method to automatically segment the images in the study area, combine with the high-resolution remote sensing images obtained in S2, and design an equal-step systematic sampling method again to visually verify the accuracy of the segmentation results. When the result accuracy reaches the preset requirements, vectorize the image segmentation results using spatial geography technology.

[0015] S8. Based on the image segmentation results of different land use types and refined vegetation types obtained in S7, use the area ratio method to statistically calculate the vegetation coverage of urban wetlands, and use geospatial technology to draw the spatial distribution map of vegetation coverage types.

[0016] Preferably, in S1, the image acquisition is achieved by means of a DJI Phantom 4 RTK equipped with an RGB camera.

[0017] Preferably, S2 is carried out through the advanced version of DJI Terra software, and the preprocessing means include automatic extraction of image data feature points, aerial triangulation processing, and image correction.

[0018] Preferably, in S3, the land use and vegetation type classification system is as follows:

[0019] (1) First - level classification: vegetation, water body, construction land, other land

[0020] (2) Second - level classification: vegetation includes arbor forest land, arbor - shrub - grass complex area, reed land, other aquatic plant areas; water body includes ponds, shoals, river channels; construction land includes houses, roads, bridges, parking lots; other land includes cultivated land.

[0021] Preferably, in step S3, using ArcGIS software, according to the selected criteria, use geospatial analysis tools to accurately delimit the boundary of the test area within the target range and vectorize it; when performing manual visual interpretation, carefully observe the characteristics of each pixel or ground object patch on the image, and judge its land use type and vegetation type according to the interpretation marks. Preferably, in step S4, use the refined classification results of land use and vegetation within the test area in S3 as the verification content; in the field, use measuring tools to accurately measure the boundary range of the land use type, and compare it with the range in the visual interpretation vector result; according to the results of accuracy verification, make targeted corrections to the visual interpretation vector result, and finally obtain a highly accurate vectorized result of land use and vegetation types that highly matches the field situation.

[0022] Preferably, in step S5, the analysis methods of spectral features and texture features are as follows:

[0023] (1) The analysis method of spectral features is the mean - variance method, and the calculation formula is as follows:

[0024]

[0025] In the formula: C Li is the brightness value of the i - th pixel within the object in the L - th band, C L is the brightness mean of a single image object in the L - th band, m is the total number of objects in the image, n is the number of pixels in the object, S 2 is the variance;

[0026] (2) The analysis method of texture features is the shape heterogeneity of the object. The shape heterogeneity index of the object is composed of two sub - heterogeneity indexes, namely the smoothness index and the compactness index, and is obtained from the shape heterogeneity increment before and after object merging; the shape heterogeneity increment before and after object merging is the weighted average of the smoothness index increment and the compactness index increment, and its expression form is as follows:

[0027] h shape =w smoothness ×h smoothness +w compactness ×h compactness

[0028] In the formula: wsmoothness With w compactness Represents the weight allocation between the two, and the sum of the two is 1; h shape Represents the increment of shape heterogeneity; h smoothness Represents the increment of smoothness index; h compactness Represents the increment of compactness index.

[0029] Preferably, in the step S6, the parameter adjustment steps for automatically segmenting the image range by the multi-scale segmentation algorithm are as follows:

[0030] (1) Set the initial segmentation scale parameter according to the previously estimated scale range;

[0031] (2) Overlay and verify the test results with the refined land use types and their ranges obtained in S4 to determine the parameters to be adjusted;

[0032] (3) First, fix other parameters, change the segmentation scale parameter, gradually increase it from a smaller scale, and observe the change in segmentation accuracy; when the segmentation accuracy starts to decline, record the scale range at this time, and then further fine-tune other parameters within this range; if the accuracy of a certain land use type improves after increasing the spectral threshold while the accuracy of other types changes little, continue to adjust in this direction;

[0033] (4) Repeat steps (1) to (3), continuously adjust the parameter combination until the segmentation result in the test area reaches the highest accuracy, record the parameter values at this time and use them as the optimal parameter combination for the test area for subsequent image segmentation of the entire study area.

[0034] Preferably, in the step S7, using the optimal parameter combination obtained in S6, further automatically segment the images in the study area by the multi-scale segmentation method and perform accuracy verification. The calculation formula for accuracy verification is as follows:

[0035] (1) The production accuracy PA reflects the consistency between the correct classification result and the reference category, and the calculation formula is as follows:

[0036]

[0037] In the formula: a ii Is the number of correct verification sample points for each category, a ki Is the total number of verification sample points for this category, and N is the total number of sample categories;;

[0038] (2) The user accuracy UA reflects the consistency between the correct classification result and the actual classification result, and the calculation formula is as follows:

[0039]

[0040] In the formula: a iiis the correct number of various verification sample points, a ik is the actual classification number of this category;

[0041] (3) The range accuracy SA reflects the consistency between the correct range result and the actual range result of the plot, and the calculation formula is as follows:

[0042]

[0043] In the formula: b ij is the correct area range of various verification sample plot, b ik is the area range of each actual plot of this category.

[0044] Preferably, in S8, when the land and vegetation types in S7 reach the preset accuracy, extract the vector map of the vegetation area and calculate and count the area of the vegetation area, then the calculation formula of the vegetation coverage rate V is as follows:

[0045]

[0046] In the formula: A v is the vegetation coverage area, A t is the total wetland area.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1) High efficiency and flexibility. The unmanned aerial vehicle (UAV) can quickly obtain large-area wetland image data. Compared with the traditional ground survey method, it greatly improves the data collection efficiency and can complete the monitoring of the vegetation coverage of large wetlands in a relatively short time. It is not restricted by terrain and environment and can easily fly over difficult-to-reach areas in the wetland, such as swamps and islands in the middle of rivers, to obtain more comprehensive vegetation information;

[0049] 2) Cost saving. Compared with traditional satellite remote sensing or manned aircraft remote sensing, the use cost of the UAV is relatively low. The purchase, maintenance and operation costs of the UAV are all relatively low, and a large number of professional personnel are not required for operation and data processing, reducing the cost of the whole project;

[0050] 3) Quick response. It can complete data collection and processing in a short time and can timely reflect the changes in wetland vegetation. This is very important for the monitoring and management of wetland ecosystems. For example, it can timely detect the impacts of natural disasters and human damage on vegetation so as to take corresponding protection measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a schematic flow chart of the method of the present invention;

[0052] Figure 2It is an example diagram of the vectorization result of land use and vegetation types in the test area in the embodiment;

[0053] Figure 3 It is an example diagram of the image segmentation results under different shape parameters in the test area in the embodiment; among them, (a) is the original image, (b) is the segmented image (shape parameter is 0.1), (c) is the segmented image (shape parameter is 0.3), and (d) is the segmented image (shape parameter is 0.5);

[0054] Figure 4 It is an example diagram of the image segmentation results under different compactness parameters in the test area in the embodiment; among them, (a) is the original image, (b) is the segmented image (compactness parameter is 0.1), (c) is the segmented image (compactness parameter is 0.3), and (d) is the segmented image (compactness parameter is 0.5);

[0055] Figure 5 It is an example diagram of the peak values of spectral and texture features in the arbor vegetation area; among them, (a) is the image of the target patch, and (b) is the peak value diagram of the spectral and texture features of the target patch;

[0056] Figure 6 It is a multi-scale automatic segmentation result diagram of land use and vegetation types in the test area in the embodiment based on UAV-RGB;

[0057] Figure 7 It is a multi-scale automatic segmentation result diagram of the test area types in the embodiment based on UAV-RGB. Specific implementation manner

[0058] The present invention will be further described and explained below in conjunction with the accompanying drawings and specific implementation manners. The technical features of each implementation manner in the present invention can be combined correspondingly without conflict.

[0059] As Figure 1 shown, a method for automatically extracting the vegetation coverage rate of urban wetlands based on UAV-RGB provided by the present invention is as follows:

[0060] S1. Apply online satellite remote sensing image data to preliminarily delimit the target range, predict the aerial photography operation conditions, and use a drone integrated with GNSS-RTK to collect high-resolution original image data of the target urban wetland at low altitude.

[0061] In actual use, the RGB images in the present invention are obtained by using a low-cost consumer drone, DJI Phantom 4 RTK, as a flight platform and carrying a visible light camera.

[0062] S2. Automatically extract image feature points, perform aerotriangulation processing, image correction and other preprocessing on all the original image data collected in S1, and splice them into a high-resolution orthophoto map of the target urban wetland.

[0063] In actual use, a series of important operations such as automatically extracting image data feature points, performing aerotriangulation processing, and image correction are carried out through the software of DJI Terra (Advanced Edition DJI Terra), and then spliced into a high-resolution orthophoto map of the target urban wetland.

[0064] S3. Based on the orthophoto map obtained in S2, establish the first-level land use types (water bodies, roads, vegetation, construction, etc.) according to the target urban wetland, automatically extract the vegetation and further subdivide its types into arbors, arbor-shrub-grass complex layers, reeds, etc.; then combine GIS technology and computer vision technology to delimit a representative test area within the target range for manual visual interpretation, and finely vectorize the first-level land use types and vegetation types. No manual visual interpretation is carried out outside the test area.

[0065] In actual use, it is planned to take Xixi Wetland in Hangzhou City, Zhejiang Province as the classification object. According to its land use type characteristics, the specific land use and vegetation type classification system is as follows:

[0066] (1) First-level classification: vegetation, water body, construction land, other land;

[0067] (2) Second-level classification: vegetation is subdivided into arbor forest land, arbor-shrub-grass complex layer area, reed land, other aquatic plant areas; water bodies can be subdivided into ponds, shoals, river channels; construction land includes houses, roads, bridges, parking lots; other land includes cultivated land, etc.

[0068] Furthermore, this step is carried out using ArcGIS software. According to the selected criteria, use geospatial analysis tools (polygon drawing, buffer analysis, etc.) to accurately delimit the boundary of the test area within the target range and vectorize it. When performing manual visual interpretation, carefully observe the characteristics of each pixel or ground object patch on the image, and judge its belonging land use type and vegetation type according to the interpretation marks.

[0069] S4. Based on the refined classification results of land use and vegetation in the test area in S3, verify the refined land types and their ranges in the test area on the spot, conduct accuracy verification and correct the visual interpretation vector results in the test area in S3 to obtain a highly accurate vectorized result of land use and vegetation types, which is reserved for accuracy evaluation at multiple segmentation scales.

[0070] In actual use, the refined classification results of land use and vegetation in the test area in S3 (land and vegetation use types, boundary ranges) are used as the verification content. In the field, measurement tools (such as GPS devices, GNSS-RTK, etc.) are used to accurately measure the boundary ranges of land use types and compare them with the ranges in the visual interpretation vector results to ensure the accuracy of spatial position information. According to the results of accuracy verification, the visual interpretation vector results are corrected specifically, and finally, a highly accurate vectorization result of land use and vegetation types that highly coincides with the actual situation in the field is obtained.

[0071] S5. Based on the classification results corrected in S4, the objects in the test area are used as the test sample set, and then the spectral features and texture features of the sample set are analyzed, that is, the six eigenvalue features of the refined classification objects, namely the mean green value (Mean Green), mean red value (Mean Red), mean blue value (Mean Blue), brightness, shape index, and length / width ratio (Length / Width), are statistically analyzed, and the same and different values between different spectral features and texture features are extracted.

[0072] In actual use, the principles and analysis methods of spectral features and texture features are as follows:

[0073] (1) The analysis method of spectral features is the mean variance method, and its basic principle is: when the number of pure objects in the image layer increases and the spectral variation between adjacent objects increases, the mean variance of the object increases; on the contrary, when the number of mixed objects increases, the spectral variation between adjacent objects decreases, and the mean variance of the object becomes smaller. The segmentation scale corresponding to the maximum mean variance is the optimal one, and the calculation formula is as follows:

[0074]

[0075] In the formula: C Li is the brightness value of the i-th pixel in the object in the L-th band, C L is the mean brightness of a single image object in the L-th band, m is the total number of objects in the image, n is the number of pixels in the object, and S 2 is the variance;

[0076] (2) The analysis method for shape features is the shape heterogeneity of the object. The shape heterogeneity index of the object is composed of two sub - heterogeneity indexes, namely the smoothness index and the compactness index, and is obtained from the increment of shape heterogeneity before and after object merging. The so - called compactness index refers to the roundness of the object, which is used to measure the degree to which the region approaches a circle and can also be used as an index to measure the regularity of the object's shape; the smoothness index is somewhat similar to roundness, but is used to represent the smoothness of the object's shape. Whether the image is smooth or not is an index to measure whether the object is regular or irregular. The increment of shape heterogeneity before and after object merging is the weighted average of the increment of the smoothness index and the increment of the compactness index, and its expression form is as follows:

[0077] h shape = w smoothness ×h smoothness + w compactness ×h compactness

[0078] In the formula: w smoothness and w compactness represent the weight allocation between the two, and the sum of the two is 1; h shape represents the increment of shape heterogeneity; h smoothness represents the increment of the smoothness index; h compactness represents the increment of the compactness index.

[0079] S6. Based on the spectral and texture feature analysis results in S5, combined with 3S spatial geography, image processing and other technologies, estimate the optimal segmentation scale, apply the multi - scale segmentation algorithm to automatically segment the image ranges of different land uses and refined vegetation types in the test area, overlay and verify the test results with the refined land use types and their ranges obtained in S4, and use the method of controlling variables to optimize the combination of each scale parameter until the segmentation result in the test area reaches the highest accuracy, and record the optimal parameter combination.

[0080] In actual use, the main principles involved in multi - scale segmentation are as follows:

[0081] (1) Distance transformation: It is an operation for binary images (the pixels in the image have only two values, 0 and 1. 0 usually represents the background and 1 represents the target object). Its purpose is to calculate the distance from each foreground pixel (pixel with a value of 1) in the image to the nearest background pixel (pixel with a value of 0). This can be used in multi - scale segmentation to determine the boundaries of objects, calculate the distances between objects, etc., and helps to segment and analyze the objects in the image or geospatial data according to the distance information;

[0082] (2) Gray-level co-occurrence matrix: It is a method for analyzing the texture features of an image. It describes the texture information of the image by statistically counting the frequencies of pixel pairs with specific gray levels at a certain direction and distance. First, select a direction (such as horizontal, vertical, diagonal, etc.) and a distance parameter d. For each pixel (i, j) in the image, consider the pixel (i + Δi, j + Δj) at a distance of d in the selected direction, where (Δi, Δj) is the offset determined according to the selected direction. Then, count the occurrence frequencies of all pixel pairs (I(i, j), I(i + Δi, j + Δj)) in the image to form a gray-level co-occurrence matrix. The size of the matrix depends on the number of gray levels G of the image, usually a square matrix of G×G. For example, if the gray level range of the image is 0 - 255 (i.e., G = 256), the gray-level co-occurrence matrix is a 256×256 matrix. Multiple texture features can be extracted from the gray-level co-occurrence matrix, and these texture features can help distinguish different types of ground objects or different objects in the image;

[0083] (3) Classification object mean: It is the average value calculated for the attribute values (such as gray values, spectral values, etc.) of the pixels within the segmented object (such as the regions in the image or the polygon regions in the geospatial space). The calculation formula is as follows:

[0084]

[0085] In the formula: n is the number of pixels, x i is the gray value of the pixel, is the mean of the classification object;

[0086] (4) Mean variance: It is a statistic that measures the degree of dispersion of a set of data. In image processing and multi-scale segmentation, it is used to describe the distribution of pixel attribute values within the segmented object. The calculation formula is as follows:

[0087]

[0088] In the formula: n is the number of pixels, x i is the attribute value of the pixels within the segmented object, is the mean, S 2 is the variance.

[0089] In this step, the parameter adjustment method for the multi-scale segmentation algorithm to automatically segment the image range is specifically as follows:

[0090] (1) Set the initial segmentation scale parameter according to the previously estimated scale range;

[0091] (2) Overlay and verify the test results with the refined land use types and their ranges obtained in S4 to determine the parameters to be adjusted, such as the segmentation scale, spectral threshold, texture threshold, etc.;

[0092] (3) First, fix other parameters and change the segmentation scale parameter, gradually increasing it from a smaller scale, and observe the change in segmentation accuracy. When the segmentation accuracy starts to decline, record the scale range at this time, and then further fine-tune other parameters within this range. If the accuracy of a certain land use type improves after increasing the spectral threshold while the accuracy of other types changes little, then continue to adjust in this direction;

[0093] (4) Repeat the above steps (1) to (3), continuously adjust the parameter combinations until the segmentation result within the test area reaches the highest accuracy. Record in detail the parameter values such as the segmentation scale, spectral threshold, and texture threshold at this time. These parameter combinations are the optimal parameter combinations for this test area and can be used for subsequent image segmentation of the entire study area.

[0094] S7. Based on the optimal parameter combination of multi-scale segmentation obtained in S6, further use the multi-scale segmentation method to automatically segment the images in the study area. Combining with the high-resolution remote sensing images obtained in S2, design an equal-step systematic sampling method again to visually verify the accuracy of the segmentation result. When the result accuracy meets the preset requirements, vectorize the image segmentation result using spatial geographic technology.

[0095] In actual use, using the optimal parameter combination obtained in S6, further use the multi-scale segmentation method to automatically segment the images in the study area and conduct accuracy verification. The calculation formula for accuracy verification is as follows:

[0096] (1) The producer accuracy (PA) reflects the consistency between the correct classification result and the reference category, and the calculation formula is as follows:

[0097]

[0098] In the formula: PA is the producer accuracy, a ii is the number of correct points for each type of verification sample, a ki is the total number of verification sample points for this type, and N is the total number of sample categories;;

[0099] (2) The user accuracy (UA) reflects the consistency between the correct classification result and the actual classification result, and the calculation formula is as follows:

[0100]

[0101] In the formula: UA is the user accuracy, a ii is the number of correct points for each type of verification sample, a ik is the actual classification number for this type;

[0102] (3) The scope accuracy (SA) reflects the consistency between the correct range result of the plot and the actual range result, and the calculation formula is as follows:

[0103]

[0104] Where: SA is the range accuracy, b ii is the correct area range of various verification sample plots, b ik is the area range of each actual plot of this type.

[0105] S8. Based on the image segmentation results of different land use types and refined vegetation types obtained in S7, the area ratio method is used to statistically calculate the vegetation coverage (V) of urban wetlands, and the geospatial technology is used to draw the spatial distribution map of vegetation coverage types.

[0106] In actual use, when the land and vegetation types in S7 reach the preset accuracy, extract the vector map of the vegetation area and calculate the area of the vegetation area. Then, the calculation formula of the vegetation coverage rate V is as follows:

[0107]

[0108] Where: V is the vegetation coverage rate, A v is the vegetation coverage area, A t is the total wetland area.

[0109] Next, the steps and effects of the method of the present invention will be specifically described through embodiments.

[0110] Embodiment

[0111] This embodiment provides an automatic extraction method for the vegetation coverage rate of urban wetlands based on UAV-RGB. The specific steps are as follows:

[0112] S1. Use an unmanned aerial vehicle to collect image data of the target wetland ecosystem: The research area is located in Xixi Wetland, Hangzhou City, Zhejiang Province, and the research area covers an area of 1150.00 hm 2 . On November 20, 2023, a visible light camera was carried on the DJI Phantom 4 RTK as a flight platform to conduct field operations on the research area to obtain image data. The flight line parameters were set in the DJI flight control software, with a flight altitude of 120 m, a side overlap rate of 75%, and a forward overlap rate of 85%. The area of the surveyed area is approximately 1386.00 hm 2 .

[0113] S2. Use the DJI Terra software (advanced version) to perform preprocessing on the original image data of the target area collected in S1, such as automatic extraction of image feature points, aerial triangulation processing, and image correction, and then splice them into a high-resolution orthophoto image of the target urban wetland.

[0114] S3. Based on the orthophoto map obtained in S2, establish the first-level land use types (water bodies, roads, vegetation, construction, etc.) according to the target urban wetlands, automatically extract the vegetation and further subdivide its types into arbors, arbor-shrub-grass multi-layers, reeds, etc., and conduct visual interpretation of the test area using ArcGIS software.

[0115] S4. Based on the refined land use and vegetation classification results in the test area in S3, conduct on-site verification, accuracy verification, and correct the visual interpretation vector results in the test area in S3 (such as Figure 2 ).

[0116] S5. Based on the classification results corrected in S4, use the objects in the test area as the test sample set, and then analyze the spectral characteristics and texture characteristics of the sample set, that is, statistically analyze six eigenvalue of the refined classification objects, including the mean green value (Mean Green), mean red value (Mean Red), mean blue value (Mean Blue), brightness, shape index, and length-width ratio (Length / Width) (such as Figure 5 , Table 1).

[0117] Table 1 Statistical table of sample parameters

[0118]

[0119] S6. Based on the spectral and texture feature analysis results in S5, combine 3S spatial geography, image processing and other technologies to estimate the optimal segmentation scale, apply the multi-scale segmentation algorithm to automatically segment the image ranges of different land uses and refined vegetation types in the test area, overlay and verify the test results with the refined land use types and their ranges obtained in S4, and use the control variable method to optimize the parameter combinations of each scale until the segmentation results in the test area reach the highest accuracy (the optimization process is as shown in Figure 3 , Figure 4 ). In Figure 3 , (a) is the original image, and (b), (c), (d) are the segmented images obtained by adopting different shape parameters under the condition that the scale parameter and the compactness parameter are both set to 100 and 0.3. According to the result comparison in the figure, when the shape parameter is 0.3, the object segmentation is more complete than when the shape parameter is 0.1, and the edge segmentation is more fitting than when the shape parameter is 0.5. Figure 4In (a) is the original image, and (b), (c), and (d) are the segmented images obtained under the condition that both the scale parameter and the shape parameter are set to 100 and 0.3, with different compactness parameters. According to the result comparison in the figure, when the compactness parameter is 0.3, the object segmentation boundary is more clearly divided and the segmentation result is more complete. After screening various combinations, the parameter combination of scale parameter 135, shape parameter 0.3, and compactness parameter 0.3 is the optimal parameter combination.

[0120] S7. Based on the optimal parameter combination of multi-scale segmentation obtained in S6, further use the multi-scale segmentation method to automatically segment the images in the study area. Combining with the high-resolution remote sensing images obtained in S2, design the equal-step systematic sampling method again to visually verify the accuracy of the segmentation result. When the result accuracy meets the preset requirements, vectorize the image segmentation result using spatial geographic technology (such as Figure 6 ).

[0121] S8. Based on the image segmentation results of different land use types and refined vegetation types obtained in S7, use the area proportion method to statistically calculate the urban wetland vegetation coverage (V), and use geospatial technology to draw the spatial distribution map of vegetation coverage types. In the automated vegetation extraction map (such as Figure 7 ), the bright green part is the vegetation coverage part. The areas of various types of vegetation are statistically obtained: the arbor land is 397.72 hm 2 , the shrub and grassland is 140.18 hm 2 , the reed land is 24.31 hm 2 , the aquatic plants are 129.82 hm 2 . The total vegetation coverage area of Xixi Wetland is 692.03 hm 2 mu, and the vegetation coverage rate of Xixi Wetland is 60.18%.

[0122] The present invention combines unmanned aerial vehicle (UAV) near-ground remote sensing and computer technology, and is a more efficient and accurate method for automatically extracting the vegetation coverage rate of urban wetlands.

[0123] The above-described embodiments are only a preferred solution of the present invention, but they are not intended to limit the present invention. Those of ordinary skill in the relevant technical fields can still make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all technical solutions obtained by adopting equivalent replacement or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A method for automatically extracting urban wetland vegetation coverage based on UAV-RGB, characterized in that: The details are as follows: S1. Use online satellite remote sensing image data to preliminarily define the target range, predict the conditions for aerial photography, and use drones with integrated GNSS-RTK to collect high-resolution original image data of the target urban wetland at low altitude; S2, pre-processing all the original image data collected in S1, and stitching them into a high-resolution orthophoto map of the target urban wetland; S3. Based on the orthophoto map obtained in S2, establish the primary land use type according to the target urban wetland, automatically extract vegetation and further subdivide its vegetation type, and then combine GIS technology and computer vision technology to delineate the test area within the target range; perform manual visual interpretation of the test area, finely vectorize the primary land use type and vegetation type, and do not perform any manual visual interpretation outside the test area; S4. Based on the refined classification results of land use and vegetation in the test area in S3, the refined land types and their scopes in the test area are verified on the spot, the accuracy is verified and the visual interpretation vector results in the test area in S3 are corrected to obtain high-accuracy vectorization results of land use and vegetation types, and the accuracy evaluation under multiple segmentation scales is prepared; S5, based on the classification results corrected in S4, taking the objects in the test area as the test sample set, and then analyzing the spectral features and texture features of the sample set, extracting the same and different values ​​between different spectral features and texture features; S6. Based on the spectral and texture feature analysis results in S5, the optimal segmentation scale is estimated, and the multi-scale segmentation algorithm is used to automatically segment the image range of different land use and refined vegetation types in the test area. The test results are superimposed and verified with the refined land use types and their ranges obtained in S4. The control variable method is used to optimize the combination of parameters at each scale until the segmentation result in the test area reaches the highest accuracy, and the optimal parameter combination is recorded; S7. Based on the optimal multi-scale segmentation parameter combination obtained in S6, the multi-scale segmentation method is further used to automatically segment the images in the study area. Combined with the high-resolution remote sensing images obtained in S2, an equal-step systematic sampling method is designed again to visually verify the accuracy of the segmentation results. When the accuracy of the results reaches the preset requirements, the spatial geographic technology is used to vectorize the image segmentation results. S8. Based on the image segmentation results of different land use types and refined vegetation types obtained in S7, the area proportion method is used to calculate the vegetation coverage of urban wetlands, and the spatial distribution map of vegetation coverage types is drawn using geographic spatial technology.

2. The automatic extraction method of urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: In S1, image acquisition is achieved with the help of DJI Phantom 4RTK equipped with an RGB camera.

3. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: The S2 is performed using DJI Terra, an advanced version of the DJI Zhitu software. The preprocessing methods include automatic extraction of image data feature points, aerial triangulation, and image correction.

4. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: In S3, the land use and vegetation type classification system is as follows: (1) First-level classification: vegetation, water bodies, construction land, and other land; (2) Secondary classification: Vegetation includes tree forests, tree-shrub-grass multilayer areas, reed fields, and other aquatic plant areas; water bodies include ponds, shoals, and rivers; construction land includes houses, roads, bridges, and parking lots; other land includes cultivated land.

5. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: In S3, ArcGIS software is used to accurately define the boundaries of the test area within the target range based on the selected standards and geospatial analysis tools, and the boundaries are vectorized; during manual visual interpretation, the characteristics of each pixel or patch of land object on the image are carefully observed, and the land use type and vegetation type to which it belongs are determined based on the interpretation signs.

6. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: In S4, the refined classification results of land use and vegetation in the test area in S3 are used as verification content; in the field, measurement tools are used to accurately measure the boundary range of land use types and compare them with the range in the visual interpretation vector results; according to the results of the accuracy verification, the visual interpretation vector results are corrected in a targeted manner, and finally a high-accuracy land use and vegetation type vectorization result that is highly consistent with the actual situation is obtained.

7. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: In S5, the analysis method of spectral features and texture features is as follows: (1) The analysis method of spectral characteristics is the mean variance method, and the calculation formula is as follows: Where: C Li is the brightness value of the i-th pixel in the object in the L-th band, C L is the average brightness of a single image object in the Lth band, m is the total number of objects in the image, n is the number of pixels in the object, S 2 is the variance; (2) The analysis method of texture features is the shape heterogeneity of the object. The shape heterogeneity index of the object is composed of two sub-heterogeneity indices, the smoothness index and the compactness index, and is obtained by the shape heterogeneity increment before and after the object is merged. The shape heterogeneity increment before and after the object is merged is the weighted average of the smoothness index increment and the compactness index increment, and its expression is as follows: h shape =w smoothness ×h smoothness +w compactness ×h compactness Where: w smoothness With w compactness represents the weight allocation between the two, and the sum of the two is 1; h shape represents the shape heterogeneity increment; h smoothness Indicates the smoothness index increment; h compactness Indicates the compactness index increment.

8. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: In S6, the steps of adjusting parameters of the multi-scale segmentation algorithm for automatically segmenting the image range are as follows: (1) According to the previously estimated scale range, set the initial segmentation scale parameter; (2) The test results are superimposed and verified with the refined land use types and their ranges obtained in S4 to determine the parameters to be adjusted; (3) First fix other parameters and change the segmentation scale parameter, starting from a smaller scale and gradually increasing it, and observe the change in segmentation accuracy; when the segmentation accuracy begins to decrease, record the scale range at this time, and then further fine-tune other parameters within this range; if the accuracy of a certain type of land use is improved after increasing the spectral threshold, while the accuracy of other types does not change much, continue to adjust in this direction; (4) Repeat steps (1) to (3) and continuously adjust the parameter combination until the segmentation result in the test area reaches the highest accuracy. Record the parameter values ​​at this time and use them as the optimal parameter combination for the test area for subsequent image segmentation of the entire study area.

9. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1 is characterized in that: In S7, the optimal parameter combination obtained in S6 is used to further use the multi-scale segmentation method to automatically segment the image in the study area, and the accuracy verification is performed. The calculation formula for the accuracy verification is as follows: (1) Production accuracy PA reflects the consistency between the correct classification result and the reference category. The calculation formula is as follows: Where: a ii is the correct number of various verification sample points, a ki is the total number of sample points verified in this category, and N is the total number of sample categories; (2) User accuracy UA reflects the consistency between the correct classification result and the actual classification result. The calculation formula is as follows: Where: a ii is the correct number of various verification sample points, a ik is the actual number of classifications of this category; (3) Range accuracy SA reflects the consistency between the correct range result and the actual range result of the plot. The calculation formula is as follows: Where: b ij is the correct area range of each type of verification sample plot, b ik The actual area range of each plot of this type.

10. The method for automatically extracting urban wetland vegetation coverage based on UAV-RGB according to claim 1, characterized in that: In S8, when the land and vegetation types in S7 reach the preset accuracy, the vegetation area vector map is extracted, and the area of ​​the vegetation area is calculated and counted. The calculation formula of the vegetation coverage rate V is as follows: Where: A v is the vegetation coverage area, A t is the total wetland area.

Images