Rapid forest window identification and digitalization method based on visible light image of unmanned aerial vehicle

By combining high-resolution UAV visible light imagery with threshold classification and Gaussian mixture model based on spectral and texture information, the problem of low cost and high efficiency in forest gap identification was solved. Robust identification was achieved in complex terrain and high vegetation diversity environments, improving identification accuracy and efficiency.

CN118968358APending Publication Date: 2024-11-15EAST CHINA NORMAL UNIV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411026117.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Existing technologies struggle to identify forest gaps at low cost and high efficiency, especially in tropical and subtropical forests. Satellite remote sensing resolution is insufficient, airborne lidar is costly and has a significant impact from complex terrain, and existing UAV visible light image classification methods cannot accurately distinguish between regenerated vegetation and canopy.

Method used

High-resolution UAV visible light imagery is used, combined with threshold classification based on spectral and texture information. The shadow effect is eliminated by using gray-level co-occurrence matrix and Gaussian mixture model. Forest gap identification is performed using moving window and Bayesian criteria. Combined with bare land identification results, small patches are removed to improve identification accuracy.

Benefits of technology

It significantly improves the accuracy and efficiency of forest gap identification, reduces equipment costs, and can robustly identify forest gaps in complex terrain and high plant diversity environments, with an identification time of only 0.5 hours and an accuracy improvement of 3%-18%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968358A_ABST
    Figure CN118968358A_ABST
Patent Text Reader

Abstract

The invention discloses a forest window rapid identification and digitalization method based on a visible light image of an unmanned aerial vehicle. The method comprises the following steps: a) obtaining a high-resolution visible light image; b) processing the image to obtain a digital orthophoto map and a digital surface model, and eliminating a shadow effect; c) obtaining a preliminary forest window identification result based on threshold classification of spectrum and texture information, and dividing a bare land in the forest; d) in combination with the height information of the forest window and the canopy pixels, calculating forest window probabilities of all the pixels by creating a moving window and utilizing a Gaussian mixture model; and e) re-identifying all the pixels according to the forest window probability, removing small patches which do not meet the minimum area of the forest window in combination with a bare land identification result to obtain an accurate identification result of the forest window, and providing a digital product. According to the method, the forest windows in complex terrains and high-plant-diversity forests can be accurately identified, and an efficient, stable and low-cost solution is provided for forest ecology, forestry management, disaster monitoring and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of forestry, and in particular relates to a method for rapid identification and digitization of forest gaps based on visible light images of unmanned aerial vehicles. Background Art

[0002] Forest gaps are a common and frequently occurring disturbance state in forest ecosystems. They regulate the distribution of understory environment and resources (heat, water and soil nutrients) through forest gap characteristics (size, shape, direction and spatial distribution), play an important role in driving forest structure and community dynamics, and have a significant impact on the spatial distribution of plant species, community composition and forest succession. Therefore, real-time monitoring of the formation and evolution of forest gaps is of great significance for quickly assessing forest disturbance and guiding the protection and management of forest ecosystems.

[0003] Reliable forest gap identification methods are an important prerequisite for monitoring the spatiotemporal dynamics of forest gaps. Traditional ground-based methods, including manual field surveys and identification by taking hemispherical photos, are time-consuming and often inaccurate, and it is even more difficult to identify forest gaps at large spatial scales. Therefore, aerial remote sensing is expected to become the most effective tool for mapping the spatial distribution of forest gaps at different spatiotemporal scales. However, the spatial resolution of identifying forest gaps using satellite remote sensing data is too coarse to capture fine-scale forest gap characteristics. Although airborne LiDAR data has the ability to accurately identify forest gaps across scales, the high acquisition cost, large data storage volume and high technical barriers have severely restricted the acquisition and promotion of data across spatial and temporal scales. High-resolution visible light images acquired by drones are the most cost-effective data for forest gap identification applications. Current recognition methods using drone visible light images include pixel-based supervised classification, object-oriented classification and canopy height model threshold classification based on aerial photogrammetry. Specifically, supervised classification may cause forest gaps to be misidentified as closed canopies, especially pixel classification based on spectral information. Broadleaved trees pose a great challenge to object-oriented classification, as scale segmentation is very difficult in tropical and subtropical forests dominated by these functional groups. Complex terrain can significantly affect the quality of forest gap identification. Steep terrain causes a certain degree of offset in the projection of surrounding trees, affecting the accuracy of object-oriented classification in deciduous forests. In addition, steep slopes and undulating terrain can also create shadows during imaging, interfering with image stitching and thus affecting the quality of digital aerial photogrammetry data. Based on this, it is particularly important to develop an efficient, robust and low-cost technology to identify forest gaps.

[0004] According to the existing literature, CN117372710A proposed a forest gap extraction method based on Sentinel-2MSI remote sensing images. However, the spectral resolution of Sentinel-1 and Sentinel-2 data used in this method is very limited, and this technology cannot effectively identify forest gaps with a small area (<20m 2 ) forest window characteristics; CN107479065A is a method for measuring the three-dimensional structure of forest windows based on laser radar, which uses a laser radar instrument mounted on a fixed-wing aircraft or an unmanned aerial vehicle to obtain data for each laser point, including X, Y, Z coordinate data and echo intensity data, and then processes the data and identifies forest windows based on threshold classification. This technology is currently only applicable to forest window identification within a smaller scale range, has high technical requirements for users, and is expensive in instrument costs, and cannot meet the needs of long-term periodic monitoring of forest windows; CN104061904A proposes a method for quickly and accurately determining the shape and area of ​​forest windows, that is, generating orthophoto images based on drone aerial photos, and then using supervised classification to separate tree species and forest windows in the image. The supervised classification used in this method requires a large number of training samples, which is time-consuming and labor-intensive, and cannot accurately distinguish forest windows and canopies in forests with rich plant diversity, because the regenerated trees, shrubs and herbaceous plants in the forest windows cannot be distinguished from mature canopies in the spectrum, and complex terrain will also affect the imaging quality of drone images, thereby affecting the accuracy of supervised classification. Therefore, this method is not suitable for tropical and subtropical forests with rich species and high terrain heterogeneity. Summary of the invention

[0005] The purpose of the present invention is to develop a low-cost, efficient and robust method for rapid identification and digitization of forest gaps, to solve the problems that the current traditional ground survey method is time-consuming and labor-intensive, the satellite remote sensing identification of small forest gaps is insufficient in accuracy, the lidar data is costly and difficult to promote, and the existing UAV-based visible light image classification method cannot distinguish between regenerated vegetation and canopy and is easily affected by complex terrain, so as to improve the identification accuracy of forest gaps of different forest types.

[0006] The specific technical solution for achieving the purpose of the present invention is:

[0007] Step 1. Obtain high-resolution visible light images; high-resolution visible light images generally refer to those taken by fixed-wing aircraft or drones equipped with cameras. High resolution means that the ground resolution of the image should not exceed 10 cm. The camera includes three spectral bands, with the central wavelength range of red light being 780-622nm, the central wavelength of green light being 577-492nm, and the central wavelength of blue light being 450-400nm. Before performing the aerial survey mission, control points should be set up to ensure product quality, with the number of control points being no less than 5. Choose clear, windless weather with good visibility and noon when the sun is shining vertically to carry out the aerial survey mission to reduce the impact of wind and shadows on the image quality. In mountainous forests with large terrain undulations, the terrain-imitating flight mode is used to keep the flight altitude difference consistent, with the heading overlap rate being greater than or equal to 60%, and the lateral overlap rate being greater than or equal to 50%.

[0008] Step 2: Process the drone image to obtain a digital orthophoto map and a digital surface model and eliminate the shadow effect; the drone image processing uses professional photogrammetry software. Before recognition, the digital orthophoto map should be matched and compensated for the hue, intensity and saturation values ​​of the shadow area according to the optical characteristics of the forest shadow, so as to eliminate the influence of the shadow on the imaging to a certain extent. At the same time, by calculating the vegetation difference index in the visible light band, the shadow effect of the entire scene can be further reduced.

[0009] Step 3: Identify bare land in the forest based on threshold classification of spectral and texture information and obtain preliminary forest gap identification results; select the threshold for distinguishing between canopy and non-canopy by comparing the green band reflectance statistics of canopy and non-canopy and the average value of gray level co-occurrence matrix (GLCM) representing texture information; and select the threshold for identifying bare land in the forest by comparing the visible-band difference vegetation index (VDVI). Preliminary forest gap identification results are obtained based on these thresholds.

[0010]

[0011] where ρ green , red and ρ blue Represent the spectral reflectance of the green, red and blue bands respectively.

[0012] The gray-level co-occurrence matrix is ​​defined as the joint distribution probability of pixel pairs. It is a symmetric matrix that not only reflects the comprehensive information of the image grayscale in adjacent directions, adjacent intervals, and change amplitudes, but also reflects the position distribution characteristics between pixels of the same grayscale level. It is the basis for calculating texture features.

[0013] Assume P(i,j) is a digital image with a size of M×N and a gray level of Ng. The gray level co-occurrence matrix that satisfies a certain spatial relationship is:

[0014] P(i,j)=#{(x1,y1),(x2,y2)∈M×N∣f(x1,y1)=i,f(x2,y2)=j}

[0015] Where #(x) represents the number of elements in the set x, P is a matrix of Ng×Ng, if the distance between (x1,y1) and (x2,y2) is d, and the angle between the two and the horizontal axis of the coordinate is θ, then the gray-level co-occurrence matrix P(i,j,d,θ) of various spacings and angles can be obtained.

[0016] Step 4. Combine the height information of forest windows and canopy layers. On the basis of the preliminary identification results of forest windows, create a moving pixel window (7×7 pixels) to gradually judge the entire area to be identified. Set the proportion of forest window pixels in the window to be more than 50%. If this condition is not met, continue to expand the window to ensure that there are enough prior samples in the window. When the conditions are met, use the mean and variance of the height of forest window and non-forest window pixels in the window as the initial values ​​to calculate the parameters in the Gaussian mixture model (GMM), and use the proportion of forest window pixels to the total number as the prior probability. The Gaussian mixture model contains multiple normal distribution functions, and its probability density function formula is as follows:

[0017]

[0018] In the formula, z is a vector Z={z1,z2,z3,…,z n}, the vector contains c categories, and the weight of the i-th category is Φ i , the function F(z) obeys the normal distribution, μ i , σ i is the mean and variance of the i-th category; then, the expectation maximization algorithm is used to solve and update the model parameters and prior probabilities, the posterior probability is calculated by the Bayesian criterion, and the solution is repeated until convergence.

[0019] Step 5: Re-evaluate all pixels according to the final forest window probability, remove small patches, and combine the bare ground identification results to obtain accurate forest window identification results; the specific operations are as follows: perform binary classification on the layer according to the probability of forest window pixels, that is, combine forest windows and non-forest windows with bare ground identification results to obtain accurate forest window identification results, and finally, remove small patches in the 8-neighborhood area of ​​the result, and set the small patch area threshold to 5-25m 2 .

[0020] Beneficial effects of the present invention:

[0021] (1) The present invention uses visible light images from unmanned aerial vehicles as a data source for forest gap identification. The required equipment cost is low, data acquisition is simple, and the application prospects are broad, with great market demand.

[0022] (2) The present invention adopts programmed recognition throughout the entire process. After the model is built, the forest gap recognition process of the present invention only takes 0.5 hours, which significantly improves the efficiency of forest gap recognition.

[0023] (3) Based on the threshold classification of spectral and texture information to identify bare land in the forest, the present invention sets a moving pixel window, and superimposes the mean and variance of the heights of forest window and non-forest window pixels in the window as initial values ​​to calculate the true probability of the pixel being a forest window in the Gaussian mixture model. This method can effectively overcome the influence of complex terrain, high plant diversity and low accuracy of terrain data on the recognition results, significantly improve the accuracy of forest window recognition, and has good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flowchart of the rapid identification and digitization of forest gaps based on visible light images from drones in the present invention;

[0025] Figure 2 This is a schematic diagram of the layout of ground control points in Tiantong Mountain according to an embodiment of the present invention;

[0026] Figure 3 Examples of UAV visible light remote sensing data: (a) UAV aerial survey photos; (b) orthophotos; (c) digital surface models;

[0027] Figure 4 The forest gap identification results of four different methods and two verification data. The forest gaps are blue patches.

[0028] Figure 5 The spatial overlap and difference of forest gaps between the four methods and the validation data.

[0029] HSTAC: A method for rapid identification and digitization of forest gaps based on visible light images from unmanned aerial vehicles; PBSC: Pixel-based supervised classification; OBIA: Object-oriented classification; CHMC: Threshold classification based on canopy height model; ULS: Forest gap classification based on airborne lidar data; Ground survey: Field survey of forest gaps DETAILED DESCRIPTION

[0030] The present invention is further described in detail below in conjunction with embodiments, but the protection scope of the present invention is not limited to the embodiments.

[0031] Example

[0032] The present embodiment is located in the forest long-term monitoring plot in Tiantong National Forest Park, Yinzhou District, Ningbo City, Zhejiang Province. The plot area is 20ha, and the length and width are 500m and 400m respectively. The plot altitude range is 304.26-602.89m, and the altitude difference is close to 300m. Two wide deep trenches run through the entire plot from north to south, and the overall terrain presents a large heterogeneity, showing steep slopes, deep valleys and significant terrain undulations. According to the plot inventory results in 2015, there are a total of 108,656 surviving individuals of 52 families, 97 genera and 154 species. The vegetation community is a typical evergreen broad-leaved forest in the subtropical China, and the dominant families are mainly Lauraceae, Theaceae and Fagaceae.

[0033] This embodiment mainly includes five steps, and the specific process is as follows: Figure 1 .

[0034] Step 1: Obtain high-resolution visible light images from drones;

[0035] This embodiment uses a MD4-1000 quad-rotor drone, combined with a ground operating system and sensors and other accessories to form the entire drone system. The maximum payload mass of the drone system is 1.2 kg. Under ideal weather conditions at noon, clear and windless, the maximum flight time without a payload is 88 minutes, and the cruising speed is 12m / s. The sensor is equipped with a Sony NEX-5 camera, which has three bands, with central wavelengths of 450nm (blue), 520nm (green) and 660nm (red). The following steps are mainly included:

[0036] 1) Select and arrange ground control points in the sample plot and surrounding open space or relatively open areas through field survey. In this embodiment, a total of 9 ground control points are arranged, and the overall distribution of the control points is as follows: Figure 2 After the ground control points were set up, Trimble RTK was used to measure the 3D coordinates of these points with an accuracy of less than 1m.

[0037] 2) Select a suitable take-off and landing platform based on the topography and vegetation coverage of the sample site obtained during the field survey and set the flight path on the ground operating system mdCockpit.

[0038] 3) To minimize the impact of wind and shadow on the image quality, try to choose clear and windless weather with good visibility and noon when the sun is shining vertically to perform the aerial survey task. Due to the large altitude difference in the sample site (nearly 300m), this embodiment adopts the mode of terrain-simulating flight to keep the flight altitude at about 200m above the ground. During the entire aerial survey mission of the sample site, the heading overlap rate and the lateral overlap rate are maintained at 80% and 70%.

[0039] 4) After the drone takes off, the ground control system is monitored in real time, and the sensor shooting quality and flight control signals are always paid attention to. In addition, the take-off and landing time and the sudden weather changes that may occur during the aerial survey are recorded in real time. Overall, this embodiment collected 303 valid photos of the Tiantong sample site, and the original photos of the canopy collected by the drone are as follows: Figure 3 (a).

[0040] Step 2: Process the drone images to obtain digital orthophotos and digital surface models and eliminate shadows;

[0041] This embodiment uses the professional software "Pix4dmapper" widely used in drone photogrammetry to process these remote sensing images and generate corresponding remote sensing products. The specific steps are as follows:

[0042] 1) Import images and set image properties: Perform a comprehensive inspection of the images and remove images with poor quality or shadows caused by sudden changes in light.

[0043] 2) Mark control points: Import the control point coordinate file in the control point editor and select the control point coordinate system to be consistent with the output coordinate system. Open the plane editor, find the image where the control points are located, mark three control points, and perform fast aerial triangulation (aerotriangulation) processing. After the initialization processing is completed, all control points can be marked in the aerotriangulation editor.

[0044] 3) Standard aerial triangulation: After the control points are confirmed and marked, adjust the template to standard mode and check the initialization processing, point cloud and texture and digital surface model, orthophoto and index steps.

[0045] 4) Imagery evaluation: The quality of drone imagery is evaluated based on the drone imagery quality report generated in each plot. The content mainly includes images, data sets, camera optimization, matching, geographic reference errors, etc. In this embodiment, the drone imagery processing reports are all tested, and finally a digital orthophoto map and a digital surface model with a spatial resolution of about 5 cm are generated. Figure 3 (b) and Figure 3 (c). Before identification, the hue, intensity and saturation values ​​of the shadow area should be matched and compensated according to the optical characteristics of the forest shadow, so as to eliminate the influence of the shadow on the imaging to a certain extent. At the same time, the visible light band vegetation difference index is calculated from the visible light band to further reduce the shadow effect of the whole scene.

[0046] Step 3: Identify bare land in the forest and obtain preliminary forest gap identification results based on threshold classification of spectral and texture information;

[0047] By comparing the green band reflectance statistics of the canopy and non-canopy, a threshold of 65 was selected to distinguish the canopy from the non-canopy. The index value of the area to be identified was obtained based on the calculation formula of the visible light band difference vegetation index. Multiple thresholds such as 0.2, 0.3 and 0.4 were set to determine the bare forest land, and 0.2 was selected as the best threshold after comparing these thresholds. At the same time, in Envi5.3, the gray level co-occurrence matrix average value of the image of the area to be identified was obtained based on the calculation principle of the gray level co-occurrence matrix to distinguish the canopy and non-canopy, and the average values ​​of all gray level co-occurrence matrices were calculated with an interval of 0.5 in the threshold range of 7-12. Finally, 8.5 was selected as the best threshold after comparing all threshold results. The threshold range that meets the three different characteristics in all pixels at the same time is the forest window. Therefore, the preliminary identification results of the forest window were obtained based on the threshold classification method.

[0048] Step 4: Combine the height information of forest gaps and canopy layers, and calculate the probability of forest gaps for all pixels by creating a moving window and using the Gaussian mixture model; based on the preliminary identification results of forest gaps, create a moving pixel window (7×7 pixels) to gradually determine the entire area to be identified, and set the proportion of forest gap pixels in the window to be more than 50% to ensure that there are enough prior samples in the window. If this condition is not met, continue to expand the window (9×9 or 11×11 pixels). When the conditions are met, calculate the mean and variance of the height of forest gap and non-forest gap pixels in the window and use them as initial values ​​to calculate the parameters in the Gaussian mixture model, and use the proportion of forest gaps to the total number as the prior probability. Subsequently, the expectation maximization algorithm is used to solve and update the model parameters and prior probabilities, and the posterior probability is calculated by the Bayesian criterion, and the solution is repeated until convergence.

[0049] Step 5: Re-identify all pixels according to the probability of forest gaps, remove small patches, and combine the bare land identification results to obtain accurate forest gap identification results; perform binary classification on the layer according to the probability of forest gap pixels, that is, combine forest gaps and non-forest gaps with bare land identification results to obtain accurate forest gap identification results. Finally, remove small patches in the 8-neighborhood, and set the patch area threshold to 5m 2 The recognition results are as follows: Figure 4 (d).

[0050] Step 6: Based on the forest gap identification results of the ground survey and the UAV-mounted laser radar data, the accuracy of the present invention is verified. Figure 4 (a) pixel-based supervised classification, Figure 4 (b) Object-oriented classification and Figure 4 (c) The comparison of threshold classification based on the canopy height model completes the recognition efficiency verification of the present invention.

[0051] The ground survey data, including 112 canopy and 179 forest gap samples, were collected by a local vegetation expert with extensive field experience using the 50 canopy gap locations surveyed as references, e.g. Figure 4 (f) The lidar data were acquired in June 2018 from approximately 71 meters above the ground in 14 separate missions using a RIEGL mini vux-1 drone scanning device mounted on a quadrotor drone. The average point density of the lidar data is approximately 44 points / m 2 Based on Delaunay triangulation, the digital elevation model (DEM), digital surface model (DSM) and canopy height model (CHM) are interpolated to 1 meter resolution. CHM (x, y) = DSM (x, y) - DEM (x, y)

[0052] LiDAR data processing was done in R version 3.6.1 using the 'lidR' package. Based on the canopy height model, a fixed height threshold classification method suitable for a high-precision canopy height model was adopted, and the "getForestGaps" function in the R package "ForestGapR" was used to identify forest gaps, where the area threshold was set to be greater than 5 square meters and the height threshold was set to be less than 8 meters. The identification results are shown in Figure 2. Figure 4 (e). Based on the validation data, the accuracy evaluation indicators of overall accuracy, Kappa coefficient and time cost were selected to evaluate the recognition accuracy and efficiency of the four methods. At the same time, spatial distribution maps were generated in the overlapping areas of different recognition results and the ratio of forest gap area to forest gap number was calculated.

[0053]

[0054] Where p0 is the sum of the number of correctly classified samples in each category divided by the total number of samples, representing the overall classification accuracy. ∈ It is the sum of the products of the number of real samples of each category and the number of predicted samples, divided by the square of the total number of samples. If the number of pixels of the actual reference of the forest window is a1, a2, ..., a k The number of forest gap pixels predicted by the selected method are b1, b2, ..., b k , the total number of pixels is N, then

[0055] Positive Effects

[0056] Based on the ground survey data to evaluate the recognition accuracy, the invention proposes a fast forest gap recognition method based on photogrammetric height, spectrum and texture information, which performs best in recognition efficiency (96% overall accuracy compared with ground survey data and 85% accuracy compared with lidar data), recognition accuracy (3%-18% improvement compared with existing supervised classification, object-oriented classification and threshold classification based on canopy height model) and speed (saving 1-1.5 hours on the same computer). Based on further verification of lidar data, the recognition accuracy of the invention is 85% and the Kappa coefficient is 0.74, as shown in Table 1. Figure 5 It means that by superimposing the spatial distribution of different forest gap identification results, the total overlapping area of ​​forest gaps is 6078m 2 The supervised classification has the highest spatial matching degree with the present invention (9503m 2 ), followed by the present invention and object-oriented classification (8836m 2 ), and finally supervised classification and object-oriented classification (7202m 2 ). In terms of the relative quantity of spatial matching, the overlapping forest gap area ratio produced by the present invention is the highest (63%), and the relative quantity percentage is the highest at 80%, as shown in Table 2. The present invention can well avoid the inherent defects of existing supervised classification, object-oriented classification and canopy height model-based threshold classification methods, can well overcome the problems of complex terrain, high plant diversity and low terrain data accuracy, and has good robustness.

[0057] Table 1 Summary of the results of identifying forest gaps and canopies using four methods: overall accuracy, error, and time cost

[0058]

[0059] Table 2 Forest gap rapid identification method based on UAV visible light images, supervised classification method and object-oriented classification method Forest gap identification results under different spatial overlaps Area and number ratios

[0060]

Claims

1. A method for rapid identification and digitization of forest gaps based on visible light images from unmanned aerial vehicles, characterized in that: The following steps are involved: Step 1: Obtain high-resolution visible light images from drones; Step 2: Process the drone images to obtain digital orthophotos and digital surface models and eliminate shadow effects; Step 3: Based on the threshold classification of spectral and texture information, preliminary forest gap identification results are obtained and bare land in the forest is divided; Step 4: Based on the preliminary recognition results, combined with the height information of forest gaps and canopy pixels, the forest gap probability of all pixels is calculated using the Gaussian mixture model by creating a moving pixel window; Step 5: Re-identify all pixels according to the forest gap probability, remove small patches that do not meet the minimum forest gap area, and combine the bare land identification results to obtain accurate forest gap identification results; where: The specific operation of step 4 is as follows: based on the preliminary forest window binary judgment, a moving window of 7×7 pixels is created to gradually judge the entire area to be identified, and the proportion of forest window pixels in the window is set to be more than 50%. If this condition is not met, the window is further expanded to ensure that there are enough prior samples in the window; when the condition is met, the mean and variance of the heights of forest window and non-forest window pixels in the window are used as initial values ​​to calculate the parameters in the Gaussian mixture model, and the proportion of forest window pixels to the total number is used as the prior probability; the Gaussian mixture model contains multiple normal distribution functions, and its probability density function formula is as follows: In the formula, z is a vector Z={z1,z2,z3,…,z n }, the vector contains c categories, and the weight of the i-th category is Φ i , the function F(z) obeys the normal distribution, μ i , σ i is the mean and variance of the i-th category; then, the expectation maximization algorithm is used to solve and update the model parameters and prior probabilities, the posterior probability is calculated by the Bayesian criterion, and the solution is repeated until convergence.

2. The method for rapid identification and digitization of forest gaps based on visible light images from unmanned aerial vehicles according to claim 1, characterized in that: The high-resolution visible light image mentioned in step 1 refers to the image taken by a fixed-wing aircraft or an unmanned aerial vehicle equipped with a camera. High resolution means that the ground resolution of the image should not exceed 10 cm. The camera includes three spectral bands, with the central wavelength of the red light band being within the spectral range of 780-622nm, the central wavelength of the green light band being 577-492nm, and the central wavelength of the blue light band being 450-400nm. Before performing the aerial survey task, control points shall be arranged to ensure product quality, and the number of control points shall be no less than 5. Aerial survey tasks shall be carried out at noon when the weather is clear and windless, the field of vision is good, and the sun is shining vertically to reduce the impact of wind and shadows on the image quality. In mountainous forests with large terrain undulations, a terrain-simulating flight mode shall be used to keep the flight altitude difference consistent, with a heading overlap rate greater than or equal to 60%, and a lateral overlap rate greater than or equal to 50%.

3. The method for rapid identification and digitization of forest gaps based on visible light images from unmanned aerial vehicles according to claim 1, characterized in that: The step 2 specifically includes: using professional photogrammetry software to process drone images to obtain digital orthophotos and digital surface models; matching and compensating the hue, intensity and saturation values ​​of the shadow area of ​​the digital orthophotos according to the optical characteristics of forest shadows before recognition to eliminate the influence of shadows on imaging; and at the same time, further reducing the shadow effect of the entire scene by calculating the vegetation difference index in the visible light band.

4. The method for rapid identification and digitization of forest gaps based on visible light images from unmanned aerial vehicles according to claim 1, characterized in that: The thresholds described in step 3 are specifically set as follows: the threshold for distinguishing between canopy and non-canopy is selected by comparing the green band reflectance statistics of canopy and non-canopy and the average value of the gray level co-occurrence matrix GLCM representing texture information, and the threshold for identifying bare forest land is selected by comparing the visible light band difference vegetation index VDVI, and the preliminary identification result of forest gap is obtained based on the threshold; where ρ green , red and ρ blue Represent the spectral reflectance of green, red and blue bands respectively; The gray-level co-occurrence matrix is ​​defined as the joint distribution probability of pixel pairs. It is a symmetric matrix that not only reflects the comprehensive information of the image grayscale in adjacent directions, adjacent intervals, and change amplitudes, but also reflects the position distribution characteristics between pixels with the same grayscale level. It is the basis for calculating texture features. Assume P(i,j) is a digital image with a size of M×N and a gray level of Ng. The gray level co-occurrence matrix that satisfies a certain spatial relationship is: P(i,j)=#{(x1,y1),(x2,y2)∈M×N∣f(x1,y1)=i,f(x2,y2)=j} Where #(x) represents the number of elements in the set x, P is a matrix of Ng×Ng, if the distance between (x1,y1) and (x2,y2) is d, and the angle between the two and the horizontal axis of the coordinate is θ, then the gray-level co-occurrence matrix P(i,j,d,θ) of various spacings and angles can be obtained.

5. The method for rapid identification and digitization of forest gaps based on visible light images from unmanned aerial vehicles according to claim 1, characterized in that: The step 5 specifically includes: performing binary classification on the layer according to the probability of forest window pixels, that is, combining forest windows and non-forest windows with bare land identification results to obtain accurate forest window identification results; finally, removing small patches in the 8-neighborhood area of ​​the result, and setting the small patch area threshold to 5-25m 2 .

Citation Information

Patent Citations

  • Method for rapidly and accurately determining shape and area of forest gap

    CN104061904A

  • Method for measuring canopy gap three-dimensional structure based on laser radar

    CN107479065A

  • Forest window extraction method based on Sentinel-2 MSI remote sensing image

    CN117372710A

  • Visible light unmanned aerial vehicle remote sensing image forest tree species classification method based on multi-feature optimization

    CN110991335A

  • Tree crown extraction method based on unmanned aerial vehicle multi-source remote sensing

    US20230039554A1