Desert sparse arbor extraction method based on fusion feature Otsu method and multispectral aerial photography

Abstract

The invention discloses a desert sparse arbor extraction method based on a fusion feature Otsu method and multispectral aerial photography, and the method is based on an unsupervised Otsu method frame, fuses SAVI vegetation indexes of near infrared and red light bands and green light and red light difference indexes based on visible light three bands, generates a two-dimensional comprehensive feature SG, guarantees that the SG values of a shadow region and a bare soil background are lower than those of arbor, and improves the extraction efficiency of the arbor. The separability of the target is enhanced, misjudgment is avoided, and the recognition accuracy is improved. By taking the SAVI mean value minimization of the pure soil area as an optimization target, the robustness of the algorithm to the dynamic change of the soil background is improved by calculating the soil adjustment parameter L in equal step length. SG comprehensive features are used as weight factors to carry out replacement random sampling, and the probability that the arbor area is sampled is improved. By repeatedly calculating the Otsu method threshold value for M times and taking the mean value for morphological opening / closing operation, the problem that the threshold value deviates from the background due to unbalanced samples in a traditional unsupervised method is solved, and the stability and reliability of a separation result are improved so as to filter a false detection area.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing and ecological monitoring, and in particular to a sparse desert tree extraction method based on fusion feature Otsu method and multispectral aerial photography. BACKGROUND

[0002] With the rapid development of remote sensing technology and unmanned aerial vehicle platform, vegetation monitoring (especially sparse vegetation extraction) has become a core technology direction in the fields of ecological protection, desertification prevention, etc. The current industry technology development focuses on three goals: high-precision extraction, low-cost adaptation, and wide-area monitoring. A multi-technology solution based on satellite remote sensing, unmanned aerial vehicle multi / hyperspectral, machine learning, etc. has been formed. For example, patents CN118587603A, CN112287892B, and CN112861722B have formed a multi-technology system covering supervised learning (such as machine learning, deep learning), unsupervised learning (such as geostatistics threshold segmentation), and semi-supervised technology, and have achieved good results in target extraction in typical vegetation areas such as forests and grasslands.

[0003] However, in the desert scene, the spectral characteristics of desert sand and tree shadow partially overlap, making it more difficult to identify trees in desert sand. It is impossible to rely on a single feature to accurately identify them. At the same time, the proportion of trees in the desert area is extremely low. During sampling, most samples are sand, shadow, and other backgrounds. In the identification method of unsupervised segmentation, the segmentation threshold is easily biased towards the background class, often leading to misjudgment of the tree area as background. In the desert area where vegetation is sparse, the problem of missed detection is particularly significant, and it is impossible to achieve the required identification accuracy. In the supervised method, the sample labeling cost is too high, making it difficult to meet the efficient monitoring demand. Moreover, the current deep learning method mostly relies on high-performance PC devices, which cannot support real-time computing results on unmanned aerial vehicle edge devices (such as Android phones).

[0004] In addition, in the desert environment, the reflectivity of sand changes with humidity and the angle of light, causing fluctuations in spectral characteristics. Traditional vegetation monitoring methods do not have special strategies for the high reflectivity of sand and shadow interference in the desert. SUMMARY

[0005] The present application aims to provide a sparse desert tree extraction method based on fusion feature Otsu method and multispectral aerial photography, to solve the problem of lack of suitable tree identification methods in the desert environment and the problem of low tree identification accuracy and high cost in the desert environment.

[0006] To achieve the above-mentioned purpose, the sparse desert tree extraction method based on fusion feature Otsu method and multispectral aerial photography according to the present application comprises the following steps: S1, using a multi-spectral sensor-equipped unmanned aerial vehicle aerial photography to obtain multi-spectral images of the desert to be analyzed; S2, performing radiation correction on the multi-spectral images to obtain corrected multi-spectral images; S3, performing geographic registration on the corrected multi-spectral images; S4, using professional remote sensing processing software to stitch the corrected multi-spectral images and output raster data; S5, calculating SAVI vegetation index and green-red difference index respectively according to the raster data and performing normalization processing; S6, fusing the normalized SAVI vegetation index and green-red difference index to obtain a strengthened tree feature value SG; S7, performing weighted random sampling with replacement based on the SG value; S8, repeating M times, inputting the M times sampling results into the law of the large numbers to obtain M corresponding segmentation thresholds; S9, calculating the average value of all segmentation thresholds; S10, marking the pixels with a strengthened tree feature value SG greater than or equal to the average value in the stitched multi-spectral images as trees.

[0007] Further, the multi-spectral sensor includes a near-infrared sensor with a wavelength range of 750-900 nm, a red light sensor with a wavelength range of 630-680 nm, and a blue, green, and red three-band visible light sensor.

[0008] Further, the radiation correction method includes performing radiation correction on the multi-spectral images through a linear regression model established by the reflectivity of a calibration plate and the digital quantization value output by the multi-spectral sensor; automatically performing radiation correction using a radiation calibration module built into the unmanned aerial vehicle; and performing radiation correction using the dark pixel method.

[0009] Further, the raster data includes blue, green, and red three-band visible light raster data, near-infrared band reflectivity raster data, and red light band reflectivity raster data.

[0010] Further, the stitching should satisfy the plane error of the stitched image being ≤2 pixels, ensuring the spatial alignment of multi-band data.

[0011] Further, the SAVI vegetation index calculation process specifically includes selecting a bare soil area without vegetation coverage, taking the absolute value of the mean SAVI vegetation index of the bare soil area as the optimization target, traversing the available value interval of the soil adjustment parameter L, and determining the optimal soil adjustment parameter L * ; and determining the optimal soil adjustment parameter L *Substituting into the SAVI vegetation index calculation formula, the SAVI vegetation index of the S4-step mosaic image is obtained; the SAVI vegetation index is normalized by truncating negative SAVI values ​​to 0 and keeping the remaining SAVI values ​​as their original values.

[0012] Furthermore, the green light to red light difference index DN G DN represents the response intensity of the green channel in a visible light image. R The response intensity of the red channel in the visible light image is given; the normalization of the green-red light difference index adopts the inverse normalization method.

[0013] Furthermore, the SAVI vegetation index and green-red light difference index fusion method includes SAVI vegetation index and green-red light difference index multiplication fusion, weighted fusion, Manhattan distance fusion, or Euclidean fusion.

[0014] Preferably, in step S7, the weighted random sampling with replacement based on the SG value specifically involves: dividing the data interval containing the SG value into K equal intervals, and calculating the probability density of each data interval. and weight Then, weighted random sampling with replacement is performed on each data interval; the probability density of the data interval is... , The number of pixels in the Kth data interval; N is the total number of pixels in all data intervals; the weight of the data interval. , Let be the probability density of the j-th data interval.

[0015] Preferably, in step S7, the weighted random sampling with replacement based on the SG value specifically involves: calculating the weights based on the logarithmic transformation of the SG value. Weighted random sampling with replacement is performed on the sample pixels; in, Let be the SG value of the i-th pixel.

[0016] The advantages of this invention lie in its unsupervised Otsu method framework, which integrates the SAVI vegetation index (reflecting spectral differences) from the near-infrared (NIR) and red (R) bands with the green-red light difference index based on the three visible light bands (RGB) to generate a two-dimensional comprehensive feature SG. This ensures that the SG values ​​of shaded areas and bare soil backgrounds are lower than those of trees, thereby enhancing the separability of targets, avoiding false positives, and improving recognition accuracy. Minimizing the mean SAVI value in pure soil areas is the optimization objective. A soil adjustment parameter L is calculated using equal step sizes to improve the algorithm's robustness to dynamic changes in the soil background. Random sampling with replacement is performed using the SG comprehensive feature as a weighting factor to increase the probability of tree areas being sampled. By repeatedly calculating the Otsu method threshold M times and taking the average, the problem of threshold bias towards the background caused by sample imbalance in traditional unsupervised methods is improved, enhancing the stability and reliability of the separation results. Opening / closing operations are performed using the mean of the Otsu method threshold to filter out false detection areas.

[0017] The method of this invention improves the feasibility of wide-area desert detection by automatically sampling randomly without the need for manual sample labeling and by not relying on graphics computing units in the calculation process. It also improves the performance of the algorithm in terms of target separability, background robustness, computational efficiency and scene generalization, providing a high-precision, low-cost and highly adaptable technical solution for the extraction of sparse trees in deserts. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method described in this invention.

[0019] Figure 2 These are example images of visible light (a), red light band reflectance (b), and near-infrared band reflectance (c) from Example 2.

[0020] Figure 3 These are example images of SAVI image (a) and GRDI image (b) from Example 2.

[0021] Figure 4 The SAVI obtained in Example 2 after normalization norm Image (a) and GRDI norm Example image (b).

[0022] Figure 5 These are example images of the fused feature SG image (a) and the binarized tree mask image (b) used to calculate the Otsu threshold in Example 2. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1 The desert sparse tree extraction method based on the Otsu method of feature fusion and multispectral aerial photography described in this invention, such as... Figure 1 As shown, it includes the following steps: S1 employs a drone equipped with a multispectral sensor for aerial photography to obtain multispectral images of the desert to be analyzed. The drone equipped with the multispectral sensor includes at least the following sensors: a visible light imaging sensor for acquiring blue, green, and red visible light bands; a multispectral imaging sensor for the red light band (R) with a wavelength range of 630-680nm and the near-red light (NIR) band with a wavelength range of 750-900nm.

[0025] In some embodiments, a visible light imaging sensor for acquiring blue, green, and red visible light bands, and a multispectral imaging sensor sensitive to chlorophyll content in vegetation with a red-edge band wavelength range of 700-750 nm and a near-red light wavelength range of 750-900 nm can also be used. Alternatively, a wide-band sensor, such as a near-infrared and panchromatic imaging sensor, can be used, and the data can be fused to form pseudo-multispectral data.

[0026] When conducting drone aerial photography, flight parameters are set according to the characteristics of sparse trees in the desert scene. Based on experience, the flight altitude is 40~120m, ensuring that the ground resolution (GSD) is ≤5cm, ensuring that the boundary details of the canopy of a single tree (usually 2-5m) are clearly captured, with a forward overlap rate of ≥70% and a lateral overlap rate of ≥80%, to ensure the accuracy of image stitching.

[0027] S2. To eliminate sensor response differences, radiometric correction needs to be performed on the multispectral image to obtain a corrected multispectral image, ensuring the accuracy of reflectance in each band. The specific steps are as follows: First, select calibration plates. Use 1-3 square standard gray plates with known reflectance (e.g., 25%, 50%, 75%), with a recommended size of 20cm x 20cm or larger for each plate.

[0028] Secondly, before the aerial survey mission begins, place the calibration board in a flat area without shadows, and shoot at a height that is 5 times the side length of the calibration board (e.g., a 30cm×30cm calibration board corresponds to a shooting height of 150cm). Adjust the camera angle to be perpendicular to the ground, ensuring that the calibration board fills the center area of ​​the lens (to avoid edge distortion), and turn off the sensor's automatic exposure function when shooting the calibration board.

[0029] Then, a linear regression model is established using the reflectivity of the calibration plate and the digital quantization value output by the sensor to perform radiometric correction on the images of each band.

[0030] In some embodiments, radiometric calibration can be automatically performed using the built-in radiometric calibration module of the UAV, eliminating the need for a calibration board and making it suitable for rapid aerial survey scenarios. Alternatively, the dark pixel method can be used to estimate atmospheric path radiation using the darkest pixels in the image (such as water bodies or shadows), indirectly correcting reflectivity. The dark pixel method is particularly suitable for emergency monitoring without a calibration board.

[0031] S3 performs georegistration on the corrected multispectral imagery. Different methods are used for image geocorrection depending on the UAV's positioning capabilities. Specifically, it can be categorized into RTK and non-RTK models.

[0032] For RTK models, orthophotos can be generated directly using the drone's built-in RTK positioning data (accuracy ≤ 5cm), without the need for additional ground control points.

[0033] For non-RTK models, image control points are evenly distributed in the survey area in a triangular or I-shaped pattern (density 150~200m, number not less than 3). The coordinates of the image control points are measured by ground total station or GNSS receiver (accuracy ≤10cm). Geographic reference image is generated by aerial triangulation and orthorectification.

[0034] S4. Using professional remote sensing processing software (such as Pix4D, Agisoft Metashape, etc.), the corrected multispectral images are stitched together and raster data is output. Specifically, this includes visible light three-band (red, green, blue) reflectance raster data, near-infrared (NIR) and red (R) band reflectance raster data. The planar error of the stitched image is ≤2 pixels to ensure spatial alignment of the multi-band data.

[0035] S5. Based on the raster data, calculate the SAVI vegetation index and the green light-red light difference index respectively, and then normalize them.

[0036] The calculation of the SAVI vegetation index specifically includes: The human eye interprets and filters bare soil areas without vegetation cover, with the number of pixels in the bare soil area not less than 50.

[0037] With the goal of minimizing the absolute value of the mean SAVI vegetation index in bare soil areas, the optimal soil regulation parameter L is determined by iterating through the possible ranges of the soil regulation parameter L to minimize the absolute value of the mean SAVI vegetation index in bare soil areas. * ; will L *Substituting the values ​​into the SAVI vegetation index calculation formula, the SAVI vegetation index of the S4-step mosaic image is obtained. Based on experience in desert areas, an initial search interval for L, Lmin, and Lmax can be set, with recommended values ​​of Lmin=0.3 and Lmax=0.8. The search is then iterated within the initial search interval with a step size ΔL (suggested value 0.1) to determine the L value that minimizes the absolute value of the mean SAVI vegetation index in the bare soil area as the optimal soil adjustment parameter L. * If multiple L values ​​have the same absolute mean and are the smallest, the smaller L value is taken. The SAVI vegetation index calculation formula is as follows: Where NIR and R are the reflectances of near-infrared and red light, respectively, and L is the reflectance of red light. * The optimal soil conditioning parameters are as follows.

[0038] To suppress the interference of negative SAVI in vegetated areas, SAVI is normalized, as shown in the following expression: The negative values ​​of SAVI will be truncated to 0, while the remaining values ​​will remain unchanged.

[0039] The green light-red light difference index The calculation is based on the green and red light channels in the geo-registered visible light image, and the calculation formula is as follows: Among them, DN G The green channel response intensity in a visible light image, with a value ranging from 0 to 255, DN R This represents the response intensity of the red channel in a visible light image.

[0040] The normalized calculation formula for the green-red light difference index is as follows: Among them, GRDI inv For the inverted GRDI image, GRDI inv,max and GRDI inv,min GRDI inv The maximum and minimum values ​​of the image.

[0041] S6 is obtained by fusing the normalized SAVI vegetation index and the green light-red light difference index to obtain the enhanced tree characteristic value SG.

[0042] In this invention, the SAVI vegetation index and the green-red light difference index can be multiplied and fused together, i.e. Alternatively, weighted fusion, Manhattan distance fusion, Euclidean fusion, and other fusion methods can be used to combine spectral and brightness information to obtain enhanced tree characteristic values ​​SG.

[0043] The weighted fusion expression is: , where a and b are weight values, and a+b=1.

[0044] Manhattan's distance from the fusion formula is: .

[0045] The Euclidean fusion formula is: .

[0046] S7, divide the data interval containing the SG value into K equal intervals, and calculate the probability density P for each data interval. k and weight W k Then, random sampling with replacement is performed on each data interval.

[0047] To alleviate the sample imbalance problem caused by the small proportion of trees in desert areas, this invention uses SG values ​​as weights for sampling with replacement. The specific steps are as follows: the data interval containing the SG value is equally divided into K data intervals, and the number of pixels in each data interval is counted. (k=1,2,…,K). The recommended value for K is 10-50. Calculate the probability density P for each data interval. k and weight W k .

[0048] in, Let N be the number of pixels in the k-th data interval; N is the total number of pixels in all data intervals. in, Let be the probability density of the j-th data interval.

[0049] In some embodiments, the weight W can be calculated using a logarithmic transformation based on the SG value. i The specific formula is as follows: in, Let be the SG value of the i-th pixel.

[0050] Then perform sampling with replacement. If a random seed is set (e.g., seed=42), perform sampling with replacement from all pixels of the image at an n% ratio (n=10-30%, 20% recommended), with each batch of samples having ≥100 pixels.

[0051] S8, repeat M times, input the M sampling results into the big law method, and obtain M corresponding segmentation thresholds.

[0052] S9, repeated multiple times, helps to reduce the fluctuation range of the segmentation threshold, and then the average of all segmentation thresholds is calculated.

[0053] S10: For the newly acquired desert multispectral image, after processing steps S1 to S6, pixels in the stitched multispectral image with enhanced tree feature value SG greater than or equal to the average value are marked as trees, thereby realizing the extraction and identification of sparse trees in the desert.

[0054] Example 2 For drone image processing of a 30m*30m sample plot in the Mu Us Desert of Inner Mongolia, a DJI Mavic 3 Multispectral Edition industrial drone was used. This drone integrates one visible light camera and four multispectral cameras (green light, red light, red edge, and near-infrared), which can acquire the visible light three-band and red light and near-infrared data required by this invention.

[0055] The parameters for orthophoto acquisition are set as follows: flight altitude 108.4m; ground resolution 5.00cm; forward overlap 70%; lateral overlap 70%.

[0056] Three square standard gray boards with known reflectance (25%, 50%, and 75%) were used for calibration. The recommended size for each board was 20*20cm. The calibration board parameters were tested to ensure uniformity within a range of + / -1.5%, conforming to JJF1232-2009 and the American ASTM standard. The calibration board was photographed at a height of 1m, with the camera angle perpendicular to the board, and the sensor's automatic exposure function was turned off.

[0057] The drone used in this embodiment is equipped with an RTK module, with a fixed solution accuracy of 1cm+1ppm horizontally and 1.5cm+1ppm vertically, so there is no need to deploy additional image control points on the ground for georegistration.

[0058] Multispectral images were stitched together using the professional remote sensing processing software Pix4D, outputting raster data of visible light, red light, and near-infrared light, such as... Figure 2 As shown, the planar error of the stitched image is 2.8cm, which is less than 2 pixels (10cm), indicating that the multi-band data has been spatially aligned.

[0059] The absolute value of the SAVI mean of 100 bare soil pixels interpreted manually |μ L The optimization objective is to minimize L. Based on experience in desert areas, the initial search interval for L is set to [0.1, 1]. The search step size ΔL is set to 0.1 to traverse the search interval. When L equals 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0, |μ L| Corresponding to 0.157, 0.152, 0.147, 0.143, 0.140, 0.138, 0.136, 0.134, 0.132, and 0.130 respectively.

[0060] When L=1, |μ L |There is a minimum value, therefore the optimal L is determined. * The value is set to 1. Therefore, the optimal soil conditioning parameter L is determined. * =1. Let L * Substituting 1 into the SAVI calculation formula, calculate the SAVI value of the entire image, such as... Figure 3 As shown, the numerical range of SAVI is -0.021 to 0.6053.

[0061] Based on the response intensity values ​​of the green and red channels in the visible light image, the GRDI image is obtained by subtracting the red channel from the green channel and then dividing by 255. Figure 3 As shown in (b), the numerical range of GRDI is 0 to 1.

[0062] Normalize the SAVI by truncating negative SAVI to zero while keeping positive values ​​unchanged to obtain the SAVI of the entire image. norm ,like Figure 4 As shown in (a), SAVI norm The numerical range is from 0 to 1.

[0063] Inverting and normalizing the GRDI yields the GRDI. norm Images, such as Figure 4 As shown in (b), GRDI norm The numerical range is from 0 to 1.

[0064] The normalized SAVI (reflecting vegetation spectral dominance) and GRDI (reflecting the difference between green and red light reflectance) are fused by a distance operation through a product operation to obtain the characteristic SG image, such as... Figure 5 As shown in (a).

[0065] The SG value ranges from 0 to 0.9451. This range is divided into 20 equally spaced intervals, and the number of pixels N in each interval is counted. k (k=1,2,…,K), calculate the probability density P for each interval. k (The proportion of pixels in this interval to the total number of pixels N), calculate the weight W within each group. k The calculation results are shown in the table below: Random seeds are set sequentially (seed=1,2,…,30), and 60,000 pixels (approximately 20%) are extracted from the SG image. The SG values ​​of 30 batches (M=30) are input into the Otsu method to obtain the corresponding segmentation threshold T.i (i=1,2,…,30), summarizing all threshold results, the minimum threshold value is 0.3119, the maximum value is 0.3158, and the average value is 0.3147, i.e., T avg =0.3147.

[0066] Binarization segmentation is used to extract sparse trees. Based on the final threshold T... avg Image SG value ≥ T avg The pixels are labeled as trees, and the rest are background (value = 0), resulting in a masked image of sparse trees in a desert, such as... Figure 5 As shown in (b). Finally, at 900m... 2 The area of ​​trees within the quadrat is 111.01 m². 2 It accounts for 12.33% of the total sample area.

[0067] In a test set consisting of 100 bare soil areas and 50 tree areas marked manually, the accuracy of this method was 98.6%, which exceeded that of the Otsu method using the SAVI single index (98.0%) and the Otsu method using the GDRI single index (97.33%).

Claims

1. A method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography, characterized in that, Includes the following steps: S1 uses drone aerial photography equipped with a multispectral sensor to obtain multispectral images of the desert to be analyzed; S2, Perform radiometric correction on the multispectral image to obtain the corrected multispectral image; S3, georegist the corrected multispectral image; S4 uses professional remote sensing processing software to stitch together the corrected multispectral images and output raster data; S5. Based on the raster data, calculate the SAVI vegetation index and the green light-red light difference index respectively and perform normalization processing. S6, the normalized SAVI vegetation index and the green light-red light difference index are combined to obtain the enhanced tree characteristic value SG; S7, weighted random sampling with replacement based on SG value; S8, repeat M times, input the M sampling results into the big law method, and obtain M corresponding segmentation thresholds; S9, calculate the average of all segmentation thresholds; S10, pixels in the stitched multispectral image whose enhanced tree feature value SG is greater than or equal to the average value are marked as trees.

2. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography as described in claim 1, characterized in that: The multispectral sensor includes a near-infrared sensor with a wavelength range of 750-900nm, a red light sensor with a wavelength range of 630-680nm, and a visible light sensor in the blue, green, and red bands.

3. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography according to claim 1, characterized in that: The radiometric correction method includes: performing radiometric correction on multispectral images using a linear regression model established by the reflectance of a calibration plate and the digital quantization value output by a multispectral sensor; automatically performing radiometric correction using a radiometric calibration module built into the UAV; and performing radiometric correction using the dark pixel method.

4. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography according to claim 1, characterized in that: The grid data includes visible light grid data in the blue, green, and red bands, reflectance grid data in the near-infrared band, and reflectance grid data in the red band.

5. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography according to claim 1, characterized in that: The stitching should meet the requirement that the planar error of the stitched image is ≤2 pixels to ensure spatial alignment of multi-band data.

6. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography according to claim 1, characterized in that: The SAVI vegetation index calculation process specifically includes: selecting bare soil areas without vegetation cover; taking the minimum absolute value of the mean SAVI vegetation index of the bare soil areas as the optimization objective; traversing the possible value range of the soil adjustment parameter L; and determining the optimal soil adjustment parameter L. * ; will L * Substituting into the SAVI vegetation index calculation formula, the SAVI vegetation index of the S4-step mosaic image is obtained; the SAVI vegetation index is normalized by truncating negative SAVI values ​​to 0 and keeping the remaining SAVI values ​​as their original values.

7. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography according to claim 1, characterized in that: The green light-red light difference index DN G DN represents the response intensity of the green channel in a visible light image. R The response intensity of the red channel in the visible light image is given; the normalization of the green-red light difference index adopts the inverse normalization method.

8. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography according to claim 1, characterized in that: The SAVI vegetation index and green-red light difference index fusion methods include SAVI vegetation index and green-red light difference index multiplication fusion, weighted fusion, Manhattan distance fusion, or Euclidean fusion.

9. The method for extracting sparse trees in deserts based on the Otsu method of feature fusion and multispectral aerial photography according to claim 1, characterized in that, In step S7, the weighted random sampling with replacement based on the SG value specifically involves: dividing the data interval containing the SG value into K equally spaced data intervals, and calculating the probability density of each data interval. and weight Then, weighted random sampling with replacement is performed on each data interval; the probability density of the data interval is... , The number of pixels in the Kth data interval; N is the total number of pixels in all data intervals; the weight of the data interval. , Let be the probability density of the j-th data interval.

10. In the desert sparse tree extraction method based on Otsu's method and multispectral aerial photography according to claim 1, step S7, the weighted random sampling with replacement based on the SG value specifically involves: calculating the weights based on the logarithmic transformation of the SG value. Weighted random sampling with replacement is performed on the sample pixels; in, Let be the SG value of the i-th pixel.

Images