Agricultural protection unmanned aerial vehicle operation boundary identification method and system

Abstract

The application provides a work boundary identification method and system of a plant protection unmanned aerial vehicle. When an index difference value between a vegetation area and a non-vegetation area of a farmland is lower than a difficult scene threshold of the plant protection unmanned aerial vehicle, a vegetation endmember spectrum and a soil endmember spectrum are extracted from the multispectral image, each pixel in the multispectral image is subjected to spectral mixture decomposition based on the vegetation endmember spectrum and the soil endmember spectrum, and a vegetation endmember abundance of each pixel is obtained. A vegetation abundance map is constructed according to all the vegetation endmember abundances, an original vegetation index map is subjected to abundance enhancement through the vegetation abundance map, and an enhanced vegetation index map is obtained. The enhanced vegetation index map is input into a pre-trained semantic segmentation network model, a binary work area mask is output, and an outline of the work area mask is used as a boundary line coordinate of the work area. Based on the above scheme, spectral recognition of a crop boundary in a low-contrast farmland scene can be realized.

Description

Technical Field

[0001] This application relates to the field of agricultural drone technology, and more specifically, to a method and system for identifying the operational boundaries of agricultural drones. Background Technology

[0002] Agricultural drones are unmanned aerial vehicles used for the prevention and control of agricultural pests and diseases and farmland management. The equipment is equipped with multispectral imaging sensors, positioning modules and spraying systems, and flies autonomously along preset routes. Its main functions include monitoring pests and diseases in farmland crops, analyzing crop growth and carrying out precise spraying operations. It enables efficient inspection and plant protection operations of large-scale farmland from an aerial perspective.

[0003] Current agricultural drone boundary recognition primarily uses the Normalized Difference Vegetation Index (NDVI) to distinguish between crop and non-crop areas. Under normal farmland conditions with vigorous crop growth and complete cover, healthy crops exhibit high reflectivity in the near-infrared band, while bare soil shows relatively high reflectivity in the red band, resulting in significant spectral differences. Using fixed thresholds or Otsu adaptive thresholds for segmentation can yield clear operational boundaries. However, in scenarios such as before fully unfolded leaves in seedling stages, decreased red light absorption due to chlorophyll degradation during disease periods, reduced near-infrared reflectivity due to decreased leaf water content during drought stress, and simultaneous increases in red and near-infrared reflectivity after rainfall, the spectral reflectivity characteristics of crops and background tend to be similar. The vegetation index histograms exhibit a unimodal distribution or severe overlap, leading to numerous false positives and false negatives with fixed threshold segmentation. Dynamic thresholding also struggles to locate effective segmentation points due to excessively small inter-class variance. Therefore, achieving spectral recognition of crop boundaries in low-contrast farmland scenarios to improve the robustness of boundary detection in autonomous drone operations has become a significant challenge for the industry. Summary of the Invention

[0004] This application provides a method and system for identifying the operational boundary of an agricultural drone, which can realize spectral identification of crop boundaries in low-contrast farmland scenes, thereby improving the robustness of boundary detection for autonomous drone operations.

[0005] In a first aspect, this application provides a method for identifying the operational boundary of an agricultural drone, including: Collect downward-view multispectral images of farmland at the current location of the drone, calculate the vegetation index map of the multispectral image, and statistically analyze the index difference values ​​between vegetated and non-vegetated areas in the vegetation index map. When the index difference value is lower than the difficult scenario threshold of the agricultural drone, vegetation endmember spectrum and soil endmember spectrum are extracted from the multispectral image. Based on the vegetation endmember spectrum and the soil endmember spectrum, spectral mixing decomposition is performed on each pixel in the multispectral image to obtain the vegetation endmember abundance of each pixel. A vegetation abundance map is constructed based on the abundance of all vegetation endmembers. The original vegetation index map is then enhanced using the vegetation abundance map to obtain an enhanced vegetation index map. The enhanced vegetation index map is input into a pre-trained semantic segmentation network model, which outputs a binarized job area mask. The contour of the job area mask is used as the boundary coordinates of the job area.

[0006] In some embodiments, calculating the vegetation index map of the multispectral image specifically includes: Extract the spectral reflectance data of the near-infrared band and the spectral reflectance data of the red band from the multispectral image; The normalized differential vegetation index value was calculated pixel by pixel using spectral reflectance data in the near-infrared and red bands. All normalized differential vegetation index values ​​are combined into a vegetation index map of the same size as the multispectral image.

[0007] In some embodiments, statistically analyzing the index difference between vegetated and non-vegetated areas in the vegetation index map specifically includes: In the vegetation index map, pixels with index values ​​higher than a first preset threshold are marked as vegetation candidate areas; Pixels with an index value lower than a second preset threshold are marked as non-vegetation candidate regions, wherein the second preset threshold is less than or equal to the first preset threshold. The vegetation index of all pixels in the vegetation candidate area and the vegetation index of all pixels in the non-vegetation candidate area are calculated respectively, and then the index difference value between the vegetation area and the non-vegetation area is obtained.

[0008] In some embodiments, extracting vegetation endmember spectra and soil endmember spectra from the multispectral image specifically includes: The multispectral image is dimensionality reduced by projecting the original multiband data into a two-dimensional feature space. In the dimensionality-reduced feature space, the pixels at extreme positions are located and used as candidate endmembers. The extreme points with high vegetation index correspond to vegetation endmember candidates, and the extreme points with low vegetation index correspond to soil endmember candidates. The spectral reflectance values ​​of the candidate endmembers are back-projected back into the original multispectral band space to obtain the complete vegetation endmember spectral vector and soil endmember spectral vector.

[0009] In some embodiments, the spectral mixing decomposition of each pixel in the multispectral image based on the vegetation endmember spectrum and the soil endmember spectrum to obtain the vegetation endmember abundance of each pixel specifically includes: The multispectral reflectance vector of each pixel is represented as a linear combination of vegetation endmember spectra and soil endmember spectra, resulting in two endmember components. By setting the sum of the abundance of the two endmember components as a normalization constraint, the proportion coefficient of the vegetation endmember spectrum in each pixel is fitted and solved to obtain the vegetation endmember abundance of each pixel.

[0010] In some embodiments, constructing a vegetation abundance map based on the abundance of all vegetation endmembers specifically includes: Create a blank abundance map matrix with the same spatial resolution as the multispectral image; The vegetation endmember abundance value of each pixel is sequentially filled into the corresponding pixel position in the blank abundance map matrix to obtain the abundance map; Spatial smoothing filtering is applied to the abundance map to output a vegetation abundance map.

[0011] In some embodiments, enhancing the original vegetation index map using the vegetation abundance map to obtain an enhanced vegetation index map specifically includes: Align the original vegetation index map and the vegetation abundance map pixel by pixel; Using the vegetation abundance value of each pixel in the vegetation abundance map as a weighting coefficient, the vegetation index of the corresponding pixel is weighted and adjusted to obtain an enhanced vegetation index map.

[0012] Secondly, this application provides an operational boundary recognition system for agricultural drones, comprising: The acquisition module is used to acquire multispectral images of farmland from the current location of the UAV, calculate the vegetation index map of the multispectral image, and count the index difference values ​​between vegetated areas and non-vegetated areas in the vegetation index map. The processing module is used to extract vegetation endmember spectra and soil endmember spectra from the multispectral image when the index difference value is lower than the difficult scenario threshold of the agricultural drone, and to perform spectral mixing decomposition on each pixel in the multispectral image based on the vegetation endmember spectra and the soil endmember spectra to obtain the vegetation endmember abundance of each pixel. The processing module is also used to construct a vegetation abundance map based on the abundance of all vegetation endmembers, and to enhance the abundance of the original vegetation index map through the vegetation abundance map to obtain an enhanced vegetation index map. The execution module is used to input the enhanced vegetation index map into a pre-trained semantic segmentation network model, output a binarized operation area mask, and use the contour of the operation area mask as the boundary coordinates of the operation area.

[0013] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for calling and running the computer program from the memory, so that the computer device executes the above-described method for identifying the operational boundary of an agricultural drone.

[0014] Fourthly, this application provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to implement the above-mentioned method for identifying the operational boundaries of agricultural drones.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: The method and system for identifying the operational boundary of a plant protection drone provided in this application involves: acquiring a downward-looking multispectral image of farmland at the current location of the drone; calculating a vegetation index map of the multispectral image; and statistically analyzing the index difference value between vegetated and non-vegetated areas in the vegetation index map. When the index difference value is lower than the hard scene threshold of the plant protection drone, vegetation endmember spectra and soil endmember spectra are extracted from the multispectral image. Based on the vegetation endmember spectra and soil endmember spectra, spectral mixing decomposition is performed on each pixel in the multispectral image to obtain the vegetation endmember abundance of each pixel. A vegetation abundance map is constructed based on all vegetation endmember abundance values. The original vegetation index map is then enhanced using the vegetation abundance map to obtain an enhanced vegetation index map. The enhanced vegetation index map is input into a pre-trained semantic segmentation network model, which outputs a binarized operational area mask. The contour of the operational area mask is used as the boundary line coordinates of the operational area.

[0016] Therefore, in this application, the enhanced vegetation index map is input into a pre-trained semantic segmentation network model, which outputs a binarized working area mask. The contour of the working area mask is used as the boundary coordinates of the working area. First, by determining the vegetation endmember abundance, the precise proportion of vegetation components in each pixel can be obtained. The vegetation endmember abundance decomposes the mixed pixels into pure vegetation and pure soil through a linear spectral mixing model, directly quantifying the actual proportion of vegetation in the land cover composition. This removes the interference of soil background on the vegetation signal, making the crop and background pixels that originally overlapped in the vegetation index space significantly separated in the abundance space. Even if the spectral curves of crops and soil almost overlap in the original multispectral image, the abundance map obtained based on endmember decomposition can still clearly show the spatial distribution of crops, improving the problem of recognition failure in low-contrast scenes. Then, by determining the enhanced vegetation index map, the abundance-weighted... The modulated composite feature image, while preserving the physical meaning of the original vegetation index, introduces sub-pixel-level compositional information to expand inter-class separability. The enhancement process adaptively adjusts the original index using vegetation abundance values ​​as weighting coefficients. Regions with high abundance receive positive gain, further increasing the crop index, while regions with low abundance receive negative suppression, further reducing the background index. This widens the feature distance between the operational and non-operational areas. The enhanced vegetation index image shows significantly improved crop boundary contrast and sharper pixel-level classification boundaries. Even with limited training samples, the semantic segmentation network model can stably output a complete operational area mask, thus avoiding false negatives and missed detections due to blurred boundaries, comprehensively improving detection robustness in complex farmland environments. In summary, based on the above scheme, spectral recognition of crop boundaries in low-contrast farmland scenes can be achieved, thereby improving the boundary detection robustness of autonomous UAV operations. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on all the drawings without creative effort.

[0018] Figure 1 This is an exemplary flowchart of a method for identifying the operational boundary of an agricultural drone according to some embodiments of this application; Figure 2 This is a flowchart illustrating the process of determining a vegetation abundance map according to some embodiments of this application; Figure 3 This is a schematic diagram of the operational boundary identification system for agricultural drones according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a computer device for implementing a method for identifying the operational boundary of an agricultural drone, according to some embodiments of this application. Detailed Implementation

[0019] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] refer to Figure 1 The figure is an exemplary flowchart of a method for identifying the operational boundary of a plant protection drone according to some embodiments of this application. The method for identifying the operational boundary of a plant protection drone mainly includes the following steps: In step 101, a downward-looking multispectral image of farmland at the current location of the UAV is acquired, the vegetation index map of the multispectral image is calculated, and the index difference value between the vegetated area and the non-vegetated area in the vegetation index map is statistically analyzed.

[0021] It should be noted that in this application, the multispectral image is a two-dimensional digital image characterizing the distribution of reflection intensity of the farmland area directly below the UAV in different spectral bands; the current position is the spatial coordinate point of the UAV in the global positioning system or navigation coordinate system used to determine the time of image acquisition; and the UAV is an aerial mobile platform used to carry a multispectral imaging sensor and fly along a preset route to acquire farmland images.

[0022] In practice, during the autonomous operation of the agricultural drone, the onboard GPS module obtains the drone's current spatial coordinates in real time, which is its current location. A multispectral imaging sensor mounted on the drone, facing vertically downwards towards the ground, triggers an image acquisition action at the current location. At the same exposure time, the sensor records the reflected light intensity of the farmland area in the near-infrared and red light bands. The near-infrared band corresponds to the wavelength range of 760 nm to 900 nm, and the red light band corresponds to the wavelength range of 630 nm to 690 nm. The sensor converts the received analog light signals in each band into digital quantized values, generating a near-infrared grayscale image and a red light grayscale image. These two images have the same number of rows and columns, and the spatial position of each pixel corresponds one-to-one. The two single-band grayscale images are then superimposed in band order into a single multispectral image file. Each pixel in this file contains both near-infrared and red light reflectance values.

[0023] In some embodiments, calculating the vegetation index map of the multispectral image can be achieved using the following steps: Extract the spectral reflectance data of the near-infrared band and the spectral reflectance data of the red band from the multispectral image; The normalized differential vegetation index value was calculated pixel by pixel using spectral reflectance data in the near-infrared and red bands. All normalized differential vegetation index values ​​are combined into a vegetation index map of the same size as the multispectral image.

[0024] It should be noted that, in this application, the spectral reflectance data in the near-infrared band is a set of values ​​used to represent the ability of green vegetation in farmland to reflect incident light in the near-infrared band; the spectral reflectance data in the red band is a set of values ​​used to represent the absorption and reflection characteristics of vegetation and soil in farmland to incident light in the red band; the normalized differential vegetation index value is a dimensionless value that quantifies the vegetation growth density and vigor at a single pixel location; and the vegetation index map is a two-dimensional grayscale image used to visually display the distribution of vegetation index at each pixel in the entire multispectral image.

[0025] In specific implementation, firstly, extracting the near-infrared and red band spectral reflectance data from the multispectral image can be achieved as follows: the multispectral image file stores two layers of data in band order; the first layer is the original recorded value of the near-infrared band, and the second layer is the original recorded value of the red band. After processing begins, the digital quantization values ​​of all pixel positions in the near-infrared band in the multispectral image are read, and all values ​​are arranged into a two-dimensional array according to the original row and column order. Each value in this array represents the sensor response intensity of the corresponding ground location in the near-infrared band, and this set of values ​​is the near-infrared band spectral reflectance data. The digital quantization values ​​of all pixel locations in the red band of the multispectral image are read in the exact same way and arranged into a two-dimensional array of the same size as the near-infrared data. This set of values ​​represents the spectral reflectance data in the red band. The number of rows and columns in both sets of data are exactly the same, and values ​​with the same row and column indices correspond to the same pixel location in the multispectral image. Then, the normalized differential vegetation index value can be calculated pixel by pixel using the spectral reflectance data in the near-infrared and red bands. This can be achieved as follows: For the two sets of extracted spectral reflectance data, starting from the first pixel location, each pixel is processed one by one. At each pixel location, the pixel pair is extracted. The normalized difference vegetation index (NDVI) is calculated using the corresponding near-infrared and red band values ​​as follows: First, subtract the red band value from the near-infrared value to obtain the numerator; then, add the red band value to the near-infrared value to obtain the denominator; finally, divide the numerator by the denominator, and the quotient is the NDVI value for that pixel location. It should be noted that when the denominator is zero, the NDVI value for that pixel is set to zero. Following this rule, the calculation is performed sequentially for all pixel locations, resulting in a NDVI value between -1 and 1 for each pixel. After processing one pixel, the process moves to the next pixel until the entire image is processed. Once all pixels of the image have been calculated, the calculated value for each pixel position is used as the normalized differential vegetation index value corresponding to that position. Finally, combining all the normalized differential vegetation index values ​​into a vegetation index map of the same size as the multispectral image can be achieved in the following way: a new blank two-dimensional array is created, with the number of rows and columns equal to the number of rows and columns of the multispectral image. According to the spatial arrangement order of pixels in the original multispectral image, each normalized differential vegetation index value is sequentially filled into the corresponding row and column positions of the blank array. After filling, each value in the two-dimensional array represents the vegetation index level of the corresponding pixel in the original image.To facilitate subsequent processing, the values ​​in this two-dimensional array are linearly mapped to the integer range of 0 to 255. After mapping, each value corresponds to a gray level, thus forming a grayscale image. In this grayscale image, areas with higher gray values ​​represent vigorous vegetation growth, while areas with lower gray values ​​represent bare soil or sparse vegetation. This grayscale image can be used as a vegetation index map.

[0026] In some embodiments, the statistical difference in index values ​​between vegetated and non-vegetated areas in the vegetation index map can be achieved by the following steps: In the vegetation index map, pixels with index values ​​higher than a first preset threshold are marked as vegetation candidate areas; Pixels with an index value lower than a second preset threshold are marked as non-vegetation candidate regions, wherein the second preset threshold is less than or equal to the first preset threshold. The vegetation index of all pixels in the vegetation candidate area and the vegetation index of all pixels in the non-vegetation candidate area are calculated respectively, and then the index difference value between the vegetation area and the non-vegetation area is obtained.

[0027] It should be noted that, in this application, the vegetation candidate region is a set of pixel locations used to initially mark the vegetation index map as belonging to crops or green vegetation; the non-vegetation candidate region is a set of pixel locations used to initially mark the vegetation index map as belonging to non-vegetation features; and the index difference value is a single value that quantifies the overall difference in vegetation index between vegetation regions and non-vegetation regions.

[0028] In specific implementation, firstly, in the vegetation index map, marking pixels with index values ​​higher than a first preset threshold as vegetation candidate regions can be achieved in the following way: Obtain a pre-set first preset threshold, which is a value between -1 and 1, specifically determined based on the vegetation index range of standard crops in farmland, with a default setting of 0.3. Iterate through each pixel in the vegetation index map, read the vegetation index value of that pixel, and compare this value with the first preset threshold. If the value is greater than the first preset threshold, then the pixel is determined to belong to a vegetation candidate region and added to a marker array of the same size as the vegetation index map. The location corresponding to the pixel is recorded as the first marker value. If the value is less than or equal to the first preset threshold, it is not recorded in the marker array or is recorded as another value. After traversing all pixels, the locations of all pixels marked with the first marker value constitute the vegetation candidate area. Then, pixels with an index value lower than the second preset threshold are marked as non-vegetation candidate areas. The second preset threshold being less than or equal to the first preset threshold can be implemented in the following way: obtain the pre-set second preset threshold, which is also a value between -1 and 1, and its value is less than or equal to the first preset threshold. It can be set to 0 by default.1. Iterate through each pixel in the vegetation index map, read the vegetation index value of that pixel, and compare it with a second preset threshold. If the value is less than the second preset threshold, the pixel is determined to belong to a non-vegetation candidate region, and its position is recorded as a second label value in the same label array used in the first step. The second label value is different from the first label value. If the value is greater than or equal to the second preset threshold, it is not recorded as a non-vegetation candidate region in the label array. It should be noted that when the vegetation index value of a pixel is between the second preset threshold and the first preset threshold, the pixel is neither marked as a vegetation candidate region nor a non-vegetation candidate region, and is reserved as a transition region and not included in subsequent calculations. After iterating through all pixels, the positions of all pixels marked with the second label value constitute the non-vegetation candidate region. Finally, calculate the vegetation index of all pixels in the vegetation candidate region and the vegetation index of all pixels in the non-vegetation candidate region, thereby obtaining the vegetation region and the non-vegetation region. The index difference between regions can be achieved as follows: Find all pixel positions marked with the first label value in the label array. Read the corresponding vegetation index values ​​from the vegetation index map based on all positions. Summate all read values ​​to obtain a total. Simultaneously, count the total number of pixels within the vegetation candidate region. Divide the total by the total number of pixels to obtain the average vegetation index of the vegetation candidate region. This average represents the overall index level of the vegetation region. Similarly, find all pixel positions marked with the second label value in the label array. Read the corresponding vegetation index values ​​from the vegetation index map based on all positions. Summate these values ​​to obtain the total for the non-vegetation candidate region. Count the total number of pixels in this region. Divide the total by the total number of pixels to obtain the average vegetation index of the non-vegetation candidate region. This average represents the overall index level of the non-vegetation region. Subtract the average vegetation index of the non-vegetation candidate region from the average vegetation index of the vegetation candidate region. Use the difference as the index difference between the vegetation region and the non-vegetation region.

[0029] In step 102, when the index difference value is lower than the difficult scenario threshold of the plant protection drone, vegetation endmember spectra and soil endmember spectra are extracted from the multispectral image. Based on the vegetation endmember spectra and the soil endmember spectra, spectral mixing decomposition is performed on each pixel in the multispectral image to obtain the vegetation endmember abundance of each pixel.

[0030] It should be noted that in this application, when the index difference value is lower than the threshold of the difficult scenario for agricultural drones, and the spectral characteristics of crops and background soil in the farmland environment are very similar and difficult to distinguish, under normal farmland operation conditions, healthy green vegetation has high reflectivity in the near-infrared band and low reflectivity in the red band, while bare soil has a small difference in reflectivity between the two bands. Therefore, the calculated normalized difference vegetation index value will show a significant difference. However, under specific conditions, this difference will be significantly reduced. For example, in the seedling stage of rice or wheat, the crop leaves have not yet fully covered the ground, and moist soil or dark soil rich in organic matter will absorb more near-infrared light, causing the vegetation index of crops to decrease while the vegetation index of soil increases. In addition, when crops suffer from diseases, drought, or lodging, their chlorophyll content decreases, their red light absorption capacity weakens, and their near-infrared reflectivity decreases, which will also make the vegetation index value close to that of bare soil. In such challenging scenarios, directly using conventional thresholding methods to extract the crop boundary can lead to broken, misaligned, or noisy boundary lines, failing to form a complete closed region. Therefore, it is necessary to identify this low-contrast state and trigger subsequent spectral unmixing and enhancement processes to recover the submerged crop boundary features by extracting more refined vegetation abundance information. The specific threshold value for this challenging scenario is usually calibrated based on standard farmland measurement data, with a default setting of 0.15. When the average index difference between vegetated and non-vegetated areas is lower than this value, it is considered a challenging scenario.

[0031] In some embodiments, extracting vegetation endmember spectra and soil endmember spectra from the multispectral image can be achieved using the following steps: The multispectral image is dimensionality reduced by projecting the original multiband data into a two-dimensional feature space. In the dimensionality-reduced feature space, the pixels at extreme positions are located and used as candidate endmembers. The extreme points with high vegetation index correspond to vegetation endmember candidates, and the extreme points with low vegetation index correspond to soil endmember candidates. The spectral reflectance values ​​of the candidate endmembers are back-projected back into the original multispectral band space to obtain the complete vegetation endmember spectral vector and soil endmember spectral vector.

[0032] It should be noted that, in this application, the two-dimensional feature space is a low-dimensional projection space used to display multi-band spectral data in planar coordinate form after dimensionality reduction; candidate endmembers are pixels at extreme locations representing the spectral characteristics of pure ground features; vegetation endmember candidates are pixels initially identified as having the spectral characteristics of standard green vegetation; soil endmember candidates are pixels initially identified as having the spectral characteristics of standard bare soil; the vegetation endmember spectral vector is a complete sequence representing the reflectance values ​​of pure green vegetation in each multispectral band; and the soil endmember spectral vector is a complete sequence representing the reflectance values ​​of pure bare soil in each multispectral band.

[0033] In specific implementation, firstly, the multispectral image is subjected to dimensionality reduction processing, projecting the original multiband data into a two-dimensional feature space. In the dimensionality-reduced feature space, pixels at extreme positions are located and used as candidate endmembers. Extreme points with high vegetation indices correspond to vegetation endmember candidates, and extreme points with low vegetation indices correspond to soil endmember candidates. This can be achieved as follows: each pixel in the multispectral image contains reflectance values ​​from multiple bands, and all reflectance values ​​constitute a high-dimensional vector. To facilitate the search for spectral extreme points, principal component analysis can be used to reduce the dimensionality of this high-dimensional vector. Specifically, the multiband data of all pixels in the entire multispectral image are arranged into a two-dimensional table, with each row corresponding to a pixel and each column corresponding to a spectral band. The covariance matrix of all spectral band data is calculated, and the... The eigenvalues ​​and eigenvectors of the covariance matrix are determined. The two eigenvectors with the largest eigenvalues ​​are selected, and the multi-band data of each original pixel is multiplied by these two eigenvectors to obtain two new values, which are used as the x-coordinate and y-coordinate of the pixel in the two-dimensional feature space. After the above transformation, all pixels in the original image are mapped onto a planar scatter plot, which is the two-dimensional feature space. In the two-dimensional feature space, each pixel corresponds to a point. Pure ground features are located at the extreme positions of the scatter plot, i.e., the vertices on the convex hull boundary. The method for locating the pixels at the extreme positions is to calculate the convex hull of all points. The convex hull is a polygon formed by connecting the outermost points, and the vertices of the convex hull are the pixels at the extreme positions. Among all the convex hull vertices, they are further filtered according to the vegetation index value. The specific approach is as follows: for each convex hull vertex pixel, find its corresponding vegetation index value in the original multispectral image. Mark the convex hull vertex with the highest vegetation index value as a high vegetation index extreme point, which represents the spectral characteristics of pure green vegetation, and use it as a vegetation endmember candidate. Mark the convex hull vertex with the lowest vegetation index value as a low vegetation index extreme point, which represents the spectral characteristics of pure bare soil or background, and use it as a soil endmember candidate. If there are multiple extreme points with similar vegetation indices, select the point farthest from the convex hull boundary as the final candidate. This will yield the vegetation endmember candidates and soil endmember candidates.

[0034] Then, in the specific implementation, the spectral reflectance values ​​of the candidate endmembers are back-projected back into the original multispectral band space to obtain the complete vegetation endmember spectral vector and soil endmember spectral vector. This can be achieved in the following way: the vegetation endmember candidates and soil endmember candidates located in the two-dimensional feature space correspond to two specific pixels in the original multispectral image. Each candidate pixel originally stores the reflectance values ​​of each band in the original multispectral image, so there is no need to perform complex back-projection calculations. Specifically, based on the row and column indices of the vegetation endmember candidates in the image, the reflectance values ​​of the pixel position in all bands are directly read from the data of the original multispectral image. For a multispectral image that usually contains two bands, near-infrared and red light, the result read is a sequence containing two values. The first value is the reflectance of the near-infrared band, and the second value is the reflectance of the red light band. This sequence of values ​​arranged in band order is the vegetation endmember spectral vector. In the same way, based on the soil endmember spectral vector... The selected row and column indices are used to read the reflectance values ​​of the pixel location across all bands from the original multispectral image, resulting in the soil endmember spectral vector. To ensure the representativeness of the extracted endmember spectra, the two spectral vectors can be verified as follows: The near-infrared band values ​​in the vegetation endmember spectral vector are compared with the red band values; normal green vegetation should have near-infrared values ​​significantly greater than red values. The near-infrared band values ​​in the soil endmember spectral vector are compared with the red band values; standard bare soil should have values ​​that are close and have no magnitude relationship. If the verification passes, both spectral vectors are retained. If the verification fails, for example, if the near-infrared value of a candidate vegetation endmember is less than or equal to the red value, the process returns to the previous step to relocate the next extreme convex hull vertex as a candidate endmember. This yields the complete vegetation and soil endmember spectral vectors. The vegetation endmember spectral vector is used as the spectral reference for pure vegetation, and the soil endmember spectral vector is used as the spectral reference for pure soil.

[0035] In some embodiments, the vegetation endmember abundance of each pixel in the multispectral image is obtained by performing spectral mixing decomposition based on the vegetation endmember spectrum and the soil endmember spectrum, which can be achieved by the following steps: The multispectral reflectance vector of each pixel is represented as a linear combination of vegetation endmember spectra and soil endmember spectra, resulting in two endmember components. By setting the sum of the abundance of the two endmember components as a normalization constraint, the proportion coefficient of the vegetation endmember spectrum in each pixel is fitted and solved to obtain the vegetation endmember abundance of each pixel.

[0036] It should be noted that, in this application, the endmember component is used to represent the multi-band spectral composition contributed by the vegetation part and the soil part in a pixel, respectively; the vegetation endmember abundance is a dimensionless value used to represent the proportion of vegetation component in a pixel.

[0037] In practical implementation, firstly, the multispectral reflectance vector of each pixel is represented as a linear combination of vegetation endmember spectra and soil endmember spectra. The two endmember components can be obtained as follows: For each pixel in the multispectral image, its reflectance values ​​in different bands constitute a multiband reflectance vector. Each pixel's reflectance vector contains two values: the first is the reflectance in the near-infrared band, and the second is the reflectance in the red band. The previous steps have already extracted the vegetation endmember spectral vector and the soil endmember spectral vector, both of which also contain reflectance values ​​in the near-infrared and red bands respectively. The reflectance vector of any pixel in the spectral mixture decomposition can be derived from the vegetation endmember spectral vector and the soil endmember spectral vector. The pixel reflectance is obtained by multiplying each component by its respective scaling factor and then summing the results. Specifically, the pixel reflectance equals the vegetation endmember spectrum multiplied by the vegetation abundance, plus the soil endmember spectrum multiplied by the soil abundance. Here, vegetation abundance represents the proportion of vegetation components in the pixel, and soil abundance represents the proportion of soil components. The two scaling factors are the two endmember components. Following this representation, each pixel corresponds to two unknowns: the vegetation endmember component and the soil endmember component. Then, setting the sum of the abundances of the two endmember components as a normalization constraint, the scaling factor of the vegetation endmember spectrum in each pixel is fitted and solved to obtain the vegetation endmember abundance of each pixel. This can be achieved in the following way: Since each pixel only provides reflectance values ​​for two bands, and it is necessary to calculate... The solution has two unknowns: vegetation endmembers and soil endmembers. Therefore, the equation theoretically has a unique solution. To ensure the physical meaning of the solution is clear, constraints need to be added. Specifically, the sum of the vegetation endmember and soil endmember components is equal to one. This constraint means that the ground cover of a pixel is entirely composed of vegetation and soil, without considering other land cover types, and the sum of the proportions of the two components is 100%. Under the normalization constraint, the specific steps to solve for the proportion coefficient of the vegetation endmember spectrum in each pixel are as follows: The near-infrared reflectance of the pixel is expressed as an equation, where the left side of the equation is the near-infrared reflectance value of the pixel, and the right side is the near-infrared reflectance value in the vegetation endmember spectrum multiplied by the vegetation endmember. The red light reflectance of the pixel is expressed as a second equation, with the left side being the pixel's red light reflectance and the right side being the vegetation end-member reflectance multiplied by the vegetation end-member component, plus the soil end-member reflectance multiplied by the soil end-member component. Using normalization constraints, the soil end-member component is replaced by one minus the vegetation end-member component, and substituted into the two equations above. At this point, both equations contain only the vegetation end-member component as the only unknown. Theoretically, the vegetation end-member component values ​​obtained from the near-infrared equation and the red light equation should be equal, but due to measurement noise and spectral variation, the two values ​​may differ slightly.Therefore, a fitting method can be used to select the vegetation endmember component value that minimizes the sum of squared errors in the two equations as the final solution. The obtained value is the vegetation endmember abundance of that pixel. For pixels whose solution results exceed the range of 0 to 1, their vegetation endmember abundance is restricted to the range: values ​​less than 0 are set to 0, and values ​​greater than 1 are set to 1. The above complete solution process is repeated for each pixel in the multispectral image until all pixels have obtained their respective vegetation endmember abundance values, thus obtaining the vegetation endmember abundance of each pixel.

[0038] In step 103, a vegetation abundance map is constructed based on the abundance of all vegetation endmembers. The original vegetation index map is then enhanced using the vegetation abundance map to obtain an enhanced vegetation index map.

[0039] In some embodiments, a vegetation abundance map is constructed based on the abundance of all vegetation endmembers, with reference to... Figure 2 The diagram described is a flowchart illustrating the process of determining a vegetation abundance map in some embodiments of this application. In this embodiment, the determination of the vegetation abundance map can be achieved using the following steps: In step 1031, a blank abundance map matrix with the same spatial resolution as the multispectral image is created; In step 1032, the vegetation endmember abundance value of each pixel is sequentially filled into the corresponding pixel position in the blank abundance map matrix to obtain the abundance map; In step 1033, the abundance map is spatially smoothed and filtered to output a vegetation abundance map.

[0040] It should be noted that in this application, the blank abundance map matrix is ​​a two-dimensional numerical table used to temporarily store the vegetation endmember abundance values ​​of each pixel; the abundance-filled map is an intermediate image used to record the original vegetation endmember abundance values ​​of each pixel; and the vegetation abundance map is a spatial distribution image used to visually display the proportion of vegetation components in each pixel of the entire image.

[0041] In specific implementation, firstly, a blank abundance map matrix with the same spatial resolution as the multispectral image is created; then, the vegetation endmember abundance value of each pixel is sequentially filled into the corresponding pixel position in the blank abundance map matrix to obtain the abundance-filled map; finally, the abundance-filled map is spatially smoothed and filtered, and the output vegetation abundance map can be implemented in the following way: there may be cases in the abundance-filled map where the abundance value of individual pixels differs too much from that of surrounding pixels. This difference is caused by image noise and spectral unmixing error and does not represent the actual vegetation distribution change. In order to eliminate such isolated outliers, the abundance-filled map is spatially smoothed and filtered. The neighborhood average filtering method can be used: set a square sliding window with an odd number of pixels in height and width. The commonly used window size is three pixels multiplied by three pixels. Align the center of the window with each pixel in the abundance map, read the abundance values ​​of all pixels within the window's coverage area, calculate the arithmetic mean of all values, and replace the original abundance value of the center pixel with this mean. Following a left-to-right, top-to-bottom order, slide the window through each pixel in the entire abundance map. The original value of each pixel is replaced by the mean value of its neighborhood. For pixels at the image edges, their neighborhood windows may extend beyond the image boundary; in this case, only the effective pixels within the image are used to calculate the mean. After the above smoothing filtering, the value of each pixel in the resulting abundance map incorporates information from its surrounding neighborhood; this value is the vegetation abundance value. Isolated noise points are suppressed, while the overall trend of vegetation abundance variation is preserved. The image obtained after spatial smoothing filtering is used as the vegetation abundance map.

[0042] In some embodiments, the enhancement of the original vegetation index map using the vegetation abundance map to obtain an enhanced vegetation index map can be achieved through the following steps: Align the original vegetation index map and the vegetation abundance map pixel by pixel; Using the vegetation abundance value of each pixel in the vegetation abundance map as a weighting coefficient, the vegetation index of the corresponding pixel is weighted and adjusted to obtain an enhanced vegetation index map.

[0043] It should be noted that in this application, the weighting coefficient is a proportional factor used to determine the degree to which the original vegetation index of each pixel is enhanced or suppressed; the enhanced vegetation index map is a vegetation index distribution image used to improve the contrast between crops and background.

[0044] In specific implementation, firstly, the original vegetation index map and the constructed vegetation abundance map are aligned pixel-by-pixel. This can be achieved as follows: the original vegetation index map and the constructed vegetation abundance map have exactly the same number of rows and columns because they both originate from the same original multispectral image, and the spatial dimensions of the images have not been changed during processing. The specific alignment operation is as follows: confirm that the pixels in the first row and first column of both images correspond to the same ground location, and confirm that the pixels in the last column of the first row, the first column of the last row, and the last column of the last row also correspond to the same ground location. Since the spatial resolution of the two images is consistent and the pixel locations naturally correspond, no scaling, rotation, or registration transformation is required. They can be directly matched one-to-one according to the row and column indices. That is, the pixel in the i-th row and j-th column of the vegetation index map describes the same spatial location as the pixel in the i-th row and j-th column of the vegetation abundance map. After alignment, for any pixel location in the image, it is possible to... Simultaneously, the original vegetation index value and the vegetation abundance value at that location are obtained. Then, using the vegetation abundance value of each pixel in the vegetation abundance map as a weighting coefficient, the vegetation index of the corresponding pixel is weighted and adjusted to obtain the enhanced vegetation index map. This can be achieved in the following way: After completing pixel-by-pixel alignment, each pixel position is enhanced separately. For a pixel located in the i-th row and j-th column, the vegetation abundance value at that location is read from the vegetation abundance map, which is between zero and one; the vegetation index value at the same location is read from the original vegetation index map, which is between -1 and 1. The basic rule of weighted adjustment is: using the vegetation abundance value as a weighting coefficient, the original vegetation index is enhanced so that the original vegetation index is increased more at locations with higher vegetation abundance, and suppressed more at locations with lower vegetation abundance. The specific adjustment method is: first, a baseline offset is determined, which is set to 0.5. Subtracting 0.5 from the vegetation abundance value of the pixel yields a difference ranging from -0.5 to +0.5. A positive difference indicates a vegetation abundance value greater than 0.5, a negative difference indicates a vegetation abundance value less than 0.5, and a zero difference indicates a vegetation abundance value equal to 0.5. This difference is then multiplied by a preset enhancement intensity coefficient, a value greater than 0 and less than or equal to 1 (default value is 0.8), used to control the overall enhancement magnitude. This multiplication is then added to the original vegetation index value to obtain the enhanced vegetation index value for that pixel. This calculation process is repeated for every pixel in the entire image. For values ​​exceeding the normal range of the vegetation index (below -1 or above 1), these values ​​are limited to within the range: values ​​below -1 are set to -1, and values ​​above 1 are set to 1. After all pixels have been calculated, the enhanced vegetation index values ​​of each pixel are organized into a new two-dimensional image according to the original row and column order. This image is the enhanced vegetation index map.

[0045] In step 104, the enhanced vegetation index map is input into a pre-trained semantic segmentation network model, and a binarized operation area mask is output. The contour of the operation area mask is used as the boundary coordinates of the operation area.

[0046] It should be noted that the semantic segmentation network model in this application is a deep learning-based image analysis model used to classify each pixel in the input image into a predefined category. The core technical principle of this model is to automatically learn features at different scales and levels in the image through a multi-layer convolutional neural network. Shallow convolutional layers are responsible for extracting local detail features such as edges and corners, while deep convolutional layers are responsible for extracting semantic features such as the overall shape of the object and regional texture. In the encoder part, the size of the feature map is gradually reduced through stride convolution or pooling operations to expand the receptive field, enabling the model to perceive a wider range of contextual information. In the decoder part, the original resolution of the feature map is gradually restored through upsampling or deconvolution operations. At the same time, the high-resolution detail features of the corresponding layer of the encoder are fused with the features of the decoder to compensate for the spatial location information lost during downsampling. After a pixel-level classification layer, a probability distribution belonging to each category is output for each pixel, and the category with the highest probability is taken as the predicted label of the pixel. For the task of task boundary recognition, the model divides pixels into two categories: "task region" and "non-task region". The output result is a classification mask map with the same size as the input image.

[0047] In practice, the enhanced vegetation index map is first fed into a pre-trained semantic segmentation network model. The enhanced vegetation index map is a two-dimensional grayscale image, which the model adjusts to a uniform size required for internal processing (default 320 pixels by 320 pixels). The model performs forward computation layer by layer, extracting high-level semantic features through multiple convolutions and downsampling by the encoder. Then, through multiple upsampling and feature fusion by the decoder, the spatial resolution and classification result are restored pixel by pixel. The model outputs two probability values ​​for each pixel, representing the probability that the pixel belongs to the working region and the probability that it belongs to a non-working region. These two probability values ​​are compared, and the one with the larger probability is taken as the final category of the pixel. If the probability of the working region is greater than the probability of the non-working region, the pixel is assigned a value of one; otherwise, it is assigned a value of zero. After performing this judgment on all pixels, a binary image consisting of zeros and ones is obtained. The set of pixels with a value of one in this image constitutes the overall range of the working region. This binary image is the binarized working region mask. The contour is then extracted from the mask. Specifically, an eight-neighbor boundary tracking algorithm is used to scan the positions in the mask where the values ​​jump from zero to one or from one to zero. The row and column coordinate sequences of all boundary pixels are recorded. All row and column coordinates are arranged according to the original spatial order in the mask to form a closed contour line. The coordinate values ​​of all pixels on the contour line are converted into the real geographical location coordinates in the UAV navigation coordinate system. The output is the boundary line coordinate sequence of the working area, which is used for the subsequent autonomous flight and spraying control of the UAV.

[0048] In another aspect, in some embodiments, this application provides an operational boundary recognition system for agricultural drones, with reference to... Figure 3 The figure is a schematic diagram of the operation boundary recognition system of an agricultural drone according to some embodiments of this application. The operation boundary recognition system of the agricultural drone includes: a data acquisition module 201, a processing module 202, and an execution module 203, which are described below: The acquisition module 201 in this application is mainly used to acquire the farmland downward multispectral image of the current position of the UAV, calculate the vegetation index map of the multispectral image, and count the index difference value between the vegetation area and the non-vegetation area in the vegetation index map. Processing module 202, in this application, is used to extract vegetation endmember spectra and soil endmember spectra from the multispectral image when the index difference value is lower than the difficult scene threshold of the agricultural drone, and to perform spectral mixing decomposition on each pixel in the multispectral image based on the vegetation endmember spectra and the soil endmember spectra to obtain the vegetation endmember abundance of each pixel. It should be noted that the processing module 202 is also used to construct a vegetation abundance map based on the abundance of all vegetation endmembers, and to enhance the abundance of the original vegetation index map through the vegetation abundance map to obtain an enhanced vegetation index map. The execution module 203 in this application is mainly used to input the enhanced vegetation index map into a pre-trained semantic segmentation network model, output a binarized operation area mask, and use the contour of the operation area mask as the boundary line coordinates of the operation area.

[0049] The foregoing has detailed examples of the operational boundary identification method and system for agricultural drones provided in this application. It is understood that the corresponding apparatus, in order to achieve the above functions, includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0050] In some embodiments, this application also provides a computer device, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for calling and running the computer program from the memory, so that the computer device executes the above-described method for identifying the operational boundary of an agricultural drone.

[0051] In some embodiments, reference Figure 4 The dashed lines in the figure indicate that the unit or module is optional. This figure is a schematic diagram of the structure of a computer device for implementing a method for identifying the operational boundary of an agricultural drone according to an embodiment of this application. The method for identifying the operational boundary of an agricultural drone described in the above embodiments can be achieved through… Figure 4 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a memory 302 and at least one communication unit 305. The computer device may be a terminal device, a server or a chip.

[0052] Processor 301 can be a general-purpose processor or a special-purpose processor. For example, processor 301 can be a central processing unit (CPU), which can be used to control computer devices, execute software programs, and process data from software programs. The computer device may also include a communication unit 305 for inputting (receiving) and outputting (transmitting) signals.

[0053] For example, the computer device may be a chip, and the communication unit 305 may be the input and / or output circuit of the chip, or the communication unit 305 may be the communication interface of the chip, which may be a component of a terminal device, network device or other device.

[0054] For example, the computer device may be a terminal device or a server, and the communication unit 305 may be a transceiver of the terminal device or the server, or the communication unit 305 may be a transceiver circuit of the terminal device or the server.

[0055] The computer device may include one or more memories 302 storing a program 304. The program 304 can be executed by a processor 301 to generate instructions 303, causing the processor 301 to execute the method described in the above method embodiments according to the instructions 303. Optionally, the memory 302 may also store data (such as a target audit model). Optionally, the processor 301 may also read data stored in the memory 302, which may be stored at the same storage address as the program 304, or it may be stored at a different storage address than the program 304.

[0056] The processor 301 and memory 302 can be configured separately or integrated together, for example, integrated on the system on chip (SOC) of the terminal device.

[0057] It should be understood that each step of the above method embodiment can be completed by hardware logic circuits or software instructions in the processor 301. The processor 301 can be a CPU, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, such as discrete gate, transistor logic devices, or discrete hardware components.

[0058] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] For example, in some embodiments, this application also provides a computer-readable storage medium storing instructions or code that, when executed on a computer, cause the computer to implement the above-described method for identifying the operational boundaries of agricultural drones.

[0060] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to all embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0061] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if all modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include all modifications and variations.

Claims

1. A method for identifying an operation boundary of a plant protection unmanned aerial vehicle, characterized in that, Includes the following steps: Collect downward-view multispectral images of farmland at the current location of the drone, calculate the vegetation index map of the multispectral image, and statistically analyze the index difference values ​​between vegetated and non-vegetated areas in the vegetation index map. When the index difference value is lower than the difficult scenario threshold of the agricultural drone, vegetation endmember spectrum and soil endmember spectrum are extracted from the multispectral image. Based on the vegetation endmember spectrum and the soil endmember spectrum, spectral mixing decomposition is performed on each pixel in the multispectral image to obtain the vegetation endmember abundance of each pixel. A vegetation abundance map is constructed based on the abundance of all vegetation endmembers. The original vegetation index map is then enhanced using the vegetation abundance map to obtain an enhanced vegetation index map. The enhanced vegetation index map is input into a pre-trained semantic segmentation network model, which outputs a binarized job area mask. The contour of the job area mask is used as the boundary coordinates of the job area.

2. The method as described in claim 1, characterized in that, Calculating the vegetation index map of the multispectral image specifically includes: Extract the spectral reflectance data of the near-infrared band and the spectral reflectance data of the red band from the multispectral image; The normalized differential vegetation index value was calculated pixel by pixel using spectral reflectance data in the near-infrared and red bands. All normalized differential vegetation index values ​​are combined into a vegetation index map of the same size as the multispectral image.

3. The method as described in claim 1, characterized in that, The statistical differences in the vegetation index between vegetated and non-vegetated areas in the vegetation index map specifically include: In the vegetation index map, pixels with index values ​​higher than a first preset threshold are marked as vegetation candidate areas; Pixels with an index value lower than a second preset threshold are marked as non-vegetation candidate regions, wherein the second preset threshold is less than or equal to the first preset threshold. The vegetation index of all pixels in the vegetation candidate area and the vegetation index of all pixels in the non-vegetation candidate area are calculated respectively, and then the index difference value between the vegetation area and the non-vegetation area is obtained.

4. The method as described in claim 1, characterized in that, Extracting vegetation endmember spectra and soil endmember spectra from the multispectral images specifically includes: The multispectral image is dimensionality reduced by projecting the original multiband data into a two-dimensional feature space. In the dimensionality-reduced feature space, the pixels at extreme positions are located and used as candidate endmembers. The extreme points with high vegetation index correspond to vegetation endmember candidates, and the extreme points with low vegetation index correspond to soil endmember candidates. The spectral reflectance values ​​of the candidate endmembers are back-projected back into the original multispectral band space to obtain the complete vegetation endmember spectral vector and soil endmember spectral vector.

5. The method as described in claim 1, characterized in that, Based on the vegetation endmember spectrum and the soil endmember spectrum, spectral mixing decomposition is performed on each pixel in the multispectral image to obtain the vegetation endmember abundance of each pixel, specifically including: The multispectral reflectance vector of each pixel is represented as a linear combination of vegetation endmember spectra and soil endmember spectra, resulting in two endmember components. By setting the sum of the abundance of the two endmember components as a normalization constraint, the proportion coefficient of the vegetation endmember spectrum in each pixel is fitted and solved to obtain the vegetation endmember abundance of each pixel.

6. The method as described in claim 1, characterized in that, Constructing a vegetation abundance map based on all vegetation endmember abundance specifically includes: Create a blank abundance map matrix with the same spatial resolution as the multispectral image; The vegetation endmember abundance value of each pixel is sequentially filled into the corresponding pixel position in the blank abundance map matrix to obtain the abundance map; Spatial smoothing filtering is applied to the abundance map to output a vegetation abundance map.

7. The method as described in claim 1, characterized in that, The enhancement of the original vegetation index map by the vegetation abundance map to obtain the enhanced vegetation index map specifically includes: Align the original vegetation index map and the vegetation abundance map pixel by pixel; Using the vegetation abundance value of each pixel in the vegetation abundance map as a weighting coefficient, the vegetation index of the corresponding pixel is weighted and adjusted to obtain an enhanced vegetation index map.

8. A plant protection drone's operational boundary identification system, characterized in that, include: The acquisition module is used to acquire multispectral images of farmland from the current location of the UAV, calculate the vegetation index map of the multispectral image, and count the index difference values ​​between vegetated areas and non-vegetated areas in the vegetation index map. The processing module is used to extract vegetation endmember spectra and soil endmember spectra from the multispectral image when the index difference value is lower than the difficult scenario threshold of the agricultural drone, and to perform spectral mixing decomposition on each pixel in the multispectral image based on the vegetation endmember spectra and the soil endmember spectra to obtain the vegetation endmember abundance of each pixel. The processing module is also used to construct a vegetation abundance map based on the abundance of all vegetation endmembers, and to enhance the abundance of the original vegetation index map through the vegetation abundance map to obtain an enhanced vegetation index map. The execution module is used to input the enhanced vegetation index map into a pre-trained semantic segmentation network model, output a binarized operation area mask, and use the contour of the operation area mask as the boundary coordinates of the operation area.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to call and run the computer programs from the memory, so that the computer device performs the operation boundary identification method for agricultural drones according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions or code that, when executed on a computer, cause the computer to implement the agricultural drone operation boundary identification method as described in any one of claims 1 to 7.

Images