A strip intercropping soybean early stage of missing plant position unmanned aerial vehicle remote sensing extraction method

Abstract

The application belongs to the technical field of remote sensing identification, and relates to a UAV remote sensing extraction method for early missing seedling positions of strip intercropping soybeans. Visible light images are acquired by using a UAV to generate orthographic images and digital surface models, and RGB-Veg multi-channel data is constructed. In combination with an improved SegNeXt model, accurate identification of soybean seedling pixel patches is realized. Through connected domain analysis and centroid extraction, seedling centroid coordinates are obtained to solve the identification problem. Secondly, the seedling centroid point set is clustered by using K-means clustering to distinguish planting strips, and a least square method is used to fit the seeding rows to complete the numbering and sorting of the seedling centroids, construct a structured database, and convert the identification result into point element data with spatial topological relations to provide data support for missing seedling positioning. Finally, according to the seeding plant distance, a missing seedling critical distance is set as a discrimination threshold, and the missing seedling sections are discriminated row by row to further calculate the missing seedling number and extract the three-dimensional coordinates of the positions of the missing seedlings, so that the rapid positioning of the missing seedlings and the timely decision of soybean reseeding can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing identification, and more particularly to a method for unmanned aerial vehicle remote sensing extraction of early missing seedling positions of strip intercropping soybeans. BACKGROUND

[0002] Soybean emergence rate is an important factor affecting soybean yield. As a dicotyledonous plant, soybean emergence is easily affected by soil conditions and seeding depth, especially after sowing, heavy rainfall leads to high soil humidity, making it difficult for soybean seeds to break through the soil and resulting in missing seedlings, which directly affects the final yield. Therefore, timely identification of emergence and replanting are important links to ensure reasonable planting and improve yield per unit. However, the leaves of early growth soybean seedlings are small, and in the strip intercropping mode, they are mixed with corn plants. The traditional manual survey method is inefficient and cannot guide production in a timely manner to adapt to the management needs of future unmanned farms.

[0003] Although remote sensing technology has been used for crop monitoring, satellite images have a spatial resolution of decimeters, which cannot identify single soybean seedlings at the seedling stage. The use of unmanned aerial vehicle remote sensing for corn, rice and other crops has achieved many results, especially the introduction of deep learning has promoted intelligent monitoring. However, there are still challenges in identifying soybean seedlings, especially in the intercropping environment, such as weak features and poor model cross-regional generalization ability, which restricts the precision field management decision-making of strip intercropping soybeans. SUMMARY

[0004] The present application provides a method for unmanned aerial vehicle remote sensing extraction of early missing seedling positions of strip intercropping soybeans, which realizes the rapid positioning of early missing seedling positions of soybeans and provides support for timely replanting and fine field management.

[0005] A method for unmanned aerial vehicle remote sensing extraction of early missing seedling positions of strip intercropping soybeans, comprising the following steps:

[0006] S1. Use unmanned aerial photography to obtain high-resolution visible light images of the test plot, generate orthophotos and digital surface models through aerial triangulation; construct RGB-Veg multi-channel data, use the constructed RGB-Veg multi-channel data as the input of the SegNeXt model, identify soybean seedling pixel patches through the SegNeXt model and output a binary image, and then obtain the coordinate position of the centroid of each seedling through connected component analysis and centroid extraction;

[0007] S2. Based on the coordinate position of the seedling, use a clustering algorithm to separate the planting strip, use a vertical projection method to determine the number of seeding rows in each planting strip and fit the seeding rows through the least squares method, and sort and number the seedling centroids of the seeding rows to obtain structured seedling centroid spatial distribution data;

[0008] S3. According to the sowing plant spacing, the critical distance of seedling shortage is set as a judgment threshold, the distance between the mass centers of adjacent seedlings is calculated and judged for each sowing row, the seedling shortage section is determined, the number of seedling shortage is calculated, and the three-dimensional coordinates of the seedling shortage position are solved.

[0009] The unmanned aerial vehicle is used to obtain a visible light image and generate an orthographic image and a digital surface model, an RGB-Veg multi-channel data is constructed by histogram equalization processing to strengthen the seedling features, and an improved SegNeXt model is combined to realize accurate identification of soybean seedlings, the seedling coordinates are obtained by connected domain analysis and mass center extraction, and the identification difficulty caused by weak seedling features in intercropping is solved; secondly, the K-means clustering is used to separate the planting zone based on the seedling mass center, and the least square method is used to fit the sowing row, the seedling mass center is sorted by row, and a structured database containing the mass center coordinates, planting zone number and other information is constructed, the two-dimensional identification result is converted into point feature data with spatial topological relationship, and data support is provided for accurate positioning. Finally, according to the sowing plant spacing, the seedling shortage section is determined row by row, and the number and three-dimensional coordinates of the seedling shortage are calculated, so that the rapid positioning of the soybean seedling shortage and the timely decision of the soybean reseeding can be realized.

[0010] Preferably, constructing the RGB-Veg multi-channel data comprises the following steps:

[0011] The visible light vegetation index is calculated by the orthographic image, the contrast of the vegetation index image is enhanced by limiting the contrast adaptive histogram equalization, and the RGB-Veg multi-channel data is synthesized by combining the visible light image and the enhanced visible light index channel.

[0012] Preferably, the SegNeXt model comprises a main network and a decoding head, and the main network comprises a C1 stage, a C2 stage, a C3 stage and a C4 stage.

[0013] The main network is an MSCAN network, which comprises the C1 stage, the C2 stage, the C3 stage and the C4 stage in sequence, wherein the C1 stage adopts an initial convolutional layer to extract shallow features, the C2 stage adopts a dynamic convolutional layer, the C3 stage and the C4 stage both adopt a convolutional layer and a coordinate attention module, the initial feature map is obtained by performing basic feature extraction through the C1 stage, the adaptive feature map is obtained by extracting features based on the initial feature map through the C2 stage, the enhanced feature map is obtained by extracting features based on the adaptive feature map through the C3 stage, the multi-scale enhanced feature map is obtained by extracting features based on the enhanced feature map through the C4 stage, and the multi-scale enhanced feature map is taken as the input of the decoding head.

[0014] The decoding head is a LightHamHead attention module, which performs feature fusion and segmentation on the multi-scale enhanced feature map through an attention mechanism, and outputs a soybean seedling pixel patch recognition binary image.

[0015] Preferably, the total loss function of the SegNeXt model is based on a weighted sum of Dice Loss, Focal Loss and binary cross-entropy loss.

[0016] Preferably, the connected component analysis and centroid extraction comprises the following steps:

[0017] The binary map output by the SegNeXt model is labeled for connected components, and a unique label is assigned to each independent seedling pixel patch. An area threshold is set to morphologically filter the connected components, and the coordinates of each seedling are obtained by centroid extraction after filtering.

[0018] Preferably, the seedling centroid within the seeding row is determined based on the following steps:

[0019] The coordinate position of the seedling centroid is normalized, the elbow rule is used to determine the optimal number of clusters, and the cluster center is obtained by repeating K-means clustering at least twice. After clustering, the soybean seedling centroid coordinates are converted back to the original projection coordinate system through inverse normalization.

[0020] The seedling centroid within each planting zone is subjected to principal component analysis to determine the main direction and perpendicular direction of the seeding row. The number of seeding rows is automatically determined using the perpendicular projection method based on the density of the soybean centroid projection in the perpendicular direction. The seeding rows are separated by clustering, and then the least squares method is used to fit a straight line to each row of seedling points. After fitting, the seedlings are sorted and numbered by row, and the seedling centroid within the seeding row is determined.

[0021] Preferably, the calculation of the number of missing seedlings comprises the following steps:

[0022] The actual plant spacing between adjacent seedlings in each seeding row is calculated. The missing seedling discrimination threshold is set to P times the seeding plant spacing, where P is greater than 1 and less than 2. When the actual plant spacing between adjacent seedlings is greater than the missing seedling discrimination threshold, it is determined that the two seedlings are in a missing section. The number of missing seedlings in the missing section is obtained by rounding the ratio of the actual plant spacing to the seeding plant spacing and subtracting 1.

[0023] Preferably, the three-dimensional coordinates of the missing seedling position are obtained based on the following steps:

[0024] Let the start point coordinate of the missing section be , the end point coordinate be , and the number of missing seedlings be . Divide the line segment between the start point and the end point into equal parts, and calculate the planar coordinates (X, Y) of the jth missing position by linear interpolation. The elevation coordinate (Z) is directly extracted from the DSM generated from the unmanned aerial image at the corresponding position of the aforementioned (X, Y) coordinates.

[0025] The beneficial effects of the present application include:

[0026] The present application uses a UAV to obtain a visible light image and generate an orthophoto and a digital surface model, constructs an RGB-Veg multi-channel data to enhance the characteristics of seedlings, and realizes accurate identification of soybean seedling pixel patches by combining an improved SegNeXt model, and then obtains seedling coordinates through connected domain analysis and centroid extraction, solving the identification problem caused by the concealment and weak characteristics of seedlings in intercropping; secondly, based on the seedling centroid, the planting zone is separated by K-means clustering, and the seeding row is fitted by the least square method, the seedlings are sorted and numbered by row, and a structured database containing information such as planting zone number and centroid coordinates is constructed, converting the two-dimensional identification result into point feature data with spatial topological relationship, providing data support for accurate positioning; finally, according to the seeding plant distance, the missing seedling section is distinguished and the missing seedling number and three-dimensional coordinates are calculated, solving the defect that the traditional method can only count the emergence rate but cannot accurately position.

[0027] The present application realizes the rapid positioning of such soybean missing seedlings in gently sloping or sloping farmland planting through the seedling spatial database and three-dimensional coordinate interpolation, and further assists in efficient soybean reseeding decision-making; the operation mode is improved from the current manual field inspection reseeding to point-to-point accurate delivery, forming an automatic processing chain of image acquisition, feature enhancement, intelligent identification, data construction and positioning output; avoiding the problems of strong subjectivity and low efficiency of traditional manual field investigation. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0029] Figure 1 The overall flowchart provided for the embodiments of the present application.

[0030] Figure 2 The improved backbone network model structure diagram provided for the embodiments of the present application.

[0031] Figure 3 The model identification result provided for the embodiments of the present application.

[0032] Figure 4 The in-seeding row detection and missing seedling positioning effect local schematic diagram provided for the embodiments of the present application. DETAILED DESCRIPTION

[0033] In order to make the technical problems, technical schemes and beneficial effects to be solved in the present application clearer, the present application will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0034] Referring to Figure 1 As shown in the drawings, a strip intercropping soybean early missing plant position unmanned aerial vehicle remote sensing extraction method comprises the following steps:

[0035] S1. Under stable light conditions, a high-resolution visible light camera is used to collect remote sensing images (original RGB images) of the test plot by using a UAV, and a high-precision digital orthophoto map (DOM) and a digital surface model (DSM) are generated by aerial triangulation and dense matching technology.

[0036] According to the orthophoto, the ExG index is calculated, that is, the green band (G) value in the orthophoto is multiplied by 2, and then the difference between the red band (R) value and the blue band (B) value is subtracted, that is The limited contrast adaptive histogram equalization is used to effectively enhance the local contrast and suppress noise by dividing the ExG image into blocks and performing local histogram enhancement in each block, and then the R, G and B channels are spliced with the ExG vegetation index channel after local histogram enhancement, and finally an RGB-ExG* input data of four channels is constructed.

[0037] Each soybean seedling is visually interpreted and manually labeled using an orthophoto, a binary mask is generated, the data set is divided into a training set, a validation set and a test set, and an improved SegNeXt model is input for training.

[0038] The improved SegNeXt model comprises a backbone network and a decoding head;

[0039] As Figure 2 shown, the backbone network is an MSCAN network, which comprises a C1 stage, a C2 stage, a C3 stage and a C4 stage in sequence, wherein the C1 stage adopts an initial convolutional layer to extract shallow features. The C2 stage adopts a dynamic convolutional layer to improve the adaptability of the model to scale and shape changes and to cope with the uneven growth of field seedlings. The C3 stage and the C4 stage both adopt convolutional layers and coordinate attention modules to enhance the spatial position perception ability of the model. The initial feature map is obtained by basic feature extraction through the C1 stage, the adaptive feature map is obtained by feature extraction based on the initial feature map through the C2 stage, the enhanced feature map is obtained by feature extraction based on the adaptive feature map through the C3 stage, and the multi-scale enhanced feature map is obtained by feature extraction based on the enhanced feature map through the C4 stage, and the multi-scale enhanced feature map is input into the decoding head;

[0040] The decoding head is a LightHamHead attention module, which performs feature fusion and segmentation on the multi-scale enhanced feature map through an attention mechanism, and outputs a soybean seedling pixel patch recognition binary image (as shown in Figure 3 ).

[0041] The loss function of the improved SegNeXt model is as follows: ;

[0042] In the formula: is the Dice loss, is the Focal Loss, is the binary cross-entropy loss, , , is a weight coefficient, and the preferred value is , , ; represents the total loss function.

[0043] Then, a connected domain marking algorithm is used to scan the entire binary image, and a unique label is assigned to each independent seedling pixel patch, and finally a label image is obtained, and each independent group (i.e. an independent seedling) is regarded as a connected domain. The area (i.e. the total number of pixels in the region) of each connected domain is calculated, and an area threshold is set. The area of the corn seedling is much larger than that of the soybean seedling. The area of the patch is deleted, and the influence of the identified corn seedling is filtered out (since the area of the corn seedling at the same period is larger than that of the soybean seedling, an area threshold is added to filter out the misidentified corn influence when calculating the area of the connected domain). For each connected domain remaining after filtering, the two-dimensional moment is calculated. For a binary image with a size of , the th moment of a region is defined as: ;

[0044] In the formula: x: horizontal coordinate of pixel; y: vertical coordinate of pixel; p: power of horizontal coordinate x; q: power of vertical coordinate y; is the pixel value of the image at coordinates ; in a binary image, for a target region, , and for the background, .

[0045] Then, three low-order moments are calculated. The zero-order moment represents the area (total number of pixels) of the connected domain; the first-order moment is the sum of the x-coordinate values of all pixels in the region; and the first-order moment is the sum of the y-coordinate values of all pixels in the region. ​

[0046] Reuse the calculated moments to directly calculate the centroid coordinates of each connected domain : ;

[0047] Store all the calculated centroid coordinates in a list or array, which is a collection of the positions of all the identified soybean seedling centroids in the entire field.

[0048] S2. Based on the obtained collection of soybean seedling centroid positions, perform Z-score standardization on the centroid coordinates to eliminate the dimensional influence of the coordinate values; then use the elbow rule to calculate the within-cluster sum of squares for cluster numbers from 1 to N, find the point with the largest change rate as the best cluster number; use the determined best cluster number to perform K-means clustering on the standardized coordinates, repeat 3 times to avoid local optimum; perform inverse standardization on the coordinates to convert the coordinates obtained through clustering back to the original geographic coordinate system, and obtain the actual spatial positions of various planting areas.

[0049] Perform principal component analysis on the seedling centroids within the seeding row, with the first principal component direction as the main direction of the seeding row and the perpendicular direction for row separation; project the centroids onto the perpendicular direction and automatically determine the number of seeding rows based on the histogram peak value of the projection values; perform K-means clustering on the projection to assign the seedling points to different seeding rows; for each seedling point within a seeding row, use the least squares method to fit a straight line and number the straight lines, and perform in-row sorting and numbering based on the projection position of the seedling point on the fitted straight line.

[0050] The Z-score standardization formula is:

[0051] ;

[0052] The inverse Z-score standardization formula is:

[0053] ;

[0054] In the formula: , is the standardized coordinate, , is the original coordinate, is the standard deviation of all seedling X coordinates, is the standard deviation of all seedling Y coordinates, is the average value of all seedling X coordinates, is the average value of all seedling Y coordinates.

[0055] S3. For the N seedlings in a sorted seeding row, calculate the actual plant spacing between adjacent seedlings ​(i from 1 to N-1);

[0056] According to the planting spacing at the time of sowing and an acceptable tolerance range (such as ), a missing seedling discrimination threshold is set; when the actual spacing , it is considered that there is a missing seedling between the two seedlings;

[0057] Since the actual spacing may fluctuate slightly due to sowing deviation or measurement error, a rounding function is used to calculate the number of missing seedlings in each section that meets : ;

[0058] In the formula: is the rounding function;

[0059] For each identified missing seedling position, its three-dimensional coordinates can be determined by interpolation, obtained through the DSM, as follows:

[0060] Calculate the corresponding parameters of each missing seedling position: ;

[0061] The coordinate calculation of the th missing seedling position ; expanded as: ; ;

[0062] In the formula: X value of the starting coordinate of the missing section, Y value of the starting coordinate of the missing section, X value of the end coordinate, Y value of the end coordinate, number of missing seedlings.

[0063] The elevation coordinate (Z) is directly extracted from the generated high-precision digital surface model (DSM) corresponding to the elevation value at the coordinate.

[0064] Finally, the system will output a list containing the planting strip number, strip number, row number, and three-dimensional coordinates of all missing seedling points. The results can achieve rapid positioning of such soybean missing seedlings in flat or sloping farmland, thereby assisting efficient soybean replanting decisions, which can be directly used to guide automated replanting equipment or manual replanting.

[0065] Embodiment:

[0066] In this embodiment, a certain soybean modern agricultural park in southwest China is selected, which is mainly hilly terrain, belongs to subtropical monsoon climate, has mild climate, obvious seasonal characteristics and abundant rainfall throughout the year. The region has a two-crop farming system, with summer crops and autumn crops, and perennial crops such as rice, corn, wheat, soybeans, rape, and sweet potatoes. The park is a soybean-corn belt intercropping demonstration base. In this embodiment, the selected soybean-corn belt intercropping has a belt width of 2m, and the soybean planting plant distance is 11cm and the row spacing is 40cm.

[0067] Based on the above scenario, the specific implementation steps of this embodiment are as follows:

[0068] An unmanned aerial vehicle is used to carry a visible light camera, with a flight height of 12m, to obtain a visible light image with a ground resolution of 4mm; then Pix4D is used for three-dimensional reconstruction to generate a digital orthophoto map (DOM) and a digital surface model (DSM).

[0069] The ExG vegetation index is calculated, and local histogram equalization enhancement is performed to synthesize ExG and RGB images to construct an RGB-Veg four-channel dataset. In this embodiment, a total of 2664 valid samples are obtained.

[0070] Using the orthophoto map, each soybean seedling is manually labeled using the Labelme tool for visual interpretation, generating a binary mask (the seedling area is 1 and the background is 0) to ensure labeling accuracy.

[0071] The multi-channel input data and the corresponding mask annotation are uniformly cropped to 512x512 pixel samples. The 2664 512x512 pixel samples are divided into training set, validation set and test set according to the ratio of 70%, 20% and 10%. The improved SegNeXt deep learning model is used for training, with an initial learning rate of 6e-5 and an iteration number of 40000. After training, the test set is used for model performance evaluation. The results show that the average accuracy (mAcc) of soybean seedlings is 97.8%. This improved model has excellent recognition accuracy and generalization ability for soybean seedlings in a belt intercropping environment.

[0072] The binary image output by the model is labeled with a connected domain, the centroid coordinates of each seedling are calculated, and the centroid coordinate set of all seedlings is output; according to the seedling centroid, clustering is performed according to the soybean planting zone, the least squares method is used to fit the seeding row, and a seedling centroid spatial distribution database is constructed; the centroid coordinates are subjected to Z-score standardization processing to eliminate the dimensional influence of the coordinate values; the elbow rule is used to calculate the within-cluster sum of squares from 1 to N clusters, and the point with the largest change rate is taken as the optimal cluster number of the planting zone; the optimal cluster number is used to perform K-means clustering on the standardized coordinates, and the clustering is repeated three times to avoid local optimization; the coordinates are converted back to the initial projection coordinate system through reverse standardization processing; principal component analysis is performed on the seedling centroids in the planting zone, the first principal component direction is taken as the main direction of the planting zone, and the vertical direction is used for row separation; the centroid is projected onto the vertical direction, and the number of seeding rows is automatically determined based on the peak value detection of the histogram of the projection value; the projection value is subjected to K-means clustering, and the seedling points are assigned to different seeding rows; according to the projection position of the seedling points on the fitted straight line, in-row sorting and numbering are performed. The row number is marked from west to east according to the X coordinate, and the seedling centroid in the row is sorted from south to north according to the Y coordinate.

[0073] A database is constructed, and a structured data set is output. Each successfully recognized soybean seedling has three key attributes: planting zone number, seeding row number, seeding row in-sequence number, and centroid coordinate (x, y); then, according to the seeding plant spacing, the missing seedling section is determined row by row, and the missing seedling position three-dimensional coordinate is calculated; the actual plant spacing between adjacent two seedlings is calculated (i from 1 to N-1).

[0074] According to the seeding plant spacing (in this example, 11 cm), the missing seedling determination threshold is set cm, when the actual plant spacing , it is determined as a missing seedling section, for example: between the first and second seedlings in the first row, the actual plant spacing is 28.3 cm, and the number of missing seedlings is determined as ;

[0075] For each missing seedling position identified, the two-dimensional coordinate is calculated by linear interpolation, and the elevation is extracted from the DSM.

[0076] For example: the starting point is the first seedling , the ending point is the second seedling , and the number of missing seedlings is . Then:

[0077] The first missing point coordinate is: ; the second missing point coordinate is: ;

[0078] A list containing all the missing seedling points, planting strip numbers, row numbers, serial numbers, and three-dimensional coordinates is generated, and a replanting decision map is output (as shown in Figure 4

[0079] In the embodiment, orthophoto images with a spatial resolution of 4 mm are used as the reference base map, and field seedling distribution information is extracted as the verification data by visual interpretation. The results show that the seedling recognition success rate is 97.8%, and the least square fitting effect of the seeding row is good . In terms of missing seedling quantity detection, the predicted number of missing seedlings is 212, and the actual number of missing seedlings is 204, with a prediction accuracy of 96.2%; in terms of missing seedling location recognition accuracy, the root mean square error of the Euclidean distance between the predicted coordinates and the actual coordinates is 3.8 cm. The above results verify the effectiveness and reliability of the method, and show that the method has the characteristics of high precision, high efficiency, and strong practicality in soybean early missing seedling recognition, and can provide reliable technical support for precision replanting operations.

[0080] The above is merely a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.​

Claims

1. A method for unmanned aerial vehicle remote sensing extraction of early missing plant positions in strip intercropping soybeans, characterized by, The method comprises the following steps: S1. Using unmanned aerial photogrammetry to obtain high-resolution visible light images of the plot to be measured, generating orthophotos and digital surface models through aerial triangulation, constructing RGB-Veg multi-channel data, taking the constructed RGB-Veg multi-channel data as the input of the SegNeXt model, identifying soybean seedling pixel patches through the SegNeXt model and outputting a binary image, and then obtaining the centroid coordinates of each seedling through connected component analysis and centroid extraction; The construction of the RGB-Veg multi-channel data comprises the following steps: According to the orthophoto, the visible light vegetation index is calculated, and then the histogram equalization is used to enhance the contrast of the vegetation index image, and the visible light image and the enhanced visible light vegetation index are synthesized to obtain the RGB-Veg multi-channel data; The SegNeXt model comprises a backbone network and a decoding head; The backbone network is an MSCAN network, which comprises a C1 stage, a C2 stage, a C3 stage and a C4 stage in sequence, wherein the C1 stage adopts an initial convolutional layer to extract shallow features, the C2 stage adopts a dynamic convolutional layer, the C3 stage and the C4 stage both adopt a convolutional layer and a coordinate attention module, the initial feature map is obtained through the basic feature extraction of the C1 stage, the adaptive feature map is obtained through the feature extraction based on the initial feature map in the C2 stage, the enhanced feature map is obtained through the feature extraction based on the adaptive feature map in the C3 stage, the multi-scale enhanced feature map is obtained through the feature extraction based on the enhanced feature map in the C4 stage, and the multi-scale enhanced feature map is taken as the input of the decoding head; The decoding head is a LightHamHead attention module, which performs feature fusion and segmentation on the multi-scale enhanced feature map through an attention mechanism, and outputs a binary image for soybean seedling pixel patch recognition; S2. Clustering the soybean seedling centroids, determining the number of planting zones in the plot and the planting zones where the seedlings are located, determining the number of seeding rows in each planting zone and fitting the seeding rows, determining the seedling centroids contained in each seeding row, and constructing a seedling spatial distribution database; S3. Setting a missing seedling critical distance as a judgment threshold according to the seeding plant distance, calculating and judging the distance between adjacent seedling centroids for each seeding row, calculating the number of missing seedlings, and calculating the three-dimensional coordinates of the missing seedling positions.

2. The unmanned aerial vehicle remote sensing method for extracting early-stage missing plant positions of strip intercropping soybeans according to claim 1, characterized in that, The total loss function of the SegNeXt model is obtained by weighted summation based on Dice Loss, Focal Loss and binary cross-entropy loss.

3. The unmanned aerial vehicle remote sensing method for extracting early-stage missing soybean positions in strip intercropping according to claim 1, characterized in that, The connected component analysis and centroid extraction comprise the following steps: The binary image output by the SegNeXt model is labeled with connected components, and each seedling pixel patch is assigned a unique label; an area threshold is set to perform morphological filtering on the connected components, and then the centroid coordinates of each seedling are obtained through centroid extraction.

4. The unmanned aerial vehicle remote sensing method for extracting early-stage missing soybean positions in strip intercropping according to claim 1, characterized in that, The seedling centroids contained in the seeding row are determined based on the following steps: The coordinate positions of the seedling centroids are normalized, the elbow rule is used to determine the optimal cluster number of the planting zone, the cluster center coordinates are obtained by repeating K-means clustering at least twice, and the coordinates are converted to the original projection coordinate system after the clustering through inverse normalization processing; The principal component analysis is performed on the seedling centroids in the planting zone to determine the main direction and the vertical direction of the seeding row. The seedling centroids are projected onto the vertical direction according to the vertical projection method, and the number of seeding rows is automatically determined by using the peak value detection of the histogram of the projection values. The least square method is used to fit a straight line to each row of seedling points to determine the seeding row to which the seedling centroid belongs, and further determine the seedling centroids contained in the seeding row.

5. The unmanned aerial vehicle remote sensing method for extracting early-stage missing soybean positions in strip intercropping according to claim 1, characterized in that, The calculating the number of missing seedlings comprises the following steps: The actual plant spacing between two adjacent seedlings in each seeding row is calculated, and the missing seedling discrimination threshold is set as P times of the seeding plant spacing, wherein P is greater than 1 and less than 2. When the actual plant spacing between two adjacent seedlings is greater than the missing seedling discrimination threshold, it is determined that the two seedlings are in a missing section. The number of missing seedlings in the missing section is obtained by rounding off the ratio of the actual plant spacing between two adjacent seedlings to the seeding plant spacing and then subtracting 1.

6. The unmanned aerial vehicle remote sensing method for extracting early-stage missing soybean positions in strip intercropping according to claim 1, characterized in that, The three-dimensional coordinates of the missing seedling positions are obtained based on the following steps: According to the start point coordinates, the end point coordinates and the number of missing seedlings of the missing section, the missing seedling positions are evenly distributed between the start point and the end point by using linear interpolation, and the planar coordinates are output. The elevation coordinates in the three-dimensional coordinates are directly extracted from the elevation values of the corresponding positions of the planar coordinates in the DSM generated from the unmanned aerial vehicle image.

Images