Intelligent counting method based on leymus chinensis tillering number
By constructing a tiller connectivity graph using multispectral band fusion and graph segmentation algorithms, the complex topological relationships and biological characteristics in Leymus chinensis tiller counting were solved, achieving high-precision individual tiller identification and counting.
Patent Information
- Application Number
- CN202610123828.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-31
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies cannot effectively handle the complex spatial topology and biological characteristics of Leymus chinensis tillers, resulting in large errors in tiller counting results, especially under conditions of canopy closure and inter-tiller shading, where existing technologies are difficult to solve.
Enhanced feature images are generated by multispectral band fusion. Combined with morphological growth models and graph segmentation algorithms, a tiller connectivity graph is constructed, adhesion is removed, and node affiliation is determined. Skeleton analysis and temporal verification are applied to achieve high-precision tiller counting.
It achieves high-precision separation of tillers under dense growth conditions, reduces dependence on local image features, and improves the stability and accuracy of counting.
Smart Images
Figure CN122048855A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural information technology, specifically to an intelligent counting method based on the number of tillers of Leymus chinensis. Background Technology
[0002] Accurately determining the number of Leymus chinensis tillers is a fundamental requirement in grassland science research and production management. Existing technologies mainly include experience-based manual field counting and automatic identification methods based on visible light images. Manual counting is inefficient, inconsistent, and risks damaging samples. Mainstream image recognition methods typically perform thresholding on color or texture features directly, or use general contour detection and morphological operations to handle adhered regions. These methods may be effective when tillers are sparse, but in real-world scenarios with dense canopies and severe shading and adhesion between tillers, the segmentation results heavily rely on preset parameters, lack objective basis for dividing overlapping areas, and lead to significant segmentation errors in the counting results.
[0003] The current technology falls short in addressing the core issues from the perspective of the physical spatial structure and biological characteristics of tillers. General image segmentation algorithms cannot effectively express and analyze the complex spatial topological relationships between multiple tillering units, making it difficult to accurately separate and assign adhering parts. Skeleton branch point analysis based on purely geometric features cannot distinguish between genuine new tillering points and structural artifacts formed by leaf overlap, shadows, or noise, resulting in a high false positive rate in node identification. Current technology lacks a counting method capable of intelligent analysis and decision-making based on both spatial topological relationships and morphological growth patterns to achieve accurate and robust counting of Leymus chinensis tillers. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent counting method for the number of tillers of Leymus chinensis to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides an intelligent counting method based on the number of tillers of Leymus chinensis, the method comprising: Perform multispectral band fusion of Leymus chinensis canopy images to generate a spectral fusion image with enhanced features; Preliminary contour detection of tiller clusters is performed on the spectral fusion image to obtain a set of detection regions containing potential tiller targets. Based on the morphological growth model of Leymus chinensis tillers, effective tiller clusters are selected. Subpixel-level edge refinement and contour closure operations are performed on the effective tillering clusters to form independent and complete tillering unit boundaries; Based on the spatial topological relationship of tillering units, a tillering connection graph is constructed to analyze the proximity and overlap between tillers. A graph segmentation algorithm is then applied to process the tillering connection graph to remove the adhesion between tillers and assign overlapping areas to different regions. For each independent tillering unit after segmentation, skeleton extraction and branch point analysis are performed to identify the main stem and potential new tillering points of a single tiller. Based on the spatial distribution pattern of new tillering points and the consistency constraint of the growth direction of the main stem, the new tillering points are merged and false processing is performed to confirm the final tillering nodes. The total number of tillering nodes after merging is counted, and the current number of tillers is checked and corrected for temporal consistency by combining the node data of the historical growth cycle. The verified and corrected tiller count results are associated and stored with the corresponding spectral fusion image, tiller unit boundary, and node markers to form a traceable counting record.
[0006] Preferably, the multispectral band fusion of the sheepgrass canopy image specifically includes: Simultaneously acquire visible light images and at least one non-visible light image of the Leymus chinensis canopy. Perform color space decomposition on the visible light image to separate the hue and luminance components. Normalize the reflectance intensity of the non-visible light image to match its intensity range with the luminance component of the visible light image. Weighted superposition of the normalized non-visible light image and the hue component generates a composite feature layer that highlights chlorophyll activity regions and water distribution characteristics. Perform multi-scale fusion of the composite feature layer and the luminance component of the visible light image based on wavelet transform, preserving the feature details of the composite feature layer and the overall structure of the luminance component, and output the final spectral fusion image.
[0007] Preferably, the preliminary contour detection of tillering clusters in the spectral fusion image specifically includes: Adaptive local thresholding is applied to the spectral fusion image to generate a binarized image that reflects the regional contrast. Multi-scale morphological opening operations are performed on the binarized image to eliminate fine noise and smooth cluster boundaries; Extract all connected components from the morphologically processed image, and use each connected component as a preliminary detection region; Calculate a geometric descriptor for each of the preliminary detection regions, the geometric descriptor including region area, perimeter, compactness, and principal axis direction; Based on the preset area threshold range and density threshold, the preliminary detection area is initially filtered to remove areas that are obviously too large, too small, or have extremely irregular shapes, thus obtaining the set of detection areas.
[0008] Preferably, the step of selecting effective tiller clusters based on the morphological growth model of Leymus chinensis tillers specifically includes: Load a predefined Leymus chinensis tiller morphological growth model, which describes the typical aspect ratio range, outline symmetry characteristics, and texture orientation pattern of tiller clusters at different growth stages. For each region in the set of detection regions, a sub-image is cropped from its corresponding original spectral fusion image. The local gradient direction histogram of the sub-image is analyzed, and its texture dominant direction is extracted. The geometric descriptor of the region and the texture dominant direction are input into the morphological growth model for matching degree calculation. A matching degree threshold is set, and regions with a matching degree higher than the threshold are determined as candidate effective tillering clusters that meet biological characteristics. Perform contour expansion based on region growing on the candidate effective tiller clusters to ensure that the contour completely covers the boundary of the tiller clusters, and finally output the effective tiller clusters; The specific steps for constructing the morphological growth model include: Multispectral images of Leymus chinensis tillers at different growth stages were collected. The multispectral images were manually annotated to identify the actual tiller cluster boundaries, main stem positions, and tiller nodes. Morphological feature parameters of each tiller cluster were extracted from the annotated images. These morphological feature parameters included aspect ratio, contour Fourier descriptor, area-to-perimeter ratio, minimum bounding rectangle orientation angle, and local binary pattern texture features. The morphological feature parameters were associated with the corresponding biological growth stage labels of the tiller clusters, including budding stage, jointing stage, and heading stage. A multi-class support vector machine model is trained using the associated feature parameters and growth stage label data. The model can output the probability of the growth stage to which the tiller cluster belongs and the matching degree of typical features under the growth stage based on the input morphological feature parameters. The trained multi-class support vector machine model and the range of feature parameters of each growth stage learned by it are collectively defined as the predefined Leymus chinensis tiller morphological growth model.
[0009] Preferably, the subpixel-level edge refinement and contour closure operation performed on the effective tillering clusters specifically includes: Using the outline of the effective tillering cluster as the initial boundary, the normal direction of each boundary point is calculated at the corresponding position in the spectral fusion image; Image grayscale values are sampled at sub-pixel precision along the normal direction to form a grayscale distribution curve on the normal profile; Locate the inflection points and extreme points of the grayscale distribution curve and use them as the edge point positions with sub-pixel precision; connect all the newly located sub-pixel edge points to form a preliminary refined contour. Discontinuities and gaps in the preliminary refined contour are detected, and the gaps are repaired using a cubic spline interpolation algorithm based on the principle of contour curvature continuity, generating the final closed and continuous tillering unit boundary.
[0010] Preferably, the step of constructing a tiller connection diagram based on the spatial topological relationship of tillering units to analyze the proximity and overlap states between tillers specifically includes: Calculate the centroid coordinates and minimum bounding rectangle of all tillering unit boundaries. Calculate the Euclidean distance and overlap ratio between any two tillering unit boundaries. If the Euclidean distance between two tillering unit boundaries is less than a set spatial proximity threshold, then establish an edge representing the proximity relationship between the corresponding two nodes in the tillering connection graph and assign a distance weight to the edge. If two tillering unit boundaries have overlapping areas, calculate the ratio of the overlapping area to the area of the smaller tillering unit. If the ratio is greater than a set overlap threshold, then establish an edge representing the overlap relationship between the corresponding nodes in the graph and assign an overlap weight to the edge. Finally, the tillering connection graph is represented in the form of an undirected weighted graph, where nodes represent tillering units and edges represent proximity or overlap relationships and their strength.
[0011] Preferably, the application graph segmentation algorithm for processing the tillering connectivity graph specifically includes: Based on the weight of each edge in the tillering connection graph, the cluster center of all nodes in the graph is calculated. Based on the distance from the node to the cluster center and the weight of the connecting edge, the nodes in the graph are initially divided using the spectral clustering method to form multiple tillering subgraphs. For tillering subgraphs with overlapping edges, the overlapping area is regarded as the area to be assigned. Based on the geometric fit and texture continuity between the overlapping area and the contour of the adjacent tillering unit, the pixels in the overlapping area are iteratively assigned to the tillering unit with higher fit. After the assignment is completed, the boundary contour of the affected tillering unit is updated to ensure that the boundaries of all tillering units do not overlap and are independently closed.
[0012] Preferably, the process of performing skeleton extraction and branch point analysis on each segmented independent tillering unit specifically includes: Uniform sampling is performed on the closed boundary of each independent tillering unit to obtain a dense sequence of contour points; a morphological thinning algorithm is used to process the internal region of the tillering unit, iteratively eroding the boundary until a skeleton centerline with a single pixel width is obtained; all intersections and endpoints are detected on the skeleton centerline, the intersections are identified as potential branch points, and the endpoints are identified as the end of the main stem or the tip of a new tiller; starting from the endpoint of the main stem, the main stem path is traced along the skeleton centerline, and all branch points encountered on the path and their spatial positions are recorded.
[0013] Preferably, the process of merging and removing false tillers based on the spatial distribution pattern of new tillers and the consistency constraint of the main stem's growth direction specifically includes: Analyze the length and direction of the skeletal path from the base of the main stem to each branch point, and group the branch points that are close to the base of the main stem and are distributed in opposite or whorled patterns into the set of tillering points of the same growth ring; For each set of tillering points, examine the angle between the extension direction of its new tillering scaffold and the growth tangent direction of the main stem at the branching point, and remove pseudo-branching points whose angle exceeds the physiological limit threshold. For multiple closely adjacent branches originating from the same tillering point, they are merged according to the initial separation angle of their skeletons. If the initial separation angle is less than the merging threshold, they are regarded as early branches of the same tiller and counted as one tillering node. After merging and eliminating false ones, the final valid tillering nodes and their corresponding spatial coordinates are confirmed.
[0014] Preferably, the step of performing time-series consistency verification and correction on the current tiller quantity specifically includes: Retrieve historical records of tiller numbers and node positions at at least two previous growth time points for the same observation unit; Establish the spatial correspondence between the current tillering node and the historical tillering node, allowing matching within a preset spatial tolerance range. Based on the matching results, count the number of newly formed tillering nodes, persistent tillering nodes, and disappeared tillering nodes. According to the growth and development pattern of Leymus chinensis tillers, check the rationality of the number of newly formed nodes, the number of disappeared nodes, and the growth time interval. If any irregularities are found, review the results by combining the size changes and texture features of the tillering unit. Based on the verification results, incorrectly identified new or missing nodes are corrected, and the current number of tillers after time-series verification is finally output. The current number of tillers, node information, and corresponding multispectral images, fine boundaries, connectivity graphs, and skeleton information are then associated and stored.
[0015] Compared with the prior art, the beneficial effects of the present invention are: A spatial topological relationship of tillering units is constructed to form a tillering connection graph, and a graph segmentation algorithm is applied to process this graph structure. The pixel-level adhesion problem is transformed into an analysis of the mathematical relationship between nodes and connecting edges. The graph segmentation algorithm performs segmentation based on the overall proximity and overlap state between tillering units, which can directly delineate the ownership of adhesion regions at the structural relationship level. This method reduces the dependence on single local image features and avoids the problems of parameter sensitivity and unstable results in traditional pixel similarity-based segmentation methods when dealing with complex adhesions, achieving high-precision separation of individual tillers under dense growth conditions.
[0016] After obtaining geometric branch points through skeletal analysis, candidate points are merged and eliminated based on the spatial distribution patterns of new tillers and by introducing consistency constraints with the main stem's growth direction. This transforms botanical morphogenetic knowledge into quantified spatial direction consistency rules. By verifying the degree to which the spatial distribution of potential nodes matches the dominant growth trend of the main stem, geometric artifacts caused by local leaf morphology, image noise, or accidental intersections can be identified and filtered. This rule can reasonably merge multiple neighboring candidate points conforming to the same growth axis into a biologically reasonable tiller node, thus elevating the initial identification based on geometric morphology to a reliable judgment incorporating knowledge of growth patterns. Attached Figure Description
[0017] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent counting method for the number of tillers of Leymus chinensis described in this invention. Figure 2 A flowchart for performing multispectral band fusion of Leymus chinensis canopy images; Figure 3 A flowchart for performing subpixel-level edge refinement and contour closure operations on effective tillering clusters; Figure 4 A comprehensive fit analysis diagram of overlapping tillering areas of Leymus chinensis; Figure 5 A diagram illustrating the effect of branching and merging of Leymus chinensis tillers. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Please see Figure 1This invention provides an intelligent counting method for Leymus chinensis tillers. The method includes: performing multispectral band fusion of Leymus chinensis canopy images to generate a spectral fusion image with enhanced features; performing preliminary contour detection of tiller clusters on the spectral fusion image to obtain a set of detection regions containing potential tiller targets, and selecting effective tiller clusters from the set of detection regions based on the morphological growth model of Leymus chinensis tillers; performing sub-pixel-level edge refinement and contour closure operations on the selected effective tiller clusters to form independent and complete tiller unit boundaries; constructing a tiller connectivity graph based on the spatial topological relationships of these tiller units to analyze the proximity and overlap states between tillers; and then applying a graph segmentation algorithm to process the tiller connectivity graph, removing the adhesion between tillers and assigning overlapping areas to specific regions. For each segmented independent tillering unit, skeleton extraction and branch point analysis were performed to identify the main stem and potential new tillers of a single tiller. Based on the spatial distribution pattern of new tillers and the consistency constraint of the main stem's growth direction, new tillers were merged and false positives were removed to confirm the final tillering nodes. The total number of tillering nodes after merging was counted, and combined with node data from historical growth cycles, the current tillering count was checked and corrected for temporal consistency. Finally, the corrected tillering count results were associated and stored with the corresponding spectral fusion image, tillering unit boundaries, and node markers to form a traceable counting record.
[0020] In one embodiment of the present invention, see [reference] Figure 2 Simultaneously, visible light images and at least one non-visible light image of the Leymus chinensis canopy are acquired. Color space decomposition is performed on the visible light image to separate the hue and luminance components. The reflectance intensity of the non-visible light image is normalized to match its intensity range with the luminance component of the visible light image. The normalized non-visible light image and the hue component are weighted and superimposed to generate a composite feature layer highlighting chlorophyll activity regions and water distribution characteristics. This composite feature layer is then fused with the luminance component of the visible light image using wavelet transform at multiple scales, preserving the feature details of the composite feature layer and the overall structure of the luminance component, outputting the final spectral fusion image. Adaptive local thresholding is applied to the spectral fusion image to generate a binarized image reflecting regional contrast. Multi-scale morphological opening operations are performed on the binarized image to eliminate fine noise and smooth cluster boundaries. All connected components are extracted from the morphologically processed image, and each connected component is used as a preliminary detection region. A geometric descriptor for each preliminary detection region is calculated, including region area, perimeter, compactness, and principal axis direction. Based on the preset area threshold range and density threshold, the initial detection area is filtered to remove areas that are obviously too large, too small, or have extremely irregular shapes, thus obtaining the set of detection areas.
[0021] The multispectral fusion process is achieved by simultaneously acquiring visible and near-infrared images. In practice, a multispectral imaging device equipped with a standard RGB sensor and a near-infrared filter is used to acquire raw image data of the Leymus chinensis canopy vertically downwards under natural lighting conditions, ensuring precise spatial alignment between the visible and near-infrared images. Color space decomposition is then performed on the acquired visible images. Specifically, the visible images are converted from the RGB color space to the HSV color space. This conversion process separates the image information into hue, saturation, and lightness components. The hue component characterizes the color category, and the lightness component characterizes the intensity of light and dark areas.
[0022] It is understandable that for synchronously acquired non-visible light band images, in specific implementations, near-infrared band images record the reflection intensity of near-infrared light by plant tissue. The original pixel value range needs adjustment. Normalization of the reflection intensity of the near-infrared band image uses a linear stretching method, mapping the minimum and maximum pixel values to the minimum and maximum pixel values of the luminance component of the visible light band image, respectively, thus matching its intensity range with the luminance component of the visible light band image. The normalized non-visible light band image is then weighted and superimposed with the hue component separated from the visible light image. In specific implementations, the mathematical expression for weighted superposition is used to generate a composite feature layer. The pixel values of the composite feature layer are determined by the weighted sum of the hue component and the normalized near-infrared band image. This operation aims to fuse color information with near-infrared reflection features. The composite feature layer and the brightness component of the visible light band image are fused using wavelet transform at multiple scales. In a specific implementation, the composite feature layer and the brightness component are decomposed into three layers of wavelet decomposition to obtain their respective high-frequency detail sub-bands and low-frequency approximate sub-bands. The fusion rule is to take the coefficients of the brightness component in the low-frequency sub-band to maintain the overall structure, and to take the coefficients of the composite feature layer in the high-frequency sub-band to retain its feature details. Finally, wavelet inverse reconstruction is performed to output the final spectral fusion image.
[0023] In specific implementation, the initial contour detection of tillering clusters in the spectral fusion image begins with image segmentation. Adaptive local threshold segmentation is applied to the spectral fusion image. This method dynamically calculates the threshold based on the gray-level statistical characteristics of local image regions, generating a binary image reflecting regional contrast. Regions with pixel values greater than the local threshold are marked as foreground. Multi-scale morphological opening operations are then performed on the binary image. Morphological opening is understood to be a sequence of operations: erosion followed by dilation. Structural elements of different sizes are used sequentially to eliminate fine noise and smooth cluster boundaries, preventing subsequent misclassification of noise as independent regions. All connected components are extracted from the morphologically processed image. The connected component extraction algorithm scans the binary image, marking interconnected foreground pixel sets as independent regions, and each connected component as a preliminary detection region.
[0024] The geometric descriptor for each of the preliminary detection regions is calculated. In specific implementations, the geometric descriptor is calculated based on the set of pixel coordinates of the region. The region area is the total number of pixels contained within the region. The region perimeter is obtained by calculating the distance between contour pixels using a chain code algorithm. The density is calculated by the ratio of the area to the square of the perimeter. The principal axis direction is obtained by performing principal component analysis on the region's pixel coordinates to obtain the direction of its first principal component. Based on preset area threshold ranges and density thresholds, the preliminary detection regions are initially filtered. The area threshold range is used to remove excessively large foreign object shadows or excessively small soil spots that may exist in the image. The density threshold is used to filter regions with extremely irregular shapes, such as broken leaves. Regions that meet the threshold conditions are retained to obtain the set of detection regions.
[0025] In one embodiment of the present invention, a predefined Leymus chinensis tiller morphological growth model is loaded. This model describes the typical aspect ratio range, contour symmetry features, and texture orientation patterns of tiller clusters at different growth stages. For each region in the detection region set, a sub-image is cropped from its corresponding original spectral fusion image. The local gradient direction histogram of the sub-image is analyzed to extract its dominant texture direction. The geometric descriptor of the region and the dominant texture direction are input into the morphological growth model for matching degree calculation. A matching degree threshold is set, and regions with a matching degree higher than the threshold are identified as candidate effective tiller clusters that meet biological characteristics. Contour expansion based on region growth is performed on the candidate effective tiller clusters to ensure that the contour completely covers the boundary of the tiller cluster. Finally, the effective tiller clusters are output. The construction steps of the morphological growth model specifically include: acquiring multispectral images of Leymus chinensis tiller samples at different growth stages, manually annotating the multispectral images, and marking the true tiller cluster boundaries, main stem positions, and tiller nodes. Morphological feature parameters of each tiller cluster are extracted from the labeled images. These morphological feature parameters include aspect ratio, contour Fourier descriptor, area-to-perimeter ratio, minimum bounding rectangle orientation angle, and local binary pattern texture features. These morphological feature parameters are associated with the corresponding biological growth stage labels of the tiller cluster, including budding, jointing, and heading stages. A multi-class support vector machine model is trained using the associated feature parameters and growth stage label data. This model outputs the probability of the tiller cluster belonging to the growth stage and the matching degree of typical features under the growth stage based on the input morphological feature parameters. The trained multi-class support vector machine model and the range of feature parameters learned for each growth stage are collectively defined as the predefined Leymus chinensis tiller morphological growth model.
[0026] Loading a predefined Leymus chinensis tiller morphology growth model is the initial step in the screening process. This predefined model describes the typical aspect ratio range, contour symmetry characteristics, and texture orientation patterns of tiller clusters at different growth stages. For each region in the detection region set, a sub-image is cropped from its corresponding original spectral fusion image, with the bounding rectangle of that region as the core and the boundary appropriately extended. The local gradient direction histogram of the sub-image is analyzed. In specific implementation, the gradient magnitude and direction of each pixel in the sub-image within a specified neighborhood are calculated. The direction is uniformly divided into several intervals within the range of 0 to 180 degrees, and the gradient magnitude contribution is statistically analyzed to form a gradient direction histogram. The dominant texture direction is extracted as the direction interval with the largest cumulative magnitude in the histogram.
[0027] It is understood that the geometric descriptor of the region and the dominant texture direction are input into the morphological growth model for matching degree calculation. The geometric descriptor includes region area, perimeter, density, and principal axis direction. The matching degree calculation is completed internally by the morphological growth model and outputs a quantified matching score. A matching degree threshold is set, and regions with a matching degree higher than the threshold are identified as candidate effective tillering clusters that meet biological characteristics. Contour expansion based on region growth is performed on the candidate effective tillering clusters. In specific implementation, the initial boundary of the candidate effective tillering cluster is used as the seed point. Based on the gray-level similarity and spatial proximity between pixels in the spectral fusion image, adjacent pixels that meet the growth criteria are iteratively merged to ensure that the contour completely wraps the boundary of the tillering cluster. Finally, the effective tillering cluster is output.
[0028] The construction steps of the morphological growth model specifically include data acquisition, feature extraction, model training and definition. Multispectral images of Leymus chinensis tiller samples at different growth stages are acquired. In specific implementation, the growth stages cover budding, jointing, and heading stages. The multispectral images are manually labeled using image annotation tools to determine the true tiller cluster boundaries, main stem positions, and tiller nodes. Morphological feature parameters for each tiller cluster are extracted from the labeled images. These morphological feature parameters include aspect ratio, contour Fourier descriptor, area-to-perimeter ratio, minimum bounding rectangle orientation angle, and local binary pattern texture features. The morphological feature parameters are associated with the corresponding biological growth stage labels of the tiller clusters. In some embodiments, the growth stage labels are stored in numerical encoding form, with budding, jointing, and heading stages corresponding to discrete category labels.
[0029] A multi-class support vector machine (SVM) model is trained using the correlated feature parameters and growth stage label data. In practice, the multi-class SVM model employs a one-to-one strategy to construct multiple binary classifiers. The training objective is to find an optimal hyperplane that effectively separates the feature parameters of different growth stages in the feature space. The decision function of the multi-class SVM model for the input feature vector x can be used to evaluate the degree of matching with each growth stage. The mathematical expression of the decision function is: in: This represents the vector of morphological feature parameters input. It is the number of support vectors. It is a Lagrange multiplier. These are the class labels corresponding to the support vectors. Kernel functions are used to calculate similarity in the feature space. This is a bias term that enables the model to output the probability of the growth stage to which the tiller cluster belongs and the matching degree of typical features under the growth stage based on the input morphological feature parameters. It can be understood that the trained multi-class support vector machine model and the range of feature parameters learned for each growth stage are collectively defined as the predefined Leymus chinensis tiller morphological growth model.
[0030] In one embodiment of the present invention, see [reference] Figure 3 Using the outline of the effective tillering cluster as the initial boundary, the normal direction of each boundary point is calculated at the corresponding position in the spectral fusion image. Image grayscale values are sampled at sub-pixel precision along the normal direction to form a grayscale distribution curve on the normal profile. The inflection points and extreme points of the grayscale distribution curve are located and used as sub-pixel precision edge point positions. All newly located sub-pixel edge points are connected to form a preliminary refined outline. Discontinuities and gaps in the preliminary refined outline are detected, and based on the principle of outline curvature continuity, a cubic spline interpolation algorithm is used to repair the gaps, generating a final closed and continuous tillering unit boundary. The centroid coordinates and minimum bounding rectangle of all tillering unit boundaries are calculated, and the Euclidean distance and overlap area ratio between any two tillering unit boundaries are calculated. If the Euclidean distance between two tillering unit boundaries is less than a set spatial proximity threshold, an edge representing the proximity relationship is established between the corresponding two nodes in the tillering connection graph, and a distance weight is assigned to the edge. If two tillering units have overlapping areas, the proportion of the overlapping area to the area of the smaller tillering unit is calculated. If the proportion is greater than a set overlap threshold, an edge representing the overlap relationship is established between the corresponding nodes in the graph, and an overlap weight is assigned to the edge. The final tillering connection graph is represented as an undirected weighted graph, where nodes represent tillering units, and edges represent proximity or overlap relationships and their strength.
[0031] In specific implementation, the contour of the effective tillering cluster is composed of a series of pixel-level coordinate points. The normal direction of each boundary point is calculated at the corresponding position in the spectral fusion image. The normal direction of the boundary point is calculated using the cross product of vectors of adjacent boundary points. The local tangent is calculated through the coordinate difference between the previous and next boundary points, thus obtaining the normal direction. Image grayscale values are sampled along the normal direction at sub-pixel precision. Sub-pixel precision sampling means moving in the normal direction with a step size smaller than one pixel physical size. At each sampling position on the normal profile, grayscale values are obtained from the spectral fusion image through bilinear interpolation, forming a grayscale distribution curve on the normal profile. The inflection points and extreme points of the grayscale distribution curve are located. The inflection points are determined by calculating the zero point of the second derivative of the grayscale distribution curve, and the extreme points are determined by calculating the zero point of the first derivative. These are used as the sub-pixel precision edge point positions. All newly located sub-pixel edge points are connected. In specific implementation, the sub-pixel edge points are connected sequentially according to the order of the original contour points to form a preliminary refined contour. Discontinuities and gaps in the preliminary refined contour are detected. The detection of discontinuities and gaps is based on whether the Euclidean distance between adjacent sub-pixel edge points exceeds a set threshold. Based on the principle of contour curvature continuity, a cubic spline interpolation algorithm is used to repair the gaps. The cubic spline interpolation uses the coordinates of several sub-pixel edge points on both sides of the gap to fit a smooth curve to supplement the missing contour points and generate the final closed and continuous tillering unit boundary.
[0032] Calculate the centroid coordinates and minimum bounding rectangle of all tiller unit boundaries. In specific implementations, the centroid coordinates of the tiller unit boundaries are calculated by averaging the coordinates of all points on the boundary. The minimum bounding rectangle is obtained by using a rotating caliper algorithm to find the rectangle that completely encloses the tiller unit boundaries and has the smallest area. Calculate the Euclidean distance and overlap area ratio between any two tiller unit boundaries. The Euclidean distance is calculated as the straight-line distance between the centroid coordinates of the two tiller unit boundaries. Calculating the overlap area ratio requires first converting the two tiller unit boundaries into a region mask, obtaining the overlapping region pixels through a logical AND operation, and calculating the area. If the Euclidean distance between two tiller unit boundaries is less than a set spatial proximity threshold, which is set to a fixed pixel value based on the typical plant spacing of Leymus chinensis tillers, then an edge representing the proximity relationship is established between the corresponding two nodes in the tiller connection graph, and a distance weight is assigned to the edge. In some embodiments, the distance weight is the reciprocal of the Euclidean distance. If the boundaries of two tillering units overlap, the proportion of the overlapping area to the area of the smaller tillering unit is calculated. The formula for calculating the proportion is: in: Represents the proportion of overlapping area. This represents the area of the overlapping region between the boundaries of two tillering units. and These represent the area of each of the two tillering units. This indicates that the smaller of the two values is taken. If the ratio is greater than a set overlap threshold, an edge representing the overlap relationship is established between the corresponding nodes in the graph, and an overlap weight is assigned to the edge. In some embodiments, the overlap weight is directly adopted as the overlap area ratio. The final tillering connection graph is represented in the form of an undirected weighted graph, where nodes represent tillering units and edges represent proximity or overlap relationships and their strength.
[0033] In one embodiment of the present invention, the cluster center of all nodes in the tillering connection graph is calculated based on the weight of each edge. Based on the distance from the node to the cluster center and the weight of the connecting edge, a spectral clustering method is used to initially divide the nodes in the graph, forming multiple tillering subgraphs. For tillering subgraphs with overlapping edges, the overlapping area is considered as an area to be assigned. Based on the geometric fit and texture continuity between the overlapping area and the contour of adjacent tillering units, the pixels in the overlapping area are iteratively assigned to tillering units with higher fit. After assignment, the boundary contours of the affected tillering units are updated to ensure that all tillering unit boundaries are non-overlapping and independently closed. The closed boundary of each independent tillering unit is uniformly sampled to obtain a dense sequence of contour points. A morphological thinning algorithm is used to process the internal region of the tillering unit, iteratively eroding the boundary until a skeleton centerline with a single pixel width is obtained. All intersections and endpoints are detected on the skeleton centerline; intersections are identified as potential branch points, and endpoints are identified as the end of the main stem or the tip of a new tillering bud. Starting from the base of the main stem, trace the main trunk path along the center line of the skeleton and record all branch points encountered along the path and their spatial locations.
[0034] The application of graph segmentation algorithms to process tillering graphs begins with the analysis of the graph structure. Based on the weight of each edge in the tillering graph, the cluster centers of all nodes are calculated. In practice, the cluster centers are calculated using a weighted average method. The coordinates of each node are represented by the centroid coordinates of its corresponding tillering unit boundary, and the edge weights are used to adjust the contribution of adjacent nodes in the calculation. Based on the distance from the nodes to the cluster centers and the weights of the connecting edges, spectral clustering is used to initially partition the nodes in the graph. Spectral clustering first constructs the Laplacian matrix of the tillering graph, then performs eigenvalue decomposition on the Laplacian matrix, using the eigenvectors to embed the nodes into a low-dimensional space and cluster them, forming multiple tillering subgraphs. For tillering sub-images with overlapping edges, the overlapping region is considered as the region to be assigned. The overlapping region is a set of pixels extracted from the intersection of the boundary masks of overlapping tillering units. Based on the geometric fit and texture continuity between the overlapping region and the contours of adjacent tillering units, the pixels in the overlapping region are iteratively assigned to tillering units with higher fit. The geometric fit is calculated based on the shortest Euclidean distance from the pixels in the overlapping region to the contours of adjacent tillering units. The texture continuity is calculated based on the gray-level similarity between the pixels in the overlapping region and the pixels within adjacent tillering units in the spectral fusion image. The determination of higher fit is based on a comprehensive score, calculated using the following formula: in: Represents the overall compatibility score. Represents the normalized geometric distance fit. Represents normalized texture similarity. This is a coefficient that adjusts the weights of geometric and texture factors. After allocation, the boundary contours of the affected tillering units are updated, the allocated pixels are incorporated into the boundary mask of the corresponding tillering unit, and its boundary coordinates are recalculated to ensure that all tillering unit boundaries do not overlap and are independently closed.
[0035] In some embodiments, the specific parameters and intermediate data involved in the tillering connection graph node clustering process can be illustrated by tables. Referring to Table 1, the feature vector coordinates and preliminary partitioning results of some nodes in the spectral clustering process are shown.
[0036] Table 1: Clustering Data of Tillering Connection Nodes Node ID Feature vector 1 Feature vector 2 Sub-graph number 1 0.15 -0.22 1 2 0.18 -0.19 1 3 -0.31 0.05 2 4 -0.28 0.09 2 5 0.02 0.33 3 It is understandable that performing skeleton extraction and branch point analysis on each independent tillering unit after segmentation is a subsequent step. Uniform sampling is performed on the closed boundary of each independent tillering unit. In specific implementation, uniform sampling extracts coordinate points along the perimeter of the closed boundary with a fixed step size, resulting in a dense sequence of contour points. A morphological thinning algorithm is used to process the internal region of the tillering unit. The morphological thinning algorithm performs iterative erosion based on the eight-neighborhood condition of pixels, removing boundary pixels until the region width is a single pixel. Iterative erosion of the boundary continues until a skeleton centerline with a single-pixel width is obtained. All intersections and endpoints are detected on the skeleton centerline. Intersection detection is achieved by determining whether the number of connecting branches within the eight-neighborhood of the skeleton pixel is greater than or equal to 3, and endpoint detection is achieved by determining whether the number of connecting branches within the eight-neighborhood is equal to 1. Intersections are identified as potential branch points, and endpoints are identified as the end of the main stem or the tip of a new tillering bud. Starting from the base of the main stem, the main path is traced along the center line of the skeleton. The base of the main stem is usually defined as the endpoint closest to the bottom of the image or with specific morphological features. The tracing algorithm traverses the skeleton pixel connection relationship based on depth-first or breadth-first search and records all branch points encountered on the path and their spatial positions.
[0037] Optionally, during skeleton extraction, redundant short branches that may be introduced by morphological refinement can be filtered out by setting a length threshold. In some embodiments, the sampling step size of the contour point sequence can be dynamically adjusted according to the total perimeter of the tiller unit boundary to balance accuracy and computational efficiency. It can be understood that the spatial positions of branch points on the skeleton centerline are recorded as pixel coordinates, which will be used for subsequent new tiller point merging analysis. Optionally, the tracking direction of the main path can be constrained according to the overall principal axis direction of the tiller unit to improve tracking accuracy.
[0038] See Figure 4 This is a comprehensive fit analysis chart of overlapping regions in Leymus chinensis tillers. Overlapping region 3 has the highest comprehensive fit (≈0.88), followed by overlapping region 1 (≈0.82), while overlapping region 4 has the lowest fit (≈0.71). Overlapping region 2 has a texture similarity (≈0.9) much higher than its geometric distance fit (≈0.72), indicating strong texture feature matching but weak spatial location matching. Overlapping region 3 has high geometric and texture fit, making it the most clearly defined allocation region. This chart is used in the overlapping region allocation stage of Leymus chinensis tiller counting. By comprehensively evaluating the matching degree of spatial geometry and spectral texture, it determines the tiller unit to which overlapping pixels should belong, which is the core basis for achieving accurate tiller segmentation and counting.
[0039] In one embodiment of the invention, the length and direction of the skeletal path from the base of the main stem to each branch point are analyzed. Branch points that are close in length to the base of the main stem and are spatially distributed in an opposite or whorled pattern are grouped into a set of tillering points of the same growth ring. For each set of tillering points, the angle between the extension direction of the new tillering bud skeleton and the growth tangent direction of the main stem at the branch point is checked, and pseudo-branching points with angles exceeding the physiological limit threshold are removed. For multiple closely adjacent branches originating from the same tillering point, they are merged according to their initial separation angle. If the initial separation angle is less than the merging threshold, they are considered as early branches of the same tiller and counted as one tillering node. After merging and removing pseudo-branches, the final valid tillering nodes and their corresponding spatial coordinates are confirmed. The historical tillering quantity and node position records of the same observation unit at at least two previous growth time points are retrieved. The spatial correspondence between the current tillering node and the historical tillering node is established, allowing matching within a preset spatial tolerance range. Based on the matching results, the number of newly formed tillers, persistent tillers, and disappeared tillers are counted. According to the growth and development pattern of Leymus chinensis tillers, the reasonableness of the number of newly formed nodes, the number of disappeared nodes, and the growth time interval is checked. If any discrepancies are found, the size changes and texture features of the tillering units are used for further verification. Based on the verification results, misjudged newly formed or disappeared nodes are corrected. Finally, the current number of tillers, after time-series verification, is output, and the current number of tillers, node information, and corresponding multispectral images, fine boundaries, connectivity graphs, and skeleton information are associated and stored.
[0040] In practice, the skeleton path length is obtained by accumulating the Euclidean distance between all adjacent pixels on the skeleton centerline from the branch point to the endpoint of the main stem base. The skeleton path direction is obtained by calculating the tangent direction of the skeleton at the branch point. Branch points that are close in length to the main stem base and whose spatial positions show a paired or whorled distribution are grouped into the set of tillering points of the same growth ring. The criterion for judging close distance is that the length difference is less than the set length tolerance threshold. The criterion for judging paired or whorled distribution is whether the azimuth angle distribution of the branch points relative to the main stem base within the same growth ring shows an approximately symmetrical or uniformly spaced characteristic.
[0041] For each set of tillering points, examine the angle between the extension direction of the new tillering scaffold and the growth tangent direction of the main stem at the branching point. In practice, the extension direction of the new tillering scaffold is defined as the direction vector along the initial segment of the new tillering scaffold starting from the branching point, and the growth tangent direction of the main stem at the branching point is defined as the tangent direction of the main stem scaffold path at the branching point. Calculate the angle between the two direction vectors. Remove pseudo-branch points whose included angle exceeds the physiological limit threshold; that is, the judgment condition is... ,in: This represents the angle between the calculated direction of new shoot extension and the tangent direction of the main stem growth. This represents a preset physiological limit angle threshold. For multiple closely adjacent branches originating from the same tillering point, they are merged based on their initial separation angle. The initial separation angle refers to the angle between the initial directions of the skeletons of different new tillers originating from the same tillering point. If the initial separation angle is less than the merging threshold, it is considered an early branch of the same tiller and counted as one tillering node. After merging and removing false positives, the final valid tillering nodes and their corresponding spatial coordinates are confirmed. The spatial coordinates of the tillering nodes are recorded as the pixel positions of their corresponding branch points in the image.
[0042] In some embodiments, historical records of tiller quantity and node location at at least two previous growth time points of the same observation unit are retrieved. These historical records are stored in a structured data file, including timestamps, tiller quantity, and a list of coordinates for each tiller node. A spatial correspondence is established between current tiller nodes and historical tiller nodes. In specific implementations, this correspondence is established based on a nearest neighbor distance matching algorithm, allowing matching within a preset spatial tolerance range. This spatial tolerance range is set according to image resolution and the potential positional drift during Leymus chinensis growth. Based on the matching results, the number of newly formed tiller nodes, persistent tiller nodes, and disappeared tiller nodes are counted. Newly formed tiller nodes refer to nodes identified in the current image but not successfully matched in the most recent historical record. Persistent tiller nodes refer to nodes successfully matched in both the current and historical records. Disappeared tiller nodes refer to nodes present in the historical record but not successfully matched in the current image.
[0043] Based on the growth and development patterns of Leymus chinensis tillers, the rationality of the number of new and disappearing nodes and the growth time interval is checked. It can be understood that the growth and development patterns of Leymus chinensis tillers include a non-negative number of new nodes during the normal growth stage that is positively correlated with the time interval, and a very small number of disappearing nodes. If any irregularities occur, such as a negative number of new nodes or an abnormally large number of disappearing nodes, a review is conducted in conjunction with the size changes and texture features of the tillering units. The review process includes re-examining whether the size of the corresponding region of the historical tillering units marked as disappearing nodes has significantly decreased in the current image, and whether the texture features have degraded. Based on the review results, misjudged new or disappearing nodes are corrected. Correction operations include remarking nodes temporarily missed due to occlusion as persistent nodes, or deleting pseudo-nodes misjudged due to noise from the list of new nodes. Finally, the current number of tillers, after time-series verification, is output, and the current number of tillers, node information, and corresponding multispectral images, fine boundaries, connectivity graphs, and skeleton information are associated and stored.
[0044] Optionally, the spatial tolerance range can be set to a dynamic value related to the image resolution, such as a width of several pixels. It can be understood that the associated storage data structure includes a time-series index, facilitating subsequent tracing and growth analysis.
[0045] See Figure 5 This is a graph analyzing the branch merging effect of Leymus chinensis tillers. In group 1 / 2, the average initial separation angle is smaller (10° / 8°), and the number of branches decreased from 3 / 2 to 1 after merging, showing a significant merging effect. In group 3, the average initial separation angle is larger (15°), and the number of branches decreased from 4 to 2 after merging, only merging some branches with small angles. The smaller the initial separation angle, the more significant the reduction in the number of branches after merging, conforming to the rule that "if the initial separation angle is less than the merging threshold, they are merged into one tiller node." This graph is used in the Leymus chinensis tiller node merging stage to quantitatively demonstrate the impact of the "initial separation angle" on the merging effect, serving as a core basis for verifying the rationality of the merging algorithm and optimizing the merging threshold.
[0046] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for intelligent counting of tiller quantity based on Leymus chinensis, characterized in that, The method includes: Perform multispectral band fusion of Leymus chinensis canopy images to generate a spectral fusion image with enhanced features; Preliminary contour detection of tiller clusters is performed on the spectral fusion image to obtain a set of detection regions containing potential tiller targets. Based on the morphological growth model of Leymus chinensis tillers, effective tiller clusters are selected. Subpixel-level edge refinement and contour closure operations are performed on the effective tillering clusters to form independent and complete tillering unit boundaries; Based on the spatial topological relationship of tillering units, a tillering connection graph is constructed to analyze the proximity and overlap between tillers. A graph segmentation algorithm is then applied to process the tillering connection graph to remove the adhesion between tillers and assign overlapping areas to different regions. For each independent tillering unit after segmentation, skeleton extraction and branch point analysis are performed to identify the main stem and potential new tillering points of a single tiller. Based on the spatial distribution pattern of new tillering points and the consistency constraint of the growth direction of the main stem, the new tillering points are merged and false processing is performed to confirm the final tillering nodes. The total number of tillering nodes after merging is counted, and the current number of tillers is checked and corrected for temporal consistency by combining the node data of the historical growth cycle. The verified and corrected tiller count results are associated and stored with the corresponding spectral fusion image, tiller unit boundary, and node markers to form a traceable counting record.
2. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 1, characterized in that, The multispectral band fusion of the sheepgrass canopy image specifically includes: Simultaneously acquire visible light images and at least one non-visible light image of the Leymus chinensis canopy. Perform color space decomposition on the visible light image to separate the hue and luminance components. Normalize the reflectance intensity of the non-visible light image to match its intensity range with the luminance component of the visible light image. Weighted superposition of the normalized non-visible light image and the hue component generates a composite feature layer that highlights chlorophyll activity regions and water distribution characteristics. Perform multi-scale fusion of the composite feature layer and the luminance component of the visible light image based on wavelet transform, preserving the feature details of the composite feature layer and the overall structure of the luminance component, and output the final spectral fusion image.
3. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 2, characterized in that, The preliminary contour detection of tillering clusters on the spectral fusion image specifically includes: Adaptive local thresholding is applied to the spectral fusion image to generate a binarized image that reflects the regional contrast. Multi-scale morphological opening operations are performed on the binarized image to eliminate fine noise and smooth cluster boundaries; Extract all connected components from the morphologically processed image, and use each connected component as a preliminary detection region; Calculate a geometric descriptor for each of the preliminary detection regions, the geometric descriptor including region area, perimeter, compactness, and principal axis direction; Based on the preset area threshold range and density threshold, the preliminary detection area is initially filtered to remove areas that are obviously too large, too small, or have extremely irregular shapes, thus obtaining the set of detection areas.
4. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 3, characterized in that, The effective tillering clusters selected based on the morphological growth model of Leymus chinensis tillers specifically include: Load a predefined Leymus chinensis tiller morphological growth model, which describes the typical aspect ratio range, outline symmetry characteristics, and texture orientation pattern of tiller clusters at different growth stages. For each region in the set of detection regions, a sub-image is cropped from its corresponding original spectral fusion image. The local gradient direction histogram of the sub-image is analyzed, and its texture dominant direction is extracted. The geometric descriptor of the region and the texture dominant direction are input into the morphological growth model for matching degree calculation. A matching degree threshold is set, and regions with a matching degree higher than the threshold are determined as candidate effective tillering clusters that meet biological characteristics. Perform contour expansion based on region growing on the candidate effective tiller clusters to ensure that the contour completely covers the boundary of the tiller clusters, and finally output the effective tiller clusters; The specific steps for constructing the morphological growth model include: Multispectral images of Leymus chinensis tillers at different growth stages were collected. The multispectral images were manually annotated to identify the actual tiller cluster boundaries, main stem positions, and tiller nodes. Morphological feature parameters of each tiller cluster were extracted from the annotated images. These morphological feature parameters included aspect ratio, contour Fourier descriptor, area-to-perimeter ratio, minimum bounding rectangle orientation angle, and local binary pattern texture features. The morphological feature parameters were associated with the corresponding biological growth stage labels of the tiller clusters, including budding stage, jointing stage, and heading stage. A multi-class support vector machine model is trained using the associated feature parameters and growth stage label data. The model can output the probability of the growth stage to which the tiller cluster belongs and the matching degree of typical features under the growth stage based on the input morphological feature parameters. The trained multi-class support vector machine model and the range of feature parameters of each growth stage learned by it are collectively defined as the predefined Leymus chinensis tiller morphological growth model.
5. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 4, characterized in that, The subpixel-level edge refinement and contour closure operation performed on the effective tillering clusters specifically includes: Using the outline of the effective tillering cluster as the initial boundary, the normal direction of each boundary point is calculated at the corresponding position in the spectral fusion image; Image grayscale values are sampled at sub-pixel precision along the normal direction to form a grayscale distribution curve on the normal profile; Locate the inflection points and extreme points of the grayscale distribution curve and use them as the edge point positions with sub-pixel precision; connect all the newly located sub-pixel edge points to form a preliminary refined contour. Discontinuities and gaps in the preliminary refined contour are detected, and the gaps are repaired using a cubic spline interpolation algorithm based on the principle of contour curvature continuity, generating the final closed and continuous tillering unit boundary.
6. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 5, characterized in that, The construction of a tiller connection diagram based on the spatial topological relationship of tillering units to analyze the proximity and overlap states between tillers specifically includes: Calculate the centroid coordinates and minimum bounding rectangle of all tillering unit boundaries. Calculate the Euclidean distance and overlap ratio between any two tillering unit boundaries. If the Euclidean distance between two tillering unit boundaries is less than a set spatial proximity threshold, then establish an edge representing the proximity relationship between the corresponding two nodes in the tillering connection graph and assign a distance weight to the edge. If two tillering unit boundaries have overlapping areas, calculate the ratio of the overlapping area to the area of the smaller tillering unit. If the ratio is greater than a set overlap threshold, then establish an edge representing the overlap relationship between the corresponding nodes in the graph and assign an overlap weight to the edge. Finally, the tillering connection graph is represented in the form of an undirected weighted graph, where nodes represent tillering units and edges represent proximity or overlap relationships and their strength.
7. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 6, characterized in that, The application graph segmentation algorithm for processing the tillering connectivity graph specifically includes: Based on the weight of each edge in the tillering connection graph, the cluster center of all nodes in the graph is calculated. Based on the distance from the node to the cluster center and the weight of the connecting edge, the nodes in the graph are initially divided using the spectral clustering method to form multiple tillering subgraphs. For tillering subgraphs with overlapping edges, the overlapping area is regarded as the area to be assigned. Based on the geometric fit and texture continuity between the overlapping area and the contour of the adjacent tillering unit, the pixels in the overlapping area are iteratively assigned to the tillering unit with higher fit. After the assignment is completed, the boundary contour of the affected tillering unit is updated to ensure that the boundaries of all tillering units do not overlap and are independently closed.
8. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 7, characterized in that, The process of performing skeleton extraction and branch point analysis on each segmented independent tillering unit specifically includes: Uniform sampling is performed on the closed boundary of each independent tillering unit to obtain a dense sequence of contour points; a morphological thinning algorithm is used to process the internal region of the tillering unit, iteratively eroding the boundary until a skeleton centerline with a single pixel width is obtained; all intersections and endpoints are detected on the skeleton centerline, the intersections are identified as potential branch points, and the endpoints are identified as the end of the main stem or the tip of a new tiller; starting from the endpoint of the main stem, the main stem path is traced along the skeleton centerline, and all branch points encountered on the path and their spatial positions are recorded.
9. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 8, characterized in that, The process of merging and eliminating false tillers based on the spatial distribution pattern of new tillers and the consistency constraint of the main stem's growth direction specifically includes: Analyze the length and direction of the skeletal path from the base of the main stem to each branch point, and group the branch points that are close to the base of the main stem and are distributed in opposite or whorled patterns into the set of tillering points of the same growth ring; For each set of tillering points, examine the angle between the extension direction of its new tillering scaffold and the growth tangent direction of the main stem at the branching point, and remove pseudo-branching points whose angle exceeds the physiological limit threshold. For multiple closely adjacent branches originating from the same tillering point, they are merged according to the initial separation angle of their skeletons. If the initial separation angle is less than the merging threshold, they are regarded as early branches of the same tiller and counted as one tillering node. After merging and eliminating false ones, the final valid tillering nodes and their corresponding spatial coordinates are confirmed.
10. The intelligent counting method based on the number of tillers of Leymus chinensis according to claim 9, characterized in that, The specific steps for performing time-series consistency verification and correction on the current number of tillers include: Retrieve historical records of tiller numbers and node positions at at least two previous growth time points for the same observation unit; Establish the spatial correspondence between the current tillering node and the historical tillering node, allowing matching within a preset spatial tolerance range. Based on the matching results, count the number of newly formed tillering nodes, persistent tillering nodes, and disappeared tillering nodes. According to the growth and development pattern of Leymus chinensis tillers, check the rationality of the number of newly formed nodes, the number of disappeared nodes, and the growth time interval. If any irregularities are found, review the results by combining the size changes and texture features of the tillering unit. Based on the verification results, incorrectly identified new or missing nodes are corrected, and the current number of tillers after time-series verification is finally output. The current number of tillers, node information, and corresponding multispectral images, fine boundaries, connectivity graphs, and skeleton information are then associated and stored.