Phenotypic trait image recognition and extraction method for efficient breeding of corn

Abstract

The present application relates to the technical field of image recognition, and further relates to a phenotypic trait image recognition and extraction method for efficient corn breeding. The method comprises the following steps: step 1: at multiple different growth stages from the corn jointing stage to the mature stage, an unmanned aerial vehicle (UAV) carrying a multispectral sensor is used to obtain images of the same plant canopy in a test field, and normalized gray images are obtained; step 2: based on the normalized gray images, gradient operation is performed on target bands to obtain significant edge intensity, and a trait-sensitive probability field corresponding to the phase is generated; step 3: an optimal observation period is preset for each target trait, a pixel set with a probability value not lower than 0.5 and having been determined as the trait type is selected from the image corresponding to the optimal observation period, and the measurement of total projected area, average connected domain area and average leaf angle is sequentially performed to obtain a comprehensive breeding serial number value of the trait. The present application realizes high-precision probability recognition of the effective phenotypic area of grains and leaves.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to a method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding. Background Technology

[0002] In modern agricultural production systems, maize, as one of the world's major food crops, has always been a key research area in agricultural science and technology for efficient breeding. With the continuous development of molecular breeding technology and genetic improvement systems, the precise, automated identification and quantitative evaluation of phenotypic traits have gradually become a bottleneck limiting the improvement of maize breeding efficiency. In traditional maize phenotypic observation systems, manual measurement and experience-based interpretation dominate, which is not only labor-intensive, time-consuming, and lacks repeatability, but also makes it difficult to guarantee objectivity and consistency in large-scale field trials. Especially when multiple phenotypic dimensions (such as grain density, leaf morphology, and plant structure) need to be evaluated simultaneously, traditional methods struggle to generate high-throughput data support, severely restricting efficient screening at the maize population level.

[0003] To address this issue, some researchers have recently attempted to introduce computer vision and image processing techniques to perform plant identification and trait extraction from visible light images. One mainstream method is to analyze the structure of maize plants based on two-dimensional RGB images, such as extracting visual traits like leaf area and ear height through color segmentation and edge detection. While this method has a certain degree of accuracy in controlled indoor environments, it is easily affected by factors such as sunlight angle, soil background reflectivity, and leaf color variations in complex field environments, leading to increased identification errors. Furthermore, visible light images are not sensitive to structural differences in reflectivity between different maize tissues (such as leaves and kernels), making it difficult to support effective spatial segmentation and texture classification.

[0004] Another more advanced technical approach is remote sensing phenotypic identification of maize canopies based on hyperspectral or multispectral imaging systems. Hyperspectral imaging, by capturing spectral reflectance across continuous bands, can be used to extract phenotypic indicators such as plant physiological status, nitrogen content, and water content. However, hyperspectral equipment is expensive, generates massive amounts of data, and requires a highly stable platform (such as a fixed elevated system), making it difficult to widely implement in field trials. In contrast, multispectral imaging balances spectral characteristics and system cost, and has become a more feasible field phenotypic data acquisition solution in recent years. Among related technologies, some studies have used drones equipped with multispectral sensors to acquire maize canopy reflectance data, combining this with vegetation indices such as NDVI and GNDVI to monitor plant growth. Although vegetation indices have some effect in overall trend analysis, they lack the ability to characterize the morphological structure of individual plants at high resolution and are difficult to directly quantitatively correlate with breeding target traits, thus failing to provide decision support at the individual plant level at the phenotypic level. Summary of the Invention

[0005] The main objective of this invention is to provide an image recognition and extraction method for phenotypic traits in high-efficiency maize breeding. This method utilizes a drone equipped with a multispectral sensor to acquire canopy images of the same plant at different growth stages. Through normalization processing, multi-scale gradient calculation, near-infrared texture fusion, and 3D feature vector construction, it achieves high-precision probabilistic identification of effective phenotypic regions of grains and leaves. By dynamically determining the optimal observation period using time series data, a phenotypic-specific mask is generated, and the total projected area, average connected region area, and average leaf angle are quantified. Finally, a comprehensive breeding ordinal value and a single-plant high-efficiency breeding index are output. This method significantly improves the accuracy, spatiotemporal consistency, and quantification efficiency of phenotypic extraction, providing an automated, high-throughput, and quantifiable solution for large-scale field breeding screening.

[0006] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0007] A method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding, the method comprising:

[0008] Step 1: During multiple different growth stages of maize from the jointing stage to maturity, images of the canopy of the same plant in the experimental field were acquired using a drone equipped with a multispectral sensor, and flight altitude, instantaneous irradiance and bare soil reflectance were collected simultaneously; the grayscale of the target band pixels was preprocessed to obtain a normalized grayscale image.

[0009] Step 2: Based on the normalized grayscale image, perform gradient operation on the target band to obtain significant edge intensity, and fuse it with near-infrared texture; simultaneously consider the three features of brightness, significant edge intensity and near-infrared texture to determine the probability that each pixel belongs to the effective phenotypic region of grain-leaf, and generate the trait-sensitive probability field of the corresponding time phase.

[0010] Step 3: Preset the optimal observation period for each target trait. Select a set of pixels with a probability value of not less than 0.5 and identified as the trait type from the images of the corresponding phase of the optimal observation period to form a trait-specific mask. For each trait, perform the measurement of total projected area, average connected region area and average leaf angle in sequence to obtain the comprehensive breeding sequence value of the trait.

[0011] Furthermore, the method also includes: Step 4: Compare the comprehensive breeding sequence value with the target threshold of the corresponding trait in the breeding archive, and calculate the trait score according to the relative deviation; For each experimental plot, use the measured yield divided by the area of ​​the experimental plot to obtain the yield per unit area, and multiply this value by the survival rate of mature plants to form a field-level production potential coefficient; Multiply all the trait scores included in the decision by the field-level production potential coefficient one by one and sum them to obtain the single-plant high-efficiency breeding index; The higher the value of the single-plant high-efficiency breeding index, the closer the plant is to the breeding target.

[0012] Furthermore, in step 1, the chlorophyll-sensitive band with a wavelength of 710nm is selected as the target band, and the standard reflectance obtained by indoor calibration with the same band calibration plate is used as the absolute reference.

[0013] Furthermore, in step 1, based on the geometric boundaries of the experimental field, a grid-based flight path planning method is adopted to ensure that the canopy of each corn plant is photographed by at least three adjacent overlapping flight paths; in order to ensure that the images of the canopy of the same plant have spatial references that can be aligned at each growth stage, the lateral overlap is set to be no less than 75% and the lateral overlap is set to be no less than 85%.

[0014] Further, step 2 specifically includes: performing gradient operations on the normalized grayscale image using three sets of scale difference operators to obtain multi-scale gradient magnitudes; then using morphological opening and closing reconstruction to purify and synthesize them into a significant edge intensity matrix to highlight the grain outline and leaf edge; next, extracting grayscale co-occurrence texture indices from synchronized near-infrared images and linearly synthesizing them into near-infrared texture grids, and geometrically registering them to the normalized grayscale image to achieve pixel-level three-grid co-addressing of brightness, significant edge intensity, and near-infrared texture; then dynamically normalizing the three features for each pixel within a local window to construct a three-dimensional feature vector of brightness-gradient-texture and feeding it into a soft discriminant model to output the preliminary probability of the effective phenotypic region of the grain-leaf; if the on-site illumination or leaf color distribution is detected to be outside the training range, unsupervised correction of the outlier probability is performed through neighborhood consistency screening; finally, anisotropic diffusion smoothing is implemented based on the gradient direction of significant edge intensity to preserve real edges while suppressing noise.

[0015] Furthermore, the process of performing gradient operations on the normalized grayscale image using three sets of scale difference operators includes: selecting three sets of spatial scales on the grid of the same normalized grayscale image; calculating the horizontal and vertical differences for each scale using symmetric difference operators of 3x3, 5x5, and 7x7 respectively; then synthesizing the gradient magnitudes of the horizontal and vertical differences at the same scale according to the Euclidean norm to obtain a set of gradient magnitudes and recording the corresponding gradient directions; performing an opening operation on the set of gradient magnitudes to filter out high-frequency isolated noise; and then retaining the structured edges through a continuous erosion-dilation cycle, finally outputting a significant edge intensity matrix.

[0016] Further, in step 2, near-infrared band images acquired synchronously with the same time phase are read, and histogram matching is performed to ensure that their brightness distribution is consistent with the normalized grayscale image. Then, a gray-level co-occurrence matrix is ​​constructed using a 3x3 sliding window, and four texture indices—contrast, homogeneity, entropy, and uniformity—are calculated. These four texture indices are then linearly synthesized according to empirical weights to obtain a single scalar near-infrared texture grid. The near-infrared texture grid is geometrically registered to the pixel grid of the normalized grayscale image, and bilinear interpolation is used to correct sub-pixel deviations, ensuring that each pixel simultaneously possesses brightness values, significant edge intensity values, and near-infrared texture values ​​in subsequent calculations. To reduce... To mitigate data redundancy, a 1:4 pyramid downsampling ratio is applied to all three feature grids to generate a multi-layer resolution pyramid sequence. This allows for consistent probability calculations and subsequent smoothing at different spatial granularities. At the original resolution level, brightness, significant edge intensity, and near-infrared texture are extracted from each pixel, and local dynamic normalization is performed on each feature. The local dynamic normalization uses an 8x8 window, with the minimum and maximum values ​​within the window as endpoints for linear stretching. This compresses each feature value to a closed interval between zero and one while preserving relative differences, thus mitigating the impact of brightness drift between different plants and maintaining a sharp contrast in gradient and texture at leaf edges and grain surfaces.

[0017] Furthermore, in step 2, for each pixel, a three-dimensional feature vector is formed according to a fixed order of brightness, gradient, and texture, and this vector is input into a pre-trained offline soft discriminant model. This soft discriminant model adopts an improved multi-class logistic regression framework, using the effective phenotypic region label of the seed-leaf area as the positive class and the labels of other regions as the negative class, and outputs a continuous probability value between zero and one for the three-dimensional feature vector. When it is detected that the peak-valley span of the overall brightness distribution of the normalized grayscale image exceeds one standard deviation of the training distribution, the nearest neighbor consistency screening algorithm is immediately activated to evaluate the probability outlier in the local three-by-three pixel neighborhood. If the outlier exceeds the threshold, the probability of the center pixel is replaced by the mean of the neighborhood probability to achieve unsupervised adaptive correction.

[0018] Furthermore, after concatenating all pixel probabilities into a preliminary morphology-sensitive probability field, directional smoothing is performed on the preliminary morphology-sensitive probability field. The directional smoothing process uses the gradient direction provided by the significant edge intensity matrix as the main smoothing direction, and performs anisotropic diffusion on the pixel probabilities along the gradient direction. At the same time, diffusion perpendicular to the gradient direction is suppressed to preserve the true leaf edge and grain outline. The diffusion coefficient decays exponentially according to the gradient magnitude, so that the boundary blurring caused by diffusion in the leaf texture is effectively controlled. After completing the directional smoothing, the local entropy of the preliminary morphology-sensitive probability field is calculated to evaluate the non-converged region. If the local entropy exceeds a set threshold, secondary denoising based on fast median filtering is performed on this non-converged region, and the median filter kernel size increases linearly with the entropy value.

[0019] The phenotypic trait image recognition and extraction method for high-efficiency maize breeding of this invention has the following beneficial effects: it realizes continuous observation and high-throughput trait quantification of the same maize plant at multiple growth stages under field conditions. Its technical approach integrates multiple key steps, including flight path overlap planning, normalized grayscale image generation, multi-scale gradient calculation, near-infrared texture extraction, feature fusion classification, direction-guided probability smoothing, and trait-specific mask generation, forming a closed-loop process from UAV data collection to trait index output. This method achieves co-addressing of brightness, edge, and texture features at the pixel level and outputs continuous probability values ​​for each pixel through a soft discriminant model, thereby significantly improving the accuracy of grain and leaf identification and the continuity of boundary representation. Simultaneously, it dynamically adjusts the optimal observation period by combining time-series probability peaks, ensuring that the target trait is extracted during the optimal saliency window. In terms of trait quantification, by measuring the total projected area, average connected region area, and average leaf angle within the mask spatial domain, it achieves joint modeling of canopy size, structure, and posture, ultimately outputting a structured comprehensive breeding ordinal value. Compared with traditional phenotypic recognition methods based on manual measurement or single image processing, this invention has a high degree of automation, high robustness and strong repeatability, and can adapt to different environmental lighting and variety differences, greatly improving the screening efficiency and evaluation accuracy of maize populations. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart of the method for image recognition and extraction of phenotypic traits in high-efficiency maize breeding provided in an embodiment of the present invention.

[0021] Figure 2 An experimental schematic diagram illustrating the correlation analysis between comprehensive breeding sequence values ​​and measured traits provided in an embodiment of the present invention;

[0022] Figure 3 An experimental schematic diagram illustrating the distribution characteristics of the high-efficiency breeding index of a single plant provided in an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of a comparative experiment on the accuracy of sensitive probability fields for traits at different reproductive stages. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0025] refer to Figure 1A method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding, the method comprising:

[0026] Step 1: During multiple different growth stages of maize from the jointing stage to maturity, images of the canopy of the same plant in the experimental field were acquired using a drone equipped with a multispectral sensor, and flight altitude, instantaneous irradiance and bare soil reflectance were collected simultaneously; the grayscale of the target band pixels was preprocessed to obtain a normalized grayscale image.

[0027] Step 2: Based on the normalized grayscale image, perform gradient operation on the target band to obtain significant edge intensity, and fuse it with near-infrared texture; simultaneously consider the three features of brightness, significant edge intensity and near-infrared texture to determine the probability that each pixel belongs to the effective phenotypic region of grain-leaf, and generate the trait-sensitive probability field of the corresponding time phase.

[0028] Step 3: Preset the optimal observation period for each target trait. Select a set of pixels with a probability value of not less than 0.5 and identified as the trait type from the images of the corresponding phase of the optimal observation period to form a trait-specific mask. For each trait, perform the measurement of total projected area, average connected region area and average leaf angle in sequence to obtain the comprehensive breeding sequence value of the trait.

[0029] Maize has five key growth stages: jointing, tasseling, pre-grain filling, post-grain filling, and maturity. When a multispectral sensor periodically observes the canopy of the same plant in the visible and near-infrared sensitive bands, differences in chlorophyll content, water content, and the microstructure of the waxy outer layer of the kernels alter the multiple scattering paths of photons in the intercellular spaces and mesophyll sponge tissue, resulting in a dynamically changing reflectance distribution along the sensor's line of sight. This distribution, superimposed with solar incidence angle, instantaneous irradiance, and soil background diffuse reflectance, is difficult to separate from varietal genetic differences and environmental random noise without correction. Therefore, by simultaneously acquiring flight altitude, instantaneous irradiance, and bare soil reflectance, and normalizing the grayscale of the target band pixels into a normalized grayscale image, a simplified radiative transfer inversion is essentially performed at the image level. This compresses the geometric degradation and illumination response terms in the remote sensing radiometric range into controllable constants, allowing direct comparison of plant reflectance characteristics at different time points in the spectral space.

[0030] Subsequently, the significant edge intensity extracted using the multi-scale gradient operator can amplify the spectral transition from the grain surface to the leaf surface. This transition originates from differences in tissue density and surface roughness, and its amplitude exhibits monotonically piecewise variations in the wavelength dimension, thus manifesting as a gradient pattern in the spatial domain. Meanwhile, the near-infrared texture quantifies the spatial uniformity of the arrangement of mesophyll vesicles and vascular bundles, as well as the water distribution, from a second-order statistical dimension. This texture information complements the grain starch fullness. The feature fusion of brightness, significant edge intensity, and near-infrared texture is equivalent to constructing a set of local feature tensors on three mutually orthogonal feature axes: spectral, geometric, and statistical. Subsequently, a soft segmentation of the effective phenotypic region of the grain-leaf axis is completed through probability density mapping. The generated trait-sensitive probability field is essentially a Markov random field approximation of each micro-element on the canopy surface in the attribute space. Soft classification avoids the fragmentation problem of traditional hard thresholding segmentation on heterogeneous canopies.

[0031] To overcome the issue that the phenotypic significance of different traits does not always occur at the same growth stage, the system introduces a dynamic calibration mechanism based on the extreme values ​​of the probability mean of time series. This mechanism aligns with the reproductive-nutrient competition model: the phenotypic peaks of different traits correspond to different source-sink balance nodes, and the time shift reflects the overall delay or advancement of plant physiological processes due to annual temperature, precipitation, or management measures. By adjusting the optimal observation period in real time, the subsequent trait-specific mask can always cover the phase with the highest information density. Once the trait-specific mask is determined, the method uses the two-dimensional projected area as a direct measure of biomass spatial projection. The average connected domain area describes the dispersion and splitting degree of grain or leaf clusters, which corresponds to the spatial aggregation degree of grain filling consistency and leaf unfolding. The average leaf angle reflects the orientation of photosynthetic organs and their shading relationship, and is a key geometric parameter for canopy light interception efficiency. These three factors together determine the degree of matching between carbon assimilation rate and grain filling rate. The integrated breeding ordinal value maps these three measures to a unified dimension and weights them according to their genetic contribution rate. Essentially, it is a multi-scale morphological-functional coupling evaluation: the total projected area provides scale information, the average connected domain area provides structural information, and the average leaf angle provides posture information. After the three-dimensional combination, the distance between the ordinal value and the target phenotypic profile is calculated in morphological space. The smaller the distance, the higher the ordinal value, which is more in line with the composite requirements of high-efficiency maize breeding for uniform ears and grains, reasonable leaf posture, and canopy permeability.

[0032] In step 3, the system first establishes an optimal observation period index table for each target trait during the program initialization phase. This index table, based on statistical data from experimental fields over the years, maps the key growth stages from jointing to maturity to the phenotypic peaks of the target trait. When entering the online observation phase, the system calls the time label of the optimal observation period according to the index table, actively retrieves images of the canopy of the same plant under that time label, and simultaneously loads the trait-sensitive probability field for the corresponding time phase. After the probability field is loaded, the system scans the probability raster row by row, reading the probability value for each pixel. When the probability value is not less than 0.5 and the pixel has been previously identified as belonging to that trait type by the classifier, the pixel is written into the pixel set already identified as belonging to that trait type and assigned a value of one on a blank raster of the same size; the remaining pixels are assigned a value of zero. Because the probability field may generate isolated pixels in the edge transition area, the system immediately calls a 3x3 structuring element to perform a closing operation after the pixel set is written, connecting small holes and fusing scattered edge fragments to obtain a topologically continuous binary raster. The system then applies the 8-connectivity labeling algorithm to the binary raster to calculate the pixel number of all connected components and writes it into the attribute table with the pixel number as the key, finally outputting a complete and non-redundant feature-specific mask.

[0033] To ensure the accuracy of the physical scale of the total projected area, after generating a shape-specific mask, the system extracts the actual ground size represented by each pixel from the camera model based on the flight altitude, camera focal length, and sensor pixel size during image acquisition. The area of ​​each pixel assigned a value of one is then summed to obtain the total projected area. Subsequently, the system iterates through the connected components in the attribute table, counting the number of pixels contained in each connected component, and multiplies the number of pixels by the ground area per pixel to obtain the connected component area. After counting all connected component areas, connected components with an area of ​​less than twenty-five pixels are discarded as noise. The average area of ​​the remaining connected components is recalculated, and the resulting value is the average connected component area. The number of connected components is retained in the attribute table for subsequent comparisons but is not included in subsequent leaf angle measurements.

[0034] To calculate the average leaf angle, the system performs centroid-aligned second-order moment analysis for each connected component in the attribute table. The method first translates the centroid of the pixel coordinates within the connected component, then calculates the second-order central moment matrix. The principal inertial axis direction vector is obtained through singular value decomposition, and the absolute value of the angle between this direction vector and the horizontal coordinate axis is defined as the leaf angle of that connected component. The system sums the leaf angles of all retained connected components and divides by the number of connected components to obtain the average leaf angle. If the number of connected components is zero, the system, according to fault-tolerant logic, directly sets the average leaf angle to zero and warns in the log that there is no usable leaf angle information for the current time phase. In addition, the system simultaneously calculates the standard deviation of the leaf angle and writes it into the attribute table for studying canopy attitude consistency; however, this standard deviation is not included in the subsequent comprehensive breeding sequence numerical calculation.

[0035] Once the total projected area, average connected region area, and average leaf angle are all obtained, the system activates the integrated breeding sequence value calculation module. The module first retrieves historical threshold entries corresponding to the current target trait from the database. These historical thresholds include area and leaf angle thresholds, used to normalize the three measures to a closed interval between zero and one. Specifically, the ratio obtained by dividing the total projected area by the area threshold is limited to between zero and one; the ratio obtained by dividing the average connected region area by the area threshold is also limited to between zero and one; and the ratio obtained by dividing the average leaf angle by the leaf angle threshold is truncated to the range of zero and one. After normalization, the module performs a weighted sum of the three normalized values ​​according to the weights stored in the threshold entries. The sum of the weights is always one, and the result is the integrated breeding sequence value. The integrated breeding sequence value, along with the experimental field number, plant number, and growth period label, is synchronously written into the breeding database and sorted among plants in the same batch. The higher the ordinal number, the more the plant meets the breeding objective in terms of the current target trait. Breeders can directly use the ordinal results to screen for high-quality genotypes for the next round of field verification or cross-breeding design.

[0036] After completing the writing of the integrated breeding sequence values, the system records the calculation timestamp, the optimal observation period label used, the threshold version number, and the calculation log. It also archives the trait-specific mask in read-only mode to ensure that complete spatial information and measurement processes can be recovered during subsequent tracing. If breeders require visual comparisons during the decision-making stage, the system can generate dynamic time-series layers based on the trait-sensitive probability fields and trait-specific masks of the same plant at different growth stages. This displays the phenotypic trend over time, helping to determine the stability and plasticity of target traits during the growth process, thereby achieving comprehensive quantification and rapid screening of high-efficiency maize breeding indicators.

[0037] Furthermore, the method also includes: Step 4: Compare the comprehensive breeding sequence value with the target threshold of the corresponding trait in the breeding archive, and calculate the trait score according to the relative deviation; For each experimental plot, use the measured yield divided by the area of ​​the experimental plot to obtain the yield per unit area, and multiply this value by the survival rate of mature plants to form a field-level production potential coefficient; Multiply all the trait scores included in the decision by the field-level production potential coefficient one by one and sum them to obtain the single-plant high-efficiency breeding index; The higher the value of the single-plant high-efficiency breeding index, the closer the plant is to the breeding target.

[0038] The relative deviation is obtained by subtracting the target threshold from the comprehensive breeding sequence value and normalizing it according to the target threshold. This relative deviation is then multiplied by 100% and converted into a percentage format, which is then packaged as a trait score. The trait score is bound to the plant number and trait identifier in the data structure to maintain index consistency during subsequent multi-trait accumulation. Simultaneously, the system reads the measured yield records for each experimental plot according to the experimental design template, calculates the yield per unit area using the surveyed area of ​​the same plot as the divisor, and extracts the survival rate of mature plants in that plot from the growth period monitoring files. The two are multiplied to obtain the field-level production potential coefficient. The field-level production potential coefficient is stored in a table field corresponding to the experimental plot number, along with the yield station number and measurement timestamp, facilitating later traceability and error diagnosis. Once the trait score and field-level production potential coefficient are ready, the system selects the trait scores to be included in this evaluation based on a pre-set decision list. These trait scores are then multiplied one by one with the field-level production potential coefficient in order, summed, and written into the single-plant high-efficiency breeding index field. At this point, each plant generates only a unique single-plant high-efficiency breeding index, with the value range limited to zero to the theoretical upper limit. The theoretical upper limit depends on the product of the upper limit of the trait score and the upper limit of the field-level production potential coefficient. To avoid extreme values ​​misleading the selection ranking, the system calls the anomaly detection module after calculation to perform box-type tests and three-times median absolute deviation tests on the single-plant high-efficiency breeding index. If abnormally high values ​​exceeding the upper limit or abnormally low values ​​below the lower limit are found, these records are marked as pending review and highlighted in red on the user interface. For data that passes the tests, the system sorts them in descending order by the single-plant high-efficiency breeding index and generates a visual report, simultaneously presenting the plant number, trait score list, field-level production potential coefficient, and single-plant high-efficiency breeding index, allowing breeders to quickly screen out high-performing individuals or those that need to be eliminated.

[0039] Furthermore, in step 1, the chlorophyll-sensitive band with a wavelength of 710nm is selected as the target band, and the standard reflectance obtained by indoor calibration with the same band calibration plate is used as the absolute reference.

[0040] The normalized grayscale image is:

[0041]

[0042] Among them, G norm (i, j, t) represents the normalized gray value of pixel (i, j) at time t, corresponding to the target band; G raw (i, j, t) represents the original pixel grayscale values ​​acquired by the multispectral sensor; δ dark (t) represents the time-related correction value used to suppress flight dark current offset; E solar (t) represents the instantaneous irradiance at time t (unit: W / m²). 2 ); R ref (λ710 H(t) represents the standard reflectance of the indoor calibration plate at a wavelength of 710 nm; H(t) represents the flight altitude (meters) captured by the UAV. ref The preset standard shooting height; R soil (t) represents the soil reflectance of the bare soil area in the field; R 710 (i, j, t) represents the 710 nm reflectance of the canopy at the pixel location; ∈ is a stability constant to prevent the denominator from being zero; i is the x-axis; j is the y-axis.

[0043] Furthermore, in step 1, based on the geometric boundaries of the experimental field, a grid-based flight path planning method is adopted to ensure that the canopy of each corn plant is photographed by at least three adjacent overlapping flight paths; in order to ensure that the images of the canopy of the same plant have spatial references that can be aligned at each growth stage, the lateral overlap is set to be no less than 75% and the lateral overlap is set to be no less than 85%.

[0044] After track generation, the module further inserts buffer segments at the inflection points of each track to ensure that the UAV maintains a constant speed and stable attitude within the turning radius, avoiding image tilt caused by pitch and panning. To ensure that the canopy of the same plant has an alignable spatial reference at different growth stages, the system simulates and projects the generated grid-based flight paths, comparing the calculated forward overlap with a preset lower limit of 85% point by point. If the forward overlap of some track segments is found to be lower than 85%, the grid segment is regenerated by inserting intermediate tracks or shortening the track spacing until the forward overlap of the entire domain meets the requirements. Based on this, the system uses a grid coverage algorithm to count the number of times the canopy of each corn plant is covered by tracks in the instantaneous field of view. When the number is less than three adjacent tracks, a compensation track is inserted in the local area. The start and end points of the compensation track are arranged along the grid centerline to minimize the increase in total flight time caused by the addition of tracks. After all tracks and compensation tracks are completed, the system writes the track information into the task file, including latitude and longitude coordinates, altitude, heading angle, speed, and exposure trigger interval, and loads it into the UAV autopilot before flight. By combining raster-based flight path planning with high overlap constraints, the system ensures that the canopy of each maize plant is captured by at least three adjacent overlapping tracks throughout all observed growth stages. This provides accurate and redundant spatial anchor points for the same plant canopy in subsequent multi-temporal registration, ensuring the alignment accuracy of raster data such as normalized grayscale images, significant edge intensity, and near-infrared texture in the temporal dimension.

[0045] Further, step 2 specifically includes: performing gradient operations on the normalized grayscale image using three sets of scale difference operators to obtain multi-scale gradient magnitudes; then using morphological opening and closing reconstruction to purify and synthesize them into a significant edge intensity matrix to highlight the grain outline and leaf edge; next, extracting grayscale co-occurrence texture indices from synchronized near-infrared images and linearly synthesizing them into near-infrared texture grids, and geometrically registering them to the normalized grayscale image to achieve pixel-level three-grid co-addressing of brightness, significant edge intensity, and near-infrared texture; then dynamically normalizing the three features for each pixel within a local window to construct a three-dimensional feature vector of brightness-gradient-texture and feeding it into a soft discriminant model to output the preliminary probability of the effective phenotypic region of the grain-leaf; if the on-site illumination or leaf color distribution is detected to be outside the training range, unsupervised correction of the outlier probability is performed through neighborhood consistency screening; finally, anisotropic diffusion smoothing is implemented based on the gradient direction of significant edge intensity to preserve real edges while suppressing noise.

[0046] Multi-scale difference operators are first used to establish a set of gradient response fields. The theoretical basis is that the first-order differential of an image exhibits gray-level transitions at tissue boundaries. By calculating horizontal and vertical differences at fine, medium, and coarse granular levels and taking the maximum amplitude, the multi-level intensity discontinuities formed by the leaf margin, midrib, and grain shell can be completely captured. However, the original gradient field often contains high-frequency noise and isolated responses. This is because sensor readout errors and field scattering light cause local brightness abrupt changes to appear simultaneously with physical boundaries. Therefore, morphological opening and closing reconstruction is used to structurally constrain the gradient amplitudes at different scales. Opening operations remove small bright spot noise, and closing operations fill small dark holes. The reconstruction process iterates under connected domain topological constraints. The final salient edge intensity matrix no longer contains random spikes but highlights continuous grain contours and leaf margins, providing a stable geometric saliency measure for subsequent discrimination.

[0047] Alongside gradient information, near-infrared texture needs to be introduced to describe the differences in plant moisture and tissue density. The near-infrared band is sensitive to scattering from spongy mesophyll tissues, but less so to specular reflection from waxy grain surfaces. Calculating the gray-level co-occurrence matrix within a sliding window can quantify the joint spatial distribution of gray levels. Entropy reflects texture complexity, contrast reflects the magnitude of gray-level differences, and uniformity reflects pattern repeatability. Linearly synthesizing these three can make the statistical differences between the fine texture inside the leaf and the relatively smooth surface of the grain explicit. Since there are slight pose differences between near-infrared and visible light images during capture, rigid registration is required to maintain pixel-level co-location. This process relies on repeatable geometric anchor points of the same plant canopy in overlapping tracks. Least-squares matching is used to map the near-infrared texture raster to the coordinate system of the normalized grayscale image, achieving three-raster co-location.

[0048] When fusing brightness, salient edge intensity, and near-infrared texture, the global drift caused by time-varying illumination must be addressed. The theoretical basis for local dynamic normalization is the adaptive contrast mechanism in the human visual system, which recalibrates signal intensity within a limited receptive field using local minimums and maximums, allowing feature values ​​to reflect relative contrast rather than absolute brightness. After linear stretching through an 8x8 window, the three features of each pixel are compressed into a closed interval between zero and one, while preserving their relative magnitudes, providing a comparable scale for the subsequent soft-discriminative model to construct the 3D feature vector of the brightness gradient texture. The soft-discriminative model uses a continuous probability output form instead of a hard threshold; its statistical basis can be considered as a likelihood ratio estimate of the positive and negative sample distributions in the 3D feature space. The output value represents the posterior probability that a pixel belongs to the effective phenotypic region of the seed leaf. This probabilistic partitioning can provide intermediate values ​​in the leaf edge gradient zone, avoiding false boundaries caused by traditional binary classification in heterogeneous canopies.

[0049] In real-world field environments, sudden changes in light intensity or abrupt changes in leaf color can cause feature distributions to deviate from the training samples, leading to local saturation or voids in the soft-discrimination model output. Neighborhood consistency screening is based on Markov random field smoothing, assuming that the probability distribution of neighboring pixels should possess spatial consistency. The algorithm first detects the probability variance within a 3x3 neighborhood. When the variance is significantly greater than the variance of the overall image mean, the central pixel is identified as an outlier and replaced by the neighborhood mean. This unsupervised correction does not rely on additional labels and can quickly restore probability continuity in anomalous scenarios extrapolated from the model. The final preliminary probability still contains point noise and step effects, requiring anisotropic diffusion smoothing based on the directional information provided by significant edge strength. Anisotropic diffusion is based on partial differential equations, using an exponentially decaying diffusion coefficient in the gradient direction to suppress boundary blurring, and a higher diffusion coefficient in the vertical gradient direction to smooth out internal random noise. This maintains the sharp edges of grains and leaf margins while forming a coherent probability plateau within the leaf, providing a morphologically stable probability peak for optimal observation period retrieval in the time series.

[0050] The preliminary probability of the effective phenotypic region of grain-leaf is:

[0051]

[0052] Among them, P i，j，t L represents the probability (range 0-1) that pixel (i, j) belongs to the effective phenotypic region of the seed-leaf at time t; i，j，t The normalized grayscale image brightness features obtained in step 1; G edge (i, j, t) represents the significant edge intensity values ​​obtained after performing multi-scale gradient operations on the normalized image; T nir (i, j, t) represents the synthesized values ​​of texture features extracted from near-infrared images (such as a weighted sum of entropy and contrast); μ L(t), μ G (t), μ T (t) represents the global average values ​​of brightness, edge, and texture of the entire image at time t.

[0053] Furthermore, the process of performing gradient operations on the normalized grayscale image using three sets of scale difference operators includes: selecting three sets of spatial scales on the grid of the same normalized grayscale image; calculating the horizontal and vertical differences for each scale using symmetric difference operators of 3x3, 5x5, and 7x7 respectively; then synthesizing the gradient magnitudes of the horizontal and vertical differences at the same scale according to the Euclidean norm to obtain a set of gradient magnitudes and recording the corresponding gradient directions; performing an opening operation on the set of gradient magnitudes to filter out high-frequency isolated noise; and then retaining the structured edges through a continuous erosion-dilation cycle, finally outputting a significant edge intensity matrix.

[0054] The fundamental purpose of using three sets of scale difference operators in parallel within the same normalized grayscale image is to ensure that structured boundaries such as leaf margins and grain shells maintain detectable response peaks against a multi-band texture background. Symmetrical difference operators of 3x3, 5x5, and 7x7 simulate the first derivatives of the image at fine, medium, and coarse granular levels, respectively. Fine granularity is most sensitive to high-curvature details such as leaf tips and midribs; medium granularity considers leaf margins and small grain clusters; and coarse granularity provides a stable response to broad, gentle boundaries formed by grain clusters or bundled leaf sheaths. When horizontal and vertical differences are combined into gradient magnitudes using the Euclidean norm, the gradient is essentially treated as the magnitude of a two-dimensional vector, thereby simultaneously characterizing both boundary strength and direction, thus avoiding directional bias caused by unidirectional differencing. The spatial superposition of the gradient magnitude sets obtained from the three scales is equivalent to merging the multi-scale Laplacian zero-crossing of the same image into a unified coordinate system, allowing structural boundaries of different sizes to accumulate and strengthen while random noise cancels each other out. The morphological opening operation first matches and removes small bright spot noise using structuring elements, and then compensates the weakened true boundary with equal amount through an erosion-dilation cycle. This process follows the set contraction-expansion homeomorphism of mathematical morphology, ensuring that only gradient peaks above the noise threshold and continuously distributed are preserved. The resulting salient edge intensity matrix thus becomes a geometrically consistent solution under the cross-scale threshold, which can completely depict the outer ring of corn kernels and leaf edges, and maintains a very low response to unstructured high-frequency interference such as soil cracks and light patches. This provides a boundary reference map with low noise, complete directional information, and intensity adaptation for subsequent brightness-gradient-texture three-grid co-location fusion.

[0055] Further, in step 2, near-infrared band images acquired synchronously with the same time phase are read, and histogram matching is performed to ensure that their brightness distribution is consistent with the normalized grayscale image. Then, a gray-level co-occurrence matrix is ​​constructed using a 3x3 sliding window, and four texture indices—contrast, homogeneity, entropy, and uniformity—are calculated. These four texture indices are then linearly synthesized according to empirical weights to obtain a single scalar near-infrared texture grid. The near-infrared texture grid is geometrically registered to the pixel grid of the normalized grayscale image, and bilinear interpolation is used to correct sub-pixel deviations, ensuring that each pixel simultaneously possesses brightness values, significant edge intensity values, and near-infrared texture values ​​in subsequent calculations. To reduce... To mitigate data redundancy, a 1:4 pyramid downsampling ratio is applied to all three feature grids to generate a multi-layer resolution pyramid sequence. This allows for consistent probability calculations and subsequent smoothing at different spatial granularities. At the original resolution level, brightness, significant edge intensity, and near-infrared texture are extracted from each pixel, and local dynamic normalization is performed on each feature. The local dynamic normalization uses an 8x8 window, with the minimum and maximum values ​​within the window as endpoints for linear stretching. This compresses each feature value to a closed interval between zero and one while preserving relative differences, thus mitigating the impact of brightness drift between different plants and maintaining a sharp contrast in gradient and texture at leaf edges and grain surfaces.

[0056] Near-infrared (NIR) images and normalized grayscale images naturally differ in imaging spectra, exposure curves, and sensor response functions. Direct stitching would lead to inconsistent brightness scales, undermining the homology assumption in subsequent probabilistic modeling. Therefore, histogram matching is first used to remap the brightness distribution of the NIR images to the distribution space of the normalized grayscale images. The principle is to maintain the pixel order by using a monotonic mapping of the cumulative distribution function, thus achieving cross-band radiative alignment without introducing pseudo-gradients. After brightness alignment, the second-order statistical texture features of plant tissues contained in the NIR images need to be transformed into measurable quantities. Constructing a grayscale co-occurrence matrix using a 3x3 sliding window is precisely to capture the joint probability distribution of pixel grayscale pairs in their spatial neighborhood. Contrast measures the grayscale level difference to reveal texture roughness, homogeneity measures the grayscale distance to reflect pattern smoothness, entropy measures information disorder to mark tissue complexity, and uniformity measures co-occurrence concentration to characterize pattern regularity. These four parameters, from complementary dimensions, describe the microscopic inhomogeneity of mesophyll air cavities, vascular bundles, and grain endosperm in near-infrared scattering. The near-infrared texture grid obtained after linear synthesis using empirical weights retains the comprehensive texture intensity that best distinguishes between the specular surface of grains and the diffuse reflection of leaves. However, there is a slight translation and rotation between the near-infrared image and the visible light image in terms of shooting posture, requiring geometric registration to the pixel grid of the normalized grayscale image. In the registration stage, bilinear interpolation is used to continuously remap sub-pixel deviations, based on the fact that the spectral texture changes approximately linearly with space at the sub-pixel scale, without introducing high-frequency pseudo-ripples. After the three grids are co-located, the data still has high-resolution redundancy, especially when the texture inside the leaf tends to be smooth. High sampling density does not bring information gain. Therefore, a 1:4 pyramid downsampling is uniformly performed on the three feature grids, and the original resolution and the three downsampling layers are stored sequentially as a multi-layer pyramid sequence, which facilitates subsequent probability estimation and diffusion smoothing in a coarse-to-fine manner. The local dynamic normalization stage simulates the human eye's adaptive perception mechanism for local contrast. Within an 8x8 window, the minimum and maximum values ​​are used as stretching endpoints to compress brightness, salient edge intensity, and near-infrared texture into a closed interval of zero to one. This not only weakens the global drift caused by different plant angles and lighting conditions but also preserves the gradient and texture peaks that are commonly present between the grain hull and leaf edges. This provides a scale-uniform, contrast-sufficient, and noise-suppressed input for the subsequent construction of the brightness-gradient-texture 3D feature vector, fundamentally improving the discrimination stability of the soft discrimination model in complex field scenes.

[0057] Furthermore, in step 2, for each pixel, a three-dimensional feature vector is formed according to a fixed order of brightness, gradient, and texture, and this vector is input into a pre-trained offline soft discriminant model. This soft discriminant model adopts an improved multi-class logistic regression framework, using the effective phenotypic region label of the seed-leaf area as the positive class and the labels of other regions as the negative class, and outputs a continuous probability value between zero and one for the three-dimensional feature vector. When it is detected that the peak-valley span of the overall brightness distribution of the normalized grayscale image exceeds one standard deviation of the training distribution, the nearest neighbor consistency screening algorithm is immediately activated to evaluate the probability outlier in the local three-by-three pixel neighborhood. If the outlier exceeds the threshold, the probability of the center pixel is replaced by the mean of the neighborhood probability to achieve unsupervised adaptive correction.

[0058] Brightness, gradient, and texture respectively carry three complementary types of information: canopy reflectivity, geometric boundaries, and tissue microstructure. The principle of concatenating these three elements in a fixed order into a three-dimensional feature vector leverages the mutual information maximization assumption of statistical discriminant models: when features are statistically approximately orthogonal and conditionally independent of the same discriminant objective, multi-class logistic regression can construct a minimum cross-entropy hypersurface in the parameter space, achieving optimal segmentation between the effective phenotypic region of grains and leaves and other regions. The improved multi-class logistic regression embeds a feature norm regularization term and a class imbalance penalty term into the traditional log-likelihood objective function. The regularization term suppresses the singularity of the feature covariance matrix, while the penalty term offsets the bias caused by the significantly larger number of background pixels compared to positive pixels by increasing the cost of the negative class. Therefore, it can maintain stable boundary sharpness and confidence in continuous probability outputs. When the overall brightness distribution of the normalized grayscale image exhibits a peak-to-valley span exceeding one standard deviation of the training distribution, it indicates that the current lighting or leaf color conditions have deviated from the model's empirical domain. Directly relying on the model output in this case will lead to systematic bias. The design idea of ​​the nearest neighbor consistency screening algorithm comes from the spatial Markov random field: if the probability value of a pixel deviates significantly from the distribution center of its three-by-three pixel neighborhood, the pixel is likely to be distorted by local shadows, sunspots, or random noise from the sensor. By calculating the probability outlier and replacing pixels exceeding the threshold with the neighborhood probability mean, it is equivalent to applying a first-order smoothing prior to the probability map, so that the local probability field can be restored to spatial consistency. At the same time, since this operation does not depend on external labels, it belongs to the category of unsupervised adaptive bias correction. It can be triggered instantly when there is a sudden change in light or leaf color variation without retraining the model, ensuring the robustness of probability estimation in dynamic environments.

[0059] Furthermore, after concatenating all pixel probabilities into a preliminary morphology-sensitive probability field, directional smoothing is performed on the preliminary morphology-sensitive probability field. The directional smoothing process uses the gradient direction provided by the significant edge intensity matrix as the main smoothing direction, and performs anisotropic diffusion on the pixel probabilities along the gradient direction. At the same time, diffusion perpendicular to the gradient direction is suppressed to preserve the true leaf edge and grain outline. The diffusion coefficient decays exponentially according to the gradient magnitude, so that the boundary blurring caused by diffusion in the leaf texture is effectively controlled. After completing the directional smoothing, the local entropy of the preliminary morphology-sensitive probability field is calculated to evaluate the non-converged region. If the local entropy exceeds a set threshold, secondary denoising based on fast median filtering is performed on this non-converged region, and the median filter kernel size increases linearly with the entropy value.

[0060] The core idea of ​​direction-guided smoothing is to use the gradient direction revealed by the salient edge intensity matrix as the spatial prior of the probability field. Information along the tangential direction between the leaf edge and the grain outline is considered a high-value signal, and the direction orthogonal to it is considered a diffusion channel that needs to be restricted. During numerical iteration, the anisotropic diffusion equation assigns a diffusion coefficient to each pixel. This coefficient is mapped from the gradient magnitude of the salient edge intensity matrix through an exponential decay function. Therefore, the diffusion rate is significantly reduced at pixels with large gradient magnitudes, equivalent to treating the sharp boundary as a heat flow barrier to prevent probability from mixing across the leaf edge and grain shell. At pixels with small gradient magnitudes, the diffusion rate is close to constant, allowing the probability to quickly equalize within the homogeneous leaf surface, thereby suppressing fine-grained noise caused by diffusion. After diffusion iteration, to quantify the still unsteady local random residuals, a 5x5 pixel window is used as the sliding entropy measurement domain to calculate the local entropy of the preliminary morphology-sensitive probability field. High local entropy means that the probability distribution within the window is still multi-peaked or discrete, belonging to the non-converged region. For these types of regions, fast median filtering is chosen as the secondary denoising operator because it can nonlinearly suppress spike noise while preserving edge positions. Furthermore, its kernel size can be dynamically amplified proportionally to the local entropy value: higher entropy results in a larger window, indicating a greater need for spatial integration in the region; lower entropy results in a smaller window, preserving local details. In this way, direction-guided smoothing first locks in the true structure macroscopically through anisotropic diffusion, and then suppresses residual noise microscopically through entropy-driven median filtering. This ultimately yields a trait-sensitive probability field that follows both the geometric boundaries of the leaf margin and grains and is continuously smooth within the leaf surface. This provides a high-confidence, low-artifact base grid for subsequent probability thresholding and trait-specific mask generation.

[0061] The comprehensive breeding sequence number is:

[0062]

[0063] Among them, S qA is the composite breeding sequence number of the target trait q, used to evaluate the plant's performance in that trait; total,q The total projected area of ​​all connected regions in the feature-specific mask (in cm²). 2 );θ avg,q N is the average leaf angle (in degrees) of all connected regions; region，q σ represents the number of connected regions in the mask identified by the 8-connectivity marker; θ，q The standard deviation of the leaf angle values ​​reflects the consistency of the leaf angle. P represents the number of target shape pixels with a probability value not less than 0.5 in a shape-sensitive probability field. all,q This represents the total number of pixels that are classified as this trait out of all pixels.

[0064] The following example uses plant number P-2025-17 from field No. 1 of the experimental station as a sample. The field is 120m long and 80m wide. A DJIM300 UAV equipped with a multispectral camera (2.4μm pixel size, 12mm focal length, 4000×3000 resolution) was used. The flight altitude was set at 40m, corresponding to a ground resolution GSD≈0.008m, and each pixel covers an area of ​​0.000064m². 2 That is, 0.64cm 2 .

[0065] The growth period was divided into five phases: jointing stage, tasseling stage, pre-grain filling stage, post-grain filling stage, and maturity stage. The following is an example of a complete treatment during the pre-grain filling stage; the same steps were used for the other phases.

[0066] Target band 710nm original pixel grayscale value G raw =986, the dark current compensation δ for this batch dark =12. Instantaneous irradiance E at the site solar =890W·m -2 Standard reflectance R of calibration plate ref =0.28. Measured reflectance R of bare soil soil =0.14. Coarse measurement of canopy reflectance R at pixel. 710 =0.32. Calculate.

[0067]

[0068] The normalized gray level is set to 3.74, and after histogram stretching and mapping to the 0-255 range, 196 is obtained.

[0069] Perform symmetric differences of 3x3, 5x5, and 7x7 on the pixel and its neighborhood. Taking a 3x3 window as an example, the horizontal difference D... h =64, vertical difference D v =72. Gradient magnitude The values ​​obtained at the three scales were 96.6, 49.3, and 28.7, respectively. After morphological opening and closing reconstruction, the value of this pixel in the salient edge intensity matrix was 82.4, and the gradient direction was θ≈48°.

[0070] The near-infrared image pixel grayscale value is 131, which becomes 198 after histogram matching. A grayscale co-occurrence matrix is ​​constructed using a 3x3 window, yielding a contrast ratio of 4.1, homogeneity of 0.82, entropy of 1.67, and uniformity of 0.94. Empirical weights of 0.35, 0.25, 0.25, and 0.15 are used to synthesize texture intensity.

[0071] T nir =0.35×4.1+0.25×0.82+0.25×1.67+0.15×0.94=2.41

[0072] Within an 8x8 window, the minimum brightness is 158 and the maximum is 211, therefore brightness normalization is performed.

[0073]

[0074] Significant edge strengths range from a minimum of 15 to a maximum of 88, with normalized G. edge = (82.4-15) / (88-15) = 0.92. The minimum texture intensity is 1.3 and the maximum is 3.4. After normalization, T = (2.41-1.3) / (3.4-1.3) = 0.53.

[0075] The three-dimensional feature vectors are concatenated in a fixed order as (0.72, 0.92, 0.53). The logistic regression coefficient β = (-2.5, 3.8, 4.1, 1.9). Calculate the linear combination.

[0076] z=-2.5+3.8×0.72+4.1×0.92+1.9×0.53=3.16;

[0077] probability Pixels are marked as effective phenotypic regions of seeds and leaves.

[0078] The significant edge gradient direction of 48° is taken as the main diffusion direction. The diffusion coefficient is set as c = exp(-0.03 × G). edge The probability of this pixel is c = exp(-0.03 × 82.4) = 0.086. After 20 iterations, the probability decreases from 0.959 to a smoothed value of 0.944.

[0079] The local 5x5 entropy is 1.12, which is less than the threshold of 1.6, so no secondary median filtering is required.

[0080] The optimal observation period has been corrected from time series to the pre-grouting stage. Threshold 0.5. After smoothing, the probability is 0.944 ≥ 0.5, and pixels are written into a trait-specific mask. The entire mask contains 15032 pixels. Total projected area.

[0081] A total =15032 × 0.64 = 9620.5cm 2 ;

[0082] Eight-connectivity analysis retains 21 connected components, with an average connected component area of ​​[missing information].

[0083] A conn =9620.5 / 21 = 458.1cm 2 ;

[0084] For each connected component, the principal inertial axis is calculated to obtain the leaf angle set {35°, 41°, ...}, with an average leaf angle of 38°.

[0085] Archive threshold: area 9500cm² 2 The area of ​​the connected domain is 460 cm². 2 Leaf angle 42°. The three normalized values ​​are 1.013, 0.996, and 0.905 respectively. Weights are 0.4, 0.3, and 0.3. Overall ordinal value.

[0086] S=0.4×1.013+0.3×0.996+0.3×0.905=0.972;

[0087] The reference threshold was set to 1.00. The relative deviation (0.972 - 1.00) / 1.00 = -0.028, corresponding to a trait score of -2.8. A negative value at this point indicates that the area and angle are slightly below the target but within the acceptable range. The experimental plot yield was 10.8 kg, and the plot area was 15 m². 2 Yield per unit area: 720 g·m -2 The survival rate of mature plants was 0.93, and the field-level production potential coefficient was 669.6 g·m³. -2 .

[0088] The current decision table only contains this trait, therefore the high-efficiency breeding index

[0089] I = (-2.8) × 669.6 = -1875;

[0090] The median value for the experimental batch was -2030; values ​​above the median were considered good performance. Breeders can see the composite sequence value of P-2025-17 (0.972) and index (-1875) on the screening interface, and combine this with other phase results to decide whether to proceed to the next round of hybridization or multi-location trials.

[0091] Figure 2The correlation analysis results between the comprehensive breeding ordinal value calculated by the method of this invention and the field measured trait values ​​are presented in detail. The horizontal axis represents the actual trait values ​​measured manually in the field, ranging from 0 to 120, and the vertical axis represents the comprehensive breeding ordinal value calculated based on the total projected area, average connected area, and average leaf angle measurement, ranging from 0 to 100.

[0092] In the scatter plot, solid circles represent the measurement data of the total projected area, totaling 12 sample points, showing a clear linear positive correlation. Hollow circles represent the measurement data of the average connected region area, containing 5 representative sample points, and their distribution trend is basically consistent with that of the total projected area. Square markers represent the measurement data of the average leaf angle, also containing 5 sample points, showing good linear correlation characteristics.

[0093] The regression line obtained by least squares fitting runs through the entire data distribution range, showing a very strong linear correlation. The correlation coefficient R reaches 0.96, indicating a high positive correlation between the comprehensive breeding ordinal value and the measured trait value, verifying the accuracy and reliability of the method of this invention in trait quantification. The root mean square error is only 4.2, further confirming the excellent performance of the measurement accuracy.

[0094] The data distribution characteristics reveal that when the measured trait values ​​are low (0-40 range), the composite breeding sequence value is mainly distributed in the 20-40 range, demonstrating the method's sensitive identification ability for low-value traits. In the medium trait value range (40-80), the composite breeding sequence value shows a stable linear growth trend with a relatively constant slope. In the high trait value range (80-120), the composite breeding sequence value reaches a relatively high level of 60-90, but the growth rate slows slightly, which is consistent with the saturation effect characteristic commonly found in biological trait measurements.

[0095] Experimental data show that the method of the present invention can effectively capture the true variation pattern of maize phenotypic traits by accurately generating trait-specific masks and comprehensively calculating multidimensional measures, providing a reliable quantitative evaluation tool for efficient breeding.

[0096] Figure 3 This study comprehensively demonstrates the distribution characteristics and statistical patterns of the single-plant high-efficiency breeding index in the experimental population, obtained through a complete calculation process. The horizontal axis represents the distribution range of the single-plant high-efficiency breeding index, divided into eight intervals of 20 numerical units, ranging from 0-20 to 140-160. The vertical axis represents the number of plants falling into each index interval, with a maximum value of 52 plants.

[0097] The bar chart clearly shows that the high-efficiency breeding index of a single plant exhibits a typical normal distribution. In the low index range (0-40), the number of plants is relatively small, with 12 and 21 plants respectively, indicating that the proportion of plants with extremely low breeding potential is small. In the medium index range (40-80), the number of plants increases significantly, reaching 45 and 52 plants respectively, with the number of plants in the 60-80 range reaching a peak, reflecting that most plants in the population have a medium level of breeding potential.

[0098] In the higher index range (80-120), the number of plants showed a decreasing trend, with 38 and 28 plants respectively, indicating that the number of plants with high breeding potential decreased but still accounted for a considerable proportion. In the highest index range (120-160), the number of plants further decreased to 16 and 8 plants, which is consistent with the biological law that excellent germplasm resources are relatively scarce.

[0099] The fitting curve to the normal distribution shows a high degree of agreement between the actual data distribution and the theoretical normal distribution, validating the statistical rationality of the single-plant high-efficiency breeding index calculation method. The mean value of the distribution is 68.5, indicating that the overall breeding potential of the experimental population is at a moderately high level. The standard deviation is approximately 25, reflecting the degree of genetic diversity within the population.

[0100] The threshold for superior strains was set at 100 in the figure. Based on this standard, 52 plants, accounting for 24.3% of the total population, had an index value exceeding 100, providing ample candidate material for subsequent breeding selection. This proportion ensured both appropriate selection intensity and maintained sufficient genetic foundation, meeting the empirical requirements of breeding practice. The distribution characteristics indicate that the method of this invention can effectively distinguish plants with different breeding potentials, providing a reliable quantitative basis for precise breeding decisions.

[0101] Figure 4 This study demonstrates the variation in the trait-sensitive probability field recognition accuracy of maize canopy images acquired by a multispectral sensor UAV at different growth stages. The horizontal axis represents the five key growth stages of maize from the jointing stage to maturity, and the vertical axis represents the recognition accuracy value of the effective phenotypic region of the grain-leaf area, ranging from 0.6 to 1.0.

[0102] Experimental results show that the proposed trait-sensitive probability field generation method based on normalized grayscale images significantly improves recognition accuracy at different growth stages. At the jointing stage, due to the relatively simple canopy structure of maize plants and the clear boundary between leaves and stems, the recognition accuracy reaches 0.68. As the growth stage progresses, the recognition accuracy improves to 0.76 at the tasseling stage. This is mainly attributed to the multi-scale processing capability of the three sets of scale difference operators for gradient operations, which effectively captures edge features at different scales.

[0103] During the silking stage, the recognition accuracy reached a peak of 0.91, indicating that the effective phenotypic region features of the corn plant's kernel-leaf area were most significant at this stage, with the best fusion effect of brightness, salient edge intensity, and near-infrared texture. The recognition accuracy further improved to 0.94 during the grain-filling stage, indicating that the probability values ​​output by the soft discriminant model were most accurate during this period, and the unsupervised correction effect of neighborhood consistency screening on outlier probabilities was significant. The recognition accuracy slightly decreased to 0.89 during the maturity stage, due to the leaves beginning to age and turn yellow, resulting in changes in texture features, but it still remained at a relatively high level.

[0104] In contrast, traditional threshold segmentation methods show significantly lower recognition accuracy at all growth stages compared to the method of this invention. The traditional method achieves an accuracy of only 0.61 at the jointing stage, 0.68 at the tasseling stage, 0.74 at the silking stage, 0.73 at the grain-filling stage, and 0.71 at the maturity stage. Comparative data demonstrate that the method of this invention achieves a 10-20% improvement in accuracy across all growth stages, particularly at the critical silking and grain-filling stages, where accuracy improvements reach 23% and 29% respectively, fully validating the effectiveness of direction-guided smoothing and anisotropic diffusion smoothing techniques.

[0105] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding, characterized in that, The method includes: Step 1: During multiple different growth stages of maize from the jointing stage to maturity, images of the canopy of the same plant in the experimental field were acquired using a drone equipped with a multispectral sensor, and flight altitude, instantaneous irradiance and bare soil reflectance were collected simultaneously; the grayscale of the target band pixels was preprocessed to obtain a normalized grayscale image. Step 2: Based on the normalized grayscale image, perform gradient operation on the target band to obtain significant edge intensity, and fuse it with near-infrared texture; simultaneously consider the three features of brightness, significant edge intensity and near-infrared texture to determine the probability that each pixel belongs to the effective phenotypic region of grain-leaf, and generate the trait-sensitive probability field of the corresponding time phase. Step 3: Preset the optimal observation period for each target trait. Select a set of pixels with a probability value of not less than 0.5 and identified as the trait type from the images of the corresponding phase of the optimal observation period to form a trait-specific mask. For each trait, measure the total projected area, average connected area and average leaf angle in sequence to obtain the comprehensive breeding sequence value of the trait. In step 2, for each pixel, a three-dimensional feature vector is formed according to a fixed order of brightness, gradient, and texture, and this vector is input into a pre-trained offline soft discriminant model. This soft discriminant model adopts an improved multi-class logistic regression framework, using the effective phenotypic region label of the seed-leaf area as the positive class and the labels of other regions as the negative class, and outputs a continuous probability value between zero and one for the three-dimensional feature vector. When the peak-to-valley span of the overall brightness distribution of the normalized grayscale image is detected to exceed one standard deviation of the training distribution, the nearest neighbor consistency screening algorithm is immediately activated to evaluate the probability outlier in the local three-by-three pixel neighborhood. If the outlier exceeds the threshold, the probability of the center pixel is replaced with the mean of the neighborhood probability to achieve unsupervised adaptive correction. All pixel probabilities are concatenated into a shape-sensitive probability field. The comprehensive breeding sequence number is: ； in, For target traits The comprehensive breeding sequence value is used to evaluate the plant's performance on this trait; The total projected area of ​​all connected regions in the morphology-specific mask (in cm²). The average leaf angle (in degrees) of all connected regions. This represents the number of connected regions in the mask identified by the 8-connectivity marker; The standard deviation of the leaf angle values ​​reflects the consistency of the leaf angle. This represents the number of target shape pixels with a probability value of not less than 0.5 in the shape-sensitive probability field. This represents the total number of pixels that are classified as this trait out of all pixels.

2. The method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding as described in claim 1, characterized in that, The method further includes: Step 4: Compare the comprehensive breeding sequence value with the target threshold of the corresponding trait in the breeding archive, and calculate the trait score according to the relative deviation; For each experimental plot, use the measured yield divided by the area of ​​the experimental plot to obtain the yield per unit area, and multiply this value by the survival rate of mature plants to form the field-level production potential coefficient; Multiply the scores of all trait included in the decision by the field-level production potential coefficient one by one and sum them to obtain the single-plant high-efficiency breeding index; The higher the value of the single-plant high-efficiency breeding index, the closer the plant is to the breeding target.

3. The method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding as described in claim 2, characterized in that, In step 1, the chlorophyll-sensitive band with a wavelength of 710nm is selected as the target band, and the standard reflectance obtained by indoor calibration with the same band calibration plate is used as the absolute reference.

4. The method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding as described in claim 3, characterized in that, In step 1, based on the geometric boundaries of the experimental field, a grid-based flight path planning method is adopted to ensure that the canopy of each corn plant is photographed by at least three adjacent overlapping flight paths. To ensure that the images of the canopy of the same plant have spatial references that can be aligned at each growth stage, the lateral overlap is set to be no less than 75% and the lateral overlap is set to be no less than 85%.

5. The method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding as described in claim 4, characterized in that, Step 2 specifically includes: performing gradient operations on the normalized grayscale image using three sets of scale difference operators to obtain multi-scale gradient magnitudes; then using morphological opening and closing reconstruction to purify and synthesize them into a significant edge intensity matrix to highlight the grain outline and leaf edge; next, extracting grayscale co-occurrence texture indices from synchronized near-infrared images and linearly synthesizing them into near-infrared texture grids, and geometrically registering them to the normalized grayscale image to achieve pixel-level three-grid co-addressing of brightness, significant edge intensity, and near-infrared texture; then dynamically normalizing the three features for each pixel within a local window to construct a three-dimensional feature vector of brightness-gradient-texture and feeding it into a soft discriminant model to output the preliminary probability of the effective phenotypic region of the grain-leaf; if the on-site lighting or leaf color distribution is detected to be outside the training range, unsupervised correction of the outlier probability is performed through neighborhood consistency screening; finally, anisotropic diffusion smoothing is implemented based on the gradient direction of significant edge intensity to preserve real edges while suppressing noise.

6. The method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding as described in claim 5, characterized in that, The process of performing gradient operations on a normalized grayscale image using three sets of scale difference operators includes: selecting three spatial scales on the grid of the same normalized grayscale image; calculating the horizontal and vertical differences for each scale using symmetric difference operators of 3x3, 5x5, and 7x7 respectively; then synthesizing the gradient magnitudes of the horizontal and vertical differences at the same scale according to the Euclidean norm to obtain a set of gradient magnitudes and recording the corresponding gradient directions; performing an opening operation on the set of gradient magnitudes to filter out high-frequency isolated noise; and then retaining the structured edges through a continuous erosion-dilation cycle, finally outputting a significant edge intensity matrix.

7. The method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding as described in claim 6, characterized in that, In step 2, near-infrared band images acquired synchronously with the same time phase are read, and histogram matching is performed to ensure that their brightness distribution is consistent with the normalized grayscale image. Then, a grayscale co-occurrence matrix is ​​constructed using a 3x3 sliding window, and four texture indices—contrast, homogeneity, entropy, and uniformity—are calculated. These four texture indices are then linearly synthesized according to empirical weights to obtain a single scalar near-infrared texture grid. The near-infrared texture grid is geometrically registered to the pixel grid of the normalized grayscale image, and bilinear interpolation is used to correct sub-pixel deviations, ensuring that each pixel simultaneously possesses brightness, significant edge intensity, and near-infrared texture values ​​in subsequent calculations. To reduce data redundancy... Furthermore, a 1:4 pyramid downsampling was performed on all three feature grids to generate a multi-layer resolution pyramid sequence, so as to uniformly complete probability calculation and subsequent smoothing at different spatial granularities. At the original resolution level, three features, namely brightness, significant edge intensity and near-infrared texture, were extracted from each pixel and local dynamic normalization was performed on each. The local dynamic normalization adopted an eight-by-eight window, and linear stretching was performed with the minimum and maximum values ​​within the window as endpoints, so that each feature value was compressed into a closed interval of zero to one and the relative differences were preserved, so as to weaken the influence of brightness drift between different plants, while maintaining the sharp contrast of gradient and texture at the leaf edge and grain surface.

8. The method for image recognition and extraction of phenotypic traits for high-efficiency maize breeding as described in claim 1, characterized in that, After concatenating all pixel probabilities into a preliminary trait-sensitive probability field, directional guided smoothing is performed on the preliminary trait-sensitive probability field. The directional guided smoothing process uses the gradient direction provided by the significant edge intensity matrix as the main smoothing direction and performs anisotropic diffusion on the pixel probabilities along the gradient direction. At the same time, diffusion perpendicular to the gradient direction is suppressed to preserve the true leaf edge and grain outline. The diffusion coefficient decays exponentially according to the gradient magnitude, which effectively controls the boundary blurring caused by diffuse reflection in leaf texture; after completing the direction-guided smoothing, the local entropy of the preliminary trait-sensitive probability field is calculated to evaluate the non-converged region. If the local entropy exceeds the set threshold, a second denoising process based on fast median filtering is performed on the non-converged region, with the median filter kernel size increasing linearly with the entropy value.

Images