Phenotypic character image recognition and extraction method for efficient corn breeding

Abstract

The invention relates to the technical field of image recognition, and further relates to a phenotypic character image recognition and extraction method for efficient corn breeding. The method comprises the following steps: step 1, acquiring an image of the same plant canopy in a test field by using an unmanned aerial vehicle carrying a multispectral sensor in a plurality of different growth periods from a corn jointing period to a mature period, so as to obtain a normalized grayscale image; 2, performing gradient operation on a target wave band based on the normalized grayscale image to obtain significant edge intensity, and generating a character sensitive probability field of a corresponding time phase; and 3, presetting an optimal observation period for each target character, selecting a pixel set which has a probability value not less than 0.5 and is judged to be the character type from an image of a time phase corresponding to the optimal observation period, and sequentially measuring the total projection area, the average connected domain area and the average leaf angle to obtain a comprehensive breeding sequence value of the character. According to the method, high-precision probability identification of the effective phenotype area of the grain-leaf is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a phenotypic trait image recognition and extraction method for efficient corn breeding. Background Art

[0002] In the modern agricultural production system, corn is one of the main food crops in the world, and its efficient breeding has always been a key research direction in the field of agricultural science and technology. With the continuous development of molecular breeding technology and genetic improvement systems, the accurate, automated identification and quantitative evaluation of phenotypic traits have gradually become a bottleneck limiting the improvement of corn breeding efficiency. In the traditional corn phenotypic observation system, manual measurement and empirical interpretation are dominant. Not only is the work intensity high, the cycle is long, and the repeatability is poor, but it is also difficult to ensure objectivity and consistency in large-scale field trials. In particular, when multiple phenotypic dimensions (such as grain density, leaf morphology, and plant structure) need to be evaluated simultaneously, traditional methods are difficult to form high-throughput data support, which seriously restricts the efficient screening of corn populations.

[0003] To address this issue, some researchers have attempted to incorporate computer vision and image processing techniques in recent years to identify and extract plant traits from visible light images. One mainstream approach analyzes corn plant structure based on two-dimensional RGB images, for example extracting visual traits such as leaf area and ear height through color segmentation and edge detection. While such methods have a certain degree of accuracy in controlled indoor environments, they are susceptible to interference from factors such as sunlight angle, soil background reflectivity, and leaf color variation in complex field environments, leading to increased recognition errors. Furthermore, visible light images are insensitive to structural reflectivity differences between different corn tissues (such as leaves and kernels), making it difficult to support effective spatial segmentation and texture classification.

[0004] Another more advanced technical solution involves remote sensing phenotyping of corn canopies using hyperspectral or multispectral imaging systems. Hyperspectral imaging captures spectral reflectance across continuous bands and can be used to extract phenotypic indicators such as plant physiological status, nitrogen content, and moisture content. However, hyperspectral equipment is expensive, generates large amounts of data, and requires a highly stable platform (such as a fixed overhead system), making it difficult to widely deploy in field trials. In contrast, multispectral imaging, which balances spectral characteristics and system cost, has become a more feasible solution for field phenotypic data collection in recent years. In related technologies, some studies have used drones equipped with multispectral sensors to acquire corn canopy reflectance data and combined it with vegetation indices such as NDVI and GNDVI to monitor plant growth. While vegetation indices are effective for overall trend analysis, they lack the ability to provide high-resolution characterization of individual plant morphology and structure, making it difficult to directly quantitatively correlate them with target breeding traits. Consequently, phenotypic decision support at the individual plant level is not possible. Summary of the Invention

[0005] The main purpose of this invention is to provide a phenotypic trait image recognition and extraction method for efficient corn breeding. This method uses an unmanned aerial vehicle (UAV) equipped with a multispectral sensor to capture canopy images of the same plant at different growth stages. Through normalization processing, multiscale gradient calculations, near-infrared texture fusion, and three-dimensional feature vector construction, it achieves high-precision probabilistic identification of the effective phenotypic region of the grain and leaf. The optimal observation period is dynamically determined by combining time series, generating trait-specific masks, and quantifying the total projected area, average connected domain area, and average leaf angle. Ultimately, the method outputs a comprehensive breeding ordinal value and a high-efficiency breeding index for each plant. 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] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0007] A phenotypic trait image recognition and extraction method for efficient corn breeding, the method comprising:

[0008] Step 1: At various growth stages of corn, from jointing to maturity, a drone equipped with a multispectral sensor was used to capture canopy images of the same plant in the experimental field. Flight altitude, instantaneous irradiance, and bare soil reflectance were also 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, a gradient operation is performed on the target band to obtain the significant edge intensity, which is then fused with the near-infrared texture. Simultaneously, considering the brightness, significant edge intensity, and near-infrared texture features, the probability of each pixel belonging to the effective seed-leaf phenotypic region is determined, generating a trait sensitivity probability field for the corresponding time phase.

[0010] Step 3: Preset the optimal observation period for each target trait, and select a set of pixels with a probability value of not less than 0.5 and determined to be the trait type from the image of the phase corresponding to the optimal observation period to form a trait-specific mask; for each trait, perform the measurements of total projected area, average connected domain area, and average leaf angle in sequence to obtain the comprehensive breeding ordinal value of the trait.

[0011] Furthermore, the method also includes: Step 4: comparing each comprehensive breeding ordinal value with the target threshold of the corresponding trait of the same variety in the breeding file, and calculating the trait score according to the relative deviation; for each experimental plot, dividing the measured yield by the area of ​​the experimental plot to obtain the yield per unit area, and multiplying this value with the survival rate of the adult plant to form a field-level production potential coefficient; multiplying all the trait scores included in the decision by the field-level production potential coefficient one by one and summing them up to obtain the single-plant efficient breeding index; the higher the value of the single-plant efficient breeding index, the closer the plant is to the breeding target.

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

[0013] Furthermore, in step 1, a grid-type flight path planning method was used based on the geometric boundaries of the experimental field to ensure that the canopy of each corn plant was photographed by at least three adjacent overlapping flight paths. To ensure that the images of the canopy of the same plant had an alignable spatial reference during each growth period, the lateral overlap was set to no less than 75% and the heading overlap was set to no less than 85%.

[0014] Furthermore, step 2 specifically includes: performing gradient operations on the normalized grayscale image through three sets of scale difference operators to obtain multi-scale gradient amplitudes; 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; then extracting grayscale symbiotic texture indicators from the synchronized near-infrared image and linearly synthesizing them into a near-infrared texture grid, and geometrically aligning them to the normalized grayscale image to achieve pixel-level co-location of brightness, significant edge intensity and near-infrared texture; then dynamically normalizing the three features of each pixel in the local window to construct a brightness-gradient-texture three-dimensional feature vector and send it to the soft discriminant model to output the preliminary probability of the effective phenotypic area of ​​the grain-leaf; if it is detected that the on-site illumination or leaf color distribution exceeds the training range, the outlier probability is unsupervisedly corrected through neighborhood consistency screening; finally, anisotropic diffusion smoothing is performed according to the gradient direction of the significant edge intensity to retain the true edge and suppress noise.

[0015] Furthermore, the process of performing gradient operations on the normalized grayscale image through three sets of scale difference operators includes: selecting three groups of spatial scales on the grid of the same normalized grayscale image, and using three-by-three, five-by-five, and seven-by-seven symmetric difference operators to calculate the horizontal difference and vertical difference for each group of scales, and then synthesizing the gradient amplitude of the horizontal and vertical differences of the same scale according to the Euclidean norm to obtain a gradient amplitude set, and recording the corresponding gradient direction; performing an opening operation on the gradient amplitude set to filter out high-frequency isolated noise, and then retaining the structured edges through continuous erosion-dilation cycles, and finally outputting a significant edge intensity matrix.

[0016] Furthermore, in step 2, the near-infrared band image collected synchronously with the same phase is read, and histogram matching is performed to make its brightness distribution consistent with the normalized grayscale image; then a grayscale co-occurrence matrix is ​​constructed with a three-by-three sliding window, and four sets of texture indicators, namely contrast, homogeneity, entropy and uniformity, are calculated, and the four sets of texture indicators are linearly synthesized according to empirical weights to obtain a near-infrared texture grid in the form of a single scalar; the near-infrared texture grid is geometrically aligned to the pixel grid of the normalized grayscale image, and bilinear interpolation is used to correct the sub-pixel deviation to ensure that each pixel has a brightness value, a significant edge intensity value and a near-infrared texture value in subsequent calculations; in order to reduce To account for data redundancy, a one-to-four pyramid downsampling was performed on all three feature grids to generate a multi-layer resolution pyramid sequence, so that probability calculation and subsequent smoothing could be uniformly completed at different spatial granularities. At the original resolution level, three features, namely brightness, significant edge intensity, and near-infrared texture, were extracted for each pixel, and local dynamic normalization was performed on each pixel. Local dynamic normalization used an eight-by-eight window and linearly stretched it with the minimum and maximum values ​​in the window as endpoints, so that each feature value was compressed to a closed interval from zero to one while retaining relative differences, thereby weakening the impact of brightness drift between different plants and maintaining a sharp contrast between gradients and textures at the leaf edge and grain surface.

[0017] Furthermore, in step 2, for each pixel, a three-dimensional feature vector is formed in a fixed order of brightness, gradient, and texture, and the vector is input into a soft discriminant model trained offline in advance; the soft discriminant model adopts an improved multi-class logistic regression framework, with the grain-leaf effective phenotypic area label as the positive class and other area labels 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-to-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 degree in a local three-by-three pixel neighborhood; if the outlier degree exceeds the threshold, the probability of the central pixel is replaced by the neighborhood probability mean to achieve unsupervised adaptive correction.

[0018] Furthermore, after all pixel probabilities are spliced ​​into a preliminary trait sensitive probability field, direction-guided smoothing is performed on the preliminary trait sensitive probability field. The direction-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, the diffusion perpendicular to the gradient direction is suppressed to retain the true leaf edge and grain outline. The diffusion coefficient decays exponentially according to the gradient amplitude, so that the boundary blurring caused by diffusion in the leaf texture is effectively controlled. After completing the direction-guided smoothing, the local entropy of the preliminary trait sensitive probability field is calculated to evaluate the non-convergence area. If the local entropy exceeds the set threshold, secondary denoising based on fast median filtering is implemented on this non-convergence area, and the median filter kernel size increases linearly with the entropy value.

[0019] The present invention's phenotypic trait image recognition and extraction method for efficient corn breeding offers the following beneficial effects: it enables continuous observation and high-throughput trait quantification of the same corn plant across 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 probabilistic smoothing, and trait-specific mask generation, forming a fully closed-loop process from drone acquisition to trait index output. This method achieves pixel-level co-location of brightness, edge, and texture features, and outputs a continuous probability value for each pixel using a soft discriminant model, significantly improving the accuracy of kernel and leaf recognition and the continuity of boundary representation. Furthermore, the optimal observation period is dynamically adjusted based on the probability peak of the time series, ensuring that the target trait is extracted within the optimal significance window. For trait quantification, the total projected area, average connected domain area, and average leaf angle are measured within the mask space domain to achieve joint modeling of canopy scale, structure, and posture, ultimately outputting a structured, integrated breeding ordinal value. Compared with traditional phenotypic identification methods based on manual measurement or single image processing, the present invention has a high degree of automation, high robustness and strong repeatability, can adapt to different environmental lighting and variety differences, and greatly improves the screening efficiency and evaluation accuracy at the corn population level. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of a method flow for a phenotypic trait image recognition and extraction method for efficient corn breeding provided by an embodiment of the present invention;

[0021] Figure 2 Schematic diagram of an experiment for analyzing the correlation between comprehensive breeding ordinal values ​​and measured traits provided by an embodiment of the present invention;

[0022] Figure 3 Schematic diagram of the experiment of the distribution characteristics of the high-efficiency breeding index of a single plant provided by an embodiment of the present invention;

[0023] Figure 4 Schematic diagram of the experiment comparing the accuracy of sensitive probability fields of traits in different growth periods. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0025] refer to Figure 1:A phenotypic trait image recognition and extraction method for efficient corn breeding, the method comprising:

[0026] Step 1: At various growth stages of corn, from jointing to maturity, a drone equipped with a multispectral sensor was used to capture canopy images of the same plant in the experimental field. Flight altitude, instantaneous irradiance, and bare soil reflectance were also 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, a gradient operation is performed on the target band to obtain the significant edge intensity, which is then fused with the near-infrared texture. Simultaneously, considering the brightness, significant edge intensity, and near-infrared texture features, the probability of each pixel belonging to the effective seed-leaf phenotypic region is determined, generating a trait sensitivity probability field for the corresponding time phase.

[0028] Step 3: Preset the optimal observation period for each target trait, and select a set of pixels with a probability value of not less than 0.5 and determined to be the trait type from the image of the phase corresponding to the optimal observation period to form a trait-specific mask; for each trait, perform the measurements of total projected area, average connected domain area, and average leaf angle in sequence to obtain the comprehensive breeding ordinal value of the trait.

[0029] Corn has five key growth stages: jointing, tasseling, early grain filling, late grain filling, and maturity. When a multispectral sensor periodically observes the same plant canopy in the visible and near-infrared wavelengths, differences in chlorophyll content, moisture content, and the microstructural structure of the grain wax coat alter the multiple scattering paths of photons in the intercellular spaces and mesophyll spongy tissue, resulting in a reflectance distribution in the sensor's line of sight that changes dynamically with the growth stage. This distribution, combined with solar incident angle, instantaneous irradiance, and soil background diffuse reflectance, makes it difficult to separate varietal genetic differences from environmental random noise without correction. Therefore, by simultaneously collecting flight altitude, instantaneous irradiance, and bare soil reflectance, and normalizing the pixel grayscale of the target band into a normalized grayscale image, a simplified radiative transfer inversion is implemented at the image level. This compresses the geometric degradation term and light response term in the remote sensing radiation range into controllable constants, making the plant reflectance characteristics at different time sections directly comparable in spectral space.

[0030] The significant edge intensity extracted using a multiscale gradient operator can then highlight the spectral transition from the kernel to the leaf surface. This transition arises from differences in tissue density and surface roughness, and its amplitude varies monotonically and segmentally along the wavelength dimension, resulting in a gradient pattern in the spatial domain. Near-infrared texture, on the other hand, quantifies the spatial uniformity of the arrangement of mesophyll bullae and vascular bundles, as well as the water distribution, using a second-order statistical dimension. This textural information complements the starch content of the kernel. The fusion of brightness, significant edge intensity, and near-infrared texture features is equivalent to constructing a set of local feature tensors along three mutually orthogonal characteristic axes: spectral, geometric, and statistical. Subsequently, probability density mapping is used to softly segment the kernel-leaf effective phenotypic region. The resulting trait-sensitive probability field is essentially a Markov random field approximation of each canopy surface element in attribute space. Soft classification avoids the fragmentation problem of traditional hard threshold segmentation in heterogeneous canopies.

[0031] To overcome the problem that phenotypic significance periods for different traits do not always coincide with the same growth period, the system introduces a dynamic calibration mechanism based on extreme values ​​of the probability mean of time series. This concept is consistent with the reproductive-nutrient competition model: phenotypic peaks for different traits correspond to different source-sink equilibrium points. Time offsets reflect the overall delay or advance of plant physiological processes due to temperature, precipitation, or management measures in that year. By adjusting the optimal observation period in real time, subsequent trait-specific masks can consistently cover the most informative time phases. Once the trait-specific mask is determined, the method uses two-dimensional projected area as a direct measure of biomass spatial projection. The average connected domain area describes the dispersion and fragmentation of seed or leaf clusters, which corresponds to the consistency of seed filling and the spatial aggregation of leaf expansion. The average leaf angle reflects the orientation of photosynthetic organs and the relationship with shading, and is a key geometric parameter for canopy light interception efficiency. Together, these three factors determine the degree of match between carbon assimilation rate and seed loading rate. The comprehensive breeding ordinal value maps these three measures to a unified dimension and weights them according to the genetic contribution rate. In essence, 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 three-dimensional combination, the distance with the target phenotypic outline is calculated in the morphological space. The smaller the distance, the higher the ordinal value, which is more in line with the complex requirements of efficient corn breeding for uniform ears, reasonable leaf posture, and canopy transparency.

[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 is based on statistical data from experimental fields over the years and maps the key growth periods from jointing to maturity to the phenotypic significant peaks of the target trait. When entering the online observation phase, the system calls the time tag of the optimal observation period according to the index table, actively retrieves the image of the same plant canopy under the time tag, and synchronously loads the trait-sensitive probability field of the corresponding time phase. After the probability field is loaded, the system scans the probability grid line by line, reads the probability value for each pixel, and when the probability value is not less than 0.5 and the pixel has been determined to be of the trait type by the previous classifier, the pixel is written to the pixel set that has been determined to be of the trait type and assigned a value of one on a blank grid of the same size, and the remaining pixels are assigned a value of zero. Because the probability field may produce isolated pixels in the edge transition zone, the system immediately calls the three-by-three structural 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 grid. The system then applies the eight-connected labeling algorithm to the binary grid, calculates the pixel numbers of all connected domains, and writes them into the attribute table with the pixel number as the key, finally outputting a complete and non-redundant trait-specific mask.

[0033] In order to ensure the accuracy of the physical scale of the total projected area, after generating the trait-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 when the image was acquired, and accumulates and sums the area of ​​each pixel assigned a value of one to obtain the total projected area. The system then traverses the connected domains in the attribute table, counts the number of pixels contained in each connected domain, and multiplies the number of pixels by the single-pixel ground area to obtain the area of ​​the connected domain. After counting the areas of all connected domains, connected domains with an area of ​​less than 25 pixels are removed as noise, and the average area of ​​the retained connected domains is recalculated. The resulting value is the average connected domain area. The number of connected domains is retained in the attribute table for subsequent comparisons, but is not included in the subsequent leaf angle measurement.

[0034] To calculate the average leaf angle, the system performs a centroid-aligned second-order moment analysis for each connected domain in the attribute table. The method first performs a centroid translation on the pixel coordinates of the connected domain, then calculates the second-order central moment matrix, and obtains the direction vector of the principal inertia axis through singular value decomposition. The absolute value of the angle between this direction vector and the horizontal coordinate axis is defined as the leaf angle of the connected domain. The system accumulates the leaf angles of all retained connected domains and divides them by the number of connected domains to obtain the average leaf angle. If the number of connected domains is zero, the system directly sets the average leaf angle to zero according to the fault-tolerant logic and warns in the log that there is no available leaf angle information in the current phase. In addition, the system also calculates the standard deviation of the leaf angle and writes it to the attribute table for studying the consistency of canopy posture, but this standard deviation does not participate in the calculation of the next comprehensive breeding ordinal value.

[0035] Once the total projected area, average connected domain area, and average leaf angle are all obtained, the system starts the comprehensive breeding ordinal value calculation module. The module first searches the database for the historical threshold entries corresponding to the current target trait. The historical thresholds include area thresholds and leaf angle thresholds, which are used to normalize the three measurements 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 domain 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 to one. After normalization is completed, the module performs a weighted summation of the three normalized values ​​according to the weights stored in the threshold entries. The sum of the weights is always one, and the sum result is the comprehensive breeding ordinal value. The comprehensive breeding ordinal value is written into the breeding database simultaneously with the experimental field number, plant number, and growth period label, and is sorted among the plants in the same batch. The higher the ordinal value, the more the plant meets the breeding target in terms of the current target trait. Breeders can directly use the ranking results to screen high-quality genotypes for the next round of field verification or mating combination design.

[0036] After writing the comprehensive breeding ordinal value, 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 the complete spatial information and measurement process can be restored during subsequent tracing. If breeders need visual comparisons during the decision-making stage, the system can generate dynamic time series layers based on the trait sensitivity probability field and trait-specific masks of the same plant at different growth stages. This can show the change trend of the phenotype over time and help determine the stability and plasticity of the target trait during the growth process, thereby achieving comprehensive quantification and rapid screening of efficient corn breeding indicators.

[0037] Furthermore, the method also includes: Step 4: comparing each comprehensive breeding ordinal value with the target threshold of the corresponding trait of the same variety in the breeding file, and calculating the trait score according to the relative deviation; for each experimental plot, dividing the measured yield by the area of ​​the experimental plot to obtain the yield per unit area, and multiplying this value with the survival rate of the adult plant to form a field-level production potential coefficient; multiplying all the trait scores included in the decision by the field-level production potential coefficient one by one and summing them up to obtain the single-plant efficient breeding index; the higher the value of the single-plant efficient 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 ordinal value and normalizing it by the target threshold. The relative deviation is multiplied by 100% and converted to 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 when multiple traits are subsequently accumulated. 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 then extracts the adult plant survival rate for that plot from the growth period monitoring file. These two values ​​are multiplied together to obtain the plot-level productivity coefficient. The plot-level productivity coefficient is stored in the table field corresponding to the experimental plot number and is stamped with the yield station number and measurement timestamp to facilitate traceability and error diagnosis. Once all trait scores and plot-level productivity coefficients are available, the system selects the trait scores to be included in the evaluation according to a pre-set decision list. These trait scores are then multiplied by the plot-level productivity coefficient one by one in the order of their entries. The cumulative sum is then written into the single-plant efficient breeding index field. At this time, each plant only generates a unique single-plant efficient breeding index, and the numerical range is limited to zero to the theoretical upper limit. The theoretical upper limit depends on the product of the trait score upper limit and the field-level production potential coefficient upper limit. In order to prevent extreme values ​​from misleading the breeding ranking, the system calls the anomaly detection module to perform a box test and a triple median absolute deviation test on the single-plant efficient breeding index after the calculation is completed. If an abnormally high value exceeding the upper limit or an abnormally low value below the lower limit is found, these records will be marked as pending review and prompted in red on the user interface. For the data that passes the test, the system arranges the single-plant efficient breeding index in descending order and generates a visual report, which simultaneously presents the plant number, trait score list, field-level production potential coefficient, and single-plant efficient breeding index, so that breeders can quickly screen out individuals with outstanding performance or those that need to be eliminated.

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

[0040] The normalized grayscale image is:

[0041]

[0042] Among them, G norm (i, j, t) is the normalized grayscale value of pixel (i, j) at time t, corresponding to the target band; G raw (i, j, t) is the original pixel grayscale value collected by the multispectral sensor; δ dark (t) is the time-dependent correction value used to suppress the flying dark current offset; E solar (t) is the instantaneous irradiance at time t (unit: W / m 2 );R ref (λ710 ) is the standard reflectivity of the indoor calibration plate at a wavelength of 710 nm; H(t) is the flight altitude (m) of the UAV; H ref It is the preset standard shooting height; R soil (t) is the soil reflectance of the bare soil area in the field; R 710 (i, j, t) is the 710 nm reflectance of the canopy at the pixel position; ∈ is a stability constant to prevent the denominator from being zero; i is the abscissa; j is the ordinate.

[0043] Furthermore, in step 1, a grid-type flight path planning method was used based on the geometric boundaries of the experimental field to ensure that the canopy of each corn plant was photographed by at least three adjacent overlapping flight paths. To ensure that the images of the canopy of the same plant had an alignable spatial reference during each growth period, the lateral overlap was set to no less than 75% and the heading overlap was set to no less than 85%.

[0044] After the track is generated, the module further inserts a buffer segment at the turning point of each track to ensure that the drone maintains a uniform speed and stable posture within the turning radius, avoiding image tilt due to pitch and panning. In order 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-shaped route, and compares the calculated heading overlap with the preset lower limit of 85% point by point; if the heading overlap of some track segments is found to be lower than 85%, the grid of that segment is regenerated by inserting an intermediate track or shortening the track spacing until the global heading overlap meets the requirements. On this basis, the system uses a grid coverage algorithm to count the number of times the canopy of each corn plant is covered by a track in the instantaneous field of view. When the number is less than three adjacent tracks, a compensating track is inserted in the local area. The starting and ending points of the compensating track are arranged along the center line of the grid to minimize the total flight time increase caused by the added track. After all tracks and compensation tracks are completed, the system writes the track information into the mission file, including latitude and longitude coordinates, altitude, heading angle, speed, and exposure trigger interval, and loads it into the drone's autopilot before flight. By combining raster route planning with high overlap constraints, the system ensures that each corn plant canopy is overlapped by at least three adjacent tracks throughout the observed growth period. This provides accurate and redundant spatial anchor points for the same plant canopy in the 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] Furthermore, step 2 specifically includes: performing gradient operations on the normalized grayscale image through three sets of scale difference operators to obtain multi-scale gradient amplitudes; 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; then extracting grayscale symbiotic texture indicators from the synchronized near-infrared image and linearly synthesizing them into a near-infrared texture grid, and geometrically aligning them to the normalized grayscale image to achieve pixel-level co-location of brightness, significant edge intensity and near-infrared texture; then dynamically normalizing the three features of each pixel in the local window to construct a brightness-gradient-texture three-dimensional feature vector and send it to the soft discriminant model to output the preliminary probability of the effective phenotypic area of ​​the grain-leaf; if it is detected that the on-site illumination or leaf color distribution exceeds the training range, the outlier probability is unsupervisedly corrected through neighborhood consistency screening; finally, anisotropic diffusion smoothing is performed according to the gradient direction of the significant edge intensity to retain the true edge and suppress noise.

[0046] A multiscale difference operator was first used to construct a set of gradient response fields. Its theoretical basis is that the first-order differential of an image exhibits grayscale transitions at tissue boundaries. By calculating horizontal and vertical differences at fine, medium, and coarse scales and taking the maximum amplitude, the multi-level intensity discontinuity surface formed by the leaf margin, main vein, and kernel shell can be fully captured. However, the original gradient field often contains high-frequency noise and isolated responses. This is because sensor readout errors and field scattered light cause local brightness changes to appear simultaneously with physical boundaries. Therefore, morphological opening and closing reconstruction is used to impose structural constraints on the gradient amplitudes at different scales. The opening operation removes small bright spots, while the closing operation fills small dark holes. The reconstruction process is iteratively constrained by the topology of the connected domain. The resulting retained salient edge intensity matrix no longer contains random spikes but instead highlights the continuous kernel outline and leaf margin. This provides a stable geometric saliency measure for subsequent discrimination.

[0047] In parallel with gradient information, near-infrared texture needs to be introduced to describe differences in plant moisture and tissue density. The near-infrared band is sensitive to scattering from the spongy tissue of the mesophyll, but is weaker to the specular reflection of the waxy seed surface. Calculating the grayscale co-occurrence matrix within a sliding window can quantify the joint distribution of grayscale levels in space. Entropy reflects texture complexity, contrast reflects the amplitude of grayscale differences, and uniformity reflects pattern repeatability. The linear synthesis of the three can make the statistical differences between the delicate texture inside the leaf and the relatively smooth surface of the seed explicit. Due to slight posture differences between the near-infrared image and the visible light image during the shooting process, rigid registration is required to maintain pixel-level co-location. This process relies on geometric anchor points that can be repeatedly observed in overlapping tracks of the same plant canopy. The near-infrared texture grid is mapped to the coordinate system of the normalized grayscale image through least squares matching to achieve co-location of the three grids.

[0048] When fusing brightness, significant edge intensity, and near-infrared texture, global drift caused by time-varying illumination must be addressed. Local dynamic normalization is based on the adaptive contrast mechanism of the human visual system, which rescales signal intensity using local minima and maxima within a limited receptive field, ensuring that feature values ​​reflect relative contrast rather than absolute brightness. After linear stretching through an eight-by-eight window, the three features of each pixel are compressed to a closed range between zero and one while preserving their relative magnitudes. This provides a comparable scale for the subsequent soft discriminant model to construct a three-dimensional feature vector of the brightness gradient texture. The soft discriminant model uses a continuous probabilistic output rather than a hard threshold. Its statistical basis can be viewed as a likelihood ratio estimate of the distribution of positive and negative samples in the three-dimensional feature space. The output value represents the posterior probability that the pixel belongs to the effective phenotypic region of the grain leaf. This probabilistic partitioning yields intermediate values ​​within the leaf edge gradient, avoiding the spurious boundaries caused by traditional binary classification in heterogeneous canopies.

[0049] In real-world field environments, sudden changes in illumination or leaf color can cause feature distributions to deviate from the training sample, leading to local saturation or holes in the output of the soft discriminant model. Neighborhood consistency screening, based on the theoretical framework of Markov random field smoothing, assumes that the probability distributions of neighboring pixels should be spatially consistent. The algorithm first examines the probability variance within a three-by-three neighborhood. When the variance is significantly greater than the variance of the overall image mean, the central pixel is identified as an outlier and then replaced with the neighborhood mean. This unsupervised correction, which does not rely on external labels, can quickly restore probability continuity in anomalous scenarios derived from the model. The resulting preliminary probabilities still contain point-like noise and step artifacts, requiring anisotropic diffusion smoothing based on the directional information provided by significant edge strength. Anisotropic diffusion is based on partial differential equations and uses an exponentially decaying diffusion coefficient in the gradient direction to suppress boundary blurring. A higher diffusion coefficient in the direction perpendicular to the gradient smoothes internal random noise. This method preserves the sharp edges between the kernel and the leaf margin while forming a coherent probability plateau within the leaf, providing a morphologically stable probability peak for retrieving the optimal observation period in the time series.

[0050] The preliminary probability of the effective phenotypic region of grains and leaves is:

[0051]

[0052] Among them, P i，j，t is the probability that pixel (i, j) belongs to the effective seed-leaf phenotype region at time t (range 0 to 1); L i，j，t is the normalized grayscale image brightness feature obtained in step 1; G edge (i, j, t) is the significant edge intensity value obtained after performing multi-scale gradient operation on the normalized image; T nir (i, j, t) is the composite value of texture features extracted from the near-infrared image (such as the weighted sum of entropy and contrast); μ L(t), μ G (t), μ T (t) represents the global average value 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 through three sets of scale difference operators includes: selecting three groups of spatial scales on the grid of the same normalized grayscale image, and using three-by-three, five-by-five, and seven-by-seven symmetric difference operators to calculate the horizontal difference and vertical difference for each group of scales, and then synthesizing the gradient amplitude of the horizontal and vertical differences of the same scale according to the Euclidean norm to obtain a gradient amplitude set, and recording the corresponding gradient direction; performing an opening operation on the gradient amplitude set to filter out high-frequency isolated noise, and then retaining the structured edges through continuous erosion-dilation cycles, and finally outputting a significant edge intensity matrix.

[0054] The fundamental purpose of applying three sets of scale difference operators in parallel on the same normalized grayscale image is to maintain detectable response peaks for structured boundaries such as leaf edges and kernel shells against a multi-band texture background. The three-by-three, five-by-five, and seven-by-seven symmetric difference operators simulate the first-order derivatives of the image at fine, medium, and coarse grains, respectively. Fine grain is most sensitive to high-curvature details such as leaf tips and main veins, medium grain considers leaf edges and small kernel clusters, and coarse grain provides a stable response to wide, gentle boundaries formed by kernel 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 modulus of a two-dimensional vector, thereby simultaneously characterizing boundary strength and direction, thus avoiding the directional bias caused by one-way differences. The spatial superposition of the gradient magnitudes obtained from each of the three scales is equivalent to merging the multi-scale Laplace zero-crossings of the same image into a unified coordinate system, allowing structural boundaries of different sizes to be cumulatively enhanced while random noise cancels out. The morphological opening operation first matches small bright spot noise with structuring elements and removes it. Then, through an erosion-dilation cycle, the weakened true boundary is compensated by an equal amount. This process follows the set contraction-expansion homeomorphism of mathematical morphology, ensuring that only gradient peaks above the noise threshold and with a continuous distribution are retained. The resulting salient edge intensity matrix thus becomes a geometrically consistent solution across scale thresholds. It can fully depict the outer ring of corn kernels and leaf edges, while maintaining extremely 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 adaptive intensity for the subsequent co-location fusion of brightness, gradient, and texture grids.

[0055] Furthermore, in step 2, the near-infrared band image collected synchronously with the same phase is read, and histogram matching is performed to make its brightness distribution consistent with the normalized grayscale image; then a grayscale co-occurrence matrix is ​​constructed with a three-by-three sliding window, and four sets of texture indicators, namely contrast, homogeneity, entropy and uniformity, are calculated, and the four sets of texture indicators are linearly synthesized according to empirical weights to obtain a near-infrared texture grid in the form of a single scalar; the near-infrared texture grid is geometrically aligned to the pixel grid of the normalized grayscale image, and bilinear interpolation is used to correct the sub-pixel deviation to ensure that each pixel has a brightness value, a significant edge intensity value and a near-infrared texture value in subsequent calculations; in order to reduce To account for data redundancy, a one-to-four pyramid downsampling was performed on all three feature grids to generate a multi-layer resolution pyramid sequence, so that probability calculation and subsequent smoothing could be uniformly completed at different spatial granularities. At the original resolution level, three features, namely brightness, significant edge intensity, and near-infrared texture, were extracted for each pixel, and local dynamic normalization was performed on each pixel. Local dynamic normalization used an eight-by-eight window and linearly stretched it with the minimum and maximum values ​​in the window as endpoints, so that each feature value was compressed to a closed interval from zero to one while retaining relative differences, thereby weakening the impact of brightness drift between different plants and maintaining a sharp contrast between gradients and textures at the leaf edge and grain surface.

[0056] Near-infrared images and normalized grayscale images exhibit natural differences in imaging spectra, exposure curves, and sensor response functions. Directly stitching these images together would result in inconsistent brightness scales, undermining the homology assumption used in subsequent probabilistic modeling. Therefore, the brightness distribution of the near-infrared image is first remapped to the distribution space of the normalized grayscale image via histogram matching. This approach maintains pixel ordering through a monotonic mapping of the cumulative distribution function, thereby achieving cross-band radiometric alignment without introducing spurious gradients. After brightness alignment, the second-order statistical texture features of plant tissues contained in the near-infrared image need to be converted into measurable quantities. A gray-level co-occurrence matrix is ​​constructed using a three-by-three sliding window to capture the joint probability distribution of pixel grayscale pairs within their spatial neighborhood. Contrast measures grayscale level differences, revealing texture roughness; homogeneity measures grayscale distances, reflecting pattern smoothness; entropy measures information disorder, signaling tissue complexity; and evenness measures co-occurrence concentration, characterizing pattern regularity. These four complementary dimensions describe the microscopic heterogeneity of mesophyll air cavities, vascular bundles, and grain endosperm in near-infrared scattering. The near-infrared texture grid, obtained through linear synthesis according to empirical weights, retains the comprehensive texture intensity that best distinguishes the specular surface of grains from the diffuse reflection of leaves. However, the near-infrared image and the visible light image exhibit slight translation and rotation in their shooting posture, necessitating geometric registration to the pixel grid of the normalized grayscale image. During the registration phase, bilinear interpolation is used to continuously remap subpixel deviations. This is based on the principle that the spatial variation of spectral texture is approximately linear at the subpixel scale, thus avoiding the introduction of high-frequency artifacts. Even with the three grids co-located, the data still suffers from high-resolution redundancy. This is especially true when the internal texture of leaves tends to be smooth, as high sampling density does not provide information gain. Therefore, a unified one-to-four pyramid downsampling is performed on the three feature grids. The original resolution and the three downsampling layers are sequentially stored as a multi-layer pyramid sequence, facilitating subsequent probability estimation and diffusion smoothing in a coarse-guided fine manner. The local dynamic normalization stage simulates the human eye's adaptive perception mechanism of local contrast. In an eight-by-eight window, the brightness, significant edge intensity, and near-infrared texture are compressed to a closed interval from zero to one with the minimum and maximum values ​​as stretching endpoints. This not only weakens the global drift caused by different plant angles and illumination, but also retains the gradient and texture peaks that are commonly found between the grain shell and the leaf edge. It provides a scale-uniform, contrast-sufficient, and noise-suppressed input for the subsequent construction of the brightness-gradient-texture three-dimensional feature vector, fundamentally improving the discrimination stability of the soft discriminant model in complex field scenes.

[0057] Furthermore, in step 2, for each pixel, a three-dimensional feature vector is formed in a fixed order of brightness, gradient, and texture, and the vector is input into a soft discriminant model trained offline in advance; the soft discriminant model adopts an improved multi-class logistic regression framework, with the grain-leaf effective phenotypic area label as the positive class and other area labels 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-to-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 degree in a local three-by-three pixel neighborhood; if the outlier degree exceeds the threshold, the probability of the central pixel is replaced by the neighborhood probability mean to achieve unsupervised adaptive correction.

[0058] Brightness, gradient, and texture carry three complementary pieces of information: canopy reflectance, geometric boundaries, and tissue microstructure, respectively. The principle behind concatenating these three features in a fixed order into a three-dimensional feature vector is based on the mutual information maximization assumption of statistical discriminant models. When the features are statistically approximately orthogonal and conditionally independent with respect to the same discriminant target, multi-class logistic regression can construct a minimum cross-entropy hypersurface in parameter space, achieving optimal segmentation between the effective seed-leaf phenotypic region and other regions. The improved multi-class logistic regression embeds a feature norm regularization term and a class imbalance penalty term in the traditional log-likelihood objective function. The regularization term suppresses singularities in the feature covariance matrix, while the penalty term compensates for the skewness caused by the large number of background pixels compared to positive pixels by increasing the cost of negative classes. This approach maintains both edge sharpness and confidence stability in the continuous probability output. When the overall brightness distribution of the normalized grayscale image exhibits peaks and valleys 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. Direct reliance on the model output in such cases can lead to systematic bias. The design concept of the nearest neighbor consistency screening algorithm is derived 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, then the pixel is likely to be distorted by local shadows, sun spots, or sensor random noise; by calculating the probability outlier and replacing the over-threshold pixels with the neighborhood probability mean, it is equivalent to applying a first-order smoothing prior on the probability map, so that the local probability field can be restored to spatial consistency. At the same time, since this operation does not rely on external labels, it belongs to the category of unsupervised adaptive correction. It can be triggered immediately when there is a sudden change in illumination or leaf color alienation without retraining the model, ensuring the robustness of probability estimation in dynamic environments.

[0059] Furthermore, after all pixel probabilities are spliced ​​into a preliminary trait sensitive probability field, direction-guided smoothing is performed on the preliminary trait sensitive probability field. The direction-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, the diffusion perpendicular to the gradient direction is suppressed to retain the true leaf edge and grain outline. The diffusion coefficient decays exponentially according to the gradient amplitude, so that the boundary blurring caused by diffusion in the leaf texture is effectively controlled. After completing the direction-guided smoothing, the local entropy of the preliminary trait sensitive probability field is calculated to evaluate the non-convergence area. If the local entropy exceeds the set threshold, secondary denoising based on fast median filtering is implemented on this non-convergence area, 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 a spatial prior for the probability field. Information along the tangential direction of the leaf margin and kernel outline is considered a high-value signal, while directions orthogonal to these directions are considered diffusion channels that need to be restricted. During numerical iterations, the anisotropic diffusion equation assigns a diffusion coefficient to each pixel. This coefficient is derived from the gradient magnitude of the salient edge intensity matrix through an exponential decay function. Consequently, the diffusion rate is significantly reduced at pixels with large gradient magnitudes, effectively treating a sharp boundary as a heat flow barrier, preventing cross-probability mixing between the leaf margin and kernel shell. At pixels with small gradient magnitudes, the diffusion rate approaches a constant, allowing probabilities to quickly equalize within the homogeneous leaf surface, thereby suppressing fine-grained noise caused by diffusion. After the diffusion iterations are complete, to quantify any local random residuals that have not yet reached steady state, the local entropy of the preliminary trait-sensitive probability field is calculated using a five-by-five pixel window as a sliding entropy measurement domain. High local entropy indicates that the probability distribution within the window is still multimodal or discrete, indicating a non-convergent region. For these regions, fast median filtering was chosen as the secondary denoising operator because it can nonlinearly suppress spike noise while preserving edge locations, and its kernel size can be dynamically scaled proportionally to the local entropy value: higher entropy values ​​lead to larger windows, indicating that the region requires stronger spatial integration; lower entropy values ​​lead to smaller windows, preserving local details. In this way, direction-guided smoothing first locks in the true structure at the macro level through anisotropic diffusion, and then suppresses residual noise at the micro level through entropy-driven median filtering. The result is a trait-sensitive probability field that adheres to the geometric boundaries of leaf edges and kernels while remaining continuous and 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 ordinal value is:

[0062]

[0063] Among them, S qis the comprehensive breeding ordinal value of the target trait q, which is used to evaluate the performance of the plant in this trait; A total,q is the total projected area of ​​all connected regions in the trait-specific mask (in cm 2 );θ avg,q is the average leaf angle of all connected regions (in degrees); N region，q is the number of connected regions identified by eight-connected markers in the mask; σ θ，q is the standard deviation of leaf angle values, reflecting the consistency of leaf angle; is the number of target trait pixels with a probability value not less than 0.5 in the trait sensitive probability field; P all,q is the total number of pixels classified as this trait among all pixels.

[0064] The following example uses plant number P-2025-17 from Plot 1 of the experimental station as the sample. The plot is 120 m long and 80 m wide. A DJIM300 drone equipped with a multispectral camera (pixel size 2.4 μm, focal length 12 mm, resolution 4000 × 3000) was used. The flight altitude was set at 40 m, corresponding to a ground resolution GSD of ≈ 0.008 m, and each pixel covered an area of ​​0.000064 m. 2 That is 0.64cm 2 .

[0065] Five phases were selected for the growth period: jointing, tasseling, early grain filling, late grain filling, and maturity. Below is an example of a complete treatment for the early grain filling phase; the same steps apply to the remaining phases.

[0066] The original pixel gray value G in the target band 710nm raw =986, dark current compensation δ for this batch dark = 12. On-site instantaneous irradiance E solar =890W·m -2 , standard reflectivity R of the calibration plate ref =0.28. Bare soil reflectivity measured R soil =0.14. Rough measurement of canopy reflectance at pixel 710 =0.32. Calculation

[0067]

[0068] The normalized grayscale is 3.74, which is mapped to the 0-255 range by histogram stretching to obtain 196.

[0069] Perform three-by-three, five-by-five, and seven-by-seven symmetric differences on the pixel and its neighborhood. Taking a three-by-three window as an example, the horizontal difference D h =64, vertical differential D v =72. Gradient amplitude 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 θ≈48°.

[0070] The grayscale value of the near-infrared image pixel is 131, which becomes 198 after histogram matching. The grayscale co-occurrence matrix is ​​constructed with a three-by-three window, and the contrast is 4.1, homogeneity is 0.82, entropy is 1.67, and uniformity is 0.94. The texture intensity is synthesized using empirical weights of 0.35, 0.25, 0.25, and 0.15.

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

[0072] In the eight-by-eight window, the minimum brightness is 158 and the maximum is 211, so the brightness is normalized.

[0073]

[0074] The minimum significant edge strength is 15 and the maximum is 88. After normalization, G edge =(82.4-15) / (88-15)=0.92. The texture intensity is 1.3 at the minimum and 3.4 at the maximum. 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 to (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 the kernel-leaf effective phenotypic region.

[0078] The significant edge gradient direction of 48° is used as the main diffusion direction. The diffusion coefficient is set as c = exp(-0.03×G edge ). For this pixel, 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 five-by-five entropy is 1.12, which is less than the threshold of 1.6, and no secondary median filtering is required.

[0080] The optimal observation period has been corrected from the time series to the early filling period. The threshold is 0.5. After smoothing, the probability is 0.944 ≥ 0.5, and the pixels are written into the trait-specific mask. The entire plant mask contains 15,032 pixels. Total projected area

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

[0082] The eight-connectedness analysis retained 21 connected domains, with an average connected domain area

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

[0084] The main inertia axis of each connected domain is calculated to obtain a leaf angle set of {35°, 41°, …}, with an average leaf angle of 38°.

[0085] Archive threshold: area 9500cm 2 , the connected domain area is 460cm 2 , leaf angle 42°. The three normalized values ​​are 1.013, 0.996, and 0.905. The weights are 0.4, 0.3, and 0.3. The comprehensive ordinal value

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

[0087] The reference threshold is 1.00. The relative deviation is (0.972-1.00) / 1.00 = -0.028, corresponding to a trait score of -2.8. Negative values ​​indicate that the area and angle are slightly below the target but within the acceptable range. The yield of the test plot is 10.8 kg, and the plot area is 15 m 2 , yield per unit area 720g·m -2 The survival rate of mature plants was 0.93, and the field-level productivity coefficient was 669.6 g·m -2 .

[0088] The current decision table only contains this trait, so the efficient breeding index

[0089] I = (-2.8) × 669.6 = -1875;

[0090] The median value of the test batch was -2030, with results above the median considered to be good performance. Breeders can see P-2025-17's comprehensive ordinal value of 0.972 and index of -1875 on the screening interface and, based on other phase results, decide whether to proceed to the next round of hybridization or multi-site testing.

[0091] Figure 2The results of the correlation analysis between the comprehensive breeding ordinal values ​​calculated using the present method and the trait values ​​measured in the field are shown in detail. The horizontal axis represents the actual trait values ​​measured 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 domain area, and average leaf angle measurement, ranging from 0 to 100.

[0092] The solid circles in the scatter plot represent the total projected area trait, which includes 12 sample points and exhibits a clear linear positive correlation. The hollow circles represent the average connected domain area trait, which includes five representative sample points and shows a similar distribution trend to the total projected area trait. The square marks represent the average leaf angle trait, which also includes five sample points and exhibits a good linear correlation.

[0093] The regression line obtained by least squares fitting runs through the entire data distribution, demonstrating a strong linear correlation. The correlation coefficient (R) reached 0.96, indicating a high positive correlation between the comprehensive breeding ordinal values ​​and the measured trait values, validating the accuracy and reliability of the present method for trait quantification. The root mean square error (RMS) was only 4.2, further confirming the excellent measurement accuracy.

[0094] The data distribution shows that when measured trait values ​​are low (range 0-40), the integrated breeding ordinal values ​​are primarily distributed between 20 and 40, demonstrating the method's sensitive identification of low-value traits. In the medium trait value range (40-80), the integrated breeding ordinal values ​​exhibit a stable linear growth trend with a relatively constant slope. In the high trait value range (80-120), the integrated breeding ordinal values ​​reach a high level of 60-90, but the growth rate slows slightly, consistent with the saturation effect commonly seen in biological trait measurements.

[0095] Experimental data show that the method of the present invention can effectively capture the true variation patterns of corn phenotypic traits through the precise generation of trait-specific masks and the comprehensive calculation of multidimensional measurements, providing a reliable quantitative evaluation tool for efficient breeding.

[0096] Figure 3 This comprehensive analysis demonstrates the distribution characteristics and statistical patterns of the high-efficiency breeding index for each plant in the experimental population, derived through a comprehensive calculation process. The horizontal axis represents the distribution range of the high-efficiency breeding index for each plant, divided into 20 equally spaced intervals, ranging from 0-20 to 140-160, for a total of eight intervals. The vertical axis represents the number of plants falling within each index interval, with a maximum of 52 plants.

[0097] The histogram clearly shows that the high-efficiency breeding index of each plant exhibits a typical normal distribution. In the low index range (0-40), the number of plants is relatively small, at 12 and 21, 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, respectively, with the number of plants in the 60-80 range reaching a peak, reflecting that the majority of plants in the population have medium breeding potential.

[0098] In the higher index range (80-120), the number of plants decreased, reaching 38 and 28, 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 decreased further, to 16 and 8, respectively, consistent with the biological principle of the relative scarcity of high-quality germplasm resources.

[0099] The normal distribution fitting curve shows that the actual data distribution closely matches the theoretical normal distribution, validating the statistical validity of the method for calculating the high-efficiency breeding index for each individual plant. The mean value of the distribution is 68.5, indicating that the overall breeding potential of the experimental population is above average. The standard deviation is approximately 25, reflecting the level of genetic diversity within the population.

[0100] The threshold for excellent lines set in the figure is 100. Based on this standard, 52 plants with an index value exceeding 100, representing 24.3% of the total population, provide ample candidate material for subsequent breeding selection. This proportion ensures both appropriate selection intensity and a sufficient genetic base, meeting the empirical requirements of breeding practice. The distribution characteristics demonstrate that the present method can effectively distinguish plants with different breeding potential, providing a reliable quantitative basis for precise breeding decisions.

[0101] Figure 4 This study demonstrates the changing patterns in trait sensitivity probability field recognition accuracy across different growth stages using maize canopy imagery acquired by a multispectral drone. The horizontal axis represents the five key growth stages of maize, from jointing to maturity, while the vertical axis represents the recognition accuracy of the effective kernel-leaf phenotypic region, ranging from 0.6 to 1.0.

[0102] Experimental results demonstrate that the proposed method for generating trait-sensitive probability fields based on normalized grayscale images significantly improves recognition accuracy across different growth stages. During the jointing stage, due to the relatively simple canopy structure of corn plants and the clear boundary between leaves and stalks, the recognition accuracy reached 0.68. As the growth period progressed, the recognition accuracy increased to 0.76 during the tasseling stage. This is primarily due to the multi-scale processing capabilities of the three scale difference operators for gradient operations, which effectively capture edge features at different scales.

[0103] Recognition accuracy peaked at 0.91 during the silking stage, when the kernel-leaf effective phenotypic region of the corn plant was most prominent. The fusion of brightness, significant edge intensity, and near-infrared texture achieved the best results. 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 that neighborhood consistency screening was effective in unsupervised correction of outlier probabilities. Recognition accuracy declined slightly to 0.89 during the maturity stage, due to changes in texture and aging of the leaves. However, it remained at a high level.

[0104] In comparison, the recognition accuracy of the traditional threshold segmentation method was significantly lower than that of the method of the present invention at all growth stages. The recognition accuracy of the traditional method was 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 showed that the method of the present invention achieved a 10-20% improvement in accuracy at all growth stages, especially in the critical silking and grain filling stages, with accuracy improvements of 23% and 29%, respectively. This fully verifies the effectiveness of the directional-guided smoothing and anisotropic diffusion smoothing techniques.

[0105] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A phenotypic trait image recognition and extraction method for efficient corn breeding, characterized in that: The method comprises: Step 1: At various growth stages of corn, from jointing to maturity, a drone equipped with a multispectral sensor was used to capture canopy images of the same plant in the experimental field. Flight altitude, instantaneous irradiance, and bare soil reflectance were also 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, a gradient operation is performed on the target band to obtain the significant edge intensity, which is then fused with the near-infrared texture. Simultaneously, considering the brightness, significant edge intensity, and near-infrared texture features, the probability of each pixel belonging to the effective seed-leaf phenotypic region is determined, generating a trait sensitivity probability field for the corresponding time phase. Step 3: Preset the optimal observation period for each target trait, and select a set of pixels with a probability value of not less than 0.5 and determined to be the trait type from the image of the phase corresponding to the optimal observation period to form a trait-specific mask; for each trait, perform the measurements of total projected area, average connected domain area, and average leaf angle in sequence to obtain the comprehensive breeding ordinal value of the trait.

2. The phenotypic trait image recognition and extraction method for efficient corn breeding according to claim 1, wherein: The method also includes: step 4: comparing each comprehensive breeding ordinal value with the target threshold of the corresponding trait of the same line in the breeding file, and calculating the trait score according to the relative deviation; for each experimental plot, dividing the measured yield by the area of ​​the experimental plot to obtain the yield per unit area, and multiplying this value with the survival rate of the adult plant to form a field-level production potential coefficient; multiplying all the trait scores included in the decision by the field-level production potential coefficient one by one and then summing them to obtain a 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 phenotypic trait image recognition and extraction method for efficient corn breeding according to claim 2, wherein: In step 1, the chlorophyll-sensitive band with a wavelength of 710 nm is selected as the target band, and the standard reflectance obtained by indoor calibration of the calibration plate in the same band is used as the absolute reference.

4. The phenotypic trait image recognition and extraction method for efficient corn breeding according to claim 3, wherein: In step 1, a grid-type flight path planning method was used based on the geometric boundaries of the experimental field to ensure that the canopy of each corn plant was photographed by at least three adjacent tracks overlapping each other. To ensure that the images of the same plant canopy had an alignable spatial reference at each growth stage, the lateral overlap was set to no less than 75% and the heading overlap was set to no less than 85%.

5. The phenotypic trait image recognition and extraction method for efficient corn breeding according to claim 4, wherein: Step 2 specifically includes: performing gradient operations on the normalized grayscale image through three sets of scale difference operators to obtain multi-scale gradient amplitudes; 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 margin; then extracting grayscale symbiotic texture indicators from the synchronized near-infrared image and linearly synthesizing them into a near-infrared texture grid, which is geometrically aligned to the normalized grayscale image to achieve pixel-level co-location of brightness, significant edge intensity and near-infrared texture; then dynamically normalizing the three features for each pixel in the local window to construct a brightness-gradient-texture three-dimensional feature vector and send it to the soft discriminant model to output the preliminary probability of the effective phenotypic area of ​​the grain-leaf; if it is detected that the on-site illumination or leaf color distribution exceeds the training range, the outlier probability is unsupervisedly corrected through neighborhood consistency screening; finally, anisotropic diffusion smoothing is performed according to the gradient direction of the significant edge intensity to retain the true edge and suppress noise.

6. The phenotypic trait image recognition and extraction method for efficient corn breeding according to 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 sets of spatial scales on the grid of the same normalized grayscale image, and calculating the horizontal and vertical differences using three-by-three, five-by-five, and seven-by-seven symmetric difference operators for each scale, respectively; then synthesizing the gradient amplitudes of the horizontal and vertical differences at the same scale according to the Euclidean norm to obtain a gradient amplitude set, and recording the corresponding gradient directions; performing an opening operation on the gradient amplitude set to filter out high-frequency isolated noise, and then retaining structured edges through continuous erosion-dilation cycles, and finally outputting a significant edge intensity matrix.

7. The phenotypic trait image recognition and extraction method for efficient corn breeding according to claim 6, wherein: In step 2, the near-infrared band image collected synchronously with the same phase is read, and histogram matching is performed to make its brightness distribution consistent with the normalized grayscale image; then a grayscale co-occurrence matrix is ​​constructed with a three-by-three sliding window, and four sets of texture indicators, namely contrast, homogeneity, entropy and uniformity, are calculated. The four sets of texture indicators are linearly synthesized according to empirical weights to obtain a near-infrared texture grid in the form of a single scalar; the near-infrared texture grid is geometrically aligned to the pixel grid of the normalized grayscale image, and bilinear interpolation is used to correct sub-pixel deviations to ensure that each pixel has brightness value, significant edge intensity value and near-infrared texture value in subsequent calculations; in order to reduce data redundancy, In addition, a one-to-four pyramid downsampling method was performed on the three feature grids to generate a multi-layer resolution pyramid sequence, so that probability calculation and subsequent smoothing can be completed uniformly at different spatial granularities. At the original resolution level, three features, namely brightness, significant edge intensity and near-infrared texture, were extracted for each pixel, and local dynamic normalization was performed on each pixel. Local dynamic normalization used an eight-by-eight window and linearly stretched it with the minimum and maximum values ​​in the window as endpoints, so that each feature value was compressed to a closed interval from zero to one and the relative difference was retained, so as to weaken the influence of brightness drift between different plants and maintain the sharp contrast of gradient and texture at the leaf edge and grain surface.

8. The phenotypic trait image recognition and extraction method for efficient corn breeding according to claim 7, wherein: In step 2, for each pixel, a three-dimensional feature vector is formed in a fixed order of brightness, gradient, and texture, and the vector is input into a soft discriminant model trained offline in advance; the soft discriminant model adopts an improved multi-class logistic regression framework, with the grain-leaf effective phenotypic area label as the positive class and other area labels 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-to-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 degree in a local three-by-three pixel neighborhood; if the outlier degree exceeds the threshold, the probability of the central pixel is replaced by the neighborhood probability mean to achieve unsupervised adaptive correction.

9. The phenotypic trait image recognition and extraction method for efficient corn breeding according to claim 8, characterized in that: After all pixel probabilities are spliced ​​into a preliminary trait sensitivity probability field, direction-guided smoothing is performed on the preliminary trait sensitivity probability field. The direction-guided smoothing process uses the gradient direction provided by the significant edge intensity matrix as the main smoothing direction, and anisotropically diffuses 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 kernel outline. The diffusion coefficient decays exponentially according to the gradient amplitude, effectively controlling the boundary blurring caused by diffusion on the leaf texture. After completing the direction-guided smoothing, the local entropy of the preliminary trait-sensitive probability field is calculated to evaluate the non-convergent area. If the local entropy exceeds the set threshold, secondary denoising based on fast median filtering is implemented on the non-convergent region, and the size of the median filter kernel increases linearly with the entropy value.

Images