A method, device and product for monitoring the number of tillers of winter wheat

Abstract

The application discloses a winter wheat tiller number monitoring method, device and product, relates to the field of agricultural monitoring, and comprises the following steps: acquiring ground measured tiller data and unmanned aerial vehicle hyperspectral wheat canopy hyperspectral images; performing near-infrared spectrum self-adaptive enhancement pretreatment on the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral images; performing self-adaptive image segmentation on the spectrum-enhanced images based on an error pixel correction method and a polarization segmentation method; extracting feature data from a mask image containing crop information and constructing a multi-angle atlas feature library; screening target features from the atlas feature library by using a multi-strategy fusion optimization grey wolf search algorithm; constructing and training a winter wheat tiller number prediction model by using a support vector machine (SVM) model and the target features and the ground measured tiller data; and performing tiller number inversion on a research area by using the winter wheat tiller number prediction model. The application can improve the accuracy and efficiency of winter wheat tiller number monitoring.

Description

Technical Field

[0001] This application relates to the field of agricultural monitoring, and in particular to a method, equipment and product for monitoring the tiller number of winter wheat. Background Technology

[0002] In China, wheat, as one of the main food crops, accounts for approximately 20% of total grain production. However, with the continuous development of agricultural production and the ongoing population growth, farmers and agricultural practitioners are facing greater production pressure and challenges in resource utilization efficiency. How to effectively improve wheat yield per unit area and the economic return on agricultural inputs is a key issue in modern agricultural development. Fertilization, as a basic agricultural management measure, plays a crucial role in promoting crop growth and increasing yield. However, judging from the current state of fertilizer application in Chinese agriculture, fertilizer use per unit of cultivated land ranks among the highest in the world. Excessive expansion of nutrient input into farmland poses a dual challenge to economic efficiency and the environment. Among these, the prominent shortcomings of low fertilizer conversion rates and indiscriminate extensive farming practices exacerbate the inflation of agricultural production costs, damage to soil ecological balance, and hinder sustainable agricultural development.

[0003] Tiller number, as a crucial indicator for assessing the growth status and yield potential of winter wheat, is directly related to crop biomass accumulation and ear formation, and also indicates the crop's nutrient and water requirements. Rapid and accurate monitoring of wheat tiller dynamics is essential for optimizing variable-rate fertilization strategies, improving fertilizer use efficiency, and ensuring the maximization of crop yield potential. Traditional tiller number monitoring methods mainly rely on manual field surveys and sample statistics. While this method provides intuitive tiller data, it suffers from significant drawbacks such as being time-consuming and labor-intensive, prone to data bias, and lacking timeliness.

[0004] Therefore, in order to solve the problems of existing technologies, there is an urgent need to provide a method for monitoring the number of tillers in winter wheat, so as to improve the accuracy and efficiency of monitoring the number of tillers in winter wheat. Summary of the Invention

[0005] The purpose of this application is to provide a method, equipment, and product for monitoring the number of tillers in winter wheat, which can improve the accuracy and efficiency of monitoring the number of tillers in winter wheat.

[0006] To achieve the above objectives, this application provides the following solution:

[0007] Firstly, this application provides a method for monitoring the number of tillers in winter wheat, the method comprising:

[0008] Acquire ground-measured tillering data and UAV hyperspectral images of wheat canopy;

[0009] Near-infrared spectral adaptive enhancement preprocessing is performed on hyperspectral images of wheat canopy from UAVs. The spectral enhancement preprocessing is used to determine vegetation density index based on the near-infrared spectral matrix, and then determine vegetation density gradient level based on the vegetation density index. After that, an adaptive enhancement factor is determined based on the vegetation density gradient level, and finally, spectral enhancement is achieved based on the adaptive enhancement factor.

[0010] Adaptive image segmentation based on error pixel correction and polarization segmentation is performed on spectrally enhanced images to obtain mask images containing crop information;

[0011] Feature data is extracted from masked images containing crop information, and a multi-angle spectral feature library is constructed; the feature data includes: spectral features, texture features, color features, and shape features;

[0012] An optimized gray wolf search algorithm using multi-strategy fusion is used to filter target features from a graph feature library;

[0013] Using target features and ground-measured tillering data, a support vector machine (SVM) model was constructed and trained to predict the number of tillers in winter wheat.

[0014] The tiller number prediction model for winter wheat was used to invert the tiller number in the study area.

[0015] Optionally, acquiring ground-measured tillering data and UAV hyperspectral wheat canopy hyperspectral images specifically includes:

[0016] Hyperspectral images of wheat canopy were acquired using a hyperspectral high-resolution camera mounted on a drone.

[0017] Hyperspectral images of wheat canopy from UAVs were preprocessed and analyzed using Pix4D software. The preprocessing and analysis included image stitching, band synthesis, atmospheric correction, and geometric correction.

[0018] The pre-processed and analyzed images are processed using the ENVI remote sensing image processing platform; the image processing includes: adjusting composition, adjusting color levels, adjusting white balance, and increasing sharpness.

[0019] Optionally, the near-infrared spectral adaptive enhancement preprocessing of the UAV hyperspectral wheat canopy hyperspectral image specifically includes:

[0020] Using the formula ΔD=R nir -R red Determine vegetation density indicators;

[0021] The average value of the vegetation density index is used to determine the vegetation density gradient level;

[0022] Using formula Determine the adaptive enhancement factor;

[0023] Near-infrared statistical curves of UAV hyperspectral wheat canopy hyperspectral images are multiplied with an adaptive enhancement factor to achieve near-infrared spectral adaptive enhancement preprocessing.

[0024] Where ΔD is the vegetation density index, and R nir and R red The near-infrared reflectance at 850 nm and the red band reflectance at 650 nm are respectively, g a To accommodate the enhancement factor, p∈i×10%, i=1,2,...,5, p is used to determine the position of the starting point P in the effective enhancement process, p is a parameter whose value is inversely proportional to the vegetation density gradient level, q∈j×10%, j=6,7,8, q is used to determine the position of the inflection point Q when the enhancement intensity begins to weaken, t q t represents the x-coordinate of point Q. q = (T-pT)×q+pT, where T is the maximum value of the abscissa of the adaptive enhancement factor, and h∈(1,1.15), where h is used to determine the maximum value of the adaptive enhancement factor function or the Q-point position.

[0025] Optionally, the adaptive image segmentation based on error pixel correction and polarization segmentation of the spectrally enhanced image to obtain a mask image containing crop information specifically includes:

[0026] Using formula Determine the enhanced spectral uplift index S; where B nir and B red These are the wavelength values ​​for the near-infrared and red bands, respectively, r. nir 'Near-infrared reflectance of the spectrally enhanced image;'

[0027] Using formula Determine the correction factor c v H ndvi It is the original NDVI matrix within the sample region, where max() represents the maximum value, mean() represents the mean value, and min() represents the minimum value.

[0028] The theoretical segmentation threshold s is determined using the segmented image matrix S based on the enhanced spectral uplift index S. m ;

[0029] Using formula Achieve correction factor c v For the theoretical segmentation threshold s m Adaptive numerical adjustment is performed, and a dynamic error correction recognition index k sensitive to green pixels is constructed;

[0030] The dynamic error correction recognition index k is compared with each pixel on the segmented image matrix S, and the pixel-level recognition and classification task is determined based on the comparison result.

[0031] Optionally, the optimized gray wolf search algorithm utilizing multi-strategy fusion to filter target features from the graph feature library specifically includes:

[0032] Using formula Determine the evaluation function Fitness; where C all C represents the ability of the selected features to represent the complete feature set. true and MI true C represents the correlation and mutual information between the selected features and the true value, respectively. fac and MI fac This indicates the correlation and mutual information between the selected features;

[0033] Using formula Determine the nonlinear convergence factor Among them, a1 and a T , where are the initial and final values ​​of the convergence factor, respectively, and T is the total number of iterations.

[0034] Secondly, this application also provides a winter wheat tiller number monitoring device, the winter wheat tiller number monitoring device comprising:

[0035] The data acquisition module is used to acquire ground-measured tillering data and UAV hyperspectral images of wheat canopy;

[0036] Spectral enhancement processing is used to perform near-infrared spectral adaptive enhancement preprocessing on hyperspectral images of wheat canopy from UAVs. The spectral enhancement preprocessing is used to determine vegetation density index based on the near-infrared spectral matrix, and then determine vegetation density gradient level based on the vegetation density index. After that, an adaptive enhancement factor is determined based on the vegetation density gradient level, and finally spectral enhancement is achieved based on the adaptive enhancement factor.

[0037] The image segmentation module is used to perform adaptive image segmentation on spectrally enhanced images based on error pixel correction and polarization segmentation methods to obtain a mask image containing crop information.

[0038] The feature extraction module is used to extract feature data from a mask image containing crop information and construct a multi-angle spectral feature library; the feature data includes: spectral features, texture features, color features, and shape features;

[0039] The target selection module is used to select target features from the graph feature library using an optimized gray wolf search algorithm that integrates multiple strategies.

[0040] The prediction model determination module is used to construct and train a winter wheat tiller number prediction model using target features and ground-measured tiller data, employing a support vector machine (SVM) model.

[0041] The tiller number inversion module is used to invert the tiller number in the study area using the winter wheat tiller number prediction model.

[0042] Thirdly, this application also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the winter wheat tiller number monitoring method.

[0043] Fourthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the winter wheat tiller number monitoring method.

[0044] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0045] This application provides a method, equipment, and product for monitoring the tiller number of winter wheat. It involves near-infrared adaptive enhancement preprocessing of hyperspectral wheat canopy images from UAVs, followed by adaptive image segmentation based on error pixel correction and polarization segmentation to remove soil background interference. Spectral indices, texture indices, color indices, and shape indices are extracted from a mask image containing only crop information to construct a multi-angle spectral feature library. A multi-strategy fusion optimized gray wolf search algorithm is used to select target features from the spectral feature library that are beneficial for tiller number inversion modeling and have low computational burden. A winter wheat tiller number prediction model is constructed using the selected features, and tiller number inversion is performed across the entire study area. This application improves the accuracy and efficiency of winter wheat tiller number monitoring. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating a method for monitoring the number of tillers in winter wheat, provided as an embodiment of this application;

[0048] Figure 2 A schematic diagram of the functional expression of the adaptive enhancement factor;

[0049] Figure 3This is a schematic diagram of the near-infrared spectroscopy correction process;

[0050] Figure 4 This is a schematic diagram of the adaptive image segmentation process based on the error pixel correction method and the polarization segmentation method.

[0051] Figure 5 Frequency statistics of polarization matrix W in the study area under the conditions of highest and lowest relative vegetation cover gradient;

[0052] Figure 6 A schematic diagram illustrating the mechanism of the CIM-IGWO multi-strategy fusion gray wolf optimized search algorithm;

[0053] Figure 7 A schematic diagram of the SVM modeling inversion results based on full features and preferred features;

[0054] Figure 8 This is a map showing the distribution of winter wheat tiller numbers across the entire study area. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] In one exemplary embodiment, such as Figure 1 As shown, a method for monitoring the tiller number of winter wheat is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, it includes the following steps S101 to S107. Wherein:

[0058] S101, acquire ground-measured tillering data and UAV hyperspectral images of wheat canopy;

[0059] S102, Near-infrared spectral adaptive enhancement preprocessing is performed on the hyperspectral image of wheat canopy from the UAV; the spectral enhancement preprocessing is used to determine the vegetation density index based on the near-infrared spectral matrix, and then determine the vegetation density gradient level based on the vegetation density index; then determine the adaptive enhancement factor based on the vegetation density gradient level, and finally realize spectral enhancement based on the adaptive enhancement factor.

[0060] S103, perform adaptive image segmentation based on error pixel correction method and polarization segmentation method on the spectrally enhanced image to obtain a mask image containing crop information;

[0061] S104, extract feature data from the mask image containing crop information and construct a multi-angle spectral feature library; the feature data includes: spectral features, texture features, color features and shape features;

[0062] S105, Optimized Grey Wolf Search Algorithm with Multi-Strategy Fusion is used to filter target features from the graph feature library;

[0063] S106. Using target features and ground-measured tillering data, a support vector machine (SVM) model was constructed and trained to predict the number of tillers in winter wheat.

[0064] S107, using a winter wheat tiller number prediction model to invert the tiller number in the study area.

[0065] In an exemplary embodiment, S101 specifically includes:

[0066] S110 utilizes a hyperspectral high-definition camera mounted on a drone to acquire hyperspectral images of wheat canopy; the drone model is DJI M300 RTK; the hyperspectral high-definition camera model is Cubert FireflEYES185.

[0067] Before acquiring spectral image data, DJI GO was used to set drone waypoints and plan flight routes.

[0068] S111, The hyperspectral image of wheat canopy from the UAV is preprocessed and analyzed using Pix4D series software; the preprocessing and analysis includes: image stitching, band synthesis, atmospheric correction and geometric correction;

[0069] S112, The image processing is performed on the pre-processed and analyzed image using the ENVI remote sensing image processing platform; the image processing includes: adjusting composition, adjusting color levels, adjusting white balance, and increasing sharpness.

[0070] Specifically, the acquisition of winter wheat tiller number data includes:

[0071] Use 1m 2 The sampling survey method involves manually sampling and counting wheat seedlings to determine the number of basic seedlings and tillers. Basic seedlings represent the actual number of wheat plants within each sampling area, expressed as plants per unit sampling area. The number of tillers (average number of tillers per individual plant) represents the sum of the number of main stems and tillers on each individual plant in the sampling area, expressed as stems per plant. The counting rule for effective tillers is: the main stem and all its tillers, including first-order tillers and lower-positioned tillers of second-order tillers.

[0072] In an exemplary embodiment, S102 specifically includes:

[0073] S201, use the formula ΔD=R nir -R red The vegetation density index is determined; ΔD is used to roughly quantify the spectral upscaling effect at specific pixels. The sensitivity of ΔD in soil and vegetation pixel classification can be improved by increasing the near-infrared reflectance of specific pixels.

[0074] S202, determine the vegetation density gradient level using the average value of the vegetation density index distribution; the average value d of the vegetation density index distribution. m For: d m =mean(ΔD)=mean(R) nir -R red ); ΔD refers to the ΔD distribution matrix over a single sample region, r nir and R red These represent the near-infrared reflectance matrix and the red band reflectance matrix of the sample area, respectively.

[0075] S203, using the formula Determine the adaptive enhancement factor; the adaptive enhancement factor is to quickly achieve adaptive spectral enhancement preprocessing between sample images under different vegetation cover densities. For example... Figure 2 As shown, the functional expression of the adaptive enhancement factor is affected by the vegetation cover density level in the sample area. The impact.

[0076] S204 performs a dot product operation between the near-infrared statistical curve of the UAV hyperspectral wheat canopy hyperspectral image and the adaptive enhancement factor to achieve near-infrared spectral adaptive enhancement preprocessing.

[0077] Where ΔD is the vegetation density index, and R nir and R red The near-infrared reflectance at 850 nm and the red band reflectance at 650 nm are respectively, g a To accommodate the enhancement factor, p∈i×10%, i=1,2,...,5, p is used to determine the position of the starting point P during effective enhancement processing (falling within the suspicious pixel region), p is a parameter whose value is inversely proportional to the vegetation density gradient level, q∈j×10%, j=6,7,8, q is used to determine the position of the inflection point Q when the enhancement intensity begins to weaken, which can prevent the enhancement operation from excessively increasing the reflectivity of the drifting part at the end, t q t represents the x-coordinate of point Q. q= (T-pT)×q+pT, where T is the maximum value of the abscissa of the adaptive enhancement factor, h∈(1,1.15), and h is used to determine the maximum value of the adaptive enhancement factor function or the Q-point position. A magnification of 1.15 is sufficient to distinguish between enhanced and unenhanced spectral data. T is the maximum value of the abscissa of this enhancement factor, which depends on the number of non-repeating numerical points on the statistical curve. Figure 2 The key points P and Q together determine the enhancement factor g. a The function trend can achieve preliminary segmentation of suspicious pixel areas and weaken reflectivity drift interference.

[0078] Adaptive enhancement factor g a This is the core of the Near Infrared-Self-Adaptive Spectral Enhancement (NIR-SASE) strategy. By multiplying the sample's near-infrared statistical curve with the enhancement factor, rapid enhancement of the reflectance spectrum of each pixel in a specific sample image can be achieved. Figure 3 Give the use of g on a specific sample region a The near-infrared spectral correction process is described in Algorithm I, with details of the code execution.

[0079]

[0080] In an exemplary embodiment, S103 specifically includes:

[0081] S301, using the formula Determine the enhanced spectral uplift index S; where B nir and B red These are the wavelength values ​​for the near-infrared and red bands, respectively. Adjusting the range of the denominator to 0 to 1 can make the exponent S more manageable overall. nir 'Near-infrared reflectance of the spectrally enhanced image; the enhanced spectral uplift index S is used to describe the slope characteristics, serving as a slope characteristic index;

[0082] S302, using the formula Determine the correction factor c v H ndvi This is the original NDVI matrix within the sample region, where max() represents the maximum value, mean() represents the mean, and min() represents the minimum value; the correction factor c v Accurate correction threshold values ​​are obtained by using a combination transformation of the NDVI exponent. When c in the sample image matrix... v A value of 1 indicates that the proportion of vegetation pixels in the area is relatively small, while a value of 1 indicates that the area contains a relatively large proportion of vegetation pixels; when c vThe greater the value deviates from the threshold of 1, the more extreme the vegetation cover level in that area.

[0083] S303, using the segmented image matrix S with the enhanced spectral uplift index S to determine the theoretical segmentation threshold s. m It can achieve rapid segmentation of farmland canopy images;

[0084] S304, using the formula Achieve correction factor c v For the theoretical segmentation threshold s m Adaptive numerical adjustment is performed, and a dynamic error correction recognition index k sensitive to green pixels is constructed; when the vegetation in the specified sample area is relatively lush (c v ≥1), the median value s that is too high needs to be adjusted. m Appropriately lower the price to recover most of the losses that are slightly below the theoretical level. m The threshold for vegetation pixels, and the correction amount should be greater for areas with denser crops. However, when the proportion of bare soil within the specified sample area is larger (c u <1), the low s needs to be adjusted. m Increase it appropriately to eliminate most of those slightly higher than the theoretical value. m The soil pixels at the threshold should also be adjusted more drastically in areas where crops are sparser. Specifically, due to c v When the value is less than 1, the smaller the value, the lower the crop density in that area. Therefore, using c... v Only by using the reciprocal of the factor as a correction factor can a greater degree of correction be obtained.

[0085] In step S305, the dynamic error correction recognition index k is compared with each pixel in the segmented image matrix S, and the pixel-level recognition and classification task is determined based on the comparison result. A normalized polarization recognition matrix W is further constructed using the comparison result. In this matrix, for all pixels satisfying S≥k, their pixel values ​​are made to approach 1; while for all pixels satisfying S<k, their pixel values ​​are made to approach 0.

[0086] Figure 5 Frequency statistics of the polarization matrix W within the study area under the conditions of highest and lowest relative vegetation cover gradients are presented. Figure 5 Part (a) represents the condition of highest vegetation density. Figure 5 Part (b) is the minimum vegetation density condition. Clearly, the frequency statistics of this matrix exhibit a regular double-peak, single-trough characteristic, with the trough falling near the interval centered at 0.5. Specifically, the element distribution values ​​corresponding to S≥k cluster towards the "1" endpoint, while those corresponding to S<k cluster towards the "0" endpoint. Therefore, using the median of the normalized sample matrix interval, 0.5, as a fixed threshold allows for rapid image segmentation. Pixels in the polarization matrix greater than 0.5 are identified as crops, while the rest are identified as soil. Ultimately, only the pixels in the polarization matrix W containing only crop information and exceeding the fixed threshold are used to construct the mask image W. mask This polarization segmentation method is applicable to sample data under different vegetation density conditions and also has good universality in different sensor datasets.

[0087] The complete algorithm steps for adaptive image segmentation based on error pixel correction and polarization segmentation are given below.

[0088]

[0089]

[0090] This application employs a spectral enhancement preprocessing strategy to amplify the differences in spectral uplift characteristics between crops and soil, and determines the degree of spectral correction based on vegetation density gradients. Building upon this, a dynamic correction identification index is constructed using theoretical segmentation thresholds and correction factors to correct pixels incorrectly enhanced during processing. Furthermore, polarization is introduced into the matrix to be segmented to improve the interpretability and robustness of the segmentation threshold.

[0091] In an exemplary embodiment, the feature data extracted in S104 are as follows:

[0092] Spectral features are derived through mathematical operations using the reflectance or absorption characteristics of different bands in remote sensing image data. Common spectral indices include vegetation indices (such as NDVI) and spectral bands, which can reflect specific information about different objects or land features. Specifically, commonly used vegetation indices can reflect crop growth status, such as chlorophyll content, vegetation cover, and growth condition, thereby helping to accurately monitor and predict crop health and growth rate. In fact, near-infrared and other spectral band features have also been shown to have a highly sensitive correlation with tiller number. Therefore, this study uses nine conventional vegetation indices (see Table 1) and enhanced near-infrared reflectance extracted from the masked image after spectral enhancement preprocessing to construct a spectral index feature set. In particular, the specified feature for each sample is represented by the average of all index features in that region.

[0093] Table 1

[0094]

[0095]

[0096] Image texture features refer to the subtle textures, patches, and morphological characteristics of objects in an image. These features can capture the spatial structure of vegetation distribution, providing a description of the spatial organization and distribution pattern of vegetation. They are unaffected by light and soil type, thus holding great potential for evaluating crop growth parameters. The Gray-Level Co-occurrence Matrix (GLCM) is a statistical method for describing image texture features. It extracts texture information by analyzing the spatial relationships between pixel gray levels in an image, thereby providing a detailed description of spatial variation patterns. Theoretically, the calculation of the GLCM matrix requires selecting multiple directions to comprehensively describe the image texture features and reduce the influence of noise in specific directions. Commonly used angles include horizontal, vertical, 45-degree, and 135-degree angles. Secondly, for each selected angle, a fixed-size sliding window needs to be chosen and slid across the image from left to right and top to bottom, and the pixels within the window are statistically analyzed. Table 2 shows the calculation formulas and descriptions of the six texture indices selected from the rich GLCM feature set in this study.

[0097] To better represent the texture differences between canopy images captured by UAVs, this study selected the near-infrared band to calculate the following six texture indices. The sliding calculation window was set to 3×3 pixels, approximately 6 cm in size. During the calculation, considering the influence of texture features in different directions on vegetation structure and growth status, the average values ​​of the texture features in the 45°, 90°, 180°, and 270° directions were selected as the final values ​​of the texture features to comprehensively reflect the texture information in the images, providing a more accurate data foundation for subsequent crop growth parameter evaluation.

[0098] Table 2

[0099]

[0100] Color features are simple yet widely used visual features, typically associated with objects or scenes in an image. Compared to other image features, color features are less dependent on changes in image scale, orientation, and viewpoint, exhibiting better robustness. Global color features of an image can usually be described using methods such as color moments, color histograms, and color correlation maps. Among these, the color matrix quickly obtains color information from an image by calculating the color distribution statistics of pixels, making it very simple and efficient. Furthermore, color moments are applicable to various image types, unaffected by factors such as image scale, brightness, and contrast, demonstrating good versatility and robustness. Table 3 presents the calculation formulas and descriptions of the commonly used color moment statistical features in this study.

[0101] In this application, since color information is mainly distributed in the relatively simple low-order moments, and the first and second moments can effectively describe the overall color distribution and degree of change of the image, they can meet the needs of most application scenarios. Furthermore, the first and second moments are more robust to noise in the image, less susceptible to noise influence, and can more stably reflect the color characteristics of the image. Therefore, this application selects the mean and variance of the RGB color space as representatives of the color matrix. By calculating the mean and variance of the red (R), green (G), and blue (B) channels respectively, a color feature set consisting of six features is formed, and these features are named Rmean, Gmean, Bmean, Rstd, Gstd, and Bstd. Rmean, Gmean, and Bmean represent the color mean of each channel, reflecting the average distribution of different colors in the image; while Rstd, Gstd, and Bstd represent the color variance of each channel, used to describe the degree of change in the color distribution in the image. These color feature sets can comprehensively reflect the color information in the winter wheat growing environment, providing important reference data for winter wheat growth monitoring and inversion modeling.

[0102] Table 3

[0103]

[0104] In remote sensing image processing, shape features serve as an important supplement, playing a crucial role in describing the spatial distribution and environmental characteristics of vegetation. After image segmentation, calculating the proportion of pixels in the crop portion can measure the spatial distribution and environmental characteristics of vegetation in the sample image. This proportion can be defined as a shape feature, describing the shape and distribution of vegetation in the image, and can infer comprehensive information such as vegetation dispersion, shape complexity, and aggregation. A higher proportion of pixels in the vegetation portion indicates a relatively concentrated vegetation distribution and potentially more complex shapes; conversely, a lower proportion may indicate a more dispersed vegetation distribution and relatively simpler shapes. Therefore, this shape feature provides important information about vegetation cover, helping to understand and analyze the growth status and distribution of vegetation under different environmental conditions. Its calculation formula is shown below:

[0105]

[0106] Where, N green This represents the number of crop pixels remaining after image segmentation and masking, where M represents the total number of pixels in the sample image.

[0107] In an exemplary embodiment, S105 specifically includes:

[0108] The CMI-IGWO method, an improved version of the standard gray wolf optimization algorithm, selects the most representative and interpretable features from the feature library. From an information theory perspective, CMI-IGWO comprehensively considers the linear and nonlinear relationships between features and the target variable, guiding the gray wolves to select the optimal features. Furthermore, CMI-IGWO designs optimization schemes that help the wolf pack balance its global and local search capabilities and reduce the risk of finding local optima, considering both convergence factors and wolf pack update strategies.

[0109] When a wolf pack begins its iterative search, a reasonable evaluation function can help it gradually approach the global optimum within the search space. The CMI-IGWO optimization search algorithm, approaching the problem from a combined statistical and information theory perspective, constructs a novel evaluation criterion as the fitness function for the gray wolf optimization algorithm, such as... On the one hand, this evaluation function, combining Pearson correlation coefficient and mutual information, can not only uncover linear relationships between variables but also capture more complex nonlinear relationships, making the analysis more comprehensive. On the other hand, this evaluation function comprehensively considers the relationship between the selected features and the true values, the independence between the selected features, and the representativeness of the selected features to all features, striving to select the most representative and independent feature subset to avoid underfitting or overfitting during model training.

[0110] Among them, C all This represents the ability of the selected features to represent the complete feature set, that is, the explanatory power for all features. true and MI true These represent the correlation and mutual information between the selected features and the true value, respectively. C... fac and MI fac This indicates the correlation and mutual information between the selected features. When C all An increase in the value indicates that the specified subset of features has sufficient capacity to replace all features, proving the necessity and feasibility of feature optimization; C true With MI true The addition of this feature subset indicates that it is closely related to the true value of tillering from both linear and nonlinear perspectives, and has great potential to improve the performance of inversion modeling; while C fac With MI fac Reducing the size of the subset of features means that the selected subsets of features are almost independent of each other and do not interfere with each other, which greatly reduces the risk of feature redundancy or getting trapped in local optima.

[0111] In the Grey Wolf optimization algorithm, the convergence factor is a parameter used to control the convergence speed of the algorithm or to balance global and local search. It determines the magnitude of individual position updates during the iteration process. Typically, the convergence factor ranges from 0 to 1; the closer the value is to 1, the slower the convergence speed and the larger the individual position update magnitude, which is beneficial for escaping local optima and moving towards the global optimum. The closer the value is to 0, the faster the convergence speed and the smaller the individual position update magnitude, which is beneficial for refining the solution within the local search space. Therefore, to better balance global and local search capabilities, this study proposes a nonlinear decay factor based on a cosine function, distinct from traditional linear decay factors, such as... a1 can take the empirical value of 1, a T The acceptable experience value is 0. Figure 6 From the iterative changes of the convergence factor, in the early stages of iteration Maintaining a relatively large value for a long period; while in the later stages This allows for maintaining a relatively small convergence factor. This non-linearly decaying convergence factor enables a greater focus on global search in the early stages of the optimization process, and a greater focus on local search in the later stages, thus allowing for more precise adjustment of individual positions within the local search space.

[0112] Considering the inherent randomness of the traditional gray wolf optimization algorithm during the search process, it exhibits a bias in the selection of the initial population and the movement strategies of individuals. It tends to explore and utilize as much information as possible in the search space to find the global optimum. However, due to the influence of its search strategy and parameter settings, especially the aforementioned nonlinear correction to the convergence factor, when entering the later stages of iteration, i.e., the loss region (see...),... Figure 6 This could lead to the algorithm getting stuck in a local optimum or converging prematurely. In other words, the wolf pack's encirclement might be limited by the loss zone. The decay occurs rapidly, but if the alpha wolf does not always move towards the global optimum, the nonlinear decay will accelerate the result into a local optimum. To compensate for the loss, this study attempts to replace the weaker hunting wolves in the latter half of the pack with new random wolves in the later stages of the iteration (see...). Figure 6 The remaining wolves follow the same pattern as the Global Optimum (GWO). This provides a new hunting direction for the wolves that may have fallen into local optima in the later stages of the iteration, and also avoids local optima caused by sacrificing global search efficiency. Therefore, this strategy provides a way to refresh the wolf pack in the later stages of the algorithm, increases the diversity of the algorithm, and enables the wolf pack to better escape local optima and move towards the global optimum.

[0113] In an exemplary embodiment, the Support Vector Machine (SVM) model in S106 is favored for its excellent performance on small sample datasets, its efficient processing capability for high-dimensional data, and its strong generalization ability. Based on these characteristics, this application utilizes the SVM model to establish an estimation model based on the aforementioned 13 target features and the true value of tiller number.

[0114] In this experiment, the Gaussian radial basis function (RBF) is used as the kernel function of the SVM model, and its mathematical form is as follows: K(x i x j )=exp(-γ||x i -x j || 2 ), where x i and x j The input data consists of two sample points. γ is a parameter controlling the width of the function, which controls the complexity of the RBF kernel's feature space and the smoothness of the decision boundary. Adjusting this parameter allows for a trade-off between overfitting and underfitting biases. Additionally, parameter c is the penalty parameter in SVM, controlling the classifier's punishment for misclassification. To find the optimal combination of model parameters and minimize overfitting or underfitting, a grid search method is used. Typically, parameter c is set to range from [-10, 10], and parameter γ to range from [0, 1]. By traversing all possible parameter combinations and evaluating the model's performance using cross-validation or other evaluation metrics, the best-performing parameter combination is ultimately determined.

[0115] Based on this, the accuracy changes of the inversion model were compared under two control scenarios: optimized feature modeling and no crop pixel extraction, and no feature selection. Figure 7 As shown. Among them, Figure 7 Part (a) presents the model results based on the training set data. Figure 7Part (b) presents the model results based on the test set data. Observations show that after feature optimization using the CMI-IGWO algorithm, the R² of the training set prediction results is 0.6582, and the R² of the validation set inversion results reaches 0.7527, indicating a significant improvement in the winter wheat tiller number inversion model after feature optimization using the CMI-IGWO search algorithm. The lack of soil background removal or screening of honorary features leads to a deterioration in tiller inversion results, demonstrating that crop pixel extraction and feature optimization help improve the accuracy of the final tiller inversion model. Soil information interference severely affects the sensitivity and correlation of the extracted features to tiller number, making it difficult to retain high-quality features beneficial for tiller number inversion modeling through feature optimization later, proving the necessity and effectiveness of soil background removal. To further demonstrate the statistical significance of the model trained with optimized features, calculations showed that the p-values ​​of the T-distribution, which measures accuracy, are all less than 0.05, meaning that the model's inversion results not only perform well on the training set but also have high accuracy on the test set.

[0116] In an exemplary embodiment, S107 employs the determination coefficient R. 2 The accuracy of the model inversion results is described by the root mean square error (RMSE). 2 R measures the model's ability to interpret and fit the actual observations, and its value ranges from 0 to 1. 2 A value closer to 1 indicates that the model better explains the variance of the target variable, meaning the model has a higher degree of fit. RMSE measures the magnitude of the error in the model inversion. The smaller the RMSE, the smaller the difference between the model's predictions and the actual observed values, meaning the higher the model's inversion accuracy.

[0117]

[0118] in, y represents the model's predicted value. i This represents the actual observed values, and m is the sample size. The closer the correlation coefficient is to 1 and the smaller the root mean square error, the better the model performs.

[0119] like Figure 8 As shown, the number of tillers in the entire area of ​​the two experimental fields was estimated using a pre-trained optimal inversion model, ultimately yielding the inversion distribution results of winter wheat tiller numbers across the entire study area. The high-precision tiller number inversion results not only reflect the growth status of wheat but also serve as an effective basis for guiding agricultural fertilization strategies, thereby maximizing wheat yield and quality.

[0120] This application has the following advantages:

[0121] (1) A spectral data preprocessing strategy was specifically designed before image segmentation to provide data support for subsequent pixel recognition. The design philosophy consistently revolves around the principle of amplifying the differences in spectral information of ground objects, while fully considering the interference and influence of vegetation cover on the amplification effect. By analyzing the trend and local characteristics of the statistical curves of the near-infrared spectral data of the samples, an adaptive enhancement factor was designed using mathematical functions. This enhancement factor, in its dot product operation with the original near-infrared reflectance data, can find a suitable amplification coefficient for each pixel in the sample image. This preprocessing method provides a feasible solution for enhancing spectral differences and contributes a new strategy for variable data enhancement using mathematical functions.

[0122] (2) The polarization concept is introduced to approximate binarize the segmentation matrix from a binary classification perspective. This process ensures that the set threshold has sufficient criteria to complete pixel classification during the segmentation process. The final threshold of this adaptive segmentation method can be fixed at the median of 0.5, and does not shift with environmental changes or sample differences, thus providing good interpretability as a segmentation threshold. Compared with other commonly used image thresholding segmentation methods, this method is more robust and effective in image segmentation under different environmental conditions.

[0123] (3) The designed multi-strategy fusion improved gray wolf search algorithm boldly adopts the knowledge of statistics and information theory to jointly construct evaluation indicators. Pearson correlation coefficient and mutual information are selected as two major statistical indicators to fully explore the linear and nonlinear relationships between multiple variables, helping to screen the optimal feature set with the lowest feature redundancy, best representing all features, and closest relationship to the true value. At the same time, this optimized search algorithm breaks away from the linearly decaying convergence speed of the wolf pack search range in traditional algorithms. It uses the convergence factor of nonlinear transformation to balance the tendency of global search ability and local search ability at different stages during the gray wolf hunting process, striving to quickly find the global optimum. Furthermore, it innovatively introduces a random wolf pack in the later stages of iteration to replace the less fit group, creating the possibility of jumping out of the local optimum again under the condition of tending towards local search.

[0124] In one exemplary embodiment, this application provides a winter wheat tiller number monitoring device comprising:

[0125] The data acquisition module is used to acquire ground-measured tillering data and UAV hyperspectral images of wheat canopy;

[0126] Spectral enhancement processing is used to perform near-infrared spectral adaptive enhancement preprocessing on hyperspectral images of wheat canopy from UAVs. The spectral enhancement preprocessing is used to determine vegetation density index based on the near-infrared spectral matrix, and then determine vegetation density gradient level based on the vegetation density index. After that, an adaptive enhancement factor is determined based on the vegetation density gradient level, and finally spectral enhancement is achieved based on the adaptive enhancement factor.

[0127] The image segmentation module is used to perform adaptive image segmentation on spectrally enhanced images based on error pixel correction and polarization segmentation methods to obtain a mask image containing crop information.

[0128] The feature extraction module is used to extract feature data from a mask image containing crop information and construct a multi-angle spectral feature library; the feature data includes: spectral features, texture features, color features, and shape features;

[0129] The target selection module is used to select target features from the graph feature library using an optimized gray wolf search algorithm that integrates multiple strategies.

[0130] The prediction model determination module is used to construct and train a winter wheat tiller number prediction model using target features and ground-measured tiller data, employing a support vector machine (SVM) model.

[0131] The tiller number inversion module is used to invert the tiller number in the study area using the winter wheat tiller number prediction model.

[0132] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0133] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0134] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0135] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0136] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0137] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0138] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for monitoring the tiller number of winter wheat, characterized in that, The winter wheat tiller number monitoring method comprises: Obtaining ground measured tiller data and unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image; Performing near-infrared spectrum self-adaptive enhancement preprocessing on the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image; the near-infrared spectrum self-adaptive enhancement preprocessing is used for determining a vegetation density index according to a near-infrared spectrum matrix, and then determining a vegetation density gradient level according to the vegetation density index; then, an adaptive enhancement factor is determined according to the vegetation density gradient level, and finally, spectrum enhancement is realized according to the adaptive enhancement factor; Performing adaptive image segmentation on the spectrum-enhanced image based on an error pixel correction method and a polarization segmentation method to obtain a mask image containing crop information; Feature data is extracted from the mask image containing crop information, and a multi-angle atlas feature library is constructed; the feature data comprises spectral features, texture features, color features and shape features; An optimized grey wolf search algorithm with multi-strategy fusion is used to screen target features from the atlas feature library; A support vector machine (SVM) model is used to construct and train a winter wheat tiller number prediction model by using the target features and the ground measured tiller data; The winter wheat tiller number prediction model is used to perform tiller number inversion on the research area; The near-infrared spectrum self-adaptive enhancement preprocessing on the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image specifically comprises: Using the formula determining the vegetation density index; The average value of the vegetation density index is used to determine the vegetation density gradient level; Using the formula determining an adaptive enhancement factor; Point multiplication operation is performed on the near-infrared statistical curve of the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image and the adaptive enhancement factor to realize the near-infrared spectrum self-adaptive enhancement preprocessing; wherein, R is a vegetation density index, nir R is a near-infrared reflectance at 850 nm, red R is a red band reflectance at 650 nm, h is an adaptive enhancement factor, ∈ i × 10%, i i = 1, 2,..., 5, for determining the position of the starting point P in the effective enhancement process, is a parameter whose value is inversely proportional to the vegetation density gradient level, ∈ j × 10%, j i = 6, 7, 8, for determining the position of the turning point Q when the enhancement intensity begins to weaken, represents the horizontal coordinate position of the Q point, , h is the horizontal coordinate maximum value of the adaptive enhancement factor, h ∈ (1, 1.15), h is used to determine the maximum value of the adaptive enhancement factor function or the position of the Q point.

2. The winter wheat tiller number monitoring method according to claim 1, characterized in that, The obtaining of the ground measured tiller data and the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image specifically comprises: A hyperspectral high-definition camera carried by the unmanned aerial vehicle is used to obtain the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image; Pix4D series software is used to pre-process and analyze the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image; the pre-processing and analysis comprises image stitching, band synthesis, atmospheric correction and geometric correction; An image processing platform ENVI is used to perform image processing on the image pre-processed and analyzed; the image processing comprises adjusting composition, adjusting color scale, adjusting white balance and increasing sharpness.

3. The winter wheat tiller number monitoring method according to claim 1, characterized by, The adaptive image segmentation based on the error pixel correction method and the polarization segmentation method on the spectrum-enhanced image to obtain the mask image containing crop information specifically comprises: Using the formula determining the enhanced spectral pull-up index ; wherein, and are wavelength values of the near-infrared band and the red band, respectively, is the near-infrared reflectance of the image of spectral enhancement; Using the formula determining the correction factor ; is the original NDVI matrix within the sample area, max ( ) represents the maximum value, mean ( ) represents the mean value, min ( ) represents the minimum value; Utilizing an enhanced spectral pull-up index of a segmented image matrix determining a theoretical segmentation threshold ; Using the formula Implementing the correction factor Adapting the theoretical split threshold Numerically adjusting adaptively and constructing a dynamic error correction identification index sensitive to green pixels k ; Dynamic type error correction identification index k With the segmented image matrix Each pixel is compared, and a pixel-level recognition classification task is determined according to the comparison result.

4. The winter wheat tiller number monitoring method according to claim 1, characterized by, The optimized grey wolf search algorithm with multi-strategy fusion is used to screen target features from the atlas feature library, specifically comprising: Using the formula determine the evaluation function ; where, represents the ability of the selected features to represent the complete set of features, and represent the correlation and mutual information between the selected features and the true value, respectively, and represent the correlation and mutual information between the selected features; Using the formula determining the nonlinear convergence factor ; wherein, and are the start and end values of the convergence factor, respectively, is the total number of iterations.

5. A winter wheat tiller number monitoring device for implementing the winter wheat tiller number monitoring method according to any one of claims 1 to 4, characterized by The winter wheat tiller number monitoring device comprises: A data acquisition module is configured to obtain ground measured tiller data and unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image; A spectrum enhancement processing is configured to perform near-infrared spectrum self-adaptive enhancement preprocessing on the unmanned aerial vehicle hyperspectral wheat canopy hyperspectral image; the near-infrared spectrum self-adaptive enhancement preprocessing is used for determining a vegetation density index according to a near-infrared spectrum matrix, and then determining a vegetation density gradient level according to the vegetation density index; then, an adaptive enhancement factor is determined according to the vegetation density gradient level, and finally, spectrum enhancement is realized according to the adaptive enhancement factor; An image segmentation module is configured to perform adaptive image segmentation on the spectral-enhanced image based on an error pixel correction method and a polarization segmentation method to obtain a mask image containing crop information. A feature extraction module is configured to extract feature data from the mask image containing crop information and construct a multi-angle atlas feature library; the feature data includes spectral features, texture features, color features, and shape features. A target screening module is configured to screen target features from the atlas feature library using a multi-strategy fusion optimization grey wolf search algorithm. A prediction model determination module is configured to construct and train a winter wheat tiller number prediction model using a support vector machine (SVM) model based on the target features and ground measured tiller data. A tiller number inversion module is configured to perform tiller number inversion on a study area using the winter wheat tiller number prediction model.

6. A computer device comprising: A memory and a processor to store a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the winter wheat tiller number monitoring method of any one of claims 1-4.

7. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the winter wheat tiller number monitoring method of any one of claims 1-4.

Images