A method for early warning of crop nutrition status based on ear phenotype characteristics

Abstract

The present application provides a kind of based on crop ear phenotype feature's nutrition condition early warning method, belong to the crop nutrition management technical field based on computer data processing;Through the high-throughput phenotyping robot in field cooperates to obtain hyperspectral image and high-density three-dimensional point cloud data, constructs and includes pure spectrum and fine three-dimensional form multi-modal ear feature atlas.Based on multiscale residual attention network and chemometric balance composite loss function, the method introduces the physical constraint of crop nutrient cooperation and antagonism in deep learning model, realizes the synchronous high-precision prediction of nitrogen, phosphorus and potassium content;Further combined with conformal prediction technology and nutrient balance method, generate the variable fertilization prescription map with statistical confidence guarantee.The present application not only solves the problem of weak signal extraction of ear in complex field background, but also ensures that the prediction result conforms to the biological law, provides scientific, reliable and quantifiable decision support for accurate fertilization in the middle and later stages of crop growth.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of crop nutrition management based on computer data processing, and particularly relates to a nutrition condition early warning method based on crop ear phenotype characteristics. BACKGROUND

[0002] In precision agriculture management, the nutrition condition monitoring of crop heading and grain filling period is a key link to determine yield and quality. However, most of the current crop nutrition monitoring still relies on traditional destructive sampling analysis or large-scale remote sensing monitoring, which not only has poor timeliness and high cost, but also is difficult to accurately capture the subtle physiological changes of the crop ear. Especially in the middle and late growth period of crops, the nitrogen, phosphorus and potassium contents of the ear as a sink organ are directly related to the final grain filling quality. However, the complex field environment, leaf shading, soil background interference and synergistic antagonism between multiple elements make it difficult for a single dimension monitoring method to obtain accurate, comprehensive and biologically regular nutrition information, resulting in a lack of scientific basis for fertilization decision-making, and further affecting the fertilizer utilization rate and ecological environment safety. Therefore, there is an urgent need for an intelligent nutrition early warning method that can integrate micro-spectrum and macro-morphological characteristics, follow the physiological mechanism of plants and have the ability to quantify uncertainty.

[0003] The existing crop nutrition condition recognition methods mainly include the following three kinds:

[0004] Manual sampling and chemical wet analysis: this method relies on manual collection of representative samples in the field, and after drying and crushing, the samples are determined in the laboratory by using Kjeldahl nitrogen determination method, vanadium molybdenum yellow colorimetric method, etc. Although this method has high measurement accuracy, its operation process is complicated, destructive and seriously lagging, and it cannot meet the needs of large-area field real-time monitoring and rapid decision-making, and the number of sampling points is limited, which is difficult to reflect the spatial variability within the field;

[0005] Canopy spectral remote sensing monitoring method: this method uses satellites, unmanned aerial vehicles or handheld spectrometers to obtain crop canopy spectral information, and inverses the nutrient content through vegetation index. The disadvantage is that the canopy spectrum often mixes a large amount of leaf, stem and soil background signals, and it is difficult to accurately separate the pure spectrum of the key organ of the ear from the mixed pixels; in addition, relying only on spectral information ignores the three-dimensional morphological structure of crops such as ear volume and fullness, and loses the geometric features closely related to biomass, resulting in spectral saturation phenomenon when the biomass is saturated in the middle and late growth period of crops, and the inversion accuracy is limited;

[0006] Traditional data-driven machine learning regression methods: This method directly inputs spectral data into a neural network or a statistical model for black-box training to establish a mapping relationship between the input and the nutrient content. Although it improves the prediction efficiency to some extent, it has the disadvantage that the model lacks physical interpretability and often ignores the inherent stoichiometric balance rules between nitrogen, phosphorus and potassium in crops. At the same time, existing models are mostly deterministic point predictions, which cannot quantify the uncertainty risk of the prediction results, leading to false positives in non-steady-state environments in the field and difficulty in generating high-confidence fertilization prescriptions.

[0007] In summary, the existing crop nutrition monitoring methods generally ignore the ear, lack of physical mechanism, low utilization rate of multi-modal information, and lack of risk assessment. SUMMARY

[0008] To solve the above problems, the present application provides a nutrition condition early warning method based on crop ear phenotypic characteristics, comprising the following steps:

[0009] S1, during the heading and grain filling period of crops, scanning to obtain a hyperspectral image data cube and high-density three-dimensional point cloud data of the crop canopy, and simultaneously obtaining the three-dimensional geographic coordinates of the scanning sample points;

[0010] The hyperspectral image data cube and the high-density three-dimensional point cloud data are registered using external calibration parameters, and a spatial bounding box is constructed according to the three-dimensional geographic coordinates for regional cropping to extract target ear region data containing spectral reflectance vectors and spatial coordinate vectors;

[0011] S2, in the spectral dimension, the target ear region data is subjected to ear spectral purification, and after spectral transformation and feature band selection, a nutrient characteristic spectral vector is constructed. In the spatial dimension, the high-density three-dimensional point cloud data in the target ear region data is used to mine geometric features reflecting ear fullness, and a morphological structure parameter vector is constructed. The nutrient characteristic spectral vector and the morphological structure parameter vector are processed and spliced to generate a multi-modal ear feature spectrum;

[0012] S3, inputting the multi-modal ear feature spectrum into the trained online nutrition early warning model to output the predicted values of nitrogen, phosphorus and potassium content;

[0013] S4, using the calibration data set to calculate the non-consistency score and determine the conformal calibration threshold, constructing the edge prediction interval, and when both the point prediction value and the upper bound of the edge prediction interval are lower than the lower threshold of the nutrient content at a certain growth stage, it is determined that there is a nutrient deficiency and the early warning is triggered.

[0014] Preferably, in S1, the hyperspectral image data cube and the high-density three-dimensional point cloud data are projected to a unified world coordinate system using a rotation matrix and a translation vector describing the relative position and attitude relationship between the hyperspectral imager and the lidar, and a graph fusion data containing both 256-band spectral information and spatial three-dimensional coordinates of each pixel point is generated by nearest neighbor interpolation algorithm registration; according to the recorded centimeter-level three-dimensional geographic coordinates of each scanning point, a fixed physical size of 20 cm by 20 cm space bounding box is constructed with the coordinates as the center, and the local data subset falling within the bounding box is cut out from the graph fusion data as the target ear region data; the target ear region data specifically includes two hundred and fifty-six spectral reflectance values of each pixel point in the four hundred nanometer to one thousand nanometer wavelength range, recorded as a spectral reflectance vector, and the spatial three-dimensional coordinate value of the pixel point in the world coordinate system, recorded as a spatial coordinate vector.

[0015] Preferably, in S2, in the spectral dimension, a pure ear average spectral curve is purified from the target ear region data using a full-constrained least squares linear spectral unmixing algorithm, and after double transformation, a competitive adaptive reweighted sampling algorithm is used to select key bands to construct a nutrition feature spectral vector; in the spatial dimension, an Alpha-shape algorithm is used to reconstruct a non-convex hull for the denoised three-dimensional point cloud data, core geometric indicators are calculated and morphological feature interaction dimensionality processing is performed to construct a high-dimensional morphological structure parameter vector; finally, an independent standardization strategy is used to process the above nutrition feature spectral vector and morphological structure parameter vector, and a multi-modal collaborative feature vector is generated by feature splicing, and a multi-modal ear feature graph composed of the multi-modal collaborative feature vector is output.

[0016] Preferably, the process of purifying a pure ear average spectral curve from the target ear region data using a full-constrained least squares linear spectral unmixing algorithm is as follows:

[0017] Firstly, based on the spectral reflectance vector, an endmember matrix containing four basic components is constructed. The typical spectral reflectance vectors of healthy ear, crop leaf, soil background and shadow are artificially selected from the spectral reflectance vectors corresponding to all pixel points, and the four vectors are taken as basic endmembers to construct a standard endmember matrix. Secondly, the full-constrained least squares linear spectral unmixing algorithm is used to calculate the abundance coefficient of each pixel point on the above four endmembers, and a linear spectral mixing model is constructed. The spectral reflectance vector of any pixel point in the target ear region data is modeled as the result of mixing the four types of endmember spectra in the above standard endmember matrix at a specific ratio. The optimization objective is to minimize the root mean square error between the spectral reflectance vector and the reconstructed spectrum, while imposing non-negative constraints on the abundance, i.e. the abundance coefficient of each type of endmember is not less than zero, and the sum of the abundance coefficients of the four types of endmembers is strictly equal to 1. The proportion of healthy ear component in each pixel point is calculated by iterative operation. Finally, a threshold is set, and only the pixel points with a healthy ear endmember abundance higher than the threshold are retained. The spectral reflectance mean value is obtained by arithmetically averaging the spectral reflectance vectors corresponding to these pixel points, and the pure ear average spectrum curve of the sample is generated based on the spectral reflectance mean value of these pixel points.

[0018] Preferably, the competitive adaptive reweighted sampling algorithm is used to screen the key wavebands, and the nutrient feature spectrum vector is constructed, and the specific process is as follows:

[0019] The 256 waveband spectrum data contained in the pure ear average spectrum curve after envelope removal operation and spectral second derivative transformation are selected for feature selection. A preset iteration number M is set, and the following steps are sequentially performed in each iteration process:

[0020] Firstly, a certain proportion of samples are randomly selected by Monte Carlo sampling method to establish a partial least squares regression model, and the regression coefficient absolute value corresponding to each waveband in the partial least squares regression model is extracted as the importance weight of the waveband. Secondly, the number of wavebands that should be retained in the current iteration round is calculated by using an exponential decay function with the current iteration number and the total iteration number as variables, and those wavebands with low weight ranking are forcibly removed. Thirdly, the adaptive reweighted sampling technique is used to normalize the regression coefficient weights of each waveband into selection probabilities, and the specific wavebands entering the next iteration round are randomly selected according to the probabilities. Fourthly, the root mean square error of the current waveband subset is calculated by cross-validation. Finally, after all iteration rounds are completed, the waveband subset corresponding to the minimum root mean square error is selected as the final optimal feature waveband combination, and the nutrient feature spectrum vector is constructed together with the key wavebands having the strongest correlation with the true values of nitrogen, phosphorus and potassium.

[0021] Preferably, the specific process of constructing the high-dimensional morphological structure parameter vector is as follows:

[0022] Firstly, the statistical outlier removal algorithm is used to remove the noise points in the high-density three-dimensional point cloud data, and the denoising three-dimensional point cloud data is obtained; secondly, the Alpha-shape algorithm is used for three-dimensional contour reconstruction of the denoised point cloud, a rolling ball radius parameter is set, a ball is simulated to roll outside the point cloud, and the gap that the ball cannot roll into forms a boundary to generate a non-convex hull closely fitting the concave and convex details of the ear part surface; based on the non-convex hull, three core morphological indexes are calculated in turn: firstly, the hull volume is calculated to quantitatively represent the biomass size of the ear part; secondly, the ratio of surface area to volume is calculated to reflect the tightness of the grain arrangement of the ear part; thirdly, the variance of the point cloud normal vector is calculated to represent the rough texture characteristics of the ear part surface; finally, based on the three core indexes of the hull volume, the ratio of surface area to volume and the variance of the point cloud normal vector, the polynomial expansion technique is used to calculate the quadratic interaction product term between different morphological indexes and the square term of each index itself, and a high-dimensional morphological expansion vector containing the original indexes and their nonlinear combinations is constructed, which is denoted as a morphological structure parameter vector.

[0023] Preferably, the online nutrition early warning model is a multi-task network composed of a shared feature encoder and a task-specific decoder.

[0024] The input of the shared feature encoder is a multi-modal ear feature map, and the main body is composed of an initial convolutional layer and five cascaded deep residual attention modules. Each residual attention module contains two paths: the main path uses two one-dimensional convolutional layers with convolution kernel sizes of 3 and 1 to extract local spectral and morphological context features, and is connected to a batch normalization layer and an exponential linear unit activation function in turn; the bypass path embeds a compression and excitation attention submodule, which first compresses the feature map into a channel descriptor through global average pooling, then uses two fully connected layers to reduce and then increase the dimension, and generates channel importance weights through a Sigmoid function, which are multiplied with the feature map of the main path channel by channel.

[0025] At the end of the shared feature encoder, three independent fully connected decoding branches are introduced in parallel, corresponding to the nitrogen, phosphorus and potassium prediction tasks respectively; each branch is designed as a bottleneck expansion structure, containing a fully connected layer with a neuron number decreasing from 256 to 64 to compress the features, a Dropout layer, and a linear output layer, finally outputting the three nutrient content prediction values of the crop ear in parallel.

[0026] Preferably, the training target of the online nutrition warning model is to minimize a composite objective function dynamically weighted by regression accuracy loss and physical consistency loss; the first part of the composite objective function is a multi-task robust regression loss, which calculates the Huber loss between the predicted value and the true value for nitrogen, phosphorus and potassium tasks, which combines the advantages of mean square error and absolute error, and sets a hyperparameter d as a threshold, when the prediction error is less than the threshold, a quadratic function is used for punishment to ensure the derivability, and when the error is greater than the threshold, a linear function is used for punishment to reduce the interference of outliers on the gradient; the second part of the composite objective function is a stoichiometric ratio consistency loss, which is based on the stoichiometric balance principle of plant nutrition to construct the element ratio matrix of the predicted value and the true value.

[0027] Preferably, the conformal calibration threshold is specifically: the uncertainty of the model is statistically calibrated by using the reserved calibration data set, for each sample in the calibration set, the non-consistency score is calculated, which is defined as the absolute value of the model prediction error; then, the non-consistency scores of all calibration samples are sorted from small to large, and according to the preset significance level 0.1, which corresponds to the confidence of ninety percent, the quantile of the score distribution is calculated, which is determined as the conformal calibration threshold, the threshold represents the statistical upper bound of the model prediction error under the current confidence level.

[0028] Preferably, it also includes a fertilization prescription map generation based on the nutrient balance method, specifically:

[0029] First, multiply the target yield by the nutrient requirement per unit yield to obtain the total nutrient demand of the crop during the whole growth period; second, subtract the soil basic nutrient supply and the absorbed nutrient amount of the current crop from the total demand, wherein the absorbed nutrient amount is converted from the shell volume of the non-convex hull to the ear dry matter mass by using the preset ear packing density coefficient, and then the ear dry matter mass is multiplied by the nutrient content point prediction value to obtain the absorbed nutrient amount; finally, divide the difference value by the fertilizer utilization rate in the season to obtain the theoretical fertilizer requirement at the point.

[0030] After the above calculation is completed for all scanning points in the field, the Kriging interpolation algorithm is used to generate a nitrogen, phosphorus and potassium variable fertilization prescription map covering the whole field.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] (1) Ear multi-modal feature unmixing and fusion technology: In view of the problem of serious field background noise interference, a feature enhancement strategy based on full constraint least squares linear spectral unmixing and Alpha-shape algorithm is proposed. First, the pure ear spectrum is purified from the mixed pixels by using the end member unmixing algorithm, and the key nutrient bands are selected by combining with the competitive adaptive reweighted sampling; At the same time, the Alpha-shape algorithm is used to reconstruct the non-convex hull of the ear, and the high-dimensional morphological features reflecting the grain fullness such as hull volume and surface area volume ratio are extracted. By fusing spectral features and three-dimensional morphological features, the model's representation ability for crop ear nutrient status is significantly improved;

[0033] (2) Physical perception network design based on chemometrics constraint: A multi-scale residual attention network and its supporting chemometrics balance composite loss function are proposed. In the model training process, not only the regression error is minimized, but also the consistency constraint of element ratio is introduced, forcing the deep learning model to follow the inherent nutrient synergistic absorption and antagonistic competition rules of crops. This physical perception mechanism effectively solves the problem of ignoring the coupling relationship between elements in traditional models when predicting multiple tasks, and improves the generalization robustness and biological rationality of the model in different field environments;

[0034] (3) Reliability decision and prescription generation based on conformal prediction: Abandoning the traditional single point prediction mode, the conformal prediction framework is introduced to statistically calibrate the uncertainty of the model, and the edge prediction interval containing the true value is constructed. The system only triggers confident deficiency warning when both the point prediction value and the upper limit of the interval are lower than the nutrient threshold, and combines the target yield and ear dry matter estimation to generate precise variable fertilization prescription using nutrient balance method. This mechanism provides a decision boundary with statistical confidence guarantee for agricultural production, effectively reducing the risk of blind fertilization caused by model misjudgment. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 The whole technical route flowchart of the application.

[0036] Figure 2 The key flowchart of the application for multi-modal atlas fusion of ear and nutrient dataset construction.

[0037] Figure 3 The atlas construction flowchart of the application for ear spectral feature unmixing and enhancement.

[0038] Figure 4 The key architecture diagram of the application for multi-scale residual attention network.

[0039] Figure 5 The nutrient status warning and fertilization prescription generation flowchart of the application based on conformal prediction.

[0040] Figure 6Figure 1 is a multi-dimensional nutrition inversion error distribution map in the embodiment.

[0041] Figure 7 Figure 2 is a comparison chart of error between the physical constraint model and the unconstrained model in different fields in the embodiment.

[0042] Figure 8 Figure 3 is a real-time monitoring result chart of nutrition warning and fertilization decision based on conformal prediction in the embodiment. DETAILED DESCRIPTION

[0043] The present application proposes a nutrition condition warning method based on the phenotypic characteristics of crop ear, and the overall process is as shown in Figure 1 .

[0044] S1, ear multi-modal atlas fusion and nutrition data set construction: first, a push-broom hyperspectral imager and a high-resolution solid-state laser radar mounted on a field high-throughput phenotyping robot are used to perform collaborative scanning during the crop heading and grain filling period, and high-spectral image data cubes and high-density three-dimensional point cloud data of the crop canopy are synchronously acquired; at the same time, real-time dynamic carrier phase difference positioning technology is used to record the centimeter-level three-dimensional geographic coordinates of the scanning points, and gradient sampling and chemical wet analysis of physical samples are performed according to the three-dimensional geographic coordinates, and a multi-dimensional nutrition true value label containing nitrogen, phosphorus and potassium concentrations is established. Subsequently, the spatial and temporal registration of heterogeneous data is completed by using external calibration parameters, and a spatial bounding box is constructed according to the three-dimensional geographic coordinates for accurate cutting, and the target ear region data containing spectral reflectance vectors and spatial coordinate vectors are extracted. Finally, the region data and the corresponding multi-dimensional nutrition true value label are associated one by one, and a structured crop ear nutrition data set is output.

[0045] S2, ear atlas feature unmixing and enhanced atlas construction: this step performs deep feature engineering on the data set output by S1. In the spectral dimension, the pure ear average spectral curve is purified from the target ear region data by using the full-constrained least squares linear spectral unmixing algorithm, and after double transformation, the competitive adaptive reweighted sampling algorithm is used to select the key waveband, and the nutrition feature spectral vector is constructed; in the spatial dimension, the Alpha-shape algorithm is used to reconstruct the non-convex hull for the denoised three-dimensional point cloud data, the core geometric indexes are calculated, and the morphological feature interactive dimensionality processing is performed, and the high-dimensional morphological structure parameter vector is constructed. Finally, the independent standardization strategy is used to process the above-mentioned nutrition feature spectral vector and morphological structure parameter vector, and the multi-modal ear feature atlas is generated by feature splicing.

[0046] S3, online nutrition warning model construction based on stoichiometric constraints: this step uses the multi-modal spike feature map output by S2 to construct and train a multi-scale residual attention network. The network is composed of a shared feature encoder containing a deep residual attention module and three parallel task-specific decoders, aiming to map the input multi-modal spike feature map to the predicted values of nitrogen, phosphorus and potassium content. During training, a stoichiometric balance composite loss function is introduced to minimize the Huber regression loss and stoichiometric ratio consistency loss, forcing the model to learn the absolute content while following the inherent nutrient synergism and antagonism rules of crops. After training, the adaptive optimization strategy is used to solidify the parameters, and the online nutrition warning model with physical constraint characteristics is output.

[0047] S4, nutrition status warning and fertilization prescription generation based on conformal prediction: this step deploys the online nutrition warning model in the field. First, use the calibration data set to calculate the non-consistency score and determine the conformal calibration threshold to quantify the uncertainty of the model. In real-time operation, use the model to obtain the point prediction value of the field spike data, and combine the conformal calibration threshold to construct the edge prediction interval. According to the decision logic, only when the point prediction value and the upper limit of the edge prediction interval are both lower than the nutrient lower limit threshold of the specific growth stage, it is determined that there is a confident nutrient deficiency and the warning is triggered. Finally, based on the confident deficiency result, the theoretical fertilizer requirement is calculated using the nutrient balance method, and the variable fertilization prescription map is generated by Kriging interpolation to guide the precise fertilization operation.

[0048] The specific implementation process of the present application will be described in detail below in conjunction with specific embodiments.

[0049] S1, multi-modal spike atlas fusion and nutrition data set construction

[0050] This step aims to construct a high-precision atlas data set containing the external spectral morphology and internal nutrient composition of the crop spike using near-ground remote sensing technology, providing data support for the subsequent establishment of quantitative mapping models between crop spike phenotypic characteristics and nitrogen, phosphorus and potassium content. The key process is as shown in Figure 2

[0051] ​S1.1 Field ear atlas coordination with high-throughput scanning: At the heading and grain filling stage of the crop, a push-broom hyperspectral imager and a high-resolution solid-state laser radar mounted on a field high-throughput phenotyping robot are used to perform field scanning. The system sets the spectral range of the hyperspectral imager to 400-1000 nm, with a spectral resolution of 2.5 nm; the scanning frequency of the laser radar is set to 10 Hz. In the artificial auxiliary light source environment, the robot drives at a constant speed along the ridge, and synchronously obtains the hyperspectral image data cube and high-density three-dimensional point cloud data of the crop canopy. Before each scanning operation, a standard diffuse white board and a dark current correction cover are used for radiation calibration to convert the digital quantity DN value of the original image into accurate spectral reflectance data and eliminate the interference of light changes on the spectral characteristics.

[0052] S1.2 Target ear positioning and gradient sampling: Within the image acquisition field of view, the real-time dynamic carrier phase difference positioning technology is used to record the centimeter-level three-dimensional geographic coordinates of each scanning point. Specifically, a real-time dynamic positioning (RTK) mobile receiver is deployed on the phenotyping robot platform, high-precision positions are solved by receiving reference station differential signals, and hardware trigger pulses are used to establish strict time synchronization between the positioning system and the imaging system; combined with the pre-calibrated spatial eccentricity vector between the global navigation satellite system (GNSS) antenna phase center and the imaging sensor optical center, the original positioning data is corrected by rigid translation, so as to solve the accurate three-dimensional geographic coordinates of the ground projection point corresponding to the center of each image field of view in real time. A stratified random sampling method is used to select no less than two hundred scanning points as a verification set. At each verification point, a handheld chlorophyll meter is first used for non-destructive preliminary screening to ensure that the sampling sample covers the three gradient levels of nutrient deficiency, adequacy and surplus; then, ten representative ears are cut at the center of the scanning field of view as representative ear samples of the point. Finally, the representative ear samples of all scanning points are obtained.

[0053] S1.3 Nutrient component determination and multi-dimensional true value labeling: The ear samples of all scanning points collected are immediately frozen in liquid nitrogen and brought back to the laboratory for drying and crushing. Then standard chemical wet analysis is performed: Kjeldahl nitrogen determination method is used to determine the total nitrogen content, vanadium molybdenum yellow colorimetric method is used to determine the total phosphorus content, and flame photometric method is used to determine the total potassium content. The concentration data of the measured nitrogen, phosphorus and potassium elements are arranged in order to form a multi-dimensional nutrient true value label of the sample point.

[0054] S1.4 Spatial-temporal registration of atlas data and dataset output: Using the extrinsic calibration parameters, i.e. the rotation matrix and translation vector describing the relative position and pose relationship between the hyperspectral imager and the lidar, the hyperspectral image datacube and the high-density 3D point cloud data are projected to a unified world coordinate system. Through nearest neighbor interpolation algorithm registration, the atlas fusion data containing both 256-band spectral information and spatial 3D coordinates for each pixel point are generated. Finally, according to the 3D geographic coordinates of each scanning point recorded in S1.2, a fixed physical size of 20cm by 20cm space bounding box is constructed with the coordinate as the center. The local data subset falling into the bounding box is cropped from the above atlas fusion data as the target ear region data. The target ear region data specifically includes two hundred and fifty-six spectral reflectance values in the four hundred nanometer to one thousand nanometer wavelength range for each pixel point, denoted as the spectral reflectance vector, and the spatial 3D coordinate values corresponding to the pixel point in the world coordinate system, denoted as the spatial coordinate vector. The target ear region data is associated with the corresponding multi-dimensional nutrient true value label in S1.3 one by one, and the structured crop ear nutrient dataset is output.

[0055] S2, Ear atlas feature unmixing and enhanced atlas construction

[0056] This step aims to perform deep signal processing and feature engineering on the structured crop ear nutrient dataset output by S1.4, extract pure ear signals from the mixed field background, and mine deep spectral features and fine 3D morphological features sensitive to nitrogen, phosphorus and potassium nutrients. The key process of this step is shown in Figure 3

[0057] ​S2.1 Spike spectral purification based on endmember unmixing: In order to solve the problem that a small amount of leaf background and shadow noise is still contained in the spectral reflectance vector of the target spike region data output by S1.4, spectral unmixing processing is performed. First, an endmember matrix containing four types of basic components is constructed based on the spectral reflectance vector. Specifically, the typical spectral reflectance vectors of healthy spike, crop leaf, soil background and cast shadow are artificially selected from the spectral reflectance vectors corresponding to all pixel points, and these four vectors are taken as basic endmembers to construct a standard endmember matrix. Second, the abundance coefficients of each pixel point on the above four endmembers are calculated by using the fully constrained least squares linear spectral unmixing algorithm (FCLS-LSUM). Specifically, a linear spectral mixing model is constructed, and the spectral reflectance vector of any pixel point in the target spike region data is modeled as the result of mixing the spectra of the four types of endmembers in the above standard endmember matrix at a specific ratio; the optimization objective is to minimize the root mean square error between the spectral reflectance vector and the reconstructed spectrum, while imposing the non-negative constraint on the abundance, i.e. the abundance coefficient of each type of endmember is not less than zero, and the sum of the abundance is equal to one, i.e. the sum of the abundance coefficients of the four types of endmembers is strictly equal to 1, and the accurate proportion of the healthy spike component in each pixel point is solved by iterative operation. Finally, the abundance threshold is set to 0.75, only the pixel points with the abundance of healthy spike endmember higher than the threshold are retained, the spectral reflectance vectors corresponding to these pixel points are arithmetically averaged to obtain the spectral reflectance mean value, and the pure spike average spectral curve of the sample is generated based on the spectral reflectance mean value of these pixel points, thereby eliminating the interference of the environment background on the nutrition inversion.

[0058] S2.2 High-order spectral transformation and feature band screening: To eliminate baseline drift and enhance the weak nutrient absorption features, double transformation is performed on the pure ear average spectral curve extracted in S2.1. First, envelope removal operation is performed to normalize the reflectance curve to 0 to 1, so as to highlight the characteristic absorption valleys of nitrogen, phosphorus and potassium elements; secondly, the second-order differential of the spectrum is calculated to eliminate the low-frequency noise caused by uneven background light. Finally, the Competitive Adaptive Reweighted Sampling (CARS) algorithm is used to select features from the 256-band spectral data contained in the pure ear average spectral curve after the above double transformation. Specifically, the Competitive Adaptive Reweighted Sampling algorithm simulates the principle of survival of the fittest in biological evolution, sets a preset number of iterations M, and in each iteration process, the following four steps are executed in turn: First, a certain proportion of samples are randomly extracted by Monte Carlo sampling method to establish a partial least squares regression model, and the regression coefficient absolute value of each band in the partial least squares regression model is extracted as the importance weight of the band; secondly, the number of bands that should be retained in the current iteration round is calculated by using the exponential decay function with the current iteration number and the total iteration number as variables, and those bands with lower weight ranking are forcibly removed accordingly; thirdly, the adaptive reweighted sampling technique is used to normalize the regression coefficient weight of each band to a selection probability, and the specific bands entering the next iteration are randomly extracted according to the probability, to ensure that important bands have a higher probability of being retained; fourthly, the root mean square error of the current band subset is calculated through cross-validation. Finally, after all the iteration rounds are completed, the band subset corresponding to the minimum root mean square error is selected as the final optimal feature band combination, that is, the nutrient feature spectrum vector composed of a number of key bands with the strongest correlation with the true values of nitrogen, phosphorus and potassium in S1.3.

[0059] S2.3 Morphological feature calculation based on point cloud topology: This step aims to mine the geometric features reflecting the fullness of the ear from the high-density three-dimensional point cloud data in the target ear region data. First, a statistical outlier removal algorithm is used to remove stray noise points in the high-density three-dimensional point cloud data, obtaining denoised three-dimensional point cloud data. Specifically, the average distance of each point in the point cloud and its K nearest neighbors is calculated, and the mean and standard deviation of the global average distance are calculated. A distance threshold is set, and points with an average distance exceeding a certain multiple of the global mean standard deviation are determined as outliers and removed. Second, the Alpha-Shape algorithm is used to reconstruct the three-dimensional contour of the denoised point cloud. This algorithm is a non-convex hull construction technique based on the rolling ball principle. By setting a rolling ball radius parameter, a ball is simulated to roll outside the point cloud. The gap that the ball cannot roll into forms the boundary, thereby generating a non-convex hull that closely fits the surface details of the ear. Based on this non-convex hull, three core morphological indicators are calculated in turn: first, the hull volume is calculated to quantitatively represent the biomass size of the ear; second, the ratio of surface area to volume is calculated to reflect the tightness of the grain arrangement in the ear; third, the variance of the point cloud normal vector is calculated to represent the rough texture features of the ear surface. Finally, the morphological feature interaction dimensionality processing is performed. Specifically, based on the hull volume, the ratio of surface area to volume, and the variance of the point cloud normal vector, the quadratic interaction product terms between different morphological indicators and the square terms of each indicator are calculated using polynomial expansion techniques to construct a high-dimensional morphological expansion vector containing the original indicators and their nonlinear combinations, denoted as the morphological structure parameter vector.

[0060] S2.4 Feature map fusion and pre-processing set output: This step performs multi-modal fusion of spectral and morphological features. Z-Score standardization is performed on the nutrient feature spectrum vector and the morphological structure parameter vector, respectively, i.e., the arithmetic mean of each dimension is subtracted and divided by the standard deviation, mapping all feature values to a uniform distribution space with a mean of zero and a variance of one, thereby eliminating the dimension effect of heterogeneous data and achieving distribution alignment. Second, the feature splicing strategy is used to concatenate the standardized nutrient feature spectrum vector and the morphological structure parameter vector in the feature dimension, generating a multi-modal ear feature map as the direct input of the subsequent model.

[0061] S3, Online nutrition early warning model construction based on chemical metrological constraints

[0062] The step constructs and trains an online nutrition warning model containing feature encoding, attention enhancement and multi-head decoding, aiming to predict the contents of nitrogen, phosphorus and potassium through a shared backbone network, and introduce the element synergy and antagonism inherent in crops as a stoichiometric physical constraint, to solve the problem of ignoring the coupling relationship between elements in traditional single-task models, thereby significantly improving the prediction accuracy and generalization robustness of the model in complex field environments. The key network architecture of this step is shown in Figure 4

[0063] S3.1 Multi-scale residual attention network design: This step constructs a multi-task network composed of a shared feature encoder and a task-specific decoder. First, a shared feature encoder is constructed. The input of the encoder is the multi-modal spike feature map output by S2.4. The main body of the encoder consists of an initial convolutional layer and five cascaded deep residual attention modules. Each residual attention module contains two paths: the main path uses two one-dimensional convolutional layers with kernel sizes of 3 and 1 to extract local spectral and morphological context features, and is connected in turn with a batch normalization layer and an exponential linear unit activation function; the bypass path embeds a Squeeze-and-Excitation (SE) attention submodule. The submodule first compresses the feature map into a channel descriptor through global average pooling, then uses two fully connected layers to reduce and then increase the dimension, generates channel importance weights through the Sigmoid function, and finally multiplies the weights with the main path feature map channel by channel, thereby automatically suppressing background noise channels and enhancing the feature response of nitrogen, phosphorus and potassium sensitive bands. Second, at the end of the shared feature encoder, three independent fully connected decoding branches are introduced in parallel, corresponding to the nitrogen, phosphorus and potassium prediction tasks respectively. Each branch is designed as a bottleneck expansion structure, containing a fully connected layer with a neuron number decreasing from 256 to 64 to compress the features, a Dropout layer to prevent overfitting, and a linear output layer, finally outputting three nutrition content prediction values of the crop spike in parallel.

[0064] ​S3.2 Stoichiometric balance compound loss function design: The model is trained end-to-end supervisedly using the preprocessed dataset in S2.4, and the training objective is to minimize a compound objective function dynamically weighted by regression accuracy loss and physical consistency loss. The first part of the compound objective function is the multi-task robust regression loss, which calculates the Huber loss between the predicted value and the true value for the three tasks of nitrogen, phosphorus and potassium. This loss function combines the advantages of mean square error and absolute error, and sets a threshold d as the threshold. When the prediction error is less than the threshold, a quadratic function is used for punishment to ensure the derivability, and when the error is greater than the threshold, a linear function is used for punishment to reduce the interference of outliers on the gradient. The second part of the compound objective function is the stoichiometric ratio consistency loss. Based on the stoichiometric balance principle of plant nutrition, the element ratio matrix of the predicted value and the true value is constructed. Specifically, the logarithmic values of the nitrogen-phosphorus ratio, the nitrogen-potassium ratio and the phosphorus-potassium ratio in the predicted results are calculated to form a predicted ratio vector; similarly, the true ratio vector of the true value label is calculated. Then, the root mean square logarithmic error between the two vectors is calculated, which forces the model to learn the absolute content while strictly following the crop-specific nutrient synergistic absorption and antagonistic competition rules, ensuring that the output nutrient structure meets the biological rationality.

[0065] S3.3 Adaptive optimization training strategy and model solidification: This step performs iterative optimization and parameter solidification of the model. The AdamW optimizer is used for parameter update, and the weight decay strategy is introduced to constrain the model complexity through regularization. At the same time, the cosine annealing learning rate scheduling strategy is implemented, which quickly increases the learning rate in the warm-up stage at the beginning of training to jump out of the local minimum, and smoothly reduces the learning rate according to the cosine curve in the later training to find the global optimal solution. In the training process, the stoichiometric balance compound loss value of the model is monitored using the validation set, and the early stopping mechanism is implemented, i.e., when the validation set loss does not decrease for consecutive periods, the training is terminated in advance to prevent overfitting. Finally, the model weight parameters with the best validation set performance are saved, which are encapsulated as an online nutrition early warning model and delivered to S4 step for field deployment.

[0066] S4, Nutrition status early warning and fertilization prescription generation based on conformal prediction

[0067] This step applies the online nutrition early warning model to real-time field operations, and uses conformal prediction technology to construct a statistically guaranteed confidence interval for the point prediction results, and generates a variable fertilization prescription map accordingly. The process of this step is shown in Figure 5

[0068] ​S4.1 Calibration set inconsistency score and threshold calculation: Before the formal deployment of the online nutrition early warning model, the uncertainty of the model is statistically calibrated using the reserved calibration data set. For each sample in the calibration set, calculate its inconsistency score, which is defined as the absolute value of the model prediction error. Then, sort all the inconsistency scores of the calibration samples from small to large, and according to the preset significance level 0.1, which corresponds to a confidence level of ninety percent, calculate the quantile of the score distribution, which is determined as the conformal calibration threshold. This threshold represents the statistical upper bound of the model prediction error at the current confidence level.

[0069] S4.2 Online prediction and dynamic confidence interval generation: During real-time monitoring in the field, input the real-time collected and preprocessed ear data into the online nutrition early warning model to obtain the point prediction value of nitrogen, phosphorus and potassium content, denoted as the point prediction value of nutrient content. Then, use the conformal calibration threshold calculated in S4.1 to directly construct the edge prediction interval of the prediction point, i.e. the point prediction value plus or minus the conformal calibration threshold. If the point prediction value is lower than the preset lower threshold of nutrient content at a specific growth stage, and the upper bound of the above edge prediction interval is still lower than the threshold, it is determined that there is a confirmed nutrient deficiency at this point, triggering an early warning signal.

[0070] S4.3 Fertilization prescription map generation based on nutrient balance method: Based on the diagnosis results of confirmed nutrient deficiency, precise fertilization decision is made. First, obtain the expected target yield of the target plot, and calculate the theoretical fertilizer requirement of the point using the nutrient balance method formula. The specific calculation logic is as follows: first, multiply the target yield by the unit yield nutrient requirement to obtain the total nutrient demand of the crop during the whole growth period; second, subtract the soil base nutrient supply from the total demand, and further subtract the absorbed nutrient amount of the current crop, wherein the absorbed nutrient amount is converted from the hull volume of the non-convex hull calculated in S2.3 to the ear dry matter mass using the preset ear packing density coefficient, and then the ear dry matter mass is multiplied by the point prediction value of nutrient content output in S4.2; finally, divide the difference obtained by the fertilizer utilization rate in the current season to obtain the theoretical fertilizer requirement of the point. After the system traverses all the scanning points in the field and completes the above calculation, the Kriging interpolation algorithm is used to generate a nitrogen, phosphorus and potassium variable fertilization prescription map covering the whole field, which is transmitted as a direct instruction to the variable fertilization machinery to guide the execution of precise fertilization operation.

[0071] Experimental analysis:

[0072] To verify the effectiveness of the crop ear phenotype-based nutrition status early warning method proposed in the present application, based on the multi-modal ear image data collected by the field high-throughput phenotyping robot, the regression performance of the multi-scale residual attention network, the physical constraint effect of the stoichiometric balance composite loss function, and the reliability of the fertilization decision based on conformal prediction were evaluated. The experimental indicators include the nutrition inversion accuracy, the physical consistency error distribution, and the accuracy rate of the variable fertilization prescription generation.

[0073] 1. Performance experiment analysis of multi-modal nutrition inversion model

[0074] After deploying the online nutrition early warning model, the prediction performance of the model on the contents of three key nutrients, nitrogen, phosphorus, and potassium, during the crop heading and grain filling period was evaluated. The prediction effect was visualized by a multi-dimensional error heat map, where the horizontal axis represents the sampling time step in the field, and the vertical axis represents different nutrition monitoring tasks, including nitrogen content, phosphorus content, potassium content, and key ratios such as nitrogen to phosphorus ratio, and the color depth maps the relative error amplitude of the predicted value, as shown in Figure 6 .

[0075] The experimental results show that the model can effectively use multi-modal ear feature images to invert the nutrition status, and the average prediction error of each element is controlled at a low level, indicating that the deep residual attention module in the feature encoder can effectively extract spectrum and morphological features sensitive to nutrition from complex field backgrounds. In addition, the multi-task decoding structure successfully realizes the synchronous prediction of different nutrient components, verifying the feature analysis ability and robustness of the model in handling high-dimensional heterogeneous data.

[0076] 2. Analysis of stoichiometric physical constraint mechanism

[0077] To verify the physical constraint effect of the stoichiometric balance composite loss function in model training, the regression model performance with and without the introduction of the physical constraint mechanism was compared. Under the same training set and hyperparameter settings, the prediction residuals of the two types of models on different field validation sets were calculated, and the average absolute error levels of the online nutrition early warning model with physical constraints and the traditional unconstrained deep learning model in nitrogen, phosphorus, and potassium prediction were displayed in the form of a column chart, as shown in Figure 7 .

[0078] The results show that after introducing the stoichiometric ratio consistency loss, the prediction residuals of the model in each test field are significantly reduced. Especially in the areas with sparse sample distribution, the model with physical constraints can force the prediction results to follow the inherent nutrient synergistic and antagonistic rules of crops, effectively avoiding the occurrence of abnormal values that violate biological common sense. This result verifies that embedding domain knowledge in deep learning can significantly improve the generalization ability and biological interpretability of the model in non-steady-state environments.

[0079] 3. Real-time evaluation of the effect of nutrient warning and fertilization decision

[0080] To evaluate the field operation performance of the nutrient status warning and fertilization prescription generation module based on conformal prediction, the model was deployed on the field operation terminal to perform point prediction and edge prediction interval construction on the real-time collected ear data. Based on the confirmed nutrient deficiency diagnosis results, the system calculates the theoretical fertilizer requirement and generates variable fertilization instructions. The results are displayed through a line chart, with the horizontal axis representing the field travel distance and the vertical axis representing the normalized nutrient index and fertilization amount. Different monitoring and decision variables are represented by different line colors, as shown in FIG. 6. Figure 8

[0081] The experimental results show that the system can capture the subtle fluctuations of the nutrient content of the crop ear in real time and construct reliable confidence intervals using conformal calibration thresholds. Only when a confident deficiency is confirmed, the system triggers specific fertilization instructions, and the generated variable fertilization amount is highly consistent with the actual growth needs of the crop. This decision-making mechanism based on uncertainty quantification effectively reduces the phenomenon of misapplication caused by model fluctuations, achieving precision and intelligent nutrient management in the field.

[0082] In summary, the present application realizes high-precision inversion of crop ear nutrients, effective integration of physical laws, and risk quantification of fertilization decisions through multi-modal ear feature map construction, warning models based on chemical metrology constraints, and conformal prediction decision-making technology, fully verifying the effectiveness of the method in precision agriculture applications.

[0083] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

[0084] Although the specific embodiments of the present application have been described above, they are not intended to limit the scope of protection of the present application. Those skilled in the art should understand that various modifications or changes made to the technical solutions of the present application without creative labor are still within the protection scope of the present application.​

Claims

1. A method for early warning of the nutritional status of a crop based on the phenotypic characteristics of the ear, characterized in that, The method comprises the following steps: S1, during the heading and grain filling stage of the crop, a hyperspectral image data cube and high-density three-dimensional point cloud data of the crop canopy are acquired by scanning, and three-dimensional geographic coordinates of the scanning sample points are acquired at the same time; The hyperspectral image data cube and the high-density three-dimensional point cloud data are registered by using external calibration parameters, and a space bounding box is constructed according to the three-dimensional geographic coordinates to perform regional clipping, so that target ear region data containing a spectral reflectance vector and a spatial coordinate vector are extracted; S2, in the spectral dimension, the target ear region data is subjected to ear spectral purification, spectral transformation and feature band screening, and a nutrient feature spectrum vector is constructed; In the spatial dimension, the high-density three-dimensional point cloud data in the target ear region data is analyzed to obtain geometric features reflecting ear fullness, and a morphological structure parameter vector is constructed; the nutrient feature spectrum vector and the morphological structure parameter vector are processed and spliced to generate a multi-modal ear feature spectrum; The specific process of constructing the nutrient feature spectrum vector is as follows: In the spectral dimension, a pure ear average spectral curve is purified from the target ear region data by using a full-constraint least square linear spectral unmixing algorithm, and after double transformation, a competitive self-adaptive reweighted sampling algorithm is used to screen key bands to construct the nutrient feature spectrum vector; in the spatial dimension, a non-convex hull is reconstructed from the denoised three-dimensional point cloud data by using an Alpha-shape algorithm, core geometric indexes are calculated, morphological feature interaction dimensionality processing is performed, and a high-dimensional morphological structure parameter vector is constructed; finally, an independent standardization strategy is used to process the above-mentioned nutrient feature spectrum vector and the morphological structure parameter vector, and a multi-modal collaborative feature vector is generated through feature splicing, and a multi-modal ear feature spectrum composed of the multi-modal collaborative feature vector is output; The specific process of constructing the morphological structure parameter vector is as follows: Firstly, a statistical outlier removal algorithm is used to remove stray noise points in the high-density three-dimensional point cloud data to obtain denoised three-dimensional point cloud data; secondly, the denoised point cloud is reconstructed by using an Alpha-shape algorithm, a rolling ball radius parameter is set, a ball is simulated to roll outside the point cloud, and a gap that the ball cannot roll into is formed as a boundary to generate a non-convex hull that closely fits the concave and convex details of the ear surface; based on the non-convex hull, three core morphological indexes are calculated in sequence: firstly, the hull volume is calculated to quantitatively represent the biomass size of the ear; secondly, the ratio of the surface area to the volume is calculated to reflect the tightness of the grain arrangement of the ear; Thirdly, the variance of the point cloud normal vector is calculated to represent the rough texture features of the ear surface; finally, based on the three core indexes of the hull volume, the ratio of the surface area to the volume and the variance of the point cloud normal vector, a high-dimensional morphological expansion vector containing the original indexes and nonlinear combinations of the indexes is calculated by using a polynomial expansion technique, and the morphological structure parameter vector is obtained; S3, the multi-modal ear feature spectrum is input into a trained online nutrition early warning model, and prediction values of nitrogen, phosphorus and potassium contents are output. S4, calculating the inconsistency score and determining the conformal calibration threshold value using the calibration dataset, constructing the edge prediction interval, and determining that there is a nutrient deficiency and triggering the early warning when both the point prediction value and the upper limit of the edge prediction interval are lower than the specific growth stage nutrient lower limit threshold value.

2. The method for early warning of the nutritional status based on the phenotypic characteristics of the ear of the crop according to claim 1, characterized in that: In S1, the hyperspectral image data cube and the high-density three-dimensional point cloud data are projected into a unified world coordinate system by using a rotation matrix and a translation vector describing the relative position and attitude relationship between the hyperspectral imager and the laser radar, and are registered by a nearest neighbor interpolation algorithm to generate atlas fusion data containing both 256-band spectral information and spatial three-dimensional coordinates of each pixel point; according to the recorded centimeter-level three-dimensional geographic coordinates of each scanning point, a space bounding box with a fixed physical size of 20 cm by 20 cm is constructed with the coordinates as the center, and a local data subset falling within the bounding box is cropped from the atlas fusion data as the target ear region data; the target ear region data includes two hundred and fifty-six spectral reflectance values of each pixel point in the four hundred nanometer to one thousand nanometer band range, recorded as a spectral reflectance vector, and the spatial three-dimensional coordinate value of the pixel point in the world coordinate system, recorded as a spatial coordinate vector.

3. The method of early warning of the nutritional status based on the phenotypic characteristics of the ear of the crop according to claim 1, characterized in that: The specific process of purifying the pure ear average spectral curve from the target ear region data by using the full-constrained least squares linear spectral unmixing algorithm is as follows: First, based on the spectral reflectance vector, an endmember matrix containing four types of basic components is constructed, and typical spectral reflectance vectors of healthy ears, crop leaves, soil background and shadows are artificially selected from the spectral reflectance vectors corresponding to all pixel points, and the four vectors are used as basic endmembers to construct a standard endmember matrix; second, the full-constrained least squares linear spectral unmixing algorithm is used to calculate the abundance coefficient of each pixel point on the above four endmembers, and a linear spectral mixing model is constructed, modeling the spectral reflectance vector of any pixel point in the target ear region data as the result of mixing the four types of endmember spectra in the above standard endmember matrix at a specific ratio; The optimization objective is to minimize the root mean square error between the spectral reflectance vector and the reconstructed spectrum, while imposing non-negative abundance constraints, i.e. the abundance coefficient of each type of endmember is not less than zero, and the sum of the abundance coefficients of the four types of endmembers is strictly equal to 1, and the proportion of the healthy ear component in each pixel point is calculated by iterative operation; finally, a threshold is set, only the pixel points with a healthy ear endmember abundance higher than the threshold are retained, the spectral reflectance vectors corresponding to these pixel points are arithmetically averaged to obtain the spectral reflectance mean value, and the pure ear average spectral curve of the sample is generated based on the spectral reflectance mean value of these pixel points.

4. The method for early warning of the nutritional status based on the phenotypic characteristics of the ear of the crop according to claim 1, characterized in that: The specific process of screening key bands by using the competitive adaptive reweighted sampling algorithm to construct a nutrient feature spectrum vector is as follows: Feature selection is performed on the 256-band spectral data contained in the pure ear average spectral curve after envelope removal operation and spectral second derivative transformation; a preset iteration number M is set, and the following steps are performed in each iteration process: Firstly, a certain proportion of samples are randomly selected by Monte Carlo sampling method to establish a partial least squares regression model, and the absolute value of the regression coefficient corresponding to each band in the partial least squares regression model is extracted as the importance weight of the band; Secondly, the number of bands to be retained in the current iteration round is calculated by using an exponential decay function with the current iteration number and the total iteration number as variables, and the bands with low weight ranking are forcibly removed accordingly; thirdly, the adaptive reweighted sampling technique is used to normalize the regression coefficient weights of each band into selection probabilities, and the specific bands entering the next iteration round are randomly selected according to the probabilities; fourthly, the root mean square error of the current band subset is calculated through cross-validation; finally, after all the iteration rounds end, the band subset corresponding to the minimum root mean square error is selected as the final optimal feature band combination, and the nutrient feature spectrum vector is composed of the optimal feature band combination and the key bands with the strongest correlation with the true values of nitrogen, phosphorus and potassium.

5. The method for early warning of the nutritional status based on the phenotypic characteristics of the ear of the crop according to claim 1, characterized in that: The online nutrition early warning model is a multi-task network composed of a shared feature encoder and a task-specific decoder; The input of the shared feature encoder is a multi-modal ear feature spectrum, and the main body is composed of an initial convolutional layer and five cascaded deep residual attention modules. Each residual attention module contains two paths: the main path uses two one-dimensional convolutional layers with convolution kernel sizes of 3 and 1 to extract local spectral and morphological context features, and is connected with a batch normalization layer and an exponential linear unit activation function in turn; the bypass path embeds a compression and excitation attention submodule, which first compresses the feature map into a channel descriptor through global average pooling, then reduces and then increases the dimension through two fully connected layers, generates a channel importance weight through a Sigmoid function, and multiplies the weight with the feature map of the main path channel by channel; At the end of the shared feature encoder, three independent fully connected decoding branches are introduced in parallel, corresponding to the nitrogen, phosphorus and potassium prediction tasks respectively; each branch is designed as a bottleneck expansion structure, containing a fully connected layer with a neuron number decreasing from 256 to 64 to compress the features, a Dropout layer, and a linear output layer in turn, and finally outputs the three nutrient content prediction values of the crop ear in parallel.

6. The method of early warning of the nutritional status based on the phenotypic characteristics of the ear of the crop according to claim 5, characterized in that: The training target of the online nutrition early warning model is to minimize a composite objective function composed of a regression accuracy loss and a physical consistency loss dynamically weighted; the first part of the composite objective function is a multi-task robust regression loss, which calculates the Huber loss between the predicted value and the true value for the nitrogen, phosphorus and potassium tasks. This loss combines the advantages of mean square error and absolute error, and sets a threshold d as a hyperparameter. When the prediction error is less than the threshold, a quadratic function is used for punishment to ensure derivability, and when the error is greater than the threshold, a linear function is used for punishment to reduce the interference of outliers on the gradient; the second part of the composite objective function is a stoichiometric ratio consistency loss, which is based on the stoichiometric balance principle of plant nutrition to construct the element ratio matrix of the predicted value and the true value.

7. The method of early warning of the nutritional status based on the phenotypic characteristics of the ear of the crop according to claim 1, characterized by the fact that: The conformal calibration threshold is specifically: the uncertainty of the model is statistically calibrated by using the reserved calibration data set, for each sample in the calibration set, the non-consistency score is calculated, which is defined as the absolute value of the model prediction error; then, the non-consistency scores of all calibration samples are sorted from small to large, and according to the preset significance level 0.1, that is, the confidence of ninety percent, the quantile of the score distribution is calculated, which is determined as the conformal calibration threshold, the threshold represents the statistical upper bound of the model prediction error under the current confidence level.

8. The method of early warning of the nutritional status based on the phenotypic characteristics of the ear of the crop according to claim 1, characterized by the fact that: It also includes a fertilization prescription map generation based on the nutrient balance method, specifically: First, multiply the target yield by the nutrient requirement per unit yield to obtain the total nutrient demand of the crop during the whole growth period; second, subtract the soil basic nutrient supply and the current crop absorbed nutrient amount from the total demand, wherein the absorbed nutrient amount is converted from the shell volume of the non-convex hull to the ear dry matter mass by using a preset ear packing density coefficient, and then multiplied by the nutrient content point prediction value to obtain the absorbed nutrient amount; finally, divide the calculated difference by the seasonal utilization rate of the fertilizer to obtain the theoretical fertilizer supplement amount of the current scanning point. After traversing all scanning points in the field to complete the above calculation, the Kriging interpolation algorithm is used to generate nitrogen, phosphorus and potassium variable fertilization prescription maps covering the whole field.

Images