Corn near-infrared multi-environment yield and agronomic trait prediction method and system based on spectrogram and trait gating

Abstract

The invention belongs to a near-infrared spectrum phenotype analysis technology in the field of agricultural breeding, and discloses a corn near-infrared multi-environment yield prediction method and system based on spectrum structure and character adaptive gating. The method aims at solving the problems that in the prior art, the relation between spectrum segments is not modeled, the sensitivity difference of characters and environments to wavelengths is difficult to distinguish, physical priori and multi-environment constraints are lacked, and consequently the prediction precision and interpretability of complex characters are insufficient, and an efficient and low-cost tool is provided for multi-environment phenotype selection and corn breeding decision making. The method comprises the core steps of data preprocessing, multi-scale spectral feature embedding, spectrogram construction and graph attention coding, feature fusion, character adaptive wavelength gating, multi-task regression, physical constraint model training prediction and the like. The system is composed of a corresponding function module and a training and model management module, key spectrum information can be fully mined, and a high-precision result can be stably output. The method can be used for multi-environment yield and quality prediction, online near-infrared monitoring and phenotype selection decision making of crops such as corn, can also be popularized to multi-mode breeding big data analysis and spectrograph waveband optimization, and is good in interpretation and wide in application prospect.

Description

Technical Field

[0002] This invention relates to the agricultural field, particularly to near-infrared spectroscopy (NIR) analysis technology applied to maize breeding, specifically to a method and system for predicting maize phenotypic traits such as yield in multiple environments based on adaptive gating of spectral structure and traits. Background Technology

[0004] Maize is an important food and feed crop globally. Grain yield and related agronomic traits (such as tasseling-silking interval, early growth vigor, final plant height, grain dry matter content, and grain nutrient content) are core indicators in maize breeding and variety selection. With the development of genomic selection, the use of molecular markers to predict complex traits has become an important tool in modern breeding. However, genotyping is costly and requires demanding experimental conditions, posing certain application barriers in some breeding units and large-scale multi-environment trials.

[0005] Near-infrared spectroscopy (NIRS) is widely used for determining the physicochemical quality and chemical composition of agricultural products due to its high throughput, low cost, non-destructive nature, and ease of operation. In recent years, the concept of phenotypic selection has been proposed, which uses high-throughput phenotypic data such as near-infrared spectroscopy to replace molecular marker data, combined with statistical and machine learning methods similar to genomic selection, to predict and select for complex traits. Studies have shown that in crops such as maize, phenotypic prediction based on grain near-infrared spectroscopy has good predictive ability for traits such as yield in multi-environment experiments, especially in cases with large differences in material populations and complex population structures, demonstrating effects comparable to or even superior to genomic selection.

[0006] Existing phenotypic selection studies mostly employ linear mixture models or kernel-function-based statistical models, such as RR-BLUP, which use the entire wavelength as a high-dimensional independent variable input to directly predict yield or other traits through regression. While these methods can utilize the overall information of the near-infrared spectrum to some extent, they still have shortcomings in the following aspects: lack of explicit spectral structure modeling, failure to distinguish the differences in wavelength sensitivity among different traits and environments, and lack of explicit constraints on physical priors and multi-environment structures.

[0007] Therefore, it is necessary to propose a new method and system for predicting maize traits such as yield in multiple environments using near-infrared spectroscopy. This method involves explicitly modeling the spectral structure within a unified framework, introducing trait-adaptive wavelength gating and multi-environment information fusion, and combining physical prior constraints to improve the accuracy and interpretability of multi-environment and multi-trait predictions. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for predicting maize yield in multiple environments based on spectral mapping and trait gating. This addresses the problems in existing technologies, such as the lack of explicit modeling of graph structure correlations between near-infrared spectral wavelengths, the lack of adaptive wavelength selection mechanisms for different traits and environments, and the lack of incorporation of prior information such as spectral smoothness and known physical trends into the training process for multi-environment phenotypic prediction. This improves the accuracy and robustness of maize yield and related agronomic trait prediction in multiple environments, enhances the interpretability of the model, and increases its application value in practical breeding.

[0010] To achieve the above objectives, the present invention proposes the following technical solution:

[0011] Multi-environment data preprocessing: The original spectral data is preprocessed by denoising, baseline correction, scattering correction, derivative transformation and scaling to form a preprocessed spectral sequence with uniform dimensions.

[0012] Multi-scale overlapping spectral feature embedding: The preprocessed spectral sequence is input into the multi-scale overlapping convolutional feature extraction module. Multiple sets of one-dimensional convolutional kernels with different lengths and strides are used to perform overlapping slicing and feature extraction on the spectrum. The outputs of each scale convolution are aligned to the same dimension and length through linear transformation, and then weighted and fused using learnable scale weights to generate the basic spectral feature sequence.

[0013] Spectrum Construction and Graph Attention Encoding: Each wavelength position in the basic spectral feature sequence is used as a graph node, and local adjacency relationships are constructed based on the physical adjacency of wavelengths. Global correlation adjacency relationships are constructed based on the statistical correlation between wavelength intensities in multi-environment data. Local and global adjacency relationships are then fused according to weights to form a spectral graph adjacency matrix. A graph attention network is used to encode the spectral graph adjacency matrix and node features, resulting in a spectral feature sequence containing both local and long-distance spectral band correlations.

[0014] Sequence coding and multi-path feature fusion: The basic spectral feature sequence is input into the sequence encoder, and long-range dependency information across the entire spectrum is extracted using a self-attention mechanism to obtain the sequence-coded features. The spectral features and sequence-coded features are concatenated or added together along the feature dimension and fused, followed by linear transformation and nonlinear activation, to obtain a unified high-dimensional spectral representation sequence. Environmental identifiers or environmental covariates are then encoded and fused with the spectral representation sequence to explicitly incorporate multiple environmental information.

[0015] Trait-Adaptive Wavelength Gating and Multi-Trait Output: For each target trait (including but not limited to grain yield, grain dry matter content, and final plant height), an independent trait-adaptive wavelength gating module is constructed. The trait gating module utilizes global context features and local features at each wavelength position to generate a set of wavelength weight vectors between 0 and 1, representing the trait's sensitivity to different wavelength bands. The high-dimensional spectral representation sequence is weighted and aggregated according to the wavelength weights to obtain the dedicated feature vector corresponding to the trait. The trait-specific feature vector is input into the corresponding regression prediction head, outputting the predicted value of the trait under various environments.

[0016] Joint training based on physical priors and multi-task mechanisms: During the model training phase, a multi-task regression loss is constructed, and the prediction errors of all target traits are weighted by uncertainty to automatically balance the contribution of each trait in training. Smoothing and sparsity regularization are introduced for the wavelength weights of the trait-gated outputs, encouraging continuous variation of weights along the wavelength dimension and concentration in a few key spectral bands. For traits with known trends in integral absorption with specific spectral bands (such as grain dry matter content), a physical constraint loss based on the monotonic relationship between integral absorption and trait values ​​is introduced, allowing the model to follow some known physical laws during optimization. The above loss terms are linearly combined to form the total loss, and end-to-end joint training of multi-scale embedding, spectral encoding, sequence encoding, trait gating, and the prediction head is performed using gradient descent-type optimization algorithms.

[0017] Furthermore, the present invention also provides a near-infrared multi-environment yield prediction system for maize to implement the above method, the system comprising:

[0018] Data preprocessing module: used to collect near-infrared spectral and multi-trait phenotypic data of maize kernels in multi-environment field trials, and to preprocess and normalize the spectra.

[0019] Multi-scale spectral feature embedding module: used to extract and fuse multi-scale overlapping convolutional features from preprocessed spectra to generate basic spectral feature sequences.

[0020] Spectral graph construction and graph attention encoding module: used to construct the spectral graph adjacency matrix based on physical and statistical relationships, and to generate spectral graph features using a graph attention network.

[0021] Sequence Encoding and Feature Fusion Module: Used to extract full-spectrum long-range dependencies through a sequence encoder and perform multi-path fusion with spectral features and environmental information.

[0022] Trait-adaptive wavelength gating module: used to generate adaptive wavelength weights for different traits, realizing trait-level band selection and feature aggregation.

[0023] Multi-task prediction module: used to receive the specific feature vectors of each trait and output the predicted values ​​of the target trait under various environments.

[0024] Training and Model Management Module: This module is used to construct a total loss function that includes multi-task regression loss, smoothing and sparse regularization, and physical constraint loss. It is used to train, validate, and update the model, and to save and deploy the model. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall architecture of a multi-index quality prediction system for maize based on spectral structure and trait adaptive gating according to the present invention.

[0027] Figure 2 This is a flowchart illustrating a multi-index quality prediction method for maize near-infrared spectroscopy based on adaptive gating of spectral structure and traits, according to the present invention. Detailed Implementation

[0029] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0030] Example 1: System Architecture

[0031] Referring to Figure 1, this embodiment provides a near-infrared multi-environment yield prediction system for maize based on spectral and trait gating, which includes:

[0032] Data preprocessing module: performs noise filtering, baseline and scattering correction, derivative transformation and normalization on spectral data, and outputs preprocessed spectral sequences.

[0033] Preprocessed NIRS spectral vectors:

[0034]

[0035] in For spectral dimensions.

[0036] Multi-scale spectral feature embedding module: In this embodiment, the following is adopted: A group of one-dimensional convolutional kernels is used to perform multi-scale overlapping embedding of the spectrum. The kernel size of the group convolution is Step size is The output features are:

[0037]

[0038] Each Linear transformation to the same dimension And align the length by interpolation or truncation. ,get:

[0039]

[0040] Subsequently, a set of learnable scale weight vectors is introduced. After Softmax normalization, the features at different scales are weighted and summed to obtain the basic spectral feature sequence:

[0041]

[0042]

[0043] in As input for subsequent modules.

[0044] The spectral construction and graph attention encoding module constructs a local spectral adjacency matrix based on wavelength adjacency relationships. It calculates the correlation coefficients between different bands in multi-environment data and constructs a global correlation adjacency matrix based on thresholds. These two types of adjacency matrices are then fused according to a set ratio or through learnable weights to obtain the final spectral adjacency matrix. A graph attention network is then used to encode the spectrum, obtaining a spectral feature sequence containing inter-spectral relationships.

[0045] Will Each row in the graph is considered a feature of a spectral node. Let the set of nodes be denoted as . .

[0046] Local adjacency construction:

[0047] If the index difference between two nodes on the wavelength axis does not exceed the preset radius Then, local edges are established between the nodes:

[0048]

[0049] Related adjacency construction:

[0050] Calculate the correlation coefficients between wavelength positions on the training set, and construct the correlation adjacency matrix according to the threshold:

[0051]

[0052] in For the first With the Pearson correlation coefficient of spectral values ​​at wavelength, For temperature parameters, The threshold value is used.

[0053] Graph adjacency fusion: Weighted fusion of local adjacency and related adjacency.

[0054]

[0055] Graph attention encoding: For each node Let its input features be The linear transformation is as follows For pairs of nodes connected by an edge. Calculate the attention coefficient:

[0056]

[0057] After Softmax normalization, we get:

[0058]

[0059] The final spectral features are:

[0060]

[0061] All nodes Stacking forms spectral feature sequences .

[0062] Sequence encoding and feature fusion module: In this embodiment, a sequence encoder is used to... Perform deep encoding to obtain sequence features .Will Spectral features Concatenate along the feature dimension:

[0063]

[0064] Then, through a linear transformation, it is mapped to a unified dimension. , thus obtaining F.

[0065] Trait-adaptive wavelength gating module: An independent gating submodule is configured for each target trait. Each gating submodule receives a high-dimensional spectral representation and global context information, and generates a corresponding set of wavelength weights. Subsequently, the gating submodule uses these weights to perform weighted aggregation on the high-dimensional spectral representation to obtain a feature vector specific to that trait.

[0066] For fusion feature sequences Perform average pooling along the wavelength dimension:

[0067]

[0068] For indicators The context vector and local features are input into the gating network:

[0069]

[0070]

[0071] in, , , , As an indicator The corresponding gating parameters, For the Sigmoid function, As an indicator For the Weights for each wavelength position.

[0072] After aggregating the trait features, input them into the corresponding regression prediction head (e.g., a one- or multi-layer fully connected network) to obtain the predicted value of the indicator:

[0073]

[0074] Multi-task prediction module: Each target trait is configured with an independent regression prediction head, receiving its corresponding dedicated feature vector and outputting the predicted value of that trait under various environments. The model supports simultaneous prediction of multiple traits such as yield, dry matter content, and plant height, achieving joint modeling and output of multiple traits.

[0075] The training and model management module constructs a multi-task loss function, combining the regression error of each trait, an adaptive weighting term based on prediction uncertainty, a smoothing and sparse regularization term for wavelength weights, and a physical constraint term for specific traits to form the total loss. It employs stochastic gradient descent or its variants to perform end-to-end optimization of model parameters. It also provides complete interfaces for training process monitoring, model version management, and deployment, supporting efficient iteration and practical application implementation.

[0076] Assume there are a total of The nth sample, the th The true label for each sample is The predicted value is .

[0077] Uncertainty-Weighted Multi-Task Regression Loss: Introducing the noise standard deviation parameter for each indicator. Construct multi-task loss ,

[0078]

[0079] This loss function can automatically balance the relative weights of each indicator, avoiding the need for manual setting of proportions.

[0080] Wavelength-weighted smoothing and sparsity-regularized loss: To encourage smoothness of wavelength weights along the wavelength dimension and overall sparsity, the following measures are taken: Introducing smoothing loss And sparsity loss:

[0081]

[0082]

[0083] Physical prior monotonic constraint loss: For indicators with clear physical trends (such as moisture), select the spectral bands related to the indicator. Define its integral absorption: During training, construct a system that satisfies... sample pairs Assuming that theoretically this index increases with increasing integral absorption, then the physical constraint loss can be defined as follows:

[0084]

[0085] This embodiment uses the following total loss:

[0086]

[0087] in, , , These are non-negative weight hyperparameters. Minimize... Joint optimization of all network parameters.

[0088] Example 2: Method Flow

[0089] See Figure 2 The near-infrared multi-environment yield prediction method for maize based on spectral and trait gating provided in this embodiment specifically includes the following steps:

[0090] Step S1: Multi-environment data preprocessing

[0091] Based on the near-infrared spectral data of grains and trait data such as yield of different genotypes, the spectra are preprocessed to form a dataset in a unified format.

[0092] Step S2: Multi-scale overlapping spectral embedding

[0093] Each preprocessed spectrum is input into a multi-scale convolutional network to extract local features from short-wavelength, medium-length, and long-wavelength bands, respectively, and the basic spectral feature sequence is obtained by weight fusion.

[0094] Step S3: Spectrum Construction and Graph Attention Encoding

[0095] Local adjacency relationships are constructed based on wavelength physical adjacency, and global adjacency relationships are constructed based on the statistical correlation of multiple environmental samples. These are fused to obtain the spectral graph adjacency matrix, and the spectral graph feature sequence is obtained by encoding with a graph attention network.

[0096] Step S4: Sequence Encoding and Feature Fusion

[0097] The basic spectral feature sequence is input into the sequence encoder to extract global context features, which are then fused with spectral features and environmental information in the feature dimension to generate a unified high-dimensional spectral representation sequence.

[0098] Step S5: Trait-adaptive wavelength gating and prediction output

[0099] For each target trait, a wavelength weight vector is generated using the trait gating module. The high-dimensional spectral representation sequence is then weighted and aggregated to obtain a dedicated feature vector for that trait. This vector is then input into the corresponding prediction head to output the prediction result.

[0100] Step S6: Multi-task joint training

[0101] During training, multi-task regression loss is calculated for the predicted results and true values ​​of all traits; smoothing and sparsity regularization are applied to the wavelength weights; physical constraint loss is applied to traits with known physical trends; the weighted sum of each loss is used to form the total loss; and the model parameters are continuously optimized through iteration until convergence.

[0102] Example 3: Application based on multi-environment NIRS and yield multimodal datasets

[0103] In this embodiment:

[0104] Each "genotype-environment" combination is considered as a sample, with the corresponding near-infrared spectrum of the grain as input and the yield and other traits under that environment as output.

[0105] All environmental data are used together to train the spectrum and trait gating model proposed in this invention to predict traits such as yield in the target environment or for unobserved genotypes;

[0106] Using traditional phenotypic selection models such as RR-BLUP as a control, we compared the predictive capabilities of different methods in various environmental scenarios.

[0107] By analyzing the wavelength weight distribution of the trait gating module, key spectral bands that contribute significantly to traits such as yield are identified, providing a reference for the further design of simplified band spectrometers.

[0108] Example 4: System Expansion and Other Application Scenarios

[0109] The near-infrared multi-environment yield prediction method and system for maize based on spectral and trait gating proposed in this invention has good scalability and can be further applied to the following scenarios:

[0110] Crop and trait extension: By replacing or supplementing training data, it can be extended to multi-environment phenotypic prediction tasks for crops such as wheat and rice, and jointly predict and select traits such as yield, quality and resistance.

[0111] Multimodal data fusion: Based on near-infrared spectroscopy, UAV remote sensing phenotypic data, environmental sensor data, or genomic data can be further fused to construct a multimodal spectral-phenotypic prediction system, thereby further improving prediction accuracy.

[0112] Online monitoring and decision support: The system of this invention can be integrated with online near-infrared detection devices and field management systems to realize real-time yield and quality prediction in grain harvesting and storage or breeding experiments, and provide decision support for fertilizer and water management and material selection.

[0113] Sensor and spectrometer design: Based on the key band information output by the trait-gated module, the development of narrowband or multichannel spectrometers optimized for specific traits can be guided, reducing hardware costs and improving measurement efficiency.

[0114] Through the above embodiments, the present invention provides an efficient, interpretable, and applicable near-infrared maize yield prediction method and system based on spectral and trait gating, which is suitable for breeding scenarios in multiple environments and provides an important technical means for phenotypic selection and breeding material screening.

[0115] The above embodiments are preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art can make changes and modifications based on the technical solutions and core concepts of the present invention, such as replacing different types of sequence encoders and graph neural network structures, adopting different regularization forms or physical constraint methods, adjusting the combination of target traits and environmental variables, etc., and these substitutions will not depart from the technical scope of the present invention. All substitutions or equivalent modifications should be included within the scope of protection of the present invention, and the final scope of protection is determined by the claims and their equivalent technical solutions.

Claims

1. A near-infrared multi-environment yield prediction method for maize based on spectral and trait gating, characterized in that, Includes the following steps: S1 Multi-environment data preprocessing: Acquire near-infrared spectral data of maize kernels and trait data such as yield collected in multiple environments for different genotypes. Perform noise reduction, baseline correction, scattering correction, derivative transformation and scaling preprocessing on the original spectral data to obtain a preprocessed spectral sequence with uniform dimensions. S2 Multi-scale Overlapping Spectral Feature Embedding: The preprocessed spectral sequence is input into the multi-scale convolutional feature extraction module. Multiple sets of one-dimensional convolutional kernels with different kernel lengths and strides are used to perform overlapping slicing and local feature extraction on the spectrum. The convolutional outputs at each scale are fused through linear transformation and length alignment to generate a basic spectral feature sequence. S3 Spectrum Construction and Graph Attention Encoding: Each wavelength position in the basic spectral feature sequence is taken as a graph node. Local adjacency relationships are constructed based on the physical adjacency of wavelengths on the wavelength axis. Global correlation adjacency relationships are constructed based on the statistical correlation between wavelength intensities in multiple environmental data. The local adjacency relationships and global correlation adjacency relationships are fused according to preset weights or learnable weights to obtain a spectral adjacency matrix. The spectral adjacency matrix and node features are encoded using a graph attention network to obtain a spectral feature sequence. S4 Sequence Encoding and Feature Fusion: The basic spectral feature sequence is input into the sequence encoder, and long-range dependency information in the full spectrum is extracted using a self-attention mechanism to obtain sequence-encoded features; the sequence-encoded features and the spectral features are concatenated or added together in the feature dimension and fused, and a high-dimensional spectral representation sequence is obtained through linear transformation and nonlinear activation; environmental identifiers and / or environmental covariates are encoded and fused with the high-dimensional spectral representation sequence to introduce multiple environmental information; S5 Trait Adaptive Wavelength Gating and Prediction Output: An independent trait adaptive wavelength gating module is constructed for each target trait. The trait adaptive wavelength gating module receives the global context features and local features of each wavelength position of the high-dimensional spectral representation sequence, generates a wavelength weight vector between 0 and 1, and performs weighted aggregation on the high-dimensional spectral representation sequence based on the wavelength weight vector to obtain a dedicated feature vector for the target trait. The dedicated feature vector is input into the corresponding regression prediction head to output the predicted value of the target trait in each environment. S6 Multi-task Joint Training: Construct a multi-task regression loss function that simultaneously constrains multiple target traits. Based on the true trait values ​​and predicted values ​​of the samples in the training set, jointly optimize the multi-scale convolutional feature extraction module, spectral map construction and graph attention encoding module, sequence encoder, trait adaptive wavelength gating module and regression prediction head to obtain a well-trained model for maize near-infrared multi-environment yield prediction.

2. The near-infrared multi-environment yield prediction method for maize based on spectral and trait gating according to claim 1, characterized in that: The S2 multi-scale overlapping spectral feature embedding uses at least three sets of one-dimensional convolution kernels of different lengths to extract short-wavelength, mid-wavelength, and long-wavelength features respectively, and performs weighted fusion of features at each scale through learnable scale weights.

3. The near-infrared multi-environment yield prediction method for maize based on spectral and trait gating according to claim 1, characterized in that: In the S3 spectrum construction and graph attention encoding, the global relevance adjacency relationship is obtained by calculating the Pearson correlation coefficient between each wavelength on the training set and selecting relevant edges based on the correlation coefficient threshold.

4. The method for predicting maize yield in multiple environments based on spectral and trait gating according to claim 1, characterized in that: In the S4 sequence encoding and feature fusion, environmental identifiers and / or environmental covariates such as meteorology and soil are encoded into environmental feature vectors through an embedding layer or a fully connected layer, and then concatenated or added to the high-dimensional spectral representation sequence in the feature dimension.

5. The near-infrared multi-environment yield prediction method for maize based on spectral and trait gating according to claim 1, characterized in that: In the S6 multi-task joint training, the multi-task regression loss includes uncertainty-weighted loss and physical monotonic constraint loss based on integral absorption, which are used to simultaneously balance the learning difficulty of each trait and inject the physical prior relationship between the spectrum and the trait.

6. A near-infrared multi-environment yield prediction system for maize based on spectral and trait gating, characterized in that, include: The data preprocessing module acquires near-infrared spectral data and yield and other trait data of maize under different genotypes in multiple environments, and performs denoising, correction and scaling preprocessing on the spectra to generate preprocessed spectral sequences. The multi-scale spectral feature embedding module extracts and aligns the preprocessed spectral sequences using multi-scale overlapping one-dimensional convolutional features to generate basic spectral feature sequences. The spectral construction and graph attention encoding module constructs a spectral adjacency matrix based on wavelength physical adjacency and spectral statistical correlation, and uses a graph attention network to encode the basic spectral feature sequences to generate spectral feature sequences. The sequence encoding and feature fusion module encodes the basic spectral feature sequence and fuses it with spectral features and environmental encoding features to obtain a high-dimensional spectral representation sequence. The trait adaptive wavelength gating module generates wavelength weight vectors for different target traits and performs weighted aggregation on the high-dimensional spectral representation sequence to obtain the dedicated feature vectors for each trait. The multi-task prediction module receives the dedicated feature vectors for each trait and outputs the predicted values ​​of yield and other target traits under multiple environments. The training and model management module is used to construct a multi-task loss function, jointly train the above modules, and store, update, and deploy the trained model.

7. A near-infrared multi-environment yield prediction system for maize based on spectral and trait gating according to claim 6, characterized in that: The multi-scale spectral feature embedding module uses multiple sets of one-dimensional convolutional layers with different kernel lengths and strides to extract multi-scale features in parallel, and outputs a basic spectral feature sequence with uniform length and dimension through linear transformation and scale weight fusion.

8. A near-infrared multi-environment yield prediction system for maize based on spectral and trait gating according to claim 6, characterized in that: The spectrum construction and graph attention encoding module includes a local adjacency construction unit and a global related adjacency construction unit. The local adjacency construction unit generates a local adjacency matrix based on the wavelength index distance, and the global related adjacency construction unit generates a related adjacency matrix based on the inter-wavelength Pearson correlation coefficient. The adjacency fusion unit then fuses the two matrices according to preset weights or learnable weights.

9. A near-infrared multi-environment yield prediction system for maize based on spectral and trait gating according to claim 6, characterized in that: The trait-adaptive wavelength gating module includes multiple trait-gating sub-modules. Each trait-gating sub-module corresponds to a target trait and contains one or more fully connected layers and a Sigmoid activation function to output the wavelength weight vector of the target trait.

10. A near-infrared multi-environment yield prediction system for maize based on spectral and trait gating according to claim 6, characterized in that: The training and model management module is used to calculate the uncertainty-weighted multi-task regression loss and the physical monotonic constraint loss based on the true trait values ​​and predicted values ​​of the training samples, to perform end-to-end updates of the parameters of each module in the system, and to manage and call different versions of the model.

11. The application of the near-infrared multi-environment yield prediction method for maize based on spectral and trait gating as described in claim 1 in phenotypic selection for maize multi-environment breeding, characterized in that: Using multi-environment near-infrared spectral data and the method described above, the yield and related agronomic traits of different genotypes under various environments are predicted, and breeding materials are screened and varieties are evaluated accordingly.

12. The application of the near-infrared multi-environment yield prediction system for maize based on spectral and trait gating as described in claim 6 in near-infrared online monitoring and decision support for crops, characterized in that: The system is integrated with a near-infrared detection device to receive spectral data in the field or production environment in real time and output yield and quality trait prediction results for use in quality assessment and management decisions during crop production.

Images