A method and system for assessing growth health by coupling a spectral-radiative transfer algorithm with a multi-task deep learning algorithm.

By combining the spectral-radiative transfer algorithm with multi-task deep learning, the timeliness and accuracy problems of existing crop growth monitoring methods are solved, realizing intelligent and quantitative assessment of crop growth, and improving monitoring accuracy and cross-domain adaptability.

CN123070784APending Publication Date: 2026-11-10HARBIN AEROSPACE STAR DATA SYST TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611189100.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-06
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing crop growth monitoring methods suffer from poor timeliness, limited spatial coverage, strong subjectivity, and difficulty in meeting the needs of large-area, high-frequency, and refined management. Furthermore, existing remote sensing technologies are unable to fully exploit multispectral information, have limited inversion accuracy, lack the ability to model dynamic features throughout the entire growth period, and deep learning applications have not yet fully utilized the inherent correlations of crop parameters.

Method used

By combining the spectral-radiative transfer algorithm with multi-task deep learning, and through multi-source data acquisition and preprocessing, spectral mixture analysis and radiative transfer model coupling are performed to construct a multi-parameter time-series feature dataset. Small-sample transfer learning is used to optimize the multi-task deep learning network, and a long-term health assessment model is constructed to achieve intelligent and quantitative assessment of crop growth.

Benefits of technology

It generates high-quality multispectral simulation data, realizes continuous simulation of key parameters throughout the entire growth period, improves the accuracy of parameter prediction, has cross-domain generalization ability, and realizes continuous dynamic monitoring and evaluation of crop growth health with an accuracy better than 85%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070784A_ABST
    Figure CN123070784A_ABST
Patent Text Reader

Abstract

This invention relates to a crop growth health assessment method and system that couples a spectral-radiative transfer algorithm with a multi-task deep learning algorithm, belonging to the field of agricultural remote sensing and intelligent monitoring technology. It aims to provide high-quality data support for crop health assessment and growth monitoring. The invention includes: multi-source data acquisition and preprocessing; coupling spectral mixing analysis with a radiative transfer model to generate multispectral reflectance simulation information; continuous simulation of the entire growth period using a crop growth model to generate a multi-parameter time-series feature dataset; construction and training of a multi-task deep learning network; employing a three-stage transfer strategy to transfer the physical knowledge and spectral-parameter mapping relationship learned in the source domain to the target domain, improving the model's generalization ability by minimizing the maximum mean difference in feature distributions between the source and target domains, resulting in an optimized multi-task deep learning network; construction of a long-term time-series growth health assessment model; and quantitative assessment and classification of growth health. This invention achieves continuous dynamic monitoring and assessment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing and intelligent monitoring technology, specifically relating to a growth health assessment method and system that couples a spectral-radiative transfer algorithm with a multi-task deep learning algorithm. Background Technology

[0002] Crop growth health monitoring is a core component of precision agriculture management, and it is of great significance for ensuring food security, optimizing field management, and improving agricultural output efficiency. Traditional crop growth monitoring methods mainly rely on manual field surveys, which suffer from problems such as poor timeliness, limited spatial coverage, and strong subjectivity, making it difficult to meet the needs of large-scale, high-frequency, and refined agricultural management.

[0003] In recent years, the rapid development of remote sensing technology has provided new technical means for monitoring crop growth. Multispectral remote sensing data, due to its advantages such as convenient acquisition, high spatiotemporal resolution, and wide coverage, has been widely used in agricultural monitoring. However, existing crop growth assessment methods based on remote sensing technology still have the following shortcomings:

[0004] First, traditional remote sensing vegetation index methods (such as NDVI, EVI, etc.) only utilize simple algebraic combinations of limited bands, failing to fully exploit the rich spectral information contained in multispectral data. This results in limited accuracy in retrieving crop physiological and biochemical parameters, and signal saturation is particularly prone to occur in the middle and late stages of crop growth.

[0005] Second, while existing radiative transfer models (such as the PROSAIL model) can establish the physical relationship between vegetation canopy structure and canopy reflectivity, the inversion of model parameters suffers from ill-conditioned inverse problems. That is, different combinations of parameters may produce similar spectral responses, leading to significant uncertainty in the inversion results. At the same time, a single radiative transfer model cannot fully account for the spectral mixing effects of factors such as soil background and atmospheric effects.

[0006] Third, most existing crop growth assessment methods only focus on remote sensing information of a single time phase and lack the ability to model the dynamic characteristics of the entire crop growth period. They are unable to capture key turning points and cumulative effects in the crop growth process, resulting in insufficient stability and reliability of the assessment results.

[0007] Fourth, the application of deep learning technology in agricultural remote sensing is still in its early stages. Existing methods mostly adopt a single-task learning mode, which fails to fully utilize the inherent correlation and synergistic information between multiple physiological and biochemical parameters of crops, and faces challenges such as insufficient training samples and limited model generalization ability.

[0008] Therefore, there is an urgent need for an intelligent assessment method for crop growth health that can integrate physical models and data-driven methods, achieve multi-parameter collaborative inversion, and have long-term dynamic assessment capabilities. Summary of the Invention

[0009] The problem this invention aims to solve is to provide high-quality data support for crop health assessment and growth monitoring. It proposes a growth health assessment method and system that couples a spectral-radiative transfer algorithm with a multi-task deep learning algorithm.

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

[0011] A growth health assessment method coupling a spectral-radiative transfer algorithm and a multi-task deep learning algorithm includes the following steps:

[0012] S1. Multi-source data acquisition and preprocessing: Acquire multispectral remote sensing image data of the target area, and simultaneously acquire meteorological observation data and crop genetic characteristics information; acquire surface reflectance data; perform atmospheric correction, geometric correction and radiometric calibration preprocessing on the multispectral remote sensing image data;

[0013] S2. Coupling of spectral mixture analysis and radiative transfer model: The surface reflectance obtained in step S1 is decomposed into endmembers and forward simulated to generate multispectral reflectance simulation information, including the initial estimates of the abundance of each endmember, leaf area index (LAI), leaf chlorophyll content (LCC), and aboveground biomass (AGB).

[0014] S3. Using the multispectral reflectance simulation information obtained in step S2 as the true value, combined with meteorological observation data and crop genetic characteristics information, the crop growth model is used to continuously simulate the leaf area index (LAI), leaf chlorophyll content (LCC), and aboveground biomass (AGB) throughout the entire growth period to generate a multi-parameter time-series feature dataset.

[0015] S4. Constructing and training a multi-task deep learning network: Construct a multi-task deep learning network based on the spectral attention mechanism, including an input layer, a shared feature extraction module, a spectral attention module, three task-specific layers, and a terminal layer. Use the multi-parameter temporal feature dataset generated in step S3 to train the model and obtain the trained multi-task deep learning network.

[0016] S5. Small Sample Transfer Learning Optimization: The multi-task deep learning network trained in step S4 is subjected to a three-stage transfer strategy of source domain pre-training, domain adaptation fine-tuning, and target domain fine-tuning. The physical knowledge and spectrum-parameter mapping relationship learned in the source domain are transferred to the target domain. The generalization ability of the model is improved by minimizing the maximum mean difference between the feature distributions of the source domain and the target domain, and the optimized multi-task deep learning network is obtained.

[0017] S6. Construction of a long-term growth health assessment model: Based on the multi-task deep learning network optimized in step S5, and combined with the sliding time window strategy, a long-term growth health assessment model is constructed by capturing the temporal dependencies between windows through gated recurrent units.

[0018] S7. Quantitative assessment and grading of crop growth health: Using the long-term growth health assessment model constructed in step S6, the predicted values ​​of LAI, LCC and AGB are output. By integrating the predicted information of the three key parameters through the comprehensive health index model, a five-level grading standard is established to achieve quantitative assessment of crop growth health.

[0019] Furthermore, the specific implementation method of step S2 includes the following steps:

[0020] S2.1. Linear Spectral Mixture Analysis Endmember Decomposition: Pixel reflectance is decomposed into a linear combination of vegetation endmembers, soil endmembers, shading endmembers, and impermeable surface endmembers. The specific formula is as follows:

[0021]

[0022] in, To observe the spectral vector of a pixel, For the endmember spectral matrix, This is the endmember abundance vector. This represents the fitting error;

[0023] Set the condition to satisfy the nonnegative abundance constraint throughout the entire process. The sum is a constraint Upper bound constraint Then, the endmember abundance solution is encapsulated into a Bregman iterative unit in the LBnet network, and the LBnet network is used to solve for each endmember abundance and LAI, LCC, and AGB vegetation biological parameters.

[0024] S2.2. Physical simulation of PROSAIL radiative transfer model: The reflectance and transmittance of a single leaf are calculated using the PROSPECT leaf optical model, and the reflectance spectrum of the pure vegetation canopy is calculated using the SAIL canopy radiative transfer model.

[0025] S2.3. Construction of Coupled Reflectance Model: The vegetation abundance and soil reflectance information obtained from the spectral mixing analysis in step S2.1 are weighted and fused with the pure vegetation canopy reflectance spectrum in step S2.2 to construct a pixel-scale coupled reflectance model:

[0026]

[0027] in, Indicates wavelength The reflectance of vegetation biological parameters at a given location is simulated using a pixel-scale coupled model. Indicates vegetation end-member abundance. Indicates soil endogenous abundance. Indicates the abundance of shadow endons. Indicates wavelength The vegetation biological parameters at that location are the pure vegetation canopy reflectance. Indicates wavelength. Represents a set of vegetation biological parameters. This represents the soil reflectance at wavelength λ. This represents the shadow reflectivity at wavelength λ. This represents the model residual term at wavelength λ;

[0028] S2.4. LBnet Iterative Inversion: The simulated reflectance of the coupled reflectance model in step S2.3 is iteratively optimized and matched with the actual surface reflectance. Each iteration includes forward simulation calculation of residuals, calculation of cost function, unconstrained parameter update, constrained projection and dual variable update, and convergence judgment. After convergence, the initial estimates of the abundance of each endmember and vegetation biological parameters LAI, LCC, and AGB are output synchronously.

[0029] Furthermore, the specific implementation method of step S3 includes the following steps:

[0030] S3.1. The Logistic growth model is used to simulate the time dynamics of LAI;

[0031] S3.2. A piecewise function model is used to simulate the time dynamics of LCC;

[0032] S3.3. A light energy utilization model is used to simulate AGB accumulation;

[0033] S3.4. Calculate the cumulative effective accumulated temperature, and then generate a time series including the leaf area index. Time series of chlorophyll content in leaves aboveground biomass time series Accumulated effective accumulated temperature time series A multi-parameter time-series feature dataset.

[0034] Furthermore, the shared feature extraction module in step S4 includes, in sequence: first-layer two-dimensional convolution - batch normalization - activation, which extracts the input from... The first layer maps the features to 64 dimensions. The second layer, a convolution-batch normalization-activation-max pooling module, maintains 64 dimensions for local feature extraction. The third layer, a residual block, expands the features from 64 dimensions to 128 dimensions for deep feature extraction. The fourth layer, a Transformer encoder, performs two layers of self-attention temporal modeling to expand the features from 128 dimensions to 256 dimensions. The fifth layer, global average pooling, outputs a 256-dimensional shared feature vector.

[0035] The Transformer encoder includes a sine and cosine position encoder, a multi-head self-attention mechanism, a GELU-activated feedforward network, and layer normalization connected in sequence.

[0036] Furthermore, the spectral attention module in step S4 employs a channel attention structure to adaptively assign weights to the multispectral input features across different band dimensions; the calculation process is as follows:

[0037] First, the input spectral feature tensor... Global average pooling is performed, and the calculation formula is as follows:

[0038]

[0039]

[0040] in, This represents the band statistics descriptor vector of the global average pooling output. Indicates the first The global average value of each spectral band. Indicates the first At the first spatial location Characteristic values ​​of each spectral band This represents the product of the spatial dimensions of the feature maps. Indicates the number of spectral bands. Indicates global average pooling. Represents a real matrix;

[0041] The formula for calculating channel attention weights is as follows:

[0042]

[0043] in, This represents the weight matrix of the dimension-reduced fully connected layer. This represents the bias vector of the dimension-reduced fully connected layer. This represents the weight matrix of the upgraded fully connected layer. This represents the bias vector of the upgraded fully connected layer. This indicates a modified linear unit activation function. This represents the Sigmoid activation function. Represents the attention weight vector for the spectral bands;

[0044] Element-wise feature weighted calculation formula:

[0045]

[0046] in, Indicates the first Attention weights for each spectral band, Indicates the weighted number of... Band feature map.

[0047] Furthermore, the three-stage transfer strategy in step S5—source domain pre-training—domain adaptation fine-tuning—target domain fine-tuning includes the following stages:

[0048] Phase 1: Source Domain Pre-training: Under the condition of sufficient source domain data, all network parameters participate in training, and the parameters of the shared feature extraction layer are extracted as prior knowledge;

[0049] Phase 2: Domain Adaptation Fine-tuning: Freeze the fixed-ratio parameters of the shared feature extraction layer, fine-tune only the parameters of the task-specific layer and the spectral attention module, and add the domain-adaptive task loss to the loss function. ,in Indicates mission loss. This represents the adaptive loss in the MMD neighborhood. Indicates the MMD loss weighting coefficient;

[0050] Phase 3: Target Domain Fine-Tuning: Using labeled samples in the target domain, freeze the early layers and fine-tune only the final layers; employ a hierarchical learning rate decay strategy. ,in The total number of network layers. The current layer number. This is the inter-layer learning rate decay coefficient. Based on the learning rate.

[0051] Furthermore, the sliding time window in step S6 is defined as follows:

[0052]

[0053]

[0054] in, Indicates the first Temporal feature matrix within a sliding window Indicates the window length. Indicates the sliding step size. Indicates the first The input feature vector of the period, This indicates the total number of periods in the time series data. Indicates the total number of windows;

[0055] The expression for the constructed long-term growth health assessment model is as follows:

[0056]

[0057] in, This indicates a multi-task deep learning model for windows. The predicted output, This represents the hidden state output at the last time step, and HI represents the final output comprehensive health index. Window length The hidden state output vector. It is a fully connected layer. For random drop-out layers, This represents the modified linear unit activation function, where GRU stands for gated recurrent unit.

[0058] Furthermore, the expression for the comprehensive health index model in step S7 is as follows:

[0059]

[0060] in, This represents the weighting coefficient of LAI. This represents the weighting coefficient of LCC. This represents the weighting coefficient of AGB. This represents the leaf area index value predicted by the model. This represents the optimal reference value for the leaf area index at the current growth stage of the corresponding crop. This represents the leaf chlorophyll content value predicted by the model. This indicates the optimal reference value for leaf chlorophyll content at the corresponding growth stage. This represents the predicted aboveground biomass value from the model. This represents the optimal reference value for aboveground biomass at the corresponding growth stage. + + =1, ;

[0061] Based on the comprehensive health index HI, the crop growth health status is divided into five levels: excellent, good, medium, weak and poor. The threshold is determined based on the relationship between yield loss rate, the quantile of the cumulative distribution function of historical data and expert knowledge.

[0062] An evaluation system for a growth health assessment method coupling a spectral-radiative transfer algorithm and a multi-task deep learning algorithm includes a processor, a memory, and a computer program stored in the memory and run on the processor. When the computer program runs, it implements the steps of the growth health assessment method coupling the spectral-radiative transfer algorithm and the multi-task deep learning algorithm.

[0063] The beneficial effects of this invention are:

[0064] This invention discloses a growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm. Firstly, it innovatively proposes a coupling method between spectral mixing analysis and a radiative transfer model, constructing an "endmember decomposition-radiative transfer" coupled reflectance model. By first extracting the pure vegetation component spectra through endmember decomposition, and then using the PROSAIL model for forward simulation, it effectively solves the parameter inversion uncertainty problem caused by the influence of soil background mixing in traditional radiative transfer models. This coupling physical mechanism provides a reliable theoretical basis for intelligent data augmentation, enabling the generation of physically consistent and reliable multispectral simulation data and initial estimates of key parameters.

[0065] Second, based on the high-precision inversion results of the physical mechanism coupling model, this invention achieves continuous simulation of three key parameters—leaf area index (LAI), leaf chlorophyll content (LCC), and aboveground biomass (AGB)—throughout the entire growth period, generating a large-scale, high spatiotemporal resolution multi-parameter time-series feature dataset. This intelligent amplification strategy effectively overcomes the core constraints of scarce field samples and high annotation costs, providing sufficient and high-quality "near-true" samples for training deep learning models, fundamentally solving the model performance bottleneck caused by insufficient training data in traditional methods.

[0066] Third, this invention constructs a multi-task deep learning network based on a spectral attention mechanism. Through channel attention, it adaptively learns the contribution of each spectral band to the prediction of different physiological and biochemical parameters, achieving collaborative inversion of multiple parameters. The multi-task learning architecture fully utilizes the inherent physical correlation between LAI, LCC, and AGB parameters, achieving information complementarity through shared feature representations, significantly improving the prediction accuracy of each parameter compared to single-task models.

[0067] Fourth, this invention introduces a three-stage few-shot transfer learning strategy of "pre-training—fine-tuning—domain adaptation," which effectively solves the problem of insufficient labeled samples for different regions and crop types by minimizing the MMD distance between the feature distributions of the source and target domains. This strategy enables the model to have extremely strong cross-domain generalization ability, requiring only a very small number of target domain samples to achieve high-precision parameter prediction, significantly reducing the cost of field surveys and data labeling.

[0068] Fifth, this invention constructs a long-term growth health assessment model based on a sliding time window and a recurrent neural network, which can capture the temporal dynamic characteristics and key turning point information in the crop growth process, realize continuous dynamic monitoring and assessment of crop growth health, and overcome the limitations of traditional single-phase assessment methods.

[0069] Sixth, this invention enables intelligent and quantitative assessment of crop growth and health, with a comprehensive assessment accuracy of over 85%, providing reliable technical support for precision agricultural management. Attached Figure Description

[0070] Figure 1 This is a flowchart of a growth health assessment method that couples a spectral-radiative transfer algorithm with a multi-task deep learning algorithm, as described in this invention.

[0071] Figure 2 This is a diagram illustrating the architecture of the coupled simulation technology for spectral mixing analysis and radiative transfer model described in this invention.

[0072] Figure 3 This is a diagram of the multi-task deep learning network architecture based on the spectral attention mechanism described in this invention.

[0073] Figure 4 This is a graph illustrating the few-shot transfer learning and domain adaptation strategy described in this invention.

[0074] Figure 5 This is a diagram of the architecture of the long-term growth health assessment model described in this invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0076] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0077] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 -Appendix Figure 5 Detailed explanation is as follows:

[0078] Example 1:

[0079] A growth health assessment method coupling a spectral-radiative transfer algorithm and a multi-task deep learning algorithm includes the following steps:

[0080] S1. Multi-source data acquisition and preprocessing: Acquire multispectral remote sensing image data of the target area, and simultaneously acquire meteorological observation data and crop genetic characteristics information; acquire surface reflectance data; perform atmospheric correction, geometric correction and radiometric calibration preprocessing on the multispectral remote sensing image data;

[0081] Furthermore, the multispectral remote sensing imagery data in step S1 includes, but is not limited to, multispectral satellite sensor data such as Sentinel-2 MSI, Landsat-8 / 9 OLI, and HJ2A / 2B CCD, covering the visible, red-edge, and near-infrared bands. Meteorological observation data includes temperature (daily maximum / minimum / average temperature, in °C), precipitation (daily cumulative precipitation, in mm), solar radiation (daily total radiation, in MJ / m²), wind speed (daily average wind speed, in m / s), and relative humidity (daily average relative humidity, %). Data are derived from meteorological station observation records or ERA5 reanalysis data products within the study area. Crop genetic characteristics information includes varietal photothermal response parameters (vernalization temperature requirements, photoperiod sensitivity coefficient), growth stage type (early / mid / late maturing varieties), and growth habit parameters (maximum LAI, harvest index, specific leaf area, etc.).

[0082] S2. Coupling of spectral mixture analysis and radiative transfer model: The surface reflectance obtained in step S1 is decomposed into endmembers and forward simulated to generate multispectral reflectance simulation information, including the initial estimates of the abundance of each endmember, leaf area index (LAI), leaf chlorophyll content (LCC), and aboveground biomass (AGB).

[0083] The specific implementation method of step S2 includes the following steps:

[0084] S2.1. Linear Spectral Mixture Analysis Endmember Decomposition: Pixel reflectance is decomposed into a linear combination of vegetation endmembers, soil endmembers, shading endmembers, and impermeable surface endmembers. The specific formula is as follows:

[0085]

[0086] in, To observe the spectral vector of a pixel, For the endmember spectral matrix, This is the endmember abundance vector. This represents the fitting error;

[0087] Set the condition to satisfy the nonnegative abundance constraint throughout the entire process. The sum is a constraint Upper bound constraint Then, the endmember abundance solution is encapsulated into a Bregman iterative unit in the LBnet network, and the LBnet network is used to solve for each endmember abundance and LAI, LCC, and AGB vegetation biological parameters.

[0088] S2.2. Physical simulation of PROSAIL radiative transfer model: The reflectance and transmittance of a single leaf are calculated using the PROSPECT leaf optical model, and the reflectance spectrum of the pure vegetation canopy is calculated using the SAIL canopy radiative transfer model.

[0089] Furthermore, the PROSAIL radiative transfer model described in step S2.2 couples the PROSPECT leaf optical model and the SAIL canopy radiative transfer model, wherein:

[0090] The PROSPECT model considers the blade as... For a layered parallel plate structure, the weighted total uptake coefficient of each biochemical component is first calculated. ,in: Indicates the chlorophyll content of leaves. Indicates carotenoid content, Indicates the content of brown pigment. Indicates the equivalent water thickness of the blade. Indicates the dry matter content of leaves. Indicates blade structural parameters, Indicates wavelength. , , , , This represents the specific absorption coefficient (physical constant) of the corresponding component; then, the reflectance and transmittance of a single-layer blade are calculated using Fresnel's interfacial reflection law and second-order exponential integrals; finally, the Stokes recursive formula is used. Layer-by-layer acquisition Total reflectance of the layered blades and total transmittance ;

[0091] The SAIL model, based on the four-flow approximation theory, decomposes the radiation field within the canopy into direct solar radiation. Scattered downward radiation Scattered upward radiation and radiation in the direction of observation Four components are used to establish a system of coupled differential equations with the cumulative leaf area index as the independent variable:

[0092]

[0093] in: The cumulative leaf area index from top to bottom within the canopy. The extinction coefficient is the extinction coefficient of direct light. The extinction coefficient of diffuse light. The extinction coefficient is the coefficient of light in the direction of observation. The diffuse scattering coefficient is... The scattering coefficient is the diffuse scattering coefficient from the direction of observation. The scattering coefficient is the diffuse upward scattering coefficient from the observation direction. The intercept coefficient is the rate of interception from direct to diffuse downward radiation. The intercept coefficient is the upward trajectory from direct to diffuse light. The contribution coefficient of a single scattering incident directly onto the observation direction;

[0094] By combining leaf projection function, single-scatter albedo, and hotspot effect correction, the canopy directional reflectivity factor is solved under the boundary conditions of incident radiation at the top of the canopy and soil reflectivity at the bottom of the canopy. ,in, Indicates radiation from the observation direction at the top of the canopy. Indicates the direct solar irradiance at the top of the atmosphere. Indicates the irradiance scattered by the top of the atmosphere. Indicates the solar zenith angle;

[0095] S2.3. Construction of Coupled Reflectance Model: The vegetation abundance and soil reflectance information obtained from the spectral mixing analysis in step S2.1 are weighted and fused with the pure vegetation canopy reflectance spectrum in step S2.2 to construct a pixel-scale coupled reflectance model:

[0096]

[0097] in, Indicates wavelength The reflectance of vegetation biological parameters at a given location is simulated using a pixel-scale coupled model. Indicates vegetation end-member abundance. Indicates soil endogenous abundance. Indicates the abundance of shadow endons. Indicates wavelength The vegetation biological parameters at that location are the pure vegetation canopy reflectance. Indicates wavelength. Represents a set of vegetation biological parameters. This represents the soil reflectance at wavelength λ. This represents the shadow reflectivity at wavelength λ. This represents the model residual term at wavelength λ;

[0098] S2.4. LBnet Iterative Inversion: The simulated reflectance of the coupled reflectance model in step S2.3 is iteratively optimized and matched with the actual surface reflectance. Each iteration includes forward simulation calculation of residuals, calculation of cost function, unconstrained parameter update, constrained projection and dual variable update, and convergence judgment. After convergence, the initial estimates of the abundance of each endmember and vegetation biological parameters LAI, LCC, and AGB are output synchronously.

[0099] Furthermore, using the LBnet iterative inversion model, the simulated reflectance generated in step S2.3 is iteratively optimized and matched with the actual surface reflectance obtained in step S1 to obtain the optimal set of vegetation biological parameters (LAI, LCC, AGB) and the abundance of vegetation in each region.

[0100] (a) The formula for calculating the residual of the forward simulation is as follows:

[0101]

[0102] in: Represents the residual vector. This represents the actual surface reflectance obtained by S1. This indicates that the S2.3 coupling model simulates reflectivity.

[0103] The cost function is calculated as follows:

[0104]

[0105] in: Indicates abundance With vegetation biological parameter set The extent to which changes affect the simulated reflectivity This represents the weight matrix.

[0106] (b) Unconstrained update, calculated Substitute into the formula to calculate the candidates and The calculation formula is as follows:

[0107]

[0108] in: Indicates the first Layer iteration step size, Represents the identity matrix. Represents the dual auxiliary variable. Indicates the number of floors.

[0109] (c) Constrained projection + dual update, for candidate values , Perform constraint corrections and update the dual variables. .

[0110] a) The formula for calculating constrained projection is as follows:

[0111]

[0112] in: Represents the joint projection operator. Represents the abundance constraint set, Physical range of vegetation biological parameters.

[0113] b) The dual variable update formula is as follows:

[0114]

[0115] in: This represents the updated dual variable.

[0116] (d) Convergence judgment: Calculate the cost function once for each iteration layer. The iteration can be stopped if any of the following conditions are met:

[0117] a) The change in cost function between two consecutive layers is less than ;

[0118] b) The norm of parameter change is less than ;

[0119] c) Reach the maximum number of iterations.

[0120] (e) Output the final result;

[0121] a) After convergence For the abundance of local products;

[0122] b) After convergence For LAI, LCC, and AGB vegetation biological parameters;

[0123] c) Calculate the final To achieve a high degree of accuracy in fitting the target, This is considered a good fit.

[0124] Furthermore, step S2 also includes using a lower lookup table (LUT) method to assist in the calculation of the initial estimate;

[0125] Construct a large-scale parameter combination lookup table, use a KD-tree to accelerate the search, and take the first... The initial estimate is obtained by taking the inverse weighted average of the cost functions of the three optimal solutions, calculated as follows:

[0126] Cost function definition:

[0127]

[0128] LUT Construction and Search Strategies:

[0129]

[0130]

[0131] in, This represents the cost function value, which is the weighted sum of squared residuals between observed and simulated reflectances. This represents the vector of vegetation biological parameters to be inverted. This indicates the number of bands involved in the inversion. Indicates the first Weight of each band, Indicates the first Observed reflectance of each band Indicates the first Band parameters Simulated reflectivity of the lower-coupled model This indicates the total number of records in the lookup table. This represents the number of optimal solutions. This represents the final estimated value of the parameter;

[0132] S3. Using the multispectral reflectance simulation information obtained in step S2 as the true value, combined with meteorological observation data and crop genetic characteristics information, the crop growth model is used to continuously simulate the leaf area index (LAI), leaf chlorophyll content (LCC), and aboveground biomass (AGB) throughout the entire growth period to generate a multi-parameter time-series feature dataset.

[0133] Furthermore, the specific implementation method of step S3 includes the following steps:

[0134] S3.1. The Logistic growth model is used to simulate the time dynamics of LAI;

[0135] Furthermore, the change of LAI over time follows the characteristics of an S-shaped growth curve, and its dynamic process is described using the Logistic function.

[0136] The formula for calculating the Logistic growth function is as follows:

[0137]

[0138] The formula for calculating the Logistic differential equation is as follows:

[0139]

[0140] The inflection point analysis calculation formula is as follows: When hour

[0141]

[0142]

[0143] in, Indicates the first Leaf area index of the day This represents the maximum LAI value throughout the entire reproductive period. This represents the growth rate parameter. This indicates the moment when LAI reaches half of its maximum value. Represents a time variable. This represents the natural exponential function.

[0144] The parameter fitting method uses nonlinear least squares, and the formula for minimizing the sum of squared residuals between the observed and model values ​​is as follows:

[0145]

[0146] in, Let be the vector of parameters to be estimated. The initial parameter values ​​are estimated as follows:

[0147]

[0148]

[0149]

[0150] in, Indicates the number of observed samples. express LAI values ​​at each observation point The model represents the first time. Observation time The simulated LAI value, Indicates the first The annual accumulated days corresponding to each observation Indicates the date the image was acquired. express The initial estimate, This represents the initial estimate of the growth rate parameter. This represents the initial estimate of the moment of maximum growth rate.

[0151] The final output is the maximum LAI and growth rate throughout the entire reproductive period. Maximum growth rate moment The simulated value of LAI at any time ;

[0152] S3.2. A piecewise function model is used to simulate the time dynamics of LCC;

[0153] Furthermore, the leaf chlorophyll content (LCC) exhibits a three-stage change pattern of "rise-plateau-decline" throughout the entire crop growth period, which can be described by a piecewise function.

[0154]

[0155] in, Indicates the first The chlorophyll content of leaves on a given day Indicates the initial chlorophyll content. Indicates the maximum chlorophyll content. Indicates the time of growth initiation. This indicates the time when chlorophyll reaches its peak. Indicates the onset time of aging. Indicates the rate of aging and decay. Represents a time variable.

[0156] When the piecewise linear model does not fit the trend well, a dual logistic model is used to fit both the rising and falling trends simultaneously.

[0157]

[0158] in, Indicates the rate of increase during the rising phase. Indicates the time of the upward inflection point. Indicates the rate of descent. Indicates the time of the inflection point in the decline. This indicates the residual chlorophyll content.

[0159] S3.3. A light energy utilization model is used to simulate AGB accumulation;

[0160] Furthermore, the accumulation of aboveground biomass (AGB) is based on the light energy utilization efficiency (LUE) model. The core idea is that plants accumulate dry matter by intercepting photosynthetically active radiation (PAR) through photosynthesis, according to a certain conversion efficiency.

[0161] The formula for calculating light interception is as follows:

[0162]

[0163] The formula for calculating dry matter production is as follows:

[0164]

[0165] The formula for calculating aboveground biomass allocation is as follows:

[0166]

[0167] The formula for calculating cumulative dry matter is as follows:

[0168]

[0169] in, Indicates the first Photosynthetically active radiation intercepted by the canopy Indicates the first Photosynthetically active radiation reaching the Earth's surface Indicates the extinction coefficient. Indicates the first Leaf area index of the day Indicates radiation utilization efficiency. Indicates the first New dry matter added daily. Indicates from Heaven Total accumulated dry matter, Indicates the starting date for the accumulation of dry matter. Indicates the proportion of the above-ground portion. Indicates the first Biomass in the heavens and earth.

[0170] S3.4. Calculate the cumulative effective accumulated temperature, and then generate a time series including the leaf area index. Time series of chlorophyll content in leaves aboveground biomass time series Accumulated effective accumulated temperature time series A multi-parameter time-series feature dataset.

[0171] Furthermore, photosynthetically active radiation (PAR) is the energy source driving the AGB cumulative model. Surface solar radiation is estimated from sunshine duration using the Angstrom empirical formula, and 48% of it is then taken as the PAR component.

[0172]

[0173] The Angstrom formula for estimating Earth's surface solar radiation is as follows:

[0174]

[0175] The formula for calculating the top-of-atmosphere radiation is as follows:

[0176]

[0177] in, Indicates photosynthetically active radiation. This represents the total solar radiation on the Earth's surface. This refers to the top-of-atmosphere radiation. Indicates the actual sunshine duration. Indicates the number of hours available for lighting. Indicates the percentage of sunshine. Represents the solar constant. This represents the Earth-Sun distance correction factor. Indicates the angle at sunset. Indicates latitude, It represents the solar declination.

[0178] Calculation of cumulative effective accumulated temperature (GDD);

[0179] GDD is a core indicator for measuring the thermal conditions for crop growth and development, and is used as a time reference to drive phenological processes and parameter dynamics.

[0180]

[0181] The formula for calculating the average daily temperature is as follows:

[0182]

[0183] in, Indicates from arrive The cumulative effective temperature over the days Indicates the first The average daily temperature of the day Indicates the highest temperature of the day. Indicates the lowest temperature of the day. Indicates the lower limit temperature for crop growth. Indicates the upper limit temperature for crop growth. Indicates the starting date for calculating accumulated temperature.

[0184] S4. Constructing and training a multi-task deep learning network: Construct a multi-task deep learning network based on the spectral attention mechanism, including an input layer, a shared feature extraction module, a spectral attention module, three task-specific layers, and a terminal layer. Use the multi-parameter temporal feature dataset generated in step S3 to train the model and obtain the trained multi-task deep learning network.

[0185] Furthermore, the shared feature extraction module in step S4 includes, in sequence: first-layer two-dimensional convolution - batch normalization - activation, which extracts the input from... The first layer maps the features to 64 dimensions. The second layer, a convolution-batch normalization-activation-max pooling module, maintains 64 dimensions for local feature extraction. The third layer, a residual block, expands the features from 64 dimensions to 128 dimensions for deep feature extraction. The fourth layer, a Transformer encoder, performs two layers of self-attention temporal modeling to expand the features from 128 dimensions to 256 dimensions. The fifth layer, global average pooling, outputs a 256-dimensional shared feature vector.

[0186] The Transformer encoder includes a sine and cosine position encoder, a multi-head self-attention mechanism, a GELU-activated feedforward network, and layer normalization connected in sequence.

[0187] Furthermore, the location code calculation formula is as follows:

[0188]

[0189]

[0190] in, Indicates the position index in the sequence. Index representing the feature dimension This represents the hidden layer dimension of the model. This indicates the position code.

[0191] The formula for calculating multi-head self-attention is as follows:

[0192]

[0193]

[0194]

[0195]

[0196] in, Represents the input feature matrix. Indicates querying the projection matrix. Represents the key projection matrix. Represents the value projection matrix, Represents the query matrix. Represents the key matrix, Represents a value matrix, This represents the key / query dimension for each attention head. This represents the dimension of the value for each attention head. Indicates the number of heads of attention. Indicates the scaling factor. This represents the Softmax normalization function. Indicates the first The output of each attention head, This indicates a splicing operation. Indicates the output projection matrix. Indicates attention, This indicates a bullish trend.

[0197] The calculation formula for feedforward networks is as follows:

[0198]

[0199] in, This represents the first-layer weight matrix. This represents the first-level bias vector. This represents the second-layer weight matrix. This represents the second-layer bias vector. Represents the activation function of the Gaussian error linear unit. This represents the input feature matrix of the feedforward network. This represents a feedforward network.

[0200] The formula for layer normalization is as follows:

[0201]

[0202] in, Representing input features The mean vector in the last dimension, Representing input features The variance vector in the last dimension, This represents a learnable vector of scaling parameters. This represents the learnable offset parameter vector. Represents the numerical stability constant. Presentation layer normalization layer.

[0203] Furthermore, the spectral attention module in step S4 employs a channel attention structure to adaptively assign weights to the multispectral input features across different band dimensions; the calculation process is as follows:

[0204] First, the input spectral feature tensor... Global average pooling is performed, and the calculation formula is as follows:

[0205]

[0206]

[0207] in, This represents the band statistics descriptor vector of the global average pooling output. Indicates the first The global average value of each spectral band. Indicates the first At the first spatial location Characteristic values ​​of each spectral band This represents the product of the spatial dimensions of the feature maps. Indicates the number of spectral bands. Indicates global average pooling. Represents a real matrix;

[0208] The formula for calculating channel attention weights is as follows:

[0209]

[0210] in, This represents the weight matrix of the dimension-reduced fully connected layer. This represents the bias vector of the dimension-reduced fully connected layer. This represents the weight matrix of the upgraded fully connected layer. This represents the bias vector of the upgraded fully connected layer. This indicates a modified linear unit activation function. This represents the Sigmoid activation function. Represents the attention weight vector for the spectral bands;

[0211] Element-wise feature weighted calculation formula:

[0212]

[0213] in, Indicates the first Attention weights for each spectral band, Indicates the weighted number of... Band feature map.

[0214] Furthermore, the formula for calculating the multi-task weighted loss function is as follows:

[0215]

[0216] The formulas for calculating the mean squared error loss for each regression task are as follows:

[0217]

[0218] The formula for calculating the regularization term is as follows:

[0219]

[0220] The formula for calculating the uncertainty-weighted strategy is as follows:

[0221]

[0222] In the formula For learnable parameters, Equivalent weights.

[0223] in, This represents the total loss function for multiple tasks. Indicates the number of tasks. Indicates the first The loss weighting coefficient for each task. Indicates the first The loss value of each task, This indicates the number of samples in the training batch. Indicates the first The sample at the th The actual label value on each task The model represents the first The sample at the th Predicted values ​​for each task, Represents the regularization coefficient. express Regularization term, This represents the weight matrix of layer 1. Represents the weight matrix The square of the Frobenius norm, Indicates the first Learnable log-variance parameters for each task Indicates the first The observation noise variance for each task. Equivalent to Weighting factors Indicates the first The observation noise of each task.

[0224] S5. Small Sample Transfer Learning Optimization: The multi-task deep learning network trained in step S4 is subjected to a three-stage transfer strategy of source domain pre-training, domain adaptation fine-tuning, and target domain fine-tuning. The physical knowledge and spectrum-parameter mapping relationship learned in the source domain are transferred to the target domain. The generalization ability of the model is improved by minimizing the maximum mean difference between the feature distributions of the source domain and the target domain, and the optimized multi-task deep learning network is obtained.

[0225] Furthermore, the three-stage transfer strategy in step S5—source domain pre-training—domain adaptation fine-tuning—target domain fine-tuning includes the following stages:

[0226] Phase 1: Source Domain Pre-training: Under the condition of sufficient source domain data, all network parameters participate in training, and the parameters of the shared feature extraction layer are extracted as prior knowledge;

[0227] Phase 2: Domain Adaptation Fine-tuning: Freeze the fixed-ratio parameters of the shared feature extraction layer, fine-tune only the parameters of the task-specific layer and the spectral attention module, and add the domain-adaptive task loss to the loss function. ,in Indicates mission loss. This represents the adaptive loss in the MMD neighborhood. This represents the MMD loss weighting coefficient;

[0228] Phase 3: Target Domain Fine-Tuning: Using labeled samples in the target domain, freeze the early layers and fine-tune only the final layers; employ a hierarchical learning rate decay strategy. ,in The total number of network layers. The current layer number. This is the inter-layer learning rate decay coefficient. Based on the learning rate.

[0229] Furthermore, the three-stage migration strategy is elaborated as follows:

[0230] Phase 1: Source Domain Pre-training: Given sufficient source domain data, the model trained in step S4 is used as a pre-trained model, with all parameters participating in the training. Learning rate. Train for 100-300 epochs. Extract the parameters of the shared feature extraction layer as prior knowledge.

[0231] Phase 2: Domain Adaptation Fine-tuning: With only a few labeled samples (or no labeled samples) in the target domain, freeze the first 60% of the parameters of the shared feature extraction layer of the pre-trained model, and fine-tune only the parameters of the last few layers, task-specific layers, and the spectral attention mechanism module. A domain adaptation term is added to the loss function.

[0232]

[0233] Learning rate Train for 30-50 epochs.

[0234] in, Indicates mission loss. This represents the adaptive loss in the MMD neighborhood. This represents the MMD loss weighting coefficient.

[0235] Phase 3: Target Domain Fine-tuning: Using 50-200 labeled samples from the target domain, freeze the early layers and fine-tune only the final layers and the task header. Learning rate. arrive Train for 20-50 epochs. Employ a tiered learning rate decay strategy:

[0236]

[0237] in, Indicates the first Learning rate of layer Indicates the basic learning rate. Indicates the total number of network layers. Indicates the current layer number. This represents the inter-layer learning rate decay coefficient.

[0238] The learning rate is highest at the top layer and decreases progressively, so that the parameters of the early layers near the input remain almost unchanged, while the final layers near the output fully adapt to the distribution of the target domain data.

[0239] Domain Adaptive Loss Function - MMD Distance: Introducing a domain adaptive loss function, which minimizes the maximum mean difference between the feature distributions of the source and target domains.

[0240] Maximum Mean Difference (MMD) definition:

[0241] ;

[0242] The empirical formula for estimating MMD is as follows:

[0243] ;

[0244] The formula for calculating the MMD neighborhood adaptive loss in deep networks is as follows:

[0245] ;

[0246] The formula for calculating the loss of multi-layered combined MMD is as follows:

[0247] ;

[0248] The formula for calculating the Gaussian kernel function (RBF kernel) is as follows:

[0249] ;

[0250] nuclear parameters Median heuristic selection:

[0251] ;

[0252] in, Represents source domain distribution Distribution with target domain The square of the maximum mean difference between them, Indicates the distribution of source domain data. Indicates the distribution of data in the target domain. Defined in the regenerating nucleus Hilbert space Functions on Represents the regenerating nucleus Hilbert space. Representation function Distribution in the source domain The expected value of the following mathematical expression Representation function Distribution in the target domain The expected value of the following mathematical expression Indicates the number of samples in the source domain. Indicates the number of samples in the target domain. Represents the kernel function. express Source domain kernel matrix, express Target domain kernel matrix, express Cross kernel matrix, Indicates all vector, This represents the feature mapping function of a deep learning model. Represents the source domain input data set. Represents the target domain input data set. This represents the adaptive loss in the MMD neighborhood. This indicates the number of network layers used to calculate multi-layer combined MMD. Indicates the first The weighting coefficients of the layer MMD loss, Indicates the first The feature mapping function output by the layer network, This represents the bandwidth parameter of the Gaussian kernel. Indicates sample and The square of the Euclidean distance between them.

[0253] S6. Construction of a long-term growth health assessment model: Based on the multi-task deep learning network optimized in step S5, and combined with the sliding time window strategy, a long-term growth health assessment model is constructed by capturing the temporal dependencies between windows through gated recurrent units.

[0254] Furthermore, the sliding time window in step S6 is defined as follows:

[0255]

[0256]

[0257] in, Indicates the first Temporal feature matrix within a sliding window Indicates the window length. Indicates the sliding step size. Indicates the first The input feature vector of the period, This indicates the total number of periods in the time series data. Indicates the total number of windows;

[0258] The expression for the constructed long-term growth health assessment model is as follows:

[0259]

[0260] in, This indicates a multi-task deep learning model for windows. The predicted output, This represents the hidden state output at the last time step, and HI represents the final output comprehensive health index. Window length The hidden state output vector. It is a fully connected layer. For random drop-out layers, This represents the modified linear unit activation function, where GRU stands for gated recurrent unit.

[0261] Furthermore, S6.1. Sliding time window strategy;

[0262] Missing data were imputed using linear interpolation or Gaussian process regression (GPR) interpolation:

[0263]

[0264] in, and Missing time points The most recent time before or after that with an observed value.

[0265] S6.2. GRU gated loop unit;

[0266] After locally modeling the multispectral temporal features within each window, the temporal dependencies between windows are captured using GRU.

[0267] The formula for calculating the GRU update gate is as follows:

[0268]

[0269] The formula for calculating the GRU reset gate is as follows:

[0270] ;

[0271] The formula for calculating GRU candidate hidden states is as follows:

[0272] ;

[0273] The formula for calculating the GRU hidden state output is as follows:

[0274]

[0275] in, This indicates updating the gate output vector. This represents the reset gate output vector. Represents the candidate hidden state vector. Indicates the first The hidden state output vector at time step 1. Indicates the first The hidden state vector at time step 1. Indicates the first The input feature vector at time t, This represents the weight matrix used to update the hidden states in the gate to the gate control signal. This represents updating the weight matrix of the input signal to the gate. This represents the bias vector for updating the gate. This represents the weight matrix from resetting the hidden state in the door to the gate control signal. This represents the weight matrix of the input gate control signal in the reset gate. This represents the bias vector of the reset gate. This represents the weight matrix of the hidden states in the candidate state computation. This represents the weight matrix input in the candidate state calculation. This represents the bias vector used to calculate the candidate state. This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function.

[0276] The GRU can be replaced with a bidirectional GRU structure, which concatenates the forward and backward hidden states.

[0277] S6.3. Timing data stream;

[0278] Input feature vector composition:

[0279] ;

[0280] S7. Quantitative assessment and grading of crop growth health: Using the long-term growth health assessment model constructed in step S6, the predicted values ​​of LAI, LCC and AGB are output. By integrating the predicted information of the three key parameters through the comprehensive health index model, a five-level grading standard is established to achieve quantitative assessment of crop growth health.

[0281] Furthermore, the expression for the comprehensive health index model in step S7 is as follows:

[0282]

[0283] in, This represents the weighting coefficient of LAI. This represents the weighting coefficient of LCC. This represents the weighting coefficient of AGB. This represents the leaf area index value predicted by the model. This represents the optimal reference value for the leaf area index at the current growth stage of the corresponding crop. This represents the leaf chlorophyll content value predicted by the model. This indicates the optimal reference value for leaf chlorophyll content at the corresponding growth stage. This represents the predicted aboveground biomass value from the model. This represents the optimal reference value for aboveground biomass at the corresponding growth stage. + + =1, ;

[0284] Based on the comprehensive health index HI, the crop growth health status is divided into five levels: excellent, good, medium, weak and poor. The threshold is determined based on the relationship between yield loss rate, the quantile of the cumulative distribution function of historical data and expert knowledge.

[0285] Furthermore, the constraints are: + + =1, ;

[0286] The formula for calculating the bell-shaped Gaussian normalized mapping function is as follows:

[0287]

[0288] ;

[0289] The formula for calculating asymmetric bell-shaped mapping is as follows:

[0290]

[0291] in, Indicates the overall health index. This represents the leaf area index value predicted by the model. This represents the optimal reference value for the leaf area index at the current growth stage of the corresponding crop. This represents the leaf chlorophyll content value predicted by the model. This indicates the optimal reference value for leaf chlorophyll content at the corresponding growth stage. This represents the predicted aboveground biomass value from the model. This represents the optimal reference value for aboveground biomass at the corresponding growth stage. Indicates the first A normalized mapping function with parameters. Indicates the first The ratio of the predicted value of each parameter to the optimal reference value. The width parameter represents the bell-shaped Gaussian mapping function. This represents the width parameter in the lower direction of the asymmetric mapping. This represents the width parameter in the vertical direction of the asymmetric mapping. This represents the weighting coefficient of LAI. This represents the weighting coefficient of LCC. This represents the weighting coefficient of AGB.

[0292] The five-level classification criteria are shown in Table 1:

[0293] Table 1

[0294]

[0295] Example 2:

[0296] An evaluation system for a growth health assessment method that couples a spectral-radiative transfer algorithm with a multi-task deep learning algorithm includes a processor, a memory, and a computer program stored in the memory and run on the processor. When the computer program runs, it implements the steps of the growth health assessment method that couples a spectral-radiative transfer algorithm with a multi-task deep learning algorithm as described in Embodiment 1.

[0297] Example 3:

[0298] This embodiment is based on Embodiment 1 and is illustrated with the following example:

[0299] Taking a winter wheat growing area in a county in the North China Plain as the research area, the implementation steps are as follows:

[0300] Step 1: Data Acquisition

[0301] Sentinel-2 MSI multispectral remote sensing images of winter wheat throughout its entire growth period from March to June 2025 were acquired in the study area, with a temporal resolution of 5 days and a spatial resolution of 10 meters. Simultaneously, temperature, precipitation, solar radiation, and relative humidity data observed at meteorological stations in the area, as well as genetic characteristic parameters of the main wheat varieties grown in the study area, were acquired.

[0302] Step 2: Coupled Simulation of Spectral Mixture Analysis and Radiative Transfer Model

[0303] Using the LBnet spectral mixture analysis model, spectral endmembers (green vegetation, dry soil, moist soil, and shade) of typical land cover in the study area were selected, and the abundance of each endmember in the mixed pixels was retrieved using the optimal model. The vegetation endmember abundance map was extracted as vegetation cover information.

[0304] Vegetation endmember abundance and soil reflectance information were input into the PROSAIL radiative transfer model. Initial parameter ranges were set as follows: LAI 0.5-6.0, Cab 20-80 ug / cm², Cw 0.005-0.05 g / cm², Cm 0.002-0.02 g / cm², ALA 30-60 degrees, and N 1.5-2.5. A lookup table (LUT) method was used to generate... We have a simulated dataset of canopy reflectivity corresponding to a combination of parameters of magnitude, and we use a KD tree to accelerate the search for the optimal match.

[0305] The reflectance simulated by the coupled model is compared with the actual observed reflectance, and the optimal parameter combination (adaptive adjustment of damping factor and convergence threshold) is solved iteratively using the Levenberg-Marquardt optimization algorithm. ), to obtain initial estimates of LAI, LCC and AGB.

[0306] Step 3: Timing simulation of key parameters

[0307] Constrained by the inverted LAI, LCC, and AGB, combined with the daily accumulated effective temperature ( , ) and photosynthetically active radiation data ( The dynamic changes in LAI (Label Artery Index) throughout the entire growth period of winter wheat were simulated using the Logistic growth function. , , .

[0308] Simulating LCC using a piecewise function model: linear growth before joint elongation (from Increase to The index remained at a plateau during the heading-grain-filling stage, and then declined exponentially in the later stages of grain filling. Using a material distribution model ( (The aboveground allocation ratio increased from 0.6 to 0.9) Combined with the harvest index, the cumulative process of AGB was simulated.

[0309] Step 4: Training the Multi-Task Deep Learning Network

[0310] A three-task deep learning network was constructed. The shared feature extraction layer employed a two-layer Transformer encoder, each containing four attention heads, with a hidden layer dimension of 128. The three task-specific layers corresponded to the LAI, LCC, and AGB prediction tasks, respectively. Each task-specific layer contained two fully connected layers (dimensions 64-32) and one spectral attention module. The spectral attention module adaptively weighted the 13 spectral bands of Sentinel-2 (compression ratio r=4). The training consisted of 200 epochs, a batch size of 64, an initial learning rate of 0.001, and cosine annealing scheduling.

[0311] Step 5: Few-shot transfer learning

[0312] Historical labeled data from across the county was used as the source domain (approximately 5000 samples), and 30 measured sample points within the study area were selected as the target domain. The first 60% of the parameters of the shared feature extraction layer of the pre-trained model were frozen, and the task-specific layer and spectral attention module were fine-tuned. The MMD distance in the domain-adaptive loss function was calculated using a Gaussian kernel (the parameters were determined heuristically by the median).

[0313] Step 6: Construction of Long-Term Evaluation Model

[0314] The sliding time window length is set to 12 periods (approximately 60 days), with a step size of 3 periods (15 days). Within each window, a multi-task deep learning model is used to predict LAI, LCC, and AGB. Temporal dependencies between windows are modeled using a 2-layer GRU network (128 hidden dimensions).

[0315] Step 7: Growth Health Assessment

[0316] The trained model was used to assess the growth health of the study area. The weights of each parameter were determined using the AHP method: w1=0.35 (LAI), w2=0.35 (LCC), w3=0.30 (AGB), and the normalized mapping function was a bell-shaped Gaussian function (parameter=0.3).

[0317] The evaluation results show that the overall evaluation accuracy of this method for the health level of winter wheat growth reaches 87.3%, with the accuracy rates for identifying "excellent" and "poor" levels reaching 89.1% and 91.5%, respectively.

[0318] Example 4:

[0319] This embodiment is based on Embodiment 1 and is illustrated with the following example:

[0320] The study area was a rice-growing region in a city in the middle and lower reaches of the Yangtze River. Multispectral images with a temporal resolution of 3 days were acquired using fused Sentinel-2 and Landsat-8 data. Model parameters were adjusted for different rice varieties: LAImax was adjusted to 6.0-8.0, leaf structure parameter N was adjusted to 1.0-1.5, and Tbase was adjusted to 10-12C.

[0321] Using the same methodology as in Example 3, with only 20 measured sample points (small sample condition), knowledge from the source domain (winter wheat production area in Example 1) was transferred to the rice production area through three-stage transfer learning. After domain adaptation fine-tuning (MMD loss weight = 0.1) and fine-tuning, the comprehensive assessment accuracy of the rice growth health level reached 86.1%.

[0322] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0323] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A growth health assessment method coupling a spectral-radiative transfer algorithm and a multi-task deep learning algorithm, characterized in that, Includes the following steps: S1. Multi-source data acquisition and preprocessing: Acquire multispectral remote sensing image data of the target area, and simultaneously acquire meteorological observation data and crop genetic characteristics information; acquire surface reflectance data; perform atmospheric correction, geometric correction and radiometric calibration preprocessing on the multispectral remote sensing image data; S2. Coupling of spectral mixture analysis and radiative transfer model: The surface reflectance obtained in step S1 is decomposed into endmembers and forward simulated to generate multispectral reflectance simulation information, including the initial estimates of the abundance of each endmember, leaf area index (LAI), leaf chlorophyll content (LCC), and aboveground biomass (AGB). S3. Using the multispectral reflectance simulation information obtained in step S2 as the true value, combined with meteorological observation data and crop genetic characteristics information, the crop growth model is used to continuously simulate the leaf area index (LAI), leaf chlorophyll content (LCC), and aboveground biomass (AGB) throughout the entire growth period to generate a multi-parameter time-series feature dataset. S4. Constructing and training a multi-task deep learning network: Construct a multi-task deep learning network based on the spectral attention mechanism, including an input layer, a shared feature extraction module, a spectral attention module, three task-specific layers, and a terminal layer. Use the multi-parameter temporal feature dataset generated in step S3 to train the model and obtain the trained multi-task deep learning network. S5. Small Sample Transfer Learning Optimization: The multi-task deep learning network trained in step S4 is subjected to a three-stage transfer strategy of source domain pre-training, domain adaptation fine-tuning, and target domain fine-tuning. The physical knowledge and spectrum-parameter mapping relationship learned in the source domain are transferred to the target domain. The generalization ability of the model is improved by minimizing the maximum mean difference between the feature distributions of the source domain and the target domain, and the optimized multi-task deep learning network is obtained. S6. Construction of a long-term growth health assessment model: Based on the multi-task deep learning network optimized in step S5, and combined with the sliding time window strategy, a long-term growth health assessment model is constructed by capturing the temporal dependencies between windows through gated recurrent units. S7. Quantitative assessment and grading of crop growth health: Using the long-term growth health assessment model constructed in step S6, the predicted values ​​of LAI, LCC and AGB are output. By integrating the predicted information of the three key parameters through the comprehensive health index model, a five-level grading standard is established to achieve quantitative assessment of crop growth health.

2. The growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm according to claim 1, characterized in that, The specific implementation method of step S2 includes the following steps: S2.

1. Linear Spectral Mixture Analysis Endmember Decomposition: Pixel reflectance is decomposed into a linear combination of vegetation endmembers, soil endmembers, shading endmembers, and impermeable surface endmembers. The specific formula is as follows: in, To observe the spectral vector of the pixel, For the endmember spectral matrix, This is the endmember abundance vector. This represents the fitting error; Set the entire process to satisfy the nonnegative abundance constraint. The sum is a constraint Upper bound constraint Then, the endmember abundance solution is encapsulated into a Bregman iterative unit in the LBnet network, and the LBnet network is used to solve for each endmember abundance and LAI, LCC, and AGB vegetation biological parameters. S2.

2. Physical simulation of PROSAIL radiative transfer model: The reflectance and transmittance of a single leaf are calculated using the PROSPECT leaf optical model, and the reflectance spectrum of the pure vegetation canopy is calculated using the SAIL canopy radiative transfer model. S2.

3. Construction of Coupled Reflectance Model: The vegetation abundance and soil reflectance information obtained from the spectral mixing analysis in step S2.1 are weighted and fused with the pure vegetation canopy reflectance spectrum in step S2.2 to construct a pixel-scale coupled reflectance model: in, Indicates wavelength The reflectance of vegetation biological parameters at a given location is simulated using a pixel-scale coupled model. Indicates vegetation end-member abundance. Indicates soil endogenous abundance. Indicates the abundance of shadow endons. Indicates wavelength The vegetation biological parameters at that location are the pure vegetation canopy reflectance. Indicates wavelength. Represents a set of vegetation biological parameters. This represents the soil reflectance at wavelength λ. This represents the shadow reflectivity at wavelength λ. This represents the model residual term at wavelength λ; S2.

4. LBnet Iterative Inversion: The simulated reflectance of the coupled reflectance model in step S2.3 is iteratively optimized and matched with the actual surface reflectance. Each iteration includes forward simulation calculation of residuals, calculation of cost function, unconstrained parameter update, constrained projection and dual variable update, and convergence judgment. After convergence, the initial estimates of the abundance of each endmember and vegetation biological parameters LAI, LCC, and AGB are output synchronously.

3. A growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm according to claim 1 or 2, characterized in that, The specific implementation method of step S3 includes the following steps: S3.

1. The Logistic growth model is used to simulate the time dynamics of LAI; S3.

2. A piecewise function model is used to simulate the time dynamics of LCC; S3.

3. A light energy utilization model is used to simulate AGB accumulation; S3.

4. Calculate the cumulative effective accumulated temperature, and then generate a time series including the leaf area index. Time series of chlorophyll content in leaves aboveground biomass time series Accumulated effective accumulated temperature time series A multi-parameter time-series feature dataset.

4. The growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm according to claim 3, characterized in that, The shared feature extraction module in step S4 includes, in sequence: first-layer 2D convolution - batch normalization - activation, which extracts the input from... The first layer maps the features to 64 dimensions. The second layer, a convolution-batch normalization-activation-max pooling module, maintains 64 dimensions for local feature extraction. The third layer, a residual block, expands the features from 64 dimensions to 128 dimensions for deep feature extraction. The fourth layer, a Transformer encoder, performs two layers of self-attention temporal modeling to expand the features from 128 dimensions to 256 dimensions. The fifth layer, global average pooling, outputs a 256-dimensional shared feature vector. The Transformer encoder includes a sine and cosine position encoder, a multi-head self-attention mechanism, a GELU-activated feedforward network, and layer normalization connected in sequence.

5. The growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm according to claim 4, characterized in that, The spectral attention module in step S4 employs a channel attention structure to adaptively assign weights to multispectral input features across different band dimensions; the calculation process is as follows: First, the input spectral feature tensor... Global average pooling is performed, and the calculation formula is as follows: in, This represents the band statistics descriptor vector of the global average pooling output. Indicates the first The global average value of each spectral band. Indicates the first At the first spatial location Characteristic values ​​of each spectral band This represents the product of the spatial dimensions of the feature maps. Indicates the number of spectral bands. Indicates global average pooling. Represents a real matrix; The formula for calculating channel attention weights is as follows: in, This represents the weight matrix of the dimension-reduced fully connected layer. This represents the bias vector of the dimension-reduced fully connected layer. This represents the weight matrix of the upgraded fully connected layer. This represents the bias vector of the upgraded fully connected layer. This indicates a modified linear unit activation function. This represents the Sigmoid activation function. Represents the attention weight vector for the spectral bands; Element-wise feature weighted calculation formula: in, Indicates the first Attention weights for each spectral band, Indicates the weighted number of... Band feature map.

6. The growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm according to claim 5, characterized in that, The three-stage transfer learning strategy in step S5—source domain pre-training—domain adaptation fine-tuning—target domain fine-tuning includes the following stages: Phase 1: Source Domain Pre-training: Under the condition of sufficient source domain data, all network parameters participate in training, and the parameters of the shared feature extraction layer are extracted as prior knowledge; Phase 2: Domain Adaptation Fine-tuning: Freeze the fixed-ratio parameters of the shared feature extraction layer, fine-tune only the parameters of the task-specific layer and the spectral attention module, and add the domain-adaptive task loss to the loss function. ,in Indicates mission loss. This represents the adaptive loss in the MMD neighborhood. This represents the MMD loss weighting coefficient; Phase 3: Target Domain Fine-Tuning: Using labeled samples in the target domain, freeze the early layers and fine-tune only the final layers; employ a hierarchical learning rate decay strategy. ,in The total number of network layers. The current layer number. This is the inter-layer learning rate decay coefficient. Based on the learning rate.

7. The growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm according to claim 6, characterized in that, The sliding time window in step S6 is defined as follows: in, Indicates the first Temporal feature matrix within a sliding window Indicates the window length. Indicates the sliding step size. Indicates the first The input feature vector of the period, This indicates the total number of periods in the time series data. Indicates the total number of windows; The expression for the constructed long-term growth health assessment model is as follows: in, This indicates a multi-task deep learning model for windows. The predicted output, This represents the hidden state output at the last time step, and HI represents the final output comprehensive health index. Window length The hidden state output vector. It is a fully connected layer. For random drop-out layers, This represents the modified linear unit activation function, where GRU stands for gated recurrent unit.

8. The growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm according to claim 7, characterized in that, The expression for the comprehensive health index model in step S7 is as follows: in, This represents the weighting coefficient of LAI. This represents the weighting coefficient of LCC. This represents the weighting coefficient of AGB. This represents the leaf area index value predicted by the model. This represents the optimal reference value for the leaf area index at the current growth stage of the corresponding crop. This represents the leaf chlorophyll content value predicted by the model. This indicates the optimal reference value for leaf chlorophyll content at the corresponding growth stage. This represents the predicted aboveground biomass value from the model. This represents the optimal reference value for aboveground biomass at the corresponding growth stage. + + =1, ; Based on the comprehensive health index HI, the crop growth health status is divided into five levels: excellent, good, medium, weak and poor. The threshold is determined based on the relationship between yield loss rate, the quantile of the cumulative distribution function of historical data and expert knowledge.

9. An evaluation system for a growth health assessment method coupling a spectral-radiative transfer algorithm and a multi-task deep learning algorithm, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed, it implements the steps of a growth health assessment method coupled with a spectral-radiative transfer algorithm and a multi-task deep learning algorithm as described in any one of claims 1-8.