An intelligent analysis method for crop growth state based on multi-source remote sensing images

Abstract

The application provides a kind of crop growth state intelligent analysis method based on multi-source remote sensing image, belongs to the technical field of crop growth analysis based on computer data processing. First, collect multi-source remote sensing time series images covering the entire growth cycle, and construct a unified structure of multi-modal growth feature dataset. Second, design a multi-channel time series modeling structure, rely on air-based spectrum and ground-based morphological information to extract the dynamic evolution trajectory of core features such as canopy height, leaf area index and SPAD value. Then, through multi-type trajectory distribution learning model, identify the current state deviation type, and combine the diffusion matching mechanism to trace the problem and explain the analysis. Finally, based on the trajectory deviation result and management target, generate variable adjustment strategy to optimize the recommendation of irrigation frequency, fertilization intensity and other management measures. The application can be widely applied to agricultural remote sensing monitoring, intelligent agricultural machinery scheduling and precision planting guidance, and has high interpretability and practical value.

Description

Technical Field

[0001] This invention belongs to the field of crop growth analysis technology based on computer data processing, and particularly relates to an intelligent analysis method for crop growth status based on multi-source remote sensing images. Background Technology

[0002] With the continuous improvement of global agricultural intelligence, remote sensing technology has become an important tool for crop growth monitoring and precision management. Multi-source remote sensing data can provide spatiotemporal dynamic information on key indicators such as canopy structure, chlorophyll content, and coverage without direct contact with crops. However, current agricultural remote sensing monitoring still suffers from insufficient ability to fuse and model multi-source data, inaccurate temporal feature modeling, and a lack of deviation identification and interpretation mechanisms, resulting in limited practicality and intelligence in precision agricultural management. Therefore, developing an intelligent analysis method for crop growth status based on multi-source remote sensing imagery and constructing an integrated technical process from feature extraction, trajectory modeling, state deviation diagnosis to management strategy recommendation has significant scientific research value and practical significance, and can significantly improve the service capabilities of agricultural remote sensing in precision regulation and intelligent management.

[0003] Currently, the main methods for analyzing crop growth status include the following:

[0004] (1) Growth status monitoring method based on a single remote sensing indicator: This method mainly relies on a representative remote sensing parameter (such as Normalized Difference Vegetation Index NDVI, chlorophyll content index SPAD, etc.) to assess the current status of crops. By setting an indicator threshold, it determines whether crop growth is normal and assists in agricultural monitoring and disaster early warning. However, this method ignores the multidimensional synergistic characteristics of crop growth processes, and relying solely on a single parameter for judgment is easily affected by environmental noise or periodic fluctuations, resulting in a high risk of misjudgment. At the same time, its ability to interpret abnormal changes is limited, making it difficult to support precise control decisions.

[0005] (2) Growth Feature Prediction Method Based on Time Series Models: This method uses statistical regression models (such as ARIMA) or deep learning models (such as LSTM and GRU) to perform time series modeling and future trend prediction of crop growth features in remote sensing images. By learning historical change patterns, it can infer future feature values. Although this method has certain predictive capabilities, it generally relies on long-term, high-quality continuous sequences and has poor robustness to missing data and noise. At the same time, it focuses on single feature modeling and lacks the ability to jointly model spatial structural features and multimodal information, which limits the realistic portrayal of crop behavior in complex scenarios.

[0006] (3) Planting strategy recommendation method based on expert experience rules: This method relies on the experience and historical knowledge base of agricultural experts, combined with crop growth stages and measured parameters, to make suggestions on management variables such as irrigation and fertilization. It is often used in planting guidance systems and agricultural technology platforms. However, this method relies heavily on human knowledge and fixed rules, and lacks the ability to adapt to the personalized trajectory and growth status of the current plot. Its recommendation results lack a dynamic feedback mechanism supported by data, making it difficult to make fine adjustments based on the degree of prediction deviation, and it cannot generate variable optimization schemes with quantitative basis and interpretability, which affects its practical operability and promotion value.

[0007] In summary, existing methods for analyzing crop growth status have significant shortcomings in terms of information dimensions and intelligent decision support, making it difficult to achieve accurate perception and control response to complex growth processes. Summary of the Invention

[0008] To address the above problems, this invention proposes an intelligent analysis method for crop growth status based on multi-source remote sensing imagery, comprising the following steps:

[0009] S1, by deploying an air-ground collaborative multi-source remote sensing acquisition system, acquires multi-channel remote sensing image sequence data of crop growth with temporal continuity and spatial differences;

[0010] S2 inputs the remote sensing image sequence data obtained in S1 into the trained multi-source image-driven crop growth feature recognition model and outputs the time-series prediction results of crop growth features. The model includes a multi-source image feature interaction channel, a time-series evolution modeling channel, and a structural feature decoupling prediction channel, which sequentially realizes multi-source feature fusion, time evolution law extraction, and multi-task prediction of target features.

[0011] S3, based on the intelligent analysis and deviation interpretation module for crop growth status, takes the time-series prediction results of crop growth characteristics output by S2 as input, combines the constructed crop growth characteristic trajectory library, models the distribution characteristics of various typical growth trajectories based on a multi-class trajectory diffusion model, and then performs matching analysis on the current crop trajectory to determine its growth status type and quantify its deviation degree and trend from the target trajectory; outputs the corresponding crop growth type matching score and deviation index; at the same time, it outputs a natural language interpretation report of the most significant trend based on the meaning of the features.

[0012] Preferably, it also includes S4: screening key control variables based on crop growth type and deviation indicators according to matching scores, and constructing an optimization function in combination with user objectives; finally, using a Bayesian optimization algorithm to search for the optimal variable configuration and form management strategy recommendations.

[0013] Preferred methods for constructing datasets for training multi-source image-driven crop growth feature recognition models include:

[0014] Within a certain growth cycle of the crop, 1→T, a total of T uniform multi-source crop image data acquisitions are performed, resulting in remote sensing image sequence data SMC=[Mc1, Mc2, ..., Mc T ], among which, Mc t S1-3 represents the multi-source crop image data collected at time point t, where t∈[1,T]; S1-3 Crop growth feature structured annotation:

[0015] In the target plots where image data was collected, several representative crop sample points were selected, and the experimental team used field measurement methods to obtain the structured growth characteristics (Gcf) of the crops. The growth characteristic indicators include canopy height (Ht), leaf area index (Li), SPAD value (Sd) reflecting chlorophyll content, vegetation coverage (Vp), normalized difference vegetation index (Ni), red edge index (Re), and aboveground biomass estimate (Ab), i.e., Gcf = [Ht, Li, Sd, Vp, Ni, Re, Ab].

[0016] Structured growth characteristics of the crop are collected at each growth stage, and complete structured labels of crop growth characteristics GT = [Gcf1, Gcf2, ..., Gcf2] are obtained. T ]; where Gcf t Let represent the structured growth characteristics of the crop at time t, where t∈[1,T];

[0017] The remote sensing image sequence data SMC is used as input data, and the complete crop growth feature structured label GT is used as output data.

[0018] Preferably, the multi-source image feature interaction channel consists of a space-based multispectral structure analysis branch and a ground morphology enhancement observation branch;

[0019] The space-based multispectral structure analysis branch takes multi-angle multispectral images acquired by a UAV as input, first obtains high-dimensional spectral response patterns through a spectral attention layer, and then extracts spectral structure features F. spc Next, this feature is fed into the image depth sensing layer, which has a dual-branch convolutional structure to extract local texture responses and multi-scale structures respectively. Finally, these are fused to obtain the hollow structure enhancement feature F. arb ;

[0020] The ground morphology enhancement observation branch takes visible light images acquired by a ground platform as input. First, it encodes the geometric weight relationship of different observation angles through an angle attention layer to obtain the viewpoint sensitive feature F. prs Subsequently, the image depth sensing layer extracts local contours and plant conditions, outputting ground morphology enhancement features F. gdb ;

[0021] The characteristic F of the two-branch output arband F gdb Joint updates are performed through a gating fusion mechanism to form a fused multi-source complementary feature F. msc ; Integration and complementarity features F msc Subsequently, the system re-enters the spectral attention layer and the angular attention layer for interactive updates; the three-layer stacked structure of "attention layer-depth perception-gated fusion" is repeated eight times to form a deep multi-source interactive structure, and the output is the final heterogeneous perceptual representation feature F of this channel. hpr .

[0022] Preferably, the time-series remote sensing sequence evolution modeling channel is used to model the evolution of remote sensing features over time, and the input is heterogeneous sensing representation features F. hpr First, the time stamp information is introduced through a time embedding enhancement layer, and the time-enhanced feature F is output. ten ;

[0023] Subsequently, this feature F ten Short-term dynamic modeling is performed using a TCN causal convolutional layer with a kernel size of 3, and the local temporal response feature F is output using the ReLU activation function. lre Feature F lre Input a lightweight Transformer encoder for long-term dependency modeling, and output temporal context-aware features F. tsp ;

[0024] The entire "TCN-Lightweight Transformer" structure is repeatedly stacked 5 times to obtain the output temporal evolution feature representation F of the final temporal remote sensing sequence evolution modeling channel. sec .

[0025] Preferably, the structural feature decoupling prediction channel is represented by temporal evolution features F. sec The input is used to achieve multi-task prediction of the structural growth characteristics of the target crop;

[0026] First, the interdependencies between features are modeled using a multi-head attention mechanism to obtain the decoupled representation F. der Then, the input is fed into a fully connected decoupling layer to separate the feature channels between different tasks; after that, it is fed into two "multilayer perceptron-ReLU activation" structures for nonlinear modeling, and outputs the preliminary predicted features for each task.

[0027] Finally, regularization is achieved through a Dropout layer, followed by a ReLU activation function, and finally fed into a multi-task prediction head to output predicted values ​​for each crop growth feature, forming a complete crop growth feature prediction result (GT). pre .

[0028] Preferably, the specific processing steps of the intelligent crop growth status analysis and deviation interpretation module include:

[0029] S31, Multi-class standard trajectory distribution modeling based on diffusion model: The crop growth feature trajectory library divides the samples into 5 crop growth trajectory subsets S1, S2, S3, S4, and S5, corresponding to high yield, low yield, drought stress, developmental arrest, and premature aging abnormality, respectively; For each type of trajectory subset, a diffusion model is used for training to learn the characteristic evolution and distribution law of that type of growth trajectory.

[0030] S32, Matching analysis between the current trajectory and the standard trajectory distribution:

[0031] The predicted crop growth characteristics GT pre As input to five trajectory diffusion models, five trajectory reconstruction results GT1, GT2, ..., GT5 are generated;

[0032] Calculate and predict the original crop growth trajectory GT pre The mean squared error of the five trajectory reconstruction results GT1, GT2, ..., GT5 is taken as the explicit error. The cosine similarity between the two is calculated as the implicit distance. Finally, the matching score is defined as the implicit distance divided by the explicit distance, and the larger the matching score, the higher the similarity after reconstruction, i.e., the predicted crop growth characteristic GT pre The more closely the trajectory data distribution matches this category;

[0033] Five matching scores were calculated as Score1, Score2, ..., Score5. The growth state type Cls of the current plot was determined based on the maximum matching score, and the matching score sequence was retained as a reliability reference.

[0034] S33, Quantitative Analysis of Deviation Trends and Characteristic Differences:

[0035] The best matching category Cls in the basis determination is used to randomly sample Trv trajectories in its corresponding trajectory diffusion model. Then, the mean trajectory GT of this category is generated by averaging all sampled trajectories step by step. Trl This is used as a statistical representative of the distribution of this type of trajectory;

[0036] The current predicted crop growth trajectory GT pre With mean trajectory GT Trl Time alignment is performed, and the difference vector is calculated step-by-step. The mean absolute error over the entire time series is then calculated as the total deviation magnitude index, Las. GT The deviation from the trend direction index Dir is obtained based on the monotonicity of the difference vector in the time series dimension. GT ;

[0037] S34 constructs a time-step heatmap based on the difference sequence, and outputs a natural language interpretation report of the most significant trend by combining the meaning of features.

[0038] Preferably, the diffusion model in S31 adopts a Unet structure, including an Encoder and a Decoder part;

[0039] The network structure of the Encoder part includes: 1D convolution with 3 kernels and stride of 1 + ReLU + layer normalization; 1D convolution with 3 kernels and stride of 2 + ReLU + Dropout; trajectory attention module, adding self-attention mechanism for position embedding and temporal embedding; stacked 2-layer Transformer encoders, used to capture long-distance temporal dependencies and stage patterns.

[0040] The Decoder part adopts the same network structure as the Encoder, using skip connections and upsampling operations to recover high-resolution trajectory representation and output the fully reconstructed trajectory.

[0041] Preferably, in step S4, a set of standard control variables SC for crop regulation is first defined, and then the current growth state type Cls and the total deviation index Las are calculated based on the output of step S3. GT Dir indicator GT A variable correlation scoring mechanism was constructed. This scoring mechanism analyzes the response relationship between each control variable and growth characteristics under the current growth state type Cls, and combines it with the deviation index Las. GT and Dir GT The significance of the influence between the variables and the corresponding variables is ranked, and the set of K key control variables KC with the greatest regulatory potential for the current abnormal state is selected.

[0042] Based on the determined set of key control variables KC, and combined with the user-defined target state, a multi-objective optimization function Loss is constructed. mof (KC):

[0043] If the objective is steady-state development, the objective function Loss mof (KC) is defined as minimizing the average absolute deviation between the predicted trajectory and the mean trajectory, ensuring that the crop state is as close as possible to the evolution process of the target trajectory.

[0044] If the objective is to maximize output, then an empirical correlation model between key growth characteristics and output is used to measure the potential output level, and the objective function Loss is applied. mof (KC) is defined as maximizing output level;

[0045] If the objective is risk aversion, then the total deviation index Las is used. GTand the Dir deviation indicator GT objective function Loss mof (KC) is defined as maximizing the future deviation magnitude and the decrease in the deviation trend. That is, by adjusting the key variable KC, the deviation of the predicted trajectory from the corresponding standard trajectory in the future period is significantly reduced.

[0046] Preferably, based on the selected set of key control variables KC and the set optimization objective function Loss mof (KC) employs a Bayesian optimization algorithm to adaptively adjust the range of control variables, simulates and generates multiple control schemes, and uses the objective function Loss mof (KC) evaluates the merits of each option in improving crop condition; finally, it outputs the optimal combination of control variables, KC. best It generates specific crop management strategy recommendations, including: recommended value ranges for each key control variable; recommended adjustment ranges; and recommended adjustment periods.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] (1) Multi-source heterogeneous remote sensing fusion growth feature modeling structure: A multi-channel growth feature modeling framework that integrates airborne multispectral images and ground morphology observation data was constructed. By combining temporal embedding, convolutional network and lightweight Transformer structure, high-precision and robust prediction of key crop features such as canopy height, leaf area index and SPAD value was achieved, breaking through the problem of limited accuracy of traditional single-source data temporal modeling.

[0049] (2) Intelligent analysis mechanism for growth state deviation based on multi-trajectory diffusion learning: This invention establishes trajectory distribution models for different types of growth trajectories (such as high yield, premature aging, drought stress, etc.) and introduces a trajectory matching attention mechanism to automatically identify the deviation type and occurrence time of the current trajectory without relying on human experience, which significantly improves the accuracy and interpretability of crop growth anomaly identification.

[0050] (3) Design of interpretable strategy recommendation module for variable regulation: Based on the prediction bias and trajectory attribution analysis results, this invention proposes a variable regulation scheme generation method based on Bayesian optimization, which outputs a structured regulation form containing operable suggestions such as fertilization timing and irrigation frequency, thereby enhancing the model's ability to be implemented and its practicality in the context of precision agricultural management. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the overall implementation process of the present invention.

[0052] Figure 2This is a diagram illustrating the overall framework of a multi-source image-driven crop growth feature recognition model.

[0053] Figure 3 Flowchart for the intelligent analysis and deviation interpretation module of crop growth status.

[0054] Figure 4 This is a comparison chart of the predicted trajectory of canopy height (cm) in the examples.

[0055] Figure 5 This is a comparison chart of leaf area index prediction trajectories in the examples.

[0056] Figure 6 This is a comparison chart of the predicted trajectories of SPAD values ​​in the examples.

[0057] Figure 7 This is a graph showing the matching score and deviation trend analysis between the current trajectory and the standard trajectory in the example. Detailed Implementation

[0058] This invention proposes an intelligent analysis method for crop growth status based on multi-source remote sensing imagery. First, multi-source remote sensing time-series images covering the entire growth cycle are collected to construct a unified multimodal growth feature dataset. Second, a multi-channel time-series modeling structure is designed, relying on spatial spectral and ground morphology information to collaboratively extract the dynamic evolution trajectories of core features such as canopy height, leaf area index, and SPAD value. Then, a multi-type trajectory distribution learning model is used to identify the current state deviation type, and a diffusion matching mechanism is combined for problem tracing and interpretation analysis. Finally, based on the trajectory deviation results and management objectives, variable adjustment strategies are generated to optimize and recommend management measures such as irrigation frequency and fertilization intensity. This invention can be widely applied to scenarios such as agricultural remote sensing monitoring, intelligent agricultural machinery scheduling, and precision planting guidance, and has high interpretability and practical value.

[0059] The specific implementation process of the present invention will be further described below with reference to specific embodiments.

[0060] The overall process of this embodiment is as follows: Figure 1 As shown:

[0061] S1, Multi-source temporal remote sensing image acquisition and dataset construction: By deploying a multi-source remote sensing acquisition system that coordinates air and ground, multi-channel remote sensing image data with temporal continuity and spatial differences are acquired. Combined with on-site manual annotation, a crop image-growth feature paired dataset and a set of crop growth feature trajectories are constructed to provide a data foundation for subsequent crop growth feature prediction modeling and trajectory status analysis.

[0062] S2, Design of a multi-source image-driven crop growth feature recognition model: Taking the remote sensing image sequence acquired in S1 as input, three channels are designed: multi-source image feature interaction, temporal evolution modeling, and structural feature decoupling prediction. These channels sequentially realize multi-source feature fusion, temporal evolution law extraction, and multi-task prediction of target features. This model can automatically identify the structural growth characteristics of crops.

[0063] S3, Design of Intelligent Analysis and Deviation Explanation Module for Crop Growth Status Based on Distribution Learning: Taking the time-series prediction results of crop growth characteristics output from stage S2 as input, combined with the growth characteristic trajectory library constructed in S1, the module models the distribution characteristics of various typical growth trajectories based on a multi-class trajectory diffusion model, and then performs matching analysis on the current crop trajectory to determine its growth status type; further, it quantifies the degree and trend of deviation from the target trajectory, and outputs the corresponding matching score, deviation index and natural language explanation report.

[0064] S4, Crop Management Strategy Optimization Suggestion Generation Module: Based on the analysis results of S3 and the user-defined goals, this module generates precise control suggestions for the current crop status. First, it selects key control variables, including management factors such as irrigation, fertilization, and light. Then, it constructs an optimization function in conjunction with the user's goals. Finally, it uses a Bayesian optimization algorithm to search for the optimal variable configuration, forming an actionable management strategy suggestion.

[0065] I. Acquisition of Multi-Source Time-Series Crop Remote Sensing Data and Construction of Crop Growth Dataset

[0066] To cover the crop growth cycle and capture regional differences and temporal dynamics, this invention collects multi-source time-series crop remote sensing data and combines it with manual annotation to construct a crop growth dataset, including a crop image-growth feature dataset and a crop growth feature trajectory dataset. Specifically, it includes the following steps:

[0067] Multi-source crop image data acquisition: This involves deploying a multi-rotor UAV platform and ground-based image acquisition equipment to obtain multi-source remote sensing image data from both aerial and ground-based observation sources. The process includes the following steps:

[0068] (1) The UAV platform is equipped with multispectral imaging equipment to conduct high spatiotemporal resolution multichannel aerial photography of the target farmland area, and the flight altitude is precisely planned. Finally, multispectral crop image data Uac=[ac1,ac2,...,aC ] is obtained from multiple angles, including red, green, blue, red edge, and near-infrared band information. N ], where ac n This represents the multispectral crop image data of the UAV under the nth shooting angle, and there are N shooting angles in total, i.e., n∈[1,N];

[0069] (2) The ground platform uses M fixed camera devices to acquire auxiliary images of visible light, Fic = [ic1, ic2, ..., ic2]. M [, used for close-range observation of crop morphology and growth status; among which ic] m Let m represent the auxiliary image acquired by the m-th fixed device, where m∈[1,N];

[0070] (3) Combine the data captured by the UAV with the data collected by the ground platform to form multi-source crop image data Mc=[Uac,Fic], and use it as the collection result of multi-source crop image data.

[0071] Remote sensing image sequence data acquisition: Based on the acquisition method of multi-source crop image data Mc, T uniform multi-source crop image data acquisitions are performed within a certain growth cycle 1→T of the crop, resulting in remote sensing image sequence data SMc=[Mc1, Mc2, ..., Mc T ], among which, Mc t Let represent the multi-source crop image data collected at time point t, where t∈[1,T]; thus ensuring dynamic observation of the crop growth process.

[0072] Structured labeling of crop growth characteristics: In the target plots where image data was collected, multiple representative crop sample points were selected, and researchers were organized to obtain the structured growth characteristics (Gcf) of the crops through field measurements. The growth characteristic indicators include canopy height (Ht), leaf area index (Li), SPAD value (Sd) reflecting chlorophyll content, vegetation coverage (Vp), normalized difference vegetation index (Ni), red edge index (Re), and aboveground biomass estimate (Ab), i.e., Gcf = [Ht, Li, Sd, VP, Ni, Re, Ab].

[0073] Structured growth characteristics of the crop are collected at each growth stage, and complete structured labels of crop growth characteristics GT = [Gcf1, Gcf2, ..., Gcf2] are obtained. T ]; where Gcf t Let represent the structured growth characteristics of crops at time t, where t∈[1,T].

[0074] Construction of crop image-growth feature dataset: Using remote sensing image sequence data SMC as input data and complete crop growth feature structured labels GT as output data, a crop image-growth feature dataset is obtained. Based on this method, a total of D datasets are collected to obtain a complete image-growth feature dataset for model training.

[0075] Construction of crop growth characteristic trajectory dataset: The complete crop growth characteristic structured label GT is used as a set of crop growth trajectories. Then, the researchers assign crop growth type labels to the trajectory based on the final yield performance of each plot, the development characteristics during the growth process and the field evaluation results. These labels include five types: high yield, low yield, drought stress, developmental arrest, and premature aging abnormality (more growth types can be subdivided to achieve more refined growth analysis in the future).

[0076] Growing characteristic structured tags Gr were collected from different plots and at different time periods within each plot. Tags were added to growth types, resulting in a large number of growth trajectories and their growth type data. Subsequently, growth trajectories of the same type were grouped into sets of the same growth type, resulting in five crop growth trajectory sets S1, S2, S3, S4, and S5 (corresponding to high yield, low yield, drought stress, developmental arrest, and premature aging abnormality, respectively). Each growth trajectory set contains a large number of growth trajectories that conform to that growth type.

[0077] Ultimately, the five sets of crop growth trajectories S1, S2, S3, S4, and S5 together constitute the crop growth feature trajectory dataset (crop growth feature trajectory library).

[0078] II. Design of a Multi-Source Image-Driven Crop Growth Feature Recognition Model

[0079] To achieve automatic identification of crop growth characteristics based on multi-source remote sensing imagery, this invention proposes a three-channel joint modeling deep learning framework. The model takes the constructed remote sensing image sequence data (SMc) as input and designs a multi-source heterogeneous image feature interaction channel, a time-series remote sensing sequence evolution modeling channel, and a structural feature decoupling prediction channel. This enables efficient extraction and prediction of crop growth structural features from multi-source heterogeneous remote sensing data. The overall model structure is as follows: Figure 2 As shown.

[0080] 1. Design of interactive channels for multi-source heterogeneous image features

[0081] To enhance the synergistic ability of multi-source remote sensing data in spatial structure and morphology representation, this channel takes as input the constructed remote sensing image sequence data (SMc), aiming to jointly mine complementary features of airborne and ground-based images. The channel consists of an airborne multispectral structure analysis branch and a ground-based morphology enhancement observation branch.

[0082] The space-based multispectral structure analysis branch takes multi-angle multispectral images acquired by a UAV as input. First, it obtains high-dimensional spectral response patterns through a spectral attention layer and extracts the spectral structure features F. spc Next, this feature is fed into the image depth sensing layer, which has a dual-branch convolutional structure to extract local texture responses and multi-scale structures respectively. Finally, these are fused to obtain the hollow structure enhancement feature F.arb ;

[0083] The ground morphology enhancement observation branch uses visible light images acquired from a ground platform as input. First, it encodes the geometric weight relationship of different observation angles through an angle attention layer to obtain the viewpoint-sensitive feature F. prs Subsequently, the image depth sensing layer extracts local contours and plant conditions, outputting ground morphology enhancement features F. gdb ;

[0084] The characteristic F of the two-branch output arb and F gdb Joint updates are performed through a gating fusion mechanism to form a fused multi-source complementary feature F. msc ; Integration and complementarity features F msc Subsequently, the system re-enters the spectral attention layer and the angular attention layer for interactive updates; and this three-layer stacked structure of "attention layer-depth perception-gated fusion" is repeated a total of 8 times, forming a deep multi-source interactive structure, and the output is the final heterogeneous perceptual representation feature F of this channel. hpr .

[0085] in:

[0086] The spectral attention layer is used to highlight the responsiveness of key bands in remote sensing images and suppress redundant spectral interference. Its structure includes: first, global average pooling is performed on the input feature map to extract channel response vectors; then, these vectors are input to two multilayer perceptrons to achieve channel compression and expansion, and a spectral attention mechanism based on attention is used to focus on different spectral characteristics, thereby achieving a significant enhancement of the spectral dimension. This mechanism can explicitly model the sensitivity of different bands to crop structure and physiological characteristics, improving the accuracy of structure recognition.

[0087] An angle attention layer is used to model the impact of different observation angles on structures in ground images. Its structure includes: first, encoding the angle information of the shooting tilt angle into a vector through an angle embedding layer; then fusing this vector with image features and inputting it into a multilayer perceptron to extract angle-guided features; next, constructing a direction response map through an angle attention mechanism to complete feature weighting; and finally outputting the angle-enhanced morphological features to improve the structural robustness under multi-view conditions.

[0088] This method is used to extract multi-scale, deep-level structural semantic information while preserving local texture, thereby improving spatial perception capabilities. Its structure includes two parallel pathways:

[0089] (1) Conv5 convolutional blocks + max pooling are used to obtain coarse-scale features with a large receptive field, and the structure-aware features A are output through residual connection-layer normalization-ReLU processing.

[0090] (2) Two Conv3 convolutional blocks + max pooling are used to extract fine-grained texture information. The residual connection-layer normalization-ReLU processing is also used to output texture enhancement feature B.

[0091] Finally, feature A and feature B are concatenated along the channel dimension to form a fused deep structural texture joint feature representation.

[0092] 2. Design of time-series remote sensing sequence evolution modeling channels

[0093] This channel is used to model the evolution of remote sensing features over time. The input is the heterogeneous sensing representation feature F output from the multi-source heterogeneous image feature interaction channel. hpr First, the time stamp information is introduced through a time embedding enhancement layer, and the time-enhanced feature F is output. ten ;

[0094] Subsequently, this feature F ten Short-term dynamic modeling is performed using causal convolutional layers in a TCN (Temporal Convolutional Network) with a kernel size of 3, and the local temporal response feature F is output using the ReLU activation function. lre To enhance long-range dependency modeling capabilities, this feature is further input into a lightweight Transformer encoder for long-term dependency modeling, outputting a temporal context-aware feature F. tsp ;

[0095] The entire "TCN-Lightweight Transformer" structure is repeatedly stacked 5 times to obtain the output temporal evolution feature representation F of the final temporal remote sensing sequence evolution modeling channel. sec .

[0096] 3. Structural Feature Decoupling Prediction Channel Design

[0097] This channel uses time-series remote sensing sequence evolution modeling to represent the time-series evolution characteristics of the channel output, F. sec The input is used to achieve multi-task prediction of the structural growth characteristics of the target crop;

[0098] First, the interdependencies between features are modeled using a multi-head attention mechanism to obtain the decoupled representation F. der Then, the fully connected decoupling layer is input to achieve feature channel separation between different tasks (such as canopy height, LAI prediction, etc.); after that, it is input into two "multilayer perceptron-ReLU activation" structures for nonlinear modeling, and outputs the preliminary predicted features of each task.

[0099] Finally, a Dropout layer is used to implement regularization and improve robustness, followed by a ReLU activation function. These results are then fed into a multi-task prediction head, which outputs predicted values ​​for each crop growth feature category, forming a complete crop growth feature prediction result (GT). pre .

[0100] 4. Model Training and Optimization: Using the constructed crop image-growth feature dataset as the training set, mean squared error (MSE) is used as the loss function, and stochastic gradient descent (SGD) optimizer is selected for training. The training stops when the set threshold is reached, and the finally trained crop growth feature recognition model is obtained. This model can be directly used for crop phenotypic growth feature recognition under multi-source remote sensing images.

[0101] III. Intelligent Analysis and Deviation Interpretation Module for Crop Growth Status Based on Distribution Learning

[0102] To achieve intelligent analysis and assessment of crop growth status, this module uses multi-source imagery to drive the crop growth feature recognition model, which outputs time-series prediction results of crop growth features (GT). pre As input, combined with the constructed standard trajectory sub-library, the current growth status of crops in a plot is determined through distribution modeling and trajectory comparison analysis, and the deviation trend, difference indicators, and visual interpretation results are output. The complete processing flow of this module is as follows: Figure 3 As shown.

[0103] 1. Multi-class standard trajectory distribution modeling based on diffusion model: Based on the crop growth characteristic trajectory database, the samples are divided into 5 crop growth trajectory subsets S1, S2, S3, S4, and S5 (corresponding to high yield, low yield, drought stress, developmental arrest, and premature aging abnormality, respectively); for each type of trajectory subset, a diffusion model is used for training to learn the characteristic evolution and distribution law of that type of growth trajectory;

[0104] Specifically, the diffusion model of this invention adopts a Unet structure (including an Encoder and a Decoder). The network structure of the Encoder is as follows: (1) 1D convolution with 3 kernels and a stride of 1 + ReLU + layer normalization; (2) 1D convolution with 3 kernels and a stride of 2 + ReLU + Dropout (0.2); (3) Trajectory attention module: adding self-attention mechanisms for position embedding and temporal embedding; (4) stacking 2 layers of Transformer encoders to capture long-distance temporal dependencies and stage patterns. The Decoder adopts the network structure corresponding to the Encoder, and uses skip connections and upsampling operations to recover high-resolution trajectory representation and output the fully reconstructed trajectory.

[0105] To characterize the typical evolutionary distribution features of different types of crop growth trajectories, this invention proposes a multi-model parallel trajectory diffusion modeling strategy, which constructs an independent diffusion modeling sub-network for each trajectory type. Therefore, based on five crop growth trajectory sets S1, S2, S3, S4, and S5 (corresponding to high yield, low yield, drought stress, developmental arrest, and premature aging abnormalities, respectively), the diffusion models are trained unsupervised, allowing each model to autonomously learn the distributions of various potential trajectories. Ultimately, five independently trained trajectory diffusion models are obtained, each representing a potential distribution generator for five standard trajectory types, for subsequent matching and evaluation.

[0106] 2. Matching analysis of current trajectory and standard trajectory distribution: This step is used to determine which standard type the growth trajectory of the current plot belongs to, specifically including:

[0107] (1) The crop growth characteristics GT predicted by S2 pre As input to five trajectory diffusion models, five trajectory reconstruction results GT1, GT2, ..., GT5 are generated;

[0108] (2) Calculate and predict the original crop growth trajectory GT pre The mean squared error of the five trajectory reconstruction results GT1, GT2, ..., GT5 is taken as the explicit error. The cosine similarity between the two is calculated as the implicit distance. Finally, the matching score is defined as the implicit distance divided by the explicit distance, and the larger the matching score, the higher the similarity after reconstruction, i.e., the predicted crop growth characteristic GT pre The more closely the trajectory data distribution matches this category;

[0109] (3) Therefore, the five matching scores are calculated as Score1, Score2, ..., Score5; the growth state type Cls of the current plot is determined based on the maximum matching score, and the matching score sequence is retained as a reliability reference.

[0110] 3. Deviation Trend and Characteristic Difference Quantitative Analysis: This provides interpretable deviation indicators to assist in analyzing the sources and trends of anomalies in the current crop status. Specifically, it includes the following steps:

[0111] (1) Based on the best matching category Cls (such as drought stress category) determined by the decision, randomly sample Trl trajectories in its corresponding trajectory diffusion model to generate Trl trajectories. Then, average all sampled trajectories step by step to generate the mean trajectory GT of this category. Trl This is used as a statistical representative of the distribution of this type of trajectory;

[0112] (2) The current predicted crop growth characteristic trajectory GT pre With mean trajectory GT TrlTime alignment is performed, and the difference vector is calculated step-by-step. Furthermore, the total time-series mean absolute error (LAS) is calculated as the total deviation magnitude index. GT The deviation from the trend direction index Dir is obtained based on the monotonicity of the difference vector in the time series dimension. GT (Including trends such as "continuous deterioration" and "sharp deviation in the middle and late stages");

[0113] (3) Generate an interpretation report: Construct a time-step heatmap based on the difference sequence (to show the intensity of the time-feature dimension deviation), and generate a natural language interpretation report Nle based on the feature meaning corresponding to the most obvious trend of the difference in the time-step heatmap. GT For example, "during the grain-filling period, the chlorophyll index was significantly lower than that of the high-yield trajectory";

[0114] Therefore, the final intelligent crop growth analysis results include: the current plot's growth status type Cls, a 5-dimensional matching score vector Score1, Score2, ..., Score5, and the total deviation index Las. GT Dir indicator, deviation from trend direction GT And the Natural Language Interpretation Report Nle GT .

[0115] 4. Intelligent Analysis Result Display: The system displays the determined growth state type (Cls) and its matching score calculation results, such as "drought stress, matching score 3.42," to help users understand the basis for state classification; it also combines the total deviation index (Las). GT Dir indicator, which deviates from the trend direction GT and report Nle GT Generate concise natural language descriptions, such as: "The current crop growth status of the plot is determined to be drought-stressed. The main difference occurs in the post-heading stage, with significant deviations in chlorophyll index and canopy structure indicators. It is recommended to focus on water supply in the mid-to-late stages."

[0116] IV. Crop Management Strategy Optimization Suggestion Generation Module

[0117] To achieve precise control scheme generation based on the current crop growth status and user objectives, this invention constructs a framework for generating crop management strategy optimization suggestions, based on intelligent analysis results and combined with user-defined target crop growth states (such as steady-state development, maximizing yield, and risk avoidance). The overall process includes the following steps:

[0118] Selection of control variables: First, define a set of standard control variables SC for crop regulation, including adjustable management factors such as irrigation frequency, irrigation volume, nitrogen, phosphorus and potassium fertilizer ratio and application timing, frequency and dosage of growth regulators, light intensity regulation method, and greenhouse ventilation parameters.

[0119] Subsequently, based on the output of the current growth state type Cls and the total deviation index Las GT Dir indicator GT A variable correlation scoring mechanism was constructed. This mechanism analyzes the response relationship between each control variable and growth characteristics under the current growth state type Cls, and combines it with the deviation index Las. GT and Dir GT The significance of the influence between the variables and the corresponding variables is ranked, and the set of K key control variables KC with the greatest regulatory potential for the current abnormal state is selected for subsequent optimization strategy design.

[0120] Growth state type Cls, 5-dimensional matching score vector Score1, Score2, ..., Score5, and total deviation index Las GT Dir indicator, deviation from trend direction GT And the Natural Language Interpretation Report Nle GT .

[0121] Optimization of objective design: Based on the determined set of key control variables KC, and combined with the user-defined objective state (steady-state development, yield maximization, or risk aversion), a multi-objective optimization function Loss is constructed. mof (KC), and refine the computational logic of this optimization function:

[0122] (1) If the objective is steady-state development, the objective function Loss mof (KC) is defined as minimizing the average absolute deviation between the predicted trajectory and the mean trajectory, ensuring that the crop state is as close as possible to the evolution process of the target trajectory.

[0123] (2) If the objective is to maximize yield, then the potential output level is measured by using an empirical correlation model between key growth characteristics (including leaf area index, greenness index, etc.) and yield, and the objective function Loss is applied. mof (KC) is defined as maximizing output level;

[0124] (3) If the objective is risk aversion, then based on the total deviation magnitude indicator Las GT and the Dir deviation indicator GT objective function Loss mof (KC) is defined as maximizing the magnitude of future deviation and the decrease in the deviation trend. That is, by adjusting the key variable KC, the deviation of the predicted trajectory from the corresponding standard trajectory in the future period is significantly reduced, thereby achieving risk mitigation.

[0125] Optimization suggestion generation: based on the selected set of key control variables KC and the set optimization objective function Loss. mof(KC) employs a Bayesian optimization algorithm to adaptively adjust the range of control variables, simulates and generates multiple control schemes, and uses the objective function Loss mof (KC) evaluates the merits of each option in improving crop condition; finally, it outputs the optimal combination of control variables, KC. best It generates specific crop management strategy recommendations, including: (1) recommended value ranges for each key control variable; (2) recommended adjustment ranges; and (3) recommended adjustment periods; thereby providing users with operational and precise management recommendations.

[0126] V. Experimental Results and Analysis

[0127] To verify the effectiveness of the intelligent analysis method for crop growth status based on multi-source remote sensing imagery proposed in this invention, this experiment conducted a simulation experiment on maize crops during their growth period from week 1 to week 12. The goal was to evaluate the performance of the proposed multi-channel modeling structure in predicting key growth characteristics, identifying state deviations, and generating variable adjustment suggestions, and to verify its adaptability and interpretability in intelligent crop growth analysis scenarios.

[0128] 1. Display of Growth Trajectory Prediction Results

[0129] Three of the most representative growth indicators in the maize growth cycle—canopy height, leaf area index (LAI), and SPAD value—were selected, and feature evolution prediction models were constructed for each. This experiment selected three typical and observable indicators during maize growth: canopy height, leaf area index (LAI), and SPAD value, and constructed prediction models based on the proposed method. Figures 4 to 6 The comparison between the model output trajectory and the measured reference trajectory is shown, thus reflecting the modeling capability of the constructed time-series modeling channel and structural feature decoupling module.

[0130] Figure 4 (Canopy height trajectory) shows that the model can accurately characterize the height evolution trend of maize plants at different growth stages, especially with the smallest prediction error in weeks 6 to 10 (tasseling to grain filling stage), which verifies the advantages of temporal embedding and lightweight Transformer structure in temporal fitting during continuous growth stages. Figure 5 The (leaf area index trajectory) reveals that the model accurately captures the process of LAI from rapid rise to plateau fluctuation and then to decay, indicating that the designed decoupled layer structure has the ability to distinguish complex temporal patterns, and is especially suitable for characterizing the evolution of canopy coverage. Figure 6 The SPAD value trajectory showed a trend highly consistent with the measured changes in chlorophyll content, indicating that the collected multi-source remote sensing spectral channels have a good coupling with changes in physiological state, further confirming the effectiveness of the multi-source heterogeneous feature interaction mechanism.

[0131] The results above demonstrate that the proposed method has the ability to model multi-dimensional temporal remote sensing features, providing stable and reliable feature support for subsequent state recognition and policy generation.

[0132] 2. Deviation analysis of growth trajectory

[0133] In the state determination stage, the system uses the LAI trajectory output by the crop growth feature recognition model driven by multi-source images as input, and performs matching analysis in combination with the standard high-yield trajectory. Figure 7 The study shows the changes in deviation scores for both throughout the reproductive period.

[0134] Experiments revealed that during weeks 5 to 6 (jointing to trumpet stage), the Leaf Area Index (LAI) was significantly lower than the standard trajectory, with a deviation score exceeding the set threshold of 0.3, indicating an abnormal growth status at this stage. Combining the trajectory classification and interpretation structure of the intelligent crop growth status analysis and deviation interpretation module, the system determined that the deviation was mainly due to "insufficient irrigation in the early stages coupled with delayed nitrogen supply," thus pinpointing the source of the error and ultimately outputting a textual explanation: "The current crop deviation in this plot is characterized by a leaf area index consistently lower than the high-yield standard trajectory and insufficient canopy expansion. Emphasis should be placed on water and fertilizer management in the early and mid-stages, with supplemental irrigation and optimized nitrogen application strategies."

[0135] This experiment verified the effectiveness of the multi-diffusion model parallel modeling mechanism in the intelligent analysis and deviation interpretation module of crop growth status in growth trajectory classification and deviation interpretation, and enhanced the system's traceability and scenario adaptability.

[0136] 3. Generation of variable adjustment schemes

[0137] Based on the deviation types identified in the previous experiment and the user-defined developmental goals (steady-state growth), a Bayesian optimization strategy is used to generate variable adjustment schemes. The output includes variable names, adjustment magnitudes, and corresponding execution cycles, and a standardized form is generated for farmers to directly deploy and implement.

[0138] (1) Increase the irrigation amount by 12.5% ​​and adjust the irrigation frequency from once every 10 days to once every 7 days;

[0139] (2) Optimize the nitrogen fertilizer application structure, changing the original plan of 3 applications to 2 concentrated fertilizations in the early stage and 1 supplementary application in the later stage;

[0140] The experimental results show that the crop management strategy optimization suggestion generation module can generate management strategy suggestions that are interpretable and operable based on the state deviation type and target constraints, thereby improving the system's practicality and intelligent decision-making level.

[0141] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0142] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for intelligent analysis of crop growth status based on multi-source remote sensing imagery, characterized in that, Includes the following steps: S1, by deploying an air-ground collaborative multi-source remote sensing acquisition system, acquires multi-channel remote sensing image sequence data of crop growth with temporal continuity and spatial differences; S2 inputs the remote sensing image sequence data obtained in S1 into the trained multi-source image-driven crop growth feature recognition model, and outputs the time-series prediction results of crop growth features. The model includes a multi-source image feature interaction channel, a time-series evolution modeling channel, and a structural feature decoupling prediction channel, which sequentially realizes multi-source feature fusion, extraction of time evolution laws, and multi-task prediction of target features. The multi-source image feature interaction channel consists of a space-based multispectral structure analysis branch and a ground morphology enhancement observation branch; the space-based structure enhancement features output by the two branches... and foundation morphology enhancement features Joint updates are performed through a gating fusion mechanism, and the final output is a heterogeneous perceptual representation feature. ; The temporal remote sensing sequence evolution modeling channel is used to model the evolution of remote sensing features over time, with heterogeneous sensing representation features as input. The final output is a temporal evolution feature representation. ; The structural feature decoupling prediction channel is represented by temporal evolution features. The input is used to achieve multi-task prediction of the structural growth characteristics of the target crop; Finally, the predicted values ​​for each crop growth characteristic are output, forming a complete crop growth characteristic prediction result. ; S3, based on the intelligent analysis and deviation interpretation module for crop growth status, takes the time-series prediction results of crop growth characteristics output by S2 as input, combines the constructed crop growth characteristic trajectory library, models the distribution characteristics of various typical growth trajectories based on a multi-class trajectory diffusion model, and then performs matching analysis on the current crop trajectory to determine its current growth status type and quantify its deviation degree and trend from the target trajectory; outputs the corresponding crop growth type matching score and deviation index; simultaneously, it outputs a natural language interpretation report of the most significant trend based on the feature meaning; the specific processing steps of the intelligent analysis and deviation interpretation module for crop growth status include: S31, Multi-class standard trajectory distribution modeling based on diffusion model: The crop growth feature trajectory database divides the samples into 5 subsets of crop growth trajectories. These correspond to high yield, low yield, drought stress, developmental arrest, and premature aging abnormality, respectively. For each subset of trajectories, a diffusion model is used for training to learn the characteristic evolution and distribution patterns of that type of growth trajectory. S32, Matching analysis between the current trajectory and the standard trajectory distribution: The predicted crop growth characteristics As input to five trajectory diffusion models, five trajectory reconstruction results are generated. ; Calculate and predict the original crop growth trajectory Comparison with 5 trajectory reconstruction results The mean squared error is used as the explicit error, and the cosine similarity between the two is calculated as the implicit distance. The final matching score is defined as the implicit distance divided by the explicit distance, with a higher matching score indicating a higher similarity after reconstruction, i.e., a higher predicted crop growth characteristics. The more closely the trajectory data distribution matches this category; The calculation results of the 5 matching scores are as follows: The growth status type of the current plot is determined based on the maximum matching score. And retain the matching score sequence as a credibility reference; S33, Quantitative Analysis of Deviation Trends and Characteristic Differences: The best matching category determined by the base Random sampling is generated in its corresponding trajectory diffusion model. The sampled trajectories are then averaged step-by-step to generate the mean trajectory for that category. This is used as a statistical representative of the distribution of this type of trajectory; The current predicted crop growth trajectory with mean locus Time alignment is performed, and the difference vector is calculated step by step. The mean absolute error over the entire time series is then used as an indicator of the total deviation. The deviation from the trend direction index is obtained by judging the monotonicity of the difference vector in the time series dimension. ; S34 constructs a time-step heatmap based on the difference sequence, and outputs a natural language interpretation report of the most significant trend by combining the meaning of features.

2. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 1, characterized in that, It also includes S4: selecting key control variables based on crop growth type and deviation indicators based on matching scores, and constructing an optimization function in combination with user objectives; finally, using a Bayesian optimization algorithm to search for the optimal variable configuration and form management strategy recommendations.

3. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 1, characterized in that: Methods for constructing datasets for training multi-source image-driven crop growth feature recognition models include: During a certain growth cycle of the crop Inside, a total of Sub-uniform multi-source crop image data acquisition yields remote sensing image sequence data. ,in, Indicates the first Multi-source crop image data collected at each time point, and S1-3 Structured annotation of crop growth characteristics: Within the target plots where image data was collected, several representative crop sample points were selected, and researchers were organized to obtain the structured growth characteristics of the crops through field measurements. Growth characteristics include canopy height. Leaf area index SPAD value, which reflects chlorophyll content Vegetation coverage Normalized Difference Vegetation Index Red border index Aboveground biomass estimation ,Right now ; Structural growth characteristics of the crop are collected at each growth stage, and complete structured labels of crop growth characteristics are obtained. ;in express The structured growth characteristics of crops at specific points in time, and ; Remote sensing image sequence data As input data, complete crop growth characteristic structured labels As output data.

4. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 1, characterized in that: The space-based multispectral structure analysis branch takes multi-angle multispectral images acquired by a UAV as input, first obtains high-dimensional spectral response patterns through a spectral attention layer, and then extracts spectral structure features. Next, this feature is fed into the image depth sensing layer, which has a dual-branch convolutional structure to extract local texture responses and multi-scale structures respectively, and finally fused to obtain the hollow structure enhancement feature. ; The ground morphology enhancement observation branch takes visible light images acquired by a ground platform as input. First, it encodes the geometric weight relationship of different observation angles through an angle attention layer to obtain viewpoint sensitive features. Subsequently, the image depth sensing layer extracts local contours and plant conditions, outputting enhanced ground morphology features. ; Characteristics of two-branch output and Joint updates are performed through a gating fusion mechanism to form multi-source complementary features after fusion. ; Integration and complementarity characteristics Subsequently, the system re-enters the spectral attention layer and the angle attention layer for interactive updates; the three-layer stacked structure of "attention layer-depth perception-gated fusion" is repeated a total of 8 times to form a deep multi-source interactive structure, and the output is the final heterogeneous perceptual representation feature of this channel. .

5. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 4, characterized in that: The temporal remote sensing sequence evolution modeling channel is used to model the evolution of remote sensing features over time, with heterogeneous sensing representation features as input. First, the collected timestamp information is introduced through a time embedding enhancement layer, and then the time-enhanced features are output. ; Subsequently, this feature Short-term dynamic modeling is performed using a TCN causal convolutional layer with a kernel size of 3, and the ReLU activation function is used to output local temporal response features. ,feature Input a lightweight Transformer encoder for long-term dependency modeling and output temporal context-aware features. ; The entire "TCN-Lightweight Transformer" structure is stacked repeatedly for 5 layers to obtain the output temporal evolution feature representation of the final temporal remote sensing sequence evolution modeling channel. .

6. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 5, characterized in that: The structural feature decoupling prediction channel is represented by temporal evolution features. The input is used to achieve multi-task prediction of the structural growth characteristics of the target crop; First, the interdependencies between features are modeled using a multi-head attention mechanism to obtain a decoupled representation. Then, the fully connected decoupling layer is input to achieve feature channel separation between different tasks. Then, the two "multilayer perceptron-ReLU activation" structures are used for nonlinear modeling, and the preliminary predicted features for each task are output. Finally, regularization is achieved through a Dropout layer, followed by a ReLU activation function, and finally fed into a multi-task prediction head to output predicted values ​​for each crop growth feature, forming a complete crop growth feature prediction result. .

7. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 1, characterized in that: The diffusion model in S31 adopts a Unet structure, including an Encoder and a Decoder. The network structure of the Encoder part includes: 1D convolution with 3 kernels and stride of 1 + ReLU + layer normalization; A 1D convolution with 3 kernels and a stride of 2, plus ReLU and Dropout; a trajectory attention module, adding self-attention mechanisms for positional and temporal embedding; and a stacked 2-layer Transformer encoder to capture long-distance temporal dependencies and stage patterns. The Decoder part adopts the same network structure as the Encoder, using skip connections and upsampling operations to recover high-resolution trajectory representation and output the fully reconstructed trajectory.

8. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 2, characterized in that: In S4, a complete set of standard control variables for crop regulation is first defined. Then, based on the current growth state type output by S3... Total Deviation Index Deviation from trend indicators A variable correlation scoring mechanism is constructed; this scoring mechanism analyzes the correlation between variables in the current growth state type. Below, the response relationship between each control variable and growth characteristic is analyzed, combined with deviation indicators. and Ranking the significance of their effects on corresponding variables, we can screen out those with the greatest potential for regulating the current abnormal state. A set of key control variables ; Determining the set of key control variables Based on this, and combined with the user-defined target state, a multi-objective optimization function is constructed. : If the objective is steady-state development, the objective function is... Defined as minimizing the average absolute deviation between the predicted trajectory and the mean trajectory, ensuring that the crop state is as close as possible to the evolution process of the target trajectory; If the objective is to maximize output, then an empirical correlation model between key growth characteristics and output is used to measure the potential output level, and the objective function is... Defined as maximizing output level; If the objective is risk aversion, then the total deviation metric will be used. and deviation from trend indicators objective function Defined as maximizing the future deviation magnitude and the decrease in the deviation trend, i.e., by adjusting key variables. This significantly reduces the deviation of the predicted trajectory from the corresponding standard trajectory in future periods.

9. The intelligent analysis method for crop growth status based on multi-source remote sensing imagery as described in claim 8, characterized in that: Based on the set of key control variables obtained through screening and the set optimization objective function A Bayesian optimization algorithm is used to adaptively adjust the range of control variables, generating multiple control schemes through simulation, and then applying the objective function. The study evaluates the effectiveness of each option in improving crop condition; ultimately, it outputs the optimal combination of control variables. It generates specific crop management strategy recommendations, including: recommended value ranges for each key control variable; recommended adjustment ranges; and recommended adjustment periods.

Images