A crop yield prediction method and system based on a time-series feature algorithm model

Abstract

The application relates to the technical field of intelligent agriculture, and particularly discloses a crop yield prediction method and system based on a time sequence feature algorithm model, which comprises the following steps: processing historical time sequence data by using a cross-region pre-trained growth stage encoder to generate a benchmark stage encoding vector representing a standard phenological stage; comparing the target year data after encoding with the benchmark to quantitatively calculate the two-dimensional differences of the phenological time sequence progress and the physiological growth intensity, and generating a cooperative offset vector; performing yield prediction on a corresponding expert prediction sub-model by a hierarchical decision maker according to the vector; and constructing a structured feedback tuple by using real yield and prediction process information to cooperatively fine-tune the system and realize self-evolution. The system comprises a benchmark encoding construction module, an offset quantization module, a dynamic prediction module and a self-adaptive optimization module. The application realizes the interpretability, strong generalization and continuous self-optimization capability of yield prediction.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, specifically to a method and system for predicting crop yield based on a time-series feature algorithm model. Background Technology

[0002] Crop yield forecasting is a core component of precision agriculture, directly impacting food security strategies, agricultural production planning, and market stability. With the development of remote sensing technology, the Internet of Things, and artificial intelligence, yield forecasting models driven by time-series data have gradually replaced traditional experience-based statistical methods, becoming the mainstream in current research and application. However, existing technological solutions still face a series of deep-seated and interconnected technical bottlenecks in their progress towards practical application and intelligentization. These bottlenecks have not been systematically resolved in any single solution, severely restricting the reliability, universality, and sustainability of forecasting technologies.

[0003] First, at the level of model mechanism, there are significant semantic gaps and a lack of interpretability. Current mainstream methods, whether based on recurrent neural networks, long short-term memory networks, or Transformer architectures, focus on automatically extracting abstract statistical features or hidden states from the raw time-series curves of environmental factors and vegetation indices. Although these features have strong mathematical fitting capabilities, they are black-box vectors detached from specific agronomic biological processes. The model cannot correlate its learned high-dimensional feature peaks with agronomic phenological stages that are sensitive to water and fertilizer and have a decisive impact on final yield, such as the tillering peak and the critical grain-filling stage. This results in predictions lacking agronomical logic support. When predictions are biased, agronomists and growers cannot diagnose the root cause of the problem at a specific growth stage, thus greatly weakening the practical value of the model results and user trust.

[0004] Secondly, at the model performance level, there are severe challenges in domain adaptation and insufficient generalization ability. Existing models are typically trained and optimized on a relatively closed dataset, and their success heavily relies on the assumption that the training and prediction data follow the same distribution. However, agricultural production has extremely strong spatiotemporal heterogeneity. When the model is directly applied to a new ecological zone or a year of abnormal climate, the data distribution changes due to differences in meteorological patterns, soil conditions, or management practices, leading to a significant decline in model performance. For example, a model trained in a humid region may not be able to correctly interpret the unique vegetation index response curve under water stress when applied to a semi-arid region. This lack of adaptability to unknown patterns means that each new application scenario almost requires collecting data from scratch and retraining the model, resulting in high deployment costs and low promotion efficiency.

[0005] Finally, at the system lifecycle level, existing technologies are generally static open-loop systems, lacking self-evolution capabilities. Once current predictive models are trained and deployed, their parameters are fixed, becoming passively executed tools. Agricultural systems are dynamic and evolving complex systems; with variety iterations, cropping system adjustments, climate change trends, and the emergence of new stresses, the historical patterns upon which the models rely gradually become outdated. To maintain predictive accuracy, services must be periodically interrupted to manually collect the latest data, label it, and restart the time-consuming full model retraining. This process is not only costly to maintain but also has significant time lags. This fundamental contradiction between static characteristics and the dynamic nature of the agricultural environment is a deep-seated obstacle preventing existing yield prediction technologies from achieving long-term, stable, and low-cost services.

[0006] In summary, existing technologies have failed to effectively integrate prior agronomic knowledge to ensure interpretability, have failed to design robust generalization mechanisms that adapt to changes in the spatiotemporal domain, and have failed to construct sustainable learning systems capable of coping with dynamic environmental evolution. Therefore, there is an urgent need in this field for a novel intelligent yield prediction scheme that can integrate data features—agronomic semantics—achieve robust cross-domain prediction, and possess closed-loop self-evolutionary capabilities, in order to overcome current bottlenecks. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting crop yield based on a time-series feature algorithm model, in order to solve three major defects in the existing technology: the disconnect between model features and agronomic knowledge leading to fuzzy diagnosis, the insufficient generalization ability of a single model to complex climates, and the static and rigid nature of the system making it difficult to adapt to dynamic environments.

[0008] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0009] A method for predicting crop yield based on a time-series feature algorithm model includes the following steps:

[0010] S1. Process historical time-series data through the growth stage encoder to generate the baseline stage encoding vector;

[0011] S2. The target time-series data is processed using the growth stage encoder to obtain the target stage encoding vector, and a cooperative offset vector is generated by comparing it with the reference stage encoding vector.

[0012] S3. Input the collaborative offset vector into the hierarchical decision maker to select the expert prediction sub-model, and then output the final output prediction value;

[0013] S4. Construct a structured feedback tuple based on the actual output value and the collaborative offset vector, and perform collaborative fine-tuning on the growth stage encoder, hierarchical decision-maker, and expert prediction sub-model.

[0014] As a preferred embodiment of the present invention, S1 specifically includes:

[0015] S11. Obtain the initial time-series dataset of the target area for the entire growth cycle in multiple consecutive historical years from the remote sensing data platform and the meteorological data platform; based on the preset climate stability index and historical yield data, select the typical historical years that meet the requirements from the historical years, and extract the crop growth time-series dataset of the corresponding years.

[0016] S12. Input the crop growth time series dataset into a pre-trained growth stage encoder; the growth stage encoder processes the dataset, extracts the time series features of each typical year, and clusters and aggregates the same phenological stage features of all typical years, outputting a set of baseline stage encoding vectors for each standard phenological stage.

[0017] As a preferred embodiment of the present invention, in S12, the pre-training of the encoder during the growth stage includes:

[0018] Using unlabeled, large-scale, multi-regional crop time-series data, an encoder base is trained through self-supervised learning tasks such as mask reconstruction or time-series comparison prediction, enabling it to learn general time-series patterns of crop growth.

[0019] On the encoder base, multi-region data with phenological stage labels are used to perform supervised fine-tuning through contrastive learning, so that general features are aligned with specific phenological semantics.

[0020] As a preferred embodiment of the present invention, S2 specifically includes:

[0021] S21. Simultaneously acquire the target original time series data from the remote sensing data platform and the meteorological data platform, which are from the sowing time to the prediction time point of the target prediction year and have the same source as the crop growth time series data structure, and obtain the target time series data after the same preprocessing.

[0022] S22. Input the target time series data into the pre-trained growth stage encoder and output the target stage encoding vector corresponding to the current growth process;

[0023] S23. Match and compare the target stage encoding vector with the baseline stage encoding vector to calculate the phenological timeline offset and physiological growth intensity offset;

[0024] S24. The calculated phenological timeline offset and physiological growth intensity offset are normalized respectively, and then concatenated into a two-dimensional vector in a predetermined order as the final collaborative offset vector.

[0025] As a preferred embodiment of the present invention, S23 specifically includes:

[0026] S231. A dynamic time warping algorithm is used to non-linearly align the temporal sequence of the target stage encoding vector with the temporal sequence of the reference stage encoding vector to eliminate the overall scaling and translation differences between the two, and to determine the matching relationship between each feature point in the target stage encoding vector and the corresponding phenological stage in the reference stage encoding vector.

[0027] S232. After completing the time alignment, calculate the absolute time difference between the occurrence time of the predefined key phenological feature points in the target stage coding vector and the standard time of the corresponding feature points in the baseline stage coding vector, as the phenological time progression offset, where key phenological feature points include growth rate peak points and cumulative biomass inflection points.

[0028] S233. Within the aligned windows of the same phenological stage, extract the intra-stage feature statistics of the target stage encoding vector and the baseline stage encoding vector, such as the mean vector and slope vector; calculate the cosine distance or regularized Euclidean distance between them, which is the physiological growth intensity offset.

[0029] As a preferred embodiment of the present invention, S3 specifically includes:

[0030] S31. Input the collaborative offset vector into the hierarchical decision maker, determine the current scenario as one of several predefined typical offset modes, and output the corresponding mode identifier and model selection weight;

[0031] S32. The hierarchical decision-maker dynamically schedules from a pre-trained expert prediction sub-model library according to the pattern identifier and model selection weight; each sub-model in the expert prediction sub-model library is specifically trained for a particular typical offset pattern, such as high temperature and drought stress type, low temperature and low light delay type, and favorable weather type.

[0032] S33. The prediction results of the scheduled expert prediction sub-models are weighted and integrated according to the weights selected by the model to output the final output prediction value; wherein, the prediction of each sub-model is based on the target time series data and the collaborative offset vector.

[0033] As a preferred embodiment of the present invention, S31 specifically includes:

[0034] S311. The collaborative offset vector is concatenated with its historical evolution trend features within a preset time window to form an enhanced decision feature vector;

[0035] S312. Input the enhanced decision feature vector into the preset rule engine or lightweight classification network in the hierarchical decision maker; by comparing the phenological time progression offset and physiological growth intensity offset in the vector with the pre-stored feature threshold intervals corresponding to different typical offset patterns, identify the most matching offset pattern category.

[0036] S313. Based on the matching results, output the corresponding pattern identifier; at the same time, calculate the model selection weights assigned to one or more expert prediction sub-models according to the matching confidence or the degree of proximity to multiple threshold intervals.

[0037] As a preferred embodiment of the present invention, S32 specifically includes:

[0038] S321. Each expert prediction sub-model in the expert prediction sub-model library is independently trained using all training samples that were historically determined by the hierarchical decision-maker to be the same typical offset pattern. Each sub-model specifically learns the output formation response pattern under its corresponding offset pattern.

[0039] S322. The hierarchical decision maker calls one or more corresponding expert prediction sub-models from the expert prediction sub-model library according to the pattern identifier; when the model selection weight indicates the existence of a dominant pattern, a single dominant model is scheduled; when the weight distribution indicates a mixed pattern, multiple related models are scheduled.

[0040] S323. The model selection weights are assigned to each expert prediction sub-model that is scheduled, as coefficients for subsequent weighted integration.

[0041] As a preferred embodiment of the present invention, S4 specifically includes:

[0042] S41. After obtaining the actual output value of the target prediction year, associate and bind it with the collaborative offset vector, mode identifier and model selection weight to jointly construct a structured feedback tuple;

[0043] S42. Add the structured feedback tuple to the historical training set to perform collaborative fine-tuning of the system using incremental learning:

[0044] S421. Using the cooperative offset vector in the structured feedback tuple as a soft supervision signal, construct the encoder contrast loss; the loss function encourages the encoder in the growth stage to generate codes that are further away in the feature space during periods when the actual growth pattern differs more from the baseline pattern; based on this loss, update the parameters of the encoder in the growth stage through the backpropagation algorithm.

[0045] S422. Based on the actual output value and the final output prediction value in the structured feedback tuple, calculate the prediction loss. At the same time, based on the correlation between the decision result of the pattern identifier and the actual output error, calculate the decision confidence loss. Combine the two with weighted sums to obtain the composite loss. Using the composite loss, optimize the internal classification threshold, network parameters, and parameters of the expert prediction sub-model associated with the structured feedback tuple using the gradient descent algorithm. This enables decision-making and prediction to evolve collaboratively in the feedback.

[0046] A crop yield prediction system based on a time-series feature algorithm model, used to implement a crop yield prediction method based on a time-series feature algorithm model, includes:

[0047] The benchmark encoding construction module is used to obtain crop growth time series datasets of the target region in multiple consecutive historical typical years, and process the time series datasets through a pre-trained, cross-regional general growth stage encoder to generate a set of benchmark stage encoding vectors that represent the core features of standard phenological stages.

[0048] The offset quantization module, connected to the baseline encoding construction module, is used to acquire the target time series data of the full growth cycle of the target prediction year, extract its target stage encoding vector using the growth stage encoder, compare the target stage encoding vector with the baseline stage encoding vector, calculate the difference between them in the two dimensions of phenological time series progress and physiological growth intensity, and fuse them to generate a collaborative offset vector.

[0049] The dynamic prediction module, connected to the offset quantization module, includes a hierarchical decision-maker and an expert prediction sub-model library. The hierarchical decision-maker is used to output a mode identifier and a model selection weight based on the offset mode represented by the collaborative offset vector, and dynamically selects or combines one or more expert prediction sub-models from the expert prediction sub-model library accordingly. The selected expert prediction sub-model is used to output the final output prediction value based on the target time series data and the collaborative offset vector.

[0050] The adaptive optimization module, connected to the baseline encoding construction module, the offset quantization module, and the dynamic prediction module, is used to construct a structured feedback tuple by combining the actual yield value of the target prediction year with the collaborative offset vector, the pattern identifier, and the model selection weights after obtaining the actual yield value. The structured feedback is then used to collaboratively fine-tune the growth stage encoder, the hierarchical decision-maker, and the expert prediction sub-model to achieve iterative system updates.

[0051] Compared with the prior art, the present invention has the following advantages:

[0052] 1. By constructing a transferable agronomic semantic temporal prototype and a two-dimensional offset quantization mechanism, the model features are aligned with agronomic phenological stages, and growth abnormalities are clearly decomposed into quantitative offsets of time process and physiological intensity. This solves the problems of black box features and ambiguous diagnostic results in traditional models, and realizes interpretable and refined growth diagnosis.

[0053] 2. Through a hierarchical decision-making and dynamic scheduling expert consultation mechanism, the system can intelligently call the most suitable special prediction model according to the specific offset pattern, thus achieving targeted solutions. This overcomes the shortcomings of single static models in terms of insufficient generalization ability when dealing with unknown climate patterns or sudden disasters, and improves the prediction generalization ability in complex scenarios.

[0054] 3. By establishing a closed loop of structured feedback and collaborative fine-tuning, each prediction practice is transformed into a learning sample for system iteration, enabling model parameters to evolve dynamically with the environment. This fundamentally solves the long-standing problem of static solidification and high maintenance costs of traditional AI models after deployment, and endows the prediction system with the ability to continuously self-optimize. Attached Figure Description

[0055] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0056] Figure 1 This is a flowchart illustrating the method described in Embodiment 1 of the present invention.

[0057] Figure 2 This is a framework diagram of the system described in Embodiment 2 of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] The concepts involved in this application will first be described with reference to the accompanying drawings. It should be noted that the following descriptions of various concepts are only for the purpose of making the content of this application easier to understand and do not constitute a limitation on the scope of protection of this application; furthermore, the embodiments and features in the embodiments of this application can be combined with each other unless otherwise specified. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0060] Example 1

[0061] like Figure 1 As shown, this invention provides a method and system for predicting crop yield based on a time-series feature algorithm model, comprising the following steps:

[0062] S1. Process historical time-series data through the growth stage encoder to generate the baseline stage encoding vector; specifically including:

[0063] S11. Selection of typical historical years and construction of crop growth time series datasets, specifically:

[0064] S111. Obtain historical remote sensing data of the target area from the remote sensing data platform and meteorological observation data of the same period from the meteorological data platform. Align and merge the two types of data according to the spatiotemporal dimensions to construct an initial time series dataset covering the entire growth cycle of multiple consecutive historical years.

[0065] The initial time series dataset uses fixed time intervals as time steps, such as daily or ten-day intervals. Each time step contains multi-dimensional time series variables, including remote sensing dimensions such as normalized vegetation index, enhanced vegetation index, leaf area index inversion value, and canopy moisture content index, and meteorological dimensions such as precipitation, average temperature, cumulative effective temperature, sunshine hours, soil moisture, and total solar radiation.

[0066] S112. Based on the preset climate stability index, calculate the climate variation coefficient within the entire growth cycle of each historical year. This coefficient is a weighted fusion of the time series standard deviations of precipitation variation coefficient, temperature variation coefficient and comprehensive drought index. The weights are determined according to the climate sensitivity of the target crop.

[0067] Simultaneously, crop yield data for each historical year in the target region are retrieved from the agricultural statistics database, and the yield deviation of each year relative to the multi-year average yield is calculated. This deviation is the standardized residual between the actual yield and the trend yield.

[0068] S113. Set climate stability threshold and yield representativeness threshold. Historical years with climate variability coefficients below the climate stability threshold and absolute yield deviation less than the yield representativeness threshold are judged as having stable climate conditions and representative yields. In this way, typical historical years that meet the requirements are selected.

[0069] Multidimensional time-series variable sequences corresponding to the time ranges of these typical historical years are extracted from the initial time-series dataset to form a crop growth time-series dataset. This crop growth time-series dataset retains the complete time-step structure, with each time step labeled with the corresponding calendar date, serving as input data for the growth stage encoder in S12, and is used to subsequently generate the baseline stage encoding vector.

[0070] S12. Encoder pre-training in the growth phase and encoding vector generation in the baseline phase, specifically:

[0071] S121. The pre-training process of the growth stage encoder includes the following two stages:

[0072] The first stage involves encoder base pre-training: Large-scale, multi-regional crop time-series data covering various climate zones and crop types are collected, without requiring phenological stage labeling. A masking reconstruction task is used to mask random time steps or random variable dimensions in the input time-series data. The encoder base is trained to reconstruct the masked original values ​​based on contextual information, with the optimization objective being to minimize the mean squared error between the reconstructed and true values. Alternatively, a time-series comparison prediction task is used, applying different enhancement transformations to the time series of the same crop growth stage to generate multiple views. Views from the same growth stage are used as positive sample pairs, and views from different growth stages or regions are used as negative sample pairs. The encoder base is trained to maximize the feature similarity of positive sample pairs and minimize the feature similarity of negative sample pairs. Through these self-supervised learning tasks, the encoder base learns the general time-series pattern of crop growth, possesses the ability to extract robust time-series feature representations, and outputs a general time-series feature representation.

[0073] The second stage, supervised fine-tuning: Building upon the encoder base, multi-regional data labeled with phenological stages is used. This data includes phenological stage time points identified manually or by algorithms. A contrastive learning framework is constructed to shorten the distance between samples belonging to the same phenological stage in the feature space and widen the distance between samples belonging to different phenological stages. The optimization objective is to minimize the cross-entropy loss or marginal loss based on feature similarity. Through supervised fine-tuning, the general temporal feature representation is aligned with the specific phenological semantics, generating a growth stage encoder with phenological discrimination capabilities. This growth stage encoder is used to generate the baseline stage encoding vector in S1 and reused for the target stage encoding vector extraction in S2, ensuring that the two vectors are comparable within the same feature space.

[0074] S122. Input the crop growth time-series dataset generated in S11 into the pre-trained growth stage encoder:

[0075] The growth stage encoder first extracts time-series features from the full-cycle time-series data of each typical historical year. Through a multi-layer temporal convolutional network or Transformer encoder structure, it performs nonlinear transformation and contextual information fusion on the multi-dimensional time-series variables at each time step to generate a time-series feature sequence that covers short-term fluctuations and long-term trends. This time-series feature sequence is a continuous high-dimensional vector representation, with each vector corresponding to a time step in the original data.

[0076] Based on crop phenology, the entire growth cycle of each historically typical year is divided into several standard phenological stages, including seedling emergence, three-leaf stage, tillering stage, jointing stage, heading stage, flowering stage, grain filling stage, and maturity stage. For each standard phenological stage, the temporal features corresponding to all time steps within that stage's time window are extracted to form a subset of temporal features for that phenological stage.

[0077] Cluster analysis was performed on the temporal feature subsets of the same standard phenological stage for all historical typical years. The core feature distribution area of ​​the phenological stage was identified by density-based or partition-based methods, and outlier feature points affected by extreme weather or observation noise were removed.

[0078] The core feature points after clustering are weighted averaged or attention-weighted aggregated to generate a set of fixed-dimensional baseline stage encoding vectors for each standard phenological stage. The baseline stage encoding vectors represent the standard growth state of the phenological stage under stable climate and typical yield conditions, and serve as the reference benchmark for calculating the co-offset vector in S2.

[0079] S2. The target time-series data is processed using a growth stage encoder to obtain the target stage encoding vector, which is then compared with the baseline stage encoding vector to generate a cooperative offset vector; specifically including:

[0080] S21. Acquisition and preprocessing of target raw time series data, specifically:

[0081] S211. Synchronously acquire remote sensing observation data from the sowing date to the prediction time of the target prediction year from the remote sensing data platform, specifically including daily or ten-day satellite remote sensing image products, covering multispectral band reflectance data and derived vegetation index products.

[0082] Meteorological observation data for the target area during the same period are obtained synchronously from the meteorological data platform, including daily precipitation, average temperature, maximum temperature, minimum temperature, relative humidity, sunshine duration, wind speed and direction, and surface radiation data.

[0083] During the data acquisition process, it is strictly ensured that the temporal resolution, spatial resolution, geographic coordinate reference system, and temporal coverage of remote sensing data and meteorological data are consistent with the acquisition standards of the initial time-series dataset in S11, so as to ensure the homogeneity of data structure and the comparability of subsequent processing.

[0084] S212. The acquired remote sensing data and meteorological data are spatiotemporally aligned and fused using the same spatiotemporal grid. Spatiotemporal alignment includes resampling the remote sensing image data to the same spatial resolution as the meteorological data, projecting both types of data onto the same geographic coordinate reference system, and matching timestamps according to calendar dates to ensure that the remote sensing variables and meteorological variables at each time step correspond to the same spatiotemporal location.

[0085] The fused data constitutes the target's original time-series data, which is organized in the form of a time series. Each time step contains multidimensional variables, including remote sensing dimensions such as normalized vegetation index, enhanced vegetation index, leaf area index inversion value, canopy moisture content index, and surface temperature inversion value, and meteorological dimensions such as precipitation, average temperature, cumulative effective temperature, sunshine hours, estimated soil moisture, and total solar radiation.

[0086] S213. Perform the same preprocessing procedure on the target raw time-series data as in the crop growth time-series dataset in S11. The preprocessing procedure includes the following steps:

[0087] Outlier removal employs statistical threshold or isolated forest algorithms to identify and label outlier observations caused by cloud cover or sensor malfunction.

[0088] Missing value imputation: Linear interpolation, spline interpolation, or spatiotemporal kriging based on adjacent time series are used to fill in the missing data after removal to ensure temporal continuity.

[0089] Temporal smoothing filtering, using Savitzky-Golay filters or Whittaker smoothers to suppress high-frequency noise and preserve the trend changes in crop growth;

[0090] Variable normalization involves standardizing or range normalizing each dimension of the variable to eliminate dimensional differences and ensure that the numerical range of each variable matches the input requirements of the encoder during the growth stage.

[0091] After the above preprocessing, target time series data with the same structure as crop growth time series data is obtained. The target time series data maintains the same time step structure as S11. Each time step contains multi-dimensional time series variables that have been cleaned, filled, smoothed and normalized. These variables serve as input data for the growth stage encoder in S22 and are used to extract the target stage encoding vector.

[0092] S22. Target stage encoding vector extraction, specifically:

[0093] S221. The target time-series data obtained after preprocessing in S21 is input into the growth stage encoder generated in S12 through two-stage pre-training. The growth stage encoder first extracts features from the target time-series data step by step. Utilizing the general time-series pattern awareness capability learned by its encoder base, it performs nonlinear transformations on the multidimensional time-series variables at each time step through a multi-layer temporal convolutional network or Transformer encoder structure to capture the complex interaction relationships between variables. It then fuses the contextual information of previous and subsequent time steps through a self-attention mechanism or recurrent connections to generate a target time-series feature sequence encompassing both short-term fluctuation features and long-term trend features. This target time-series feature sequence is a continuous high-dimensional vector representation sequence, with a sequence length consistent with the number of time steps in the target time-series data. Each high-dimensional vector represents the embedding representation of the crop growth state at the corresponding time step in the feature space.

[0094] S222. Based on the same standard phenological stage division rules defined in S12, including the seedling stage, three-leaf stage, tillering stage, jointing stage, heading stage, flowering stage, grain filling stage, and maturity stage, the target temporal characteristic sequence is divided into stages.

[0095] The stage division is based on a pre-defined phenological stage discrimination criterion, which is supported by the phenological semantic alignment capability of the growth stage encoder through supervised fine-tuning of stage learning. By identifying key feature patterns representing phenological transitions in the target temporal feature sequence, the start and end positions of each standard phenological stage in the target temporal feature sequence are determined. A temporal feature subset corresponding to the standard phenological stage covered by the currently observed period is extracted. This temporal feature subset contains high-dimensional feature vectors of all time steps within the time window of that phenological stage.

[0096] S223. Perform feature aggregation on the extracted temporal feature subset. The aggregation method is the same as that used in S12 to generate the baseline stage encoding vector, that is, perform weighted averaging or weighted fusion based on attention mechanism on all high-dimensional feature vectors in the temporal feature subset to generate the target stage encoding vector corresponding to the current growth process.

[0097] The dimension of the target stage encoding vector is consistent with that of the baseline stage encoding vector generated in S12, and both are generated by the same growth stage encoder in the same feature space, ensuring direct metric comparability. This target stage encoding vector represents the crop growth status of the target predicted year within the currently observed time period, serving as input data in S23 for matching and comparing with the baseline stage encoding vector, and for calculating the phenological timeline offset and physiological growth intensity offset.

[0098] S23. Phenological-Stress Co-migration Quantization: The target stage encoding vector output from S22 is matched and compared with the baseline stage encoding vector output from S12. The phenological time series progress offset and physiological growth intensity offset are calculated respectively, providing input data for the generation of the co-migration vector in S24; specifically:

[0099] S231. Timing nonlinear alignment, specifically:

[0100] A dynamic time warping algorithm is used to construct a cumulative distance matrix between the temporal sequences of the target stage coding vector and the baseline stage coding vector. Each element of this cumulative distance matrix represents the local distance between the feature vector of the target stage coding vector at a certain time step and the feature vector of the baseline stage coding vector at a certain time step. The local distance is measured by Euclidean distance or Mahalanobis distance.

[0101] The optimal regular path from the starting point to the ending point is searched in the cumulative distance matrix using a dynamic programming algorithm. This path satisfies monotonicity and continuity constraints, ensuring the preservation of time order and no-jump matching.

[0102] Based on the optimal regularization path, the temporal sequence of the target stage coding vector is nonlinearly aligned with the temporal sequence of the reference stage coding vector. This alignment process eliminates the overall scaling and translation differences caused by differences in sowing period and temperature accumulation rate. It establishes the optimal matching relationship between each feature point in the target stage coding vector and the corresponding phenological stage in the reference stage coding vector, providing a time reference alignment basis for the offset calculation of S232 and S233.

[0103] S232. Calculation of phenological timeline offset, specifically:

[0104] After completing the temporal alignment of S231, predefined key phenological feature points in the target stage encoding vector are identified. Key phenological feature points include growth rate peak points, which are the maximum values ​​of the first derivative of the vegetation index temporal curve in a specific dimension after being mapped to the feature space by the growth stage encoder; and cumulative biomass inflection points, which are the extreme values ​​in the dimension where the second derivative of the cumulative vegetation index changes sign after mapping.

[0105] Extract the occurrence time of these key phenological feature points in the target time series, and represent it as the cumulative number of days after sowing; based on the optimal matching relationship established in S231, retrieve the standard time of the corresponding feature points in the baseline stage encoding vector, which is also represented as the cumulative number of days after sowing.

[0106] The absolute time difference between the occurrence time of key phenological feature points in the target time series and their standard time in the baseline time series is calculated and expressed in days as the phenological time series progress offset. A positive offset indicates a delay in the fertile period of the target predicted year relative to a typical historical year, while a negative offset indicates an earlier fertile period, quantifying the degree of deviation of the target predicted year from the phenological development progress dimension.

[0107] S233. Calculation of physiological growth intensity offset, specifically:

[0108] Within the same phenological stage window aligned with S231, the intra-stage feature statistics of the target stage encoding vector and the baseline stage encoding vector are extracted respectively. The intra-stage feature statistics include the mean vector of feature vectors at all time steps in the stage, representing the average growth state within the phenological stage; and the slope vector, which is obtained by time-series regression fitting of the intra-stage feature vectors, representing the growth change trend within the phenological stage.

[0109] The cosine distance between the mean vectors of the target stage encoding vector and the baseline stage encoding vector is calculated to measure the degree of deviation between them in the characteristic direction. The closer the cosine distance value is to zero, the more consistent the directions are; the closer it is to one, the more divergent the directions are. Alternatively, the Euclidean distance after variance regularization, i.e., Mahalanobis distance, is calculated to measure the comprehensive difference between the two in the characteristic amplitude, considering the variance and covariance structure of each dimension. The above distances are used as physiological growth intensity offsets. This offset quantifies the difference in physiological metabolic activity and growth vitality of the target predicted year relative to historical typical years within the same phenological stage, reflecting the variation in growth intensity caused by environmental stress factors or management interventions.

[0110] S24. Cooperative offset vector generation, specifically:

[0111] S241. Receive the phenological timeline offset calculated by S232 and the physiological growth intensity offset calculated by S233, and normalize them respectively to eliminate the differences in dimensions and numerical ranges, so as to ensure that the two-dimensional offsets have a balanced contribution weight in the subsequent fusion.

[0112] The phenological timeline offset is normalized using either min-max normalization or Z-score standardization. Min-max normalization maps the phenological timeline offset to a preset numerical range. Specifically, it determines the historical maximum absolute value of the phenological timeline offset based on historical year data, divides the current offset by this historical maximum absolute value, and compresses the result to a range of -1 to +1. Positive values ​​indicate delay, negative values ​​indicate advancement, and the absolute value indicates the degree of deviation. Z-score standardization calculates the standard score of the phenological timeline offset relative to the historical mean, representing the current year's relative position in the historical distribution.

[0113] The physiological growth intensity offset is independently normalized using a method that is the same as or adapted to the distribution characteristics of the phenological timeline offset. If the same method is used, the physiological growth intensity offset is also mapped to the negative one to positive one interval, or its standard score is calculated. If the historical distribution of the physiological growth intensity offset shows skewed characteristics, quantile normalization or Box-Cox transformation is used to make it approximately conform to a standard normal distribution before Z-score standardization to enhance the distinguishability of extreme stress conditions.

[0114] S242. The normalized phenological time-series progress offset and physiological growth intensity offset are concatenated in a predetermined order, with the phenological time-series progress offset placed in the first dimension and the physiological growth intensity offset placed in the second dimension, constructing a two-dimensional vector. This two-dimensional vector is the final co-offset vector, which simultaneously encapsulates the differences in developmental progress in the time dimension and the differences in growth status in the physiological dimension, forming a comprehensive quantitative representation of the deviation of crop growth from the typical state in the target prediction year.

[0115] The collaborative offset vector serves as the core input data for the hierarchical decision-maker in S3. It drives the hierarchical decision-maker to dynamically select or combine the most suitable model from multiple pre-trained expert prediction sub-models based on the offset pattern. The selected expert prediction sub-model then outputs the final yield prediction value based on the target time series data and the collaborative offset vector.

[0116] S3. Input the collaborative offset vector into the hierarchical decision maker to select the expert prediction sub-model, and then output the final output prediction value; specifically including:

[0117] S31. Misalignment Mode Determination and Model Selection Weight Generation: The collaborative misalignment vector generated in S24 is input into the hierarchical decision-maker. The hierarchical decision-maker, considering both the current misalignment state and historical trends, determines the current scenario as one of several predefined typical misalignment modes and outputs the corresponding mode identifier and model selection weights. This provides a decision-making basis for the dynamic scheduling of the expert prediction sub-model in S32. Specifically:

[0118] S311. The collaborative offset vector generated in S24 is concatenated with its historical evolution trend features within a preset time window to form an enhanced decision feature vector. Specifically:

[0119] A preset time window is set as several consecutive time steps prior to the current prediction time. The value sequence of the collaborative offset vector at these historical time steps is extracted back, forming the short-term historical trajectory of the collaborative offset vector. Based on this historical trajectory, the evolution trend characteristics are calculated: first-order difference feature, i.e., the difference between the collaborative offset vectors of adjacent time steps, representing the short-term rate of change; moving average feature, i.e., the mean of the collaborative offset vector within the time window, representing the medium-term trend of change; and linear fitting slope feature, i.e., the slope obtained by performing least-squares linear fitting on each dimension of the collaborative offset vector within the time window, representing the overall evolution direction.

[0120] The current value of the collaborative offset vector, the historical value sequence within a preset time window, and the extracted evolution trend features are concatenated into vectors in a fixed-dimensional order to construct an enhanced decision feature vector. This enhanced decision feature vector has a higher dimension than the original collaborative offset vector and contains both the current offset state and dynamic evolution information, providing comprehensive decision-making basis for S312.

[0121] S312. Offset pattern category identification, specifically:

[0122] The enhanced decision feature vector constructed by S311 is input into the preset rule engine or lightweight classification network in the hierarchical decision maker to identify the most matching offset pattern category.

[0123] If a rule-based engine is used, several characteristic threshold ranges for typical offset patterns are pre-stored. Each typical offset pattern corresponds to a specific set of two-dimensional threshold range combinations, including threshold ranges for phenological time-series progress offset and physiological growth intensity offset. The phenological time-series progress offset component in the enhanced decision feature vector is compared with the pre-stored phenological time-series progress offset threshold range, and the physiological growth intensity offset component is compared with the pre-stored physiological growth intensity offset threshold range. This determines which set of threshold range combinations the value of the enhanced decision feature vector falls into, thereby identifying the most matching offset pattern category.

[0124] If a lightweight classification network, such as a multilayer perceptron or a small convolutional network, is used, the historical augmentation decision feature vectors with offset pattern labels are used as training samples. The nonlinear mapping relationship from the feature space to the offset pattern category is learned through backpropagation. Forward inference is performed on the input augmentation decision feature vectors. After the network output layer is normalized by the softmax function, the posterior probability distribution of each typical offset pattern is obtained. The offset pattern category corresponding to the highest probability is taken as the best-matching offset pattern category.

[0125] S313. Pattern identifier and model selection weight output, specifically:

[0126] Based on the matching results in S312, the corresponding pattern identifier is output. This pattern identifier is a unique code for a typical offset pattern, such as HD for high temperature and drought stress type, LC for low temperature and low sunshine delay type, and FN for favorable weather type, etc., which is used in S322 to retrieve the corresponding model from the expert prediction sub-model library.

[0127] Simultaneously, model selection weights are calculated based on the matching confidence score of S312 or its proximity to multiple threshold intervals. If S312 uses a rule-based engine, the Euclidean distance between the enhanced decision feature vector and the center of the matching threshold interval is calculated as a proximity index, and the distance to the boundary of adjacent threshold intervals is calculated as a discrimination index. A normalized confidence score is calculated by combining both, serving as the model selection weight for each typical offset pattern. If S312 uses a lightweight classification network, the output posterior probability distribution is directly used as the model selection weight for each typical offset pattern.

[0128] When the maximum model selection weight exceeds the preset dominant mode threshold, it indicates that a single dominant mode exists. The mode identifier corresponding to the maximum weight is output, and the remaining weights are reset to zero or kept at a small value. When the maximum model selection weight does not exceed the dominant mode threshold, or the ratio of the second largest weight to the maximum weight exceeds the preset mixed mode threshold, it indicates that a mixed mode exists. The identifiers of multiple related modes and their normalized model selection weights are output, providing input for the multi-model scheduling of S322 and the weight allocation of S323.

[0129] S32. The hierarchical decision maker dynamically schedules one or more expert prediction sub-models from the pre-trained expert prediction sub-model library based on the pattern identifier and model selection weights output from S313, providing the model foundation for the weighted ensemble prediction in S33; specifically:

[0130] S321. The expert prediction sub-model library contains multiple expert prediction sub-models. Each expert prediction sub-model is specifically trained for a particular typical offset pattern, forming a specialized modeling capability for the response law of output under that offset pattern. The specific construction process is as follows:

[0131] Retrieve the crop growth time series dataset of historical years, the baseline stage encoding vector generated by S1 for the corresponding year, the collaborative offset vector generated by S2, and the corresponding actual yield value; using the same hierarchical decision-making process from S31 to S313, determine the offset pattern for these historical years, and aggregate all historical year samples that were determined by the hierarchical decision-maker to be the same typical offset pattern into a special training subset for that pattern.

[0132] For each typical offset pattern, a dedicated training subset is used as training data to independently train an expert prediction sub-model. The input features during training include the target time-series data generated in S21, the target stage encoding vector generated in S22, and the co-offset vector generated in S24. The prediction target is the actual yield value for the corresponding year. By optimizing yield prediction loss functions such as mean squared error or mean absolute percentage error, the expert prediction sub-model learns the nonlinear impact mechanism of phenological time-series progress offset and physiological growth intensity offset on the final yield under this specific offset pattern, including the combined effects of different stress intensities and occurrence periods. The trained expert prediction sub-models are stored in the expert prediction sub-model library, establishing a one-to-one mapping relationship with the pattern identifiers output in S313 to ensure the accuracy of subsequent scheduling.

[0133] S322. The hierarchical decision-maker receives the pattern identifier and model selection weights output from S313, and retrieves and calls the corresponding expert prediction sub-model from the expert prediction sub-model library accordingly. Specifically:

[0134] When the model selection weights output by S313 indicate the existence of a single dominant mode, i.e., the maximum weight is greater than or equal to the preset dominant mode threshold, the hierarchical decision maker only schedules the single expert prediction sub-model corresponding to the dominant mode identifier. At this time, the model is considered to be the best fit for the current crop growth deviation state.

[0135] When the model selection weight distribution indicator output by S313 is a mixed mode, that is, the maximum weight is less than the dominant mode threshold and the ratio of the second largest weight to the maximum weight exceeds the preset mixed mode threshold, the hierarchical decision maker schedules multiple expert prediction sub-models corresponding to the multiple mode identifiers with the top weight ranking, in order to cover the combined effects of multiple coexisting stress or favorable factors.

[0136] The scheduled expert prediction sub-model receives the target time series data from S21 and the collaborative offset vector from S24 as common inputs, and prepares to execute the production prediction inference.

[0137] S323. The model selection weights output from S313 are assigned to each expert prediction sub-model scheduled in S322, as coefficients for the weighted integration in S33. Specifically:

[0138] If S322 schedules a single expert prediction sub-model, then the model selection weights of the corresponding mode of that model are normalized and assigned a value of one, while the weights of the remaining unscheduled models are assigned a value of zero. In this case, the weighted integration degenerates into a single model prediction.

[0139] If S322 schedules multiple expert prediction sub-models, the original model selection weights for each scheduled model's corresponding mode are extracted, normalized to make the sum of the weights equal to one, and then allocated to each scheduled expert prediction sub-model. These normalized model selection weights represent the relative contribution of each expert prediction sub-model to the final output prediction; higher weights indicate a higher confidence level of the corresponding offset mode in the current scenario, and a greater influence of its prediction result on the final output. The allocated model selection weights are then passed to S33 along with the scheduled expert prediction sub-models for weighted integration calculation.

[0140] S33. Receive the expert prediction sub-model scheduled by S322 and the model selection weights allocated by S323, perform output prediction inference and weighted integration, and output the final output prediction value; specifically:

[0141] S331. The target time-series data generated in S21 and the collaborative offset vector generated in S24 are passed as common input data to each expert prediction sub-model scheduled by S322. Each expert prediction sub-model forms a response pattern based on the output of a specific offset pattern learned through its specialized training, and performs forward inference on the input data:

[0142] First, using the same growth stage encoder structure or shared weights as S12 and S22, feature extraction is performed on the target time series data to generate the target stage encoding vector;

[0143] Then, the target stage encoding vector and the collaborative offset vector are fused and input into the dedicated prediction network of this expert prediction sub-model;

[0144] Dedicated prediction networks are typically multi-layer fully connected networks or temporal neural networks that learn the nonlinear mapping relationship between phenological progress, growth intensity and final yield under specific offset patterns.

[0145] The network output layer generates the output prediction value of the expert prediction sub-model under the offset mode assumption, i.e., the single-mode output prediction value.

[0146] S332. Collect all single-mode output forecasts generated by the scheduled expert prediction sub-models to form a single-mode output forecast set. Receive the model selection weights assigned in S323, which correspond one-to-one with the scheduled expert prediction sub-models.

[0147] The set of single-mode output forecasts is weighted and summed according to the model selection weights: the single-mode output forecast of each expert prediction sub-model is multiplied by its corresponding model selection weight to obtain the weighted output forecast component; all weighted output forecast components are summed to obtain the weighted output forecast sum; since S323 has completed the normalization process, the weighted output forecast sum is the weighted average value, which is used as the final output output value.

[0148] S333. This final yield prediction integrates the professional judgments of each expert prediction sub-model under their respective preferred offset modes. By adaptively reflecting the matching degree between the current crop growth status and each typical offset mode through model selection weights, it achieves accurate yield prediction based on the phenological-stress co-deviation state represented by the S24 co-offset vector. This final yield prediction serves as the prediction benchmark for decision verification and model adaptive optimization in S4. Together with the subsequently obtained actual yield values, it forms a structured feedback tuple, driving the collaborative fine-tuning of the S1 growth stage encoder, the S3 hierarchical decision-maker, and the expert prediction sub-model.

[0149] S4. Construct structured feedback tuples based on actual yield values ​​and collaborative offset vectors, and perform collaborative fine-tuning of the growth stage encoder, hierarchical decision-maker, and expert prediction sub-model; specifically including:

[0150] S41. After the crop's entire growth cycle ends and harvest is completed in the target prediction year, obtain the actual yield value for that year from agricultural statistical databases, remote sensing yield inversion verification, or field yield measurement data. This actual yield value is the crop grain yield per unit area, measured in kilograms per hectare or tons per hectare. Associate and bind this actual yield value with the collaborative offset vector generated in S2, the pattern identifier generated in S31, and the model selection weights assigned in S323, constructing a structured feedback tuple. Specifically:

[0151] S411. Extract the collaborative offset vector generated in S24. This collaborative offset vector is a two-dimensional vector. The first dimension is the phenological time series progress offset calculated and normalized by S232, which represents the degree of deviation of the target prediction year from the historical typical year in terms of phenological development progress. Positive values ​​indicate delay, and negative values ​​indicate advancement.

[0152] The second dimension is the physiological growth intensity offset calculated and normalized by S233, which represents the degree of deviation of the target predicted year from the historical typical year in terms of physiological metabolic activity and growth vigor. A positive value indicates weaker growth potential, and a negative value indicates stronger growth potential.

[0153] S412. Extract the pattern identifier output by S313. This pattern identifier is the typical offset pattern category code for the hierarchical decision-maker to determine the current situation, such as HD indicating high temperature and drought stress type, LC indicating low temperature and low sunshine delay type, FN indicating favorable weather type, etc.

[0154] S413. Extract the model selection weights assigned in S323. These model selection weights are normalized weight values ​​corresponding to the pattern identifiers, representing the confidence level of the hierarchical decision-maker in the pattern's judgment result, or the relative contribution ratio of each relevant pattern under the mixed pattern.

[0155] S414. The above four data elements and the target prediction year identifier are encapsulated together into a structured feedback tuple. The field structure of the structured feedback tuple includes: year identifier field, which records the unique identifier code of the target prediction year; actual output value field, which records the actual output value of the year; collaborative offset vector field, which records the value of the two-dimensional offset vector generated in S24; pattern identifier field, which records the typical offset pattern category code output in S313; and model selection weight field, which records the normalized weight value assigned in S323.

[0156] This structured feedback tuple serves as a training sample with complete supervised information. It is used for comparative fine-tuning of the growth stage encoder in S421 and collaborative fine-tuning of the hierarchical decision-maker and expert prediction sub-model in S422. This enables the model system to learn and memorize the actual growth deviation and yield formation results of the target prediction year, driving the system to adaptively iteratively update.

[0157] S42. Add the structured feedback tuples constructed in S41 to the historical training set, and use incremental learning to collaboratively fine-tune the S1 growth stage encoder, S3 hierarchical decision maker, and expert prediction sub-model. This achieves adaptive iterative updates of the model system, improving the prediction accuracy and pattern adaptability for future years. Specifically:

[0158] S421. Extract the collaborative offset vector from the structured feedback tuple, and calculate the magnitude of this collaborative offset vector as a general quantitative indicator of the difference between the actual growth pattern and the baseline pattern. The larger the magnitude, the further the crop growth state in the target prediction year deviates from the historical typical state. Based on this quantitative indicator, construct an encoder contrastive loss function. The design goal of this loss function is: for samples with large collaborative offset vector magnitudes, enhance the encoder's ability to distinguish deviation states in the feature space during the growth stage; for samples with small collaborative offset vector magnitudes, maintain the compact clustering characteristics of the feature space. The specific implementation process is as follows:

[0159] The target time-series data associated with structured feedback tuples is re-inputted into the growth stage encoder pre-trained in S12. Forward inference generates a fine-tuned stage encoding vector, which has the same feature space dimension as the target stage encoding vector generated in S22. The baseline stage encoding vector corresponding to the current phenological stage generated in S12 is retrieved, and the Euclidean or cosine distance between the fine-tuned stage encoding vector and the baseline stage encoding vector in the feature space is calculated as the encoding distance metric.

[0160] The magnitude of the collaborative offset vector is used as a soft supervision weight to construct a weighted contrastive loss: when the magnitude is greater than the preset deviation threshold, the loss contribution of the encoding distance metric is increased to encourage the encoder to expand the feature distance of the difference samples during the growth stage; when the magnitude is less than the preset deviation threshold, the loss contribution of the encoding distance metric is reduced or a feature convergence constraint is introduced to maintain the feature compactness of typical state samples.

[0161] The gradient of the encoder's contrastive loss is passed to the parameters of the temporal convolutional network layer or Transformer encoder layer of the growth stage encoder through the backpropagation algorithm. The network weights are updated by performing stochastic gradient descent or Adam optimization algorithm, which optimizes the feature extraction and phenological semantic alignment capabilities of the growth stage encoder, making it more sensitive to the phenological-stress co-shift and more comparable to the output target stage encoding vector and the baseline stage encoding vector.

[0162] S422. Fine-tuning of the hierarchical decision-maker and expert prediction sub-model, specifically:

[0163] The actual production value and the final production forecast value output by S33 are extracted from the structured feedback tuples, and the forecast loss is calculated. The forecast loss adopts the mean squared error loss function, which is the square of the difference between the actual production value and the final production forecast value, or the Huber loss function. When the error is small, the squared loss is used to ensure smoothness, and when the error is large, the linear loss is used to reduce the impact of outliers, thus measuring the degree of numerical deviation in the production forecast.

[0164] Simultaneously, pattern identifiers are extracted from the structured feedback tuples, and the performance of the corresponding expert prediction sub-model in the current prediction is analyzed. If the deviation between the single-mode output prediction value and the actual output value generated by the expert prediction sub-model corresponding to the pattern identifier is significantly greater than that of other unselected expert prediction sub-models, it indicates that the pattern judgment of the hierarchical decision-maker is misjudged or that the specific modeling capability of the expert prediction sub-model is insufficient. Based on this correlation analysis, a decision confidence loss is constructed: for samples with correct pattern judgment and small prediction error, the output confidence of the pattern identifier in the hierarchical decision-maker is enhanced; for samples with incorrect pattern judgment or associated with high error, the output confidence of the corresponding pattern identifier is reduced or a penalty term for pattern confusion is added. The decision confidence loss adopts the cross-entropy loss function, comparing the model selection weight distribution output by S313 with the ideal weight distribution based on the actual error, or increasing the decision boundary between correct and incorrect modes through an interval-based loss function.

[0165] The prediction loss and decision confidence loss are weighted and fused using preset weighting coefficients to obtain a composite loss. These preset weighting coefficients are dynamically adjusted based on the current system optimization priorities; the weight of the prediction loss is increased when the prediction error is large, and the weight of the decision confidence loss is increased when the pattern determination is chaotic. Using this composite loss, the hierarchical decision maker and the expert prediction sub-model are simultaneously optimized using the gradient descent algorithm.

[0166] For a hierarchical decision maker, if a rule engine is used, the feature threshold interval boundary in S312 is adjusted based on the gradient direction of the composite loss, so that the judgment range of typical offset patterns is more in line with the actual production formation pattern; if a lightweight classification network is used, the weight parameters of the fully connected layer or convolutional layer of the network are updated by backpropagation, and the mapping relationship from the enhanced decision feature vector to the pattern identifier is optimized.

[0167] For the expert prediction sub-model, identify the expert prediction sub-model associated with the pattern identifier in the structured feedback tuple, i.e., the model scheduled in S322. Backpropagate the composite loss to the dedicated prediction network of this model, update the weight parameters from its input layer to its output layer, and optimize the estimation of the nonlinear response relationship between phenological progress, growth intensity, and yield under this specific offset mode. If the structured feedback tuple corresponds to a mixed mode and multiple expert prediction sub-models are scheduled, then gradient updates are performed in parallel for each relevant model, with the update magnitude proportional to the model selection weights.

[0168] Through this collaborative fine-tuning, the model judgment accuracy of the hierarchical decision-maker and the model-specific prediction capability of the expert prediction sub-model evolve synergistically under the drive of real yield feedback. The system gradually accumulates knowledge of new extreme climate events or complex stress models, realizing online evolution of model performance and long-term adaptive improvement.

[0169] Example 2

[0170] like Figure 2 As shown, a crop yield prediction system based on a time-series feature algorithm model is used to implement a crop yield prediction method based on a time-series feature algorithm model, including:

[0171] The baseline encoding construction module is used to acquire crop growth time-series datasets of the target region over multiple consecutive historical typical years, and processes the time-series datasets through a pre-trained, cross-regional growth stage encoder to generate a set of baseline stage encoding vectors representing the core features of standard phenological stages; specifically including:

[0172] The historical data acquisition subunit is used to acquire initial time-series data of continuous historical years of the target area from remote sensing and meteorological data platforms, and to select typical historical years based on climate stability and historical yield, and output crop growth time-series dataset.

[0173] The encoder processing subunit, connected to the historical data acquisition subunit, is used to input the crop growth time series dataset into the pre-trained growth stage encoder, extract and aggregate the phenological stage features of each typical year, and output the baseline stage encoding vector.

[0174] The offset quantization module, connected to the baseline encoding construction module, is used to acquire the full growth cycle time-series data of the target prediction year. It extracts the target stage encoding vector using the growth stage encoder, compares the target stage encoding vector with the baseline stage encoding vector, calculates the differences in phenological time-series progress and physiological growth intensity, and fuses them to generate a collaborative offset vector. Specifically, it includes:

[0175] The target data acquisition subunit is used to synchronously acquire the raw data of the target prediction year from the remote sensing and meteorological data platform, and output the target time series data after preprocessing.

[0176] The target coding subunit, connected to the target data acquisition subunit, is used to input the target time-series data into the growth stage encoder and output the target stage coding vector.

[0177] The offset calculation subunit is connected to the baseline coding construction module and the target coding subunit, respectively. It is used to compare the target stage coding vector with the baseline stage coding vector, calculate the phenological time series progress offset and the physiological growth intensity offset, and fuse them to generate a collaborative offset vector.

[0178] The dynamic prediction module, connected to the offset quantization module, includes a hierarchical decision-maker and an expert prediction sub-model library. The hierarchical decision-maker outputs a mode identifier and model selection weights based on the offset pattern represented by the co-offset vector, and dynamically selects or combines one or more expert prediction sub-models from the expert prediction sub-model library accordingly. The selected expert prediction sub-model outputs the final yield prediction value based on the target time-series data and the co-offset vector. Specifically, it includes:

[0179] The hierarchical decision subunit, connected to the offset quantization module, is used to receive the collaborative offset vector, match typical offset patterns by analyzing the two-dimensional offset combination relationship, and output the corresponding pattern identifier and model selection weight.

[0180] The expert model library sub-unit stores multiple expert prediction sub-models specifically trained for different typical offset patterns;

[0181] The model scheduling and prediction subunit is connected to the hierarchical decision-making subunit and the expert model library subunit, respectively. It is used to schedule the corresponding model from the expert model library subunit according to the pattern identifier and model selection weight, organize it to perform prediction and weighted integration based on the target time series data and the collaborative offset vector, and output the final output prediction value.

[0182] The adaptive optimization module, connected to the baseline encoding construction module, offset quantization module, and dynamic prediction module, is used to construct a structured feedback tuple after obtaining the actual yield value for the target prediction year, along with the collaborative offset vector, mode identifier, and model selection weights. This structured feedback is then used to collaboratively fine-tune the growth stage encoder, hierarchical decision maker, and expert prediction sub-model, enabling iterative system updates. Specifically, this includes:

[0183] The feedback tuple construction subunit is connected to the offset quantization module and the dynamic prediction module respectively. It is used to receive the collaborative offset vector, mode identifier, model selection weight and the subsequently entered real output value, and bind them to construct a structured feedback tuple.

[0184] The collaborative fine-tuning subunit is connected to the feedback tuple construction subunit and then to the growth stage encoder, hierarchical decision maker, and various expert prediction sub-models. It is used to calculate the composite loss using structured feedback tuples and to jointly optimize and update the parameters of the above parts in an incremental learning manner.

[0185] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects:

[0186] By constructing a transferable agronomic semantic benchmark and employing dual-dimensional quantitative diagnosis, this invention achieves interpretable and refined monitoring of crop growth anomalies, solving the problems of feature black boxes and diagnostic ambiguity inherent in traditional models. The growth stage encoder of this invention maps the original time-series data stream into benchmark stage encoding vectors aligned with standard phenological stages through cross-regional comparative learning, establishing a semantic bridge between model features and agronomic knowledge. Based on this, dual-dimensional offset quantization further decomposes growth anomalies into precise phenological time-series progress offsets and physiological growth intensity offsets. This fundamentally solves the semantic gap problem of poor feature interpretability and disconnection from agronomic logic in existing deep learning models. It enables the model not only to determine whether growth is abnormal, but also to diagnose whether the process is premature / delayed, or whether physiological function is enhanced / inhibited, and at which critical stage it occurs. This provides agronomists with refined diagnostic reports that can be directly used for production decisions, greatly improving the reliability and practicality of prediction results.

[0187] Through a hierarchical decision-making and dynamic scheduling expert consultation mechanism, the generalization ability and predictive robustness of the model in complex and ever-changing environments are significantly improved, overcoming the shortcomings of insufficient adaptability of a single static model. The hierarchical decision-maker and dynamic model scheduling mechanism of this invention constitute a biomimetic intelligent decision-making system. The system first identifies typical deviation patterns based on the aforementioned refined diagnostic results, and then, like an expert consultation, dynamically schedules expert predictive sub-models specializing in these abnormal patterns for prediction and integration. This architecture completely changes the predicament of traditional single models attempting to cope with all changes in a static way. Faced with different scenarios such as high temperature and drought, low temperature and insufficient sunshine, and favorable weather, the system can call upon the most professional models for targeted analysis, achieving a tailored approach. This significantly enhances the model's adaptability and predictive stability when facing complex climate patterns and sudden disasters that are unseen or rarely appear in the training data, effectively solving the industry pain point of a sharp drop in the model's generalization ability in unknown years or new urban areas.

[0188] By establishing a self-evolving closed loop of structured feedback and collaborative fine-tuning, the system possesses the ability to continuously self-optimize in dynamic agricultural environments, overcoming the limitations of traditional prediction models that are static, fixed, and have high maintenance costs. The adaptive optimization module of this invention transforms each season's prediction practice into learning resources. It encapsulates predicted values, diagnostic criteria, and decision logic into structured feedback tuples, and uses these to collaboratively fine-tune the encoder, decision-maker, and prediction model. This constitutes a complete perception-diagnosis-decision-verification-evolution closed loop, transforming the system from a tool that becomes fixed after deployment into a living entity capable of autonomous iteration as data accumulates and the environment changes. This system fundamentally solves the problems of AI models easily becoming outdated in dynamic real-world environments and the high-cost maintenance required for periodic manual data relabeling and training. It achieves long-term sustainable assurance of prediction accuracy, providing a long-term viable solution for agricultural intelligence.

[0189] The embodiments and / or implementation methods described above are merely preferred embodiments and / or implementation methods for implementing the technology of the present invention, and are not intended to limit the implementation methods of the technology of the present invention in any way. Any person skilled in the art can make some modifications or alterations to other equivalent embodiments without departing from the scope of the technical means disclosed in the present invention, but these should still be regarded as the technology or embodiments that are substantially the same as the present invention.

[0190] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. The above descriptions are only preferred embodiments of this application. It should be noted that due to the limitations of written expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of this application, and can also combine the above technical features in an appropriate manner. These improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other situations without modification, should all be considered within the scope of protection of this application.

Claims

1. A method for predicting crop yield based on a time-series feature algorithm model, characterized in that, include: Historical time-series data is processed by the growth stage encoder to generate baseline stage encoding vectors; The target stage coding vector is obtained by processing the target time-series data using the growth stage encoder, and then compared with the reference stage coding vector to generate a cooperative offset vector. The collaborative offset vector is input into the hierarchical decision maker to select the expert prediction sub-model, and then outputs the final output prediction value. A structured feedback tuple is constructed based on the actual output value and the collaborative offset vector, and the growth stage encoder, hierarchical decision maker, and expert prediction sub-model are collaboratively fine-tuned.

2. The crop yield prediction method based on a time-series feature algorithm model according to claim 1, characterized in that, The process of processing historical time-series data through a growth stage encoder to generate a baseline stage encoding vector specifically includes: The initial time-series dataset of the target area throughout the entire growth cycle in multiple consecutive historical years is obtained from remote sensing data platform and meteorological data platform; based on preset climate stability index and historical yield data, typical historical years that meet the requirements are selected from the historical years, and the crop growth time-series dataset of the corresponding years is extracted. The crop growth time series dataset is input into a pre-trained growth stage encoder; the growth stage encoder processes the dataset, extracts the time series features of each typical year, and clusters and aggregates the same phenological stage features of all typical years, outputting a set of baseline stage encoding vectors for each standard phenological stage.

3. The crop yield prediction method based on a time-series feature algorithm model according to claim 2, characterized in that, The pre-training of the growth stage encoder includes: Using unlabeled, large-scale, multi-regional crop time-series data, an encoder base is trained through self-supervised learning tasks such as mask reconstruction or time-series comparison prediction, enabling it to learn general time-series patterns of crop growth. On the encoder base, multi-region data with phenological stage labels are used to perform supervised fine-tuning through contrastive learning, so that general features are aligned with specific phenological semantics.

4. The crop yield prediction method based on a time-series feature algorithm model according to claim 3, characterized in that, The target stage encoding vector is obtained by processing the target time-series data using the growth stage encoder, and a cooperative offset vector is generated by comparing it with the reference stage encoding vector. Specifically, this includes: The target original time series data, which is from sowing to the prediction time point of the target prediction year and has the same source as the crop growth time series data structure, is synchronously acquired from the remote sensing data platform and the meteorological data platform, and the target time series data is obtained after the same preprocessing. The target time series data is input into a pre-trained growth stage encoder, and the output is a target stage encoding vector corresponding to the current growth process. The target stage encoding vector is matched and compared with the baseline stage encoding vector to calculate the phenological timeline offset and physiological growth intensity offset. The calculated phenological timeline offset and physiological growth intensity offset are normalized and then concatenated into a two-dimensional vector in a predetermined order as the final collaborative offset vector.

5. The crop yield prediction method based on a time-series feature algorithm model according to claim 4, characterized in that, The target stage encoding vector is matched and compared with the baseline stage encoding vector to calculate the phenological time series progress offset and physiological growth intensity offset, specifically including: A dynamic time warping algorithm is used to non-linearly align the temporal sequence of the target stage encoding vector with the temporal sequence of the reference stage encoding vector, and determine the matching relationship between each feature point in the target stage encoding vector and the corresponding phenological stage in the reference stage encoding vector. After completing the time alignment, the absolute time difference between the occurrence time of the predefined key phenological feature points in the target stage coding vector and the standard time of the corresponding feature points in the baseline stage coding vector is calculated as the phenological time sequence progress offset. Within the aligned windows of the same phenological stage, extract the intra-stage feature statistics of the target stage encoding vector and the reference stage encoding vector respectively; calculate the cosine distance or regularized Euclidean distance between the two, which is the physiological growth intensity offset.

6. The crop yield prediction method based on a time-series feature algorithm model according to claim 5, characterized in that, The collaborative offset vector is input into the hierarchical decision-maker to select the expert prediction sub-model, and then the final output prediction value is output, specifically including: The collaborative offset vector is input into the hierarchical decision-maker, which determines the current scenario as one of several predefined typical offset modes and outputs the corresponding mode identifier and model selection weight. The hierarchical decision-maker dynamically schedules models from a pre-trained expert prediction sub-model library based on pattern identifiers and model selection weights. The prediction results of the scheduled expert prediction sub-models are weighted and integrated according to the weights selected by the model to output the final output prediction value.

7. The crop yield prediction method based on a time-series feature algorithm model according to claim 6, characterized in that, The collaborative offset vector is input into the hierarchical decision-maker, which determines the current scenario as one of several predefined typical offset modes and outputs the corresponding mode identifier and model selection weights, specifically including: The collaborative offset vector is concatenated with its historical evolution trend features within a preset time window to form an enhanced decision feature vector; The enhanced decision feature vector is input into the preset rule engine or lightweight classification network in the hierarchical decision maker; by comparing the phenological time progression offset and physiological growth intensity offset in the vector with the pre-stored feature threshold intervals corresponding to different typical offset patterns, the most matching offset pattern category is identified. Based on the matching results, the corresponding pattern identifier is output; at the same time, the model selection weights assigned to one or more expert prediction sub-models are calculated according to the matching confidence or the degree of proximity to multiple threshold intervals.

8. The crop yield prediction method based on a time-series feature algorithm model according to claim 7, characterized in that, The hierarchical decision-maker dynamically schedules models from a pre-trained expert prediction sub-model library based on pattern identifiers and model selection weights, specifically including: Each expert prediction sub-model in the expert prediction sub-model library is independently trained using all training samples that were historically identified by the hierarchical decision-maker as the same typical offset pattern. Each sub-model specifically learns the output formation response pattern under its corresponding offset pattern. The hierarchical decision maker calls one or more corresponding expert prediction sub-models from the expert prediction sub-model library according to the pattern identifier; when the model selection weight indicates the existence of a dominant pattern, a single dominant model is scheduled; when the weight distribution indicates a mixed pattern, multiple related models are scheduled. The model selection weights are assigned to each expert prediction sub-model that is scheduled, and used as coefficients for subsequent weighted integration.

9. A method for predicting crop yield based on a time-series feature algorithm model according to claim 8, characterized in that, The step of constructing structured feedback tuples based on the actual output value and the collaborative offset vector, and then collaboratively fine-tuning the growth stage encoder, hierarchical decision-maker, and expert prediction sub-model, specifically includes: After obtaining the actual output value for the target prediction year, it is associated and bound with the collaborative offset vector, pattern identifier, and model selection weights to jointly construct a structured feedback tuple. The structured feedback tuples are added to the historical training set to collaboratively fine-tune the system using incremental learning. The co-offset vector in the structured feedback tuple is used as a soft supervision signal to construct the encoder contrast loss. The loss function encourages the encoder in the growth stage to generate codes that are further away in the feature space during periods when the actual growth pattern differs more from the baseline pattern. Based on this loss, the parameters of the encoder in the growth stage are updated through the backpropagation algorithm. Based on the actual output value and the final output prediction value in the structured feedback tuple, the prediction loss is calculated. At the same time, based on the correlation between the decision result of the pattern identifier and the actual output error, the decision confidence loss is calculated. The two are weighted and fused to obtain the composite loss. Using the composite loss, the internal classification threshold, network parameters, and parameters of the expert prediction sub-model associated with the structured feedback tuple are simultaneously optimized using the gradient descent algorithm.

10. A crop yield prediction system based on a time-series feature algorithm model, characterized in that, A method for predicting crop yield based on a time-series feature algorithm model as described in any one of claims 1-9 includes: The benchmark encoding construction module is used to obtain crop growth time series datasets of the target region in multiple consecutive historical typical years, and process the time series datasets through a pre-trained, cross-regional general growth stage encoder to generate a set of benchmark stage encoding vectors that represent the core features of standard phenological stages. The offset quantization module, connected to the baseline encoding construction module, is used to acquire the target time series data of the full growth cycle of the target prediction year, extract its target stage encoding vector using the growth stage encoder, compare the target stage encoding vector with the baseline stage encoding vector, calculate the difference between them in the two dimensions of phenological time series progress and physiological growth intensity, and fuse them to generate a collaborative offset vector. The dynamic prediction module, connected to the offset quantization module, includes a hierarchical decision-maker and an expert prediction sub-model library. The hierarchical decision-maker is used to output a mode identifier and a model selection weight based on the offset mode represented by the collaborative offset vector, and dynamically selects or combines one or more expert prediction sub-models from the expert prediction sub-model library accordingly. The selected expert prediction sub-model is used to output the final output prediction value based on the target time series data and the collaborative offset vector. The adaptive optimization module, which is connected to the baseline encoding construction module, the offset quantization module, and the dynamic prediction module, is used to construct a structured feedback tuple by combining the actual yield value of the target prediction year with the collaborative offset vector, the pattern identifier, and the model selection weights after obtaining the actual yield value. The structured feedback is then used to collaboratively fine-tune the growth stage encoder, the hierarchical decision-maker, and the expert prediction sub-model.

Images