Winter wheat growth period prediction and intelligent management method and system based on multi-source data

Abstract

The invention belongs to the technical field of agricultural management, and particularly relates to a winter wheat growth period prediction and intelligent management method and system based on multi-source data, and the method comprises the steps: generating a continuous time sequence through multi-source heterogeneous data, and sequentially carrying out the space-time modeling and dynamic feature interaction processing of continuous time sequence features, so as to predict the growth period of winter wheat; wherein in the spatio-temporal modeling stage, the feature expression ability is enhanced from three aspects of time periodicity, seasonal gating and geospatial dependence; in the dynamic feature interaction stage, the characterization capacity of the time-varying interaction relation between winter wheat time sequence variables is effectively improved through energy dynamic coupling; after refined prediction of the daily growth process of winter wheat is completed, a key growth stage turning point is extracted, a prediction result is explained in a natural language form, and personalized agricultural management suggestions are generated to facilitate guidance of agricultural operation.

Description

Technical Field

[0001] The present invention belongs to the field of agricultural management technology, and in particular relates to a method and system for predicting the growth period of winter wheat and intelligently managing it based on multi-source data. Background Art

[0002] The growth process of winter wheat is influenced by multiple environmental factors, including temperature, precipitation, sunlight, and accumulated heat. Accurately predicting key growth stages (such as sowing, jointing, heading, grain filling, and maturity) is crucial for formulating effective field management strategies. In recent years, machine learning and deep learning technologies have been widely used in agricultural forecasting, achieving significant progress in improving forecast accuracy. However, the following limitations and shortcomings remain:

[0003] 1. Insufficient multi-source data integration: Most current winter wheat growing period prediction methods rely on a single or limited set of data sources. For example, they use only meteorological data (temperature, precipitation, etc.) or remote sensing data (such as NDVI and EVI) for modeling. This approach ignores the fact that the growing period is influenced by multiple ecological and environmental factors.

[0004] 2. Process model parameters are highly sensitive and difficult to migrate regionally: Although existing crop process models have good explanatory power for physiological processes, they are highly dependent on initial parameter settings and local conditions, making it difficult to quickly adapt to data conditions in different regions or years, limiting their promotion and application.

[0005] 3. Feature weights are static and time series perception is weak:

[0006] Traditional models (such as random forests) use fixed feature importance, ignoring the fact that the contribution of features at different reproductive stages varies over time. Static feature modeling fails to capture this dynamic change, resulting in unstable predictions and weak generalization.

[0007] 4. Prediction results are disconnected from agricultural recommendations: Most studies are limited to numerical predictions. For example, they give the jointing period as the 85th day, but do not further combine the prediction results to generate personalized agricultural recommendations (such as whether topdressing or pesticides are needed). As a result, the prediction results are difficult to directly guide actual agricultural operations. Summary of the Invention

[0008] In order to solve the problems raised in the background technology, the present invention provides a method and system for winter wheat growth period prediction and intelligent management based on multi-source data.

[0009] The technical solutions of the present invention are as follows:

[0010] The present invention provides a winter wheat growth period prediction and intelligent management method based on multi-source data, comprising:

[0011] S1: Obtain meteorological data from winter wheat sowing to the forecast stage and calculate daily temperature contribution values;

[0012] According to the day length, preset photoperiod and photoperiod sensitivity coefficient, the daily temperature contribution value is adjusted and accumulated day by day to obtain the daily accumulated temperature value;

[0013] S2: Obtain observation data, geographic data, remote sensing data, and soil data from the winter wheat sowing period to the predicted stage, combine them with daily accumulated temperature values, and align them on a daily time scale to form a continuous time series;

[0014] S3: Extract features from the continuous time series features in the time dimension to generate seasonal adjustment vectors, which are then fused with the continuous time series features to obtain seasonal adjustment features.

[0015] In the spatial dimension, the continuous time series features use geographic data to construct spatial embedding vectors, which are then fused with seasonal adjustment features through dot product attention processing to obtain spatiotemporal merged features.

[0016] S4: The spatiotemporal merged features are processed through bilinear interaction to generate a time interaction energy map. The absolute time difference matrix composed of all time step differences is calculated and combined with the preset temperature coefficient parameters. An exponential decay function is applied to obtain the time gating weight. The time gating weight is multiplied by the time interaction energy map and applied to the spatiotemporal merged features to generate a gated fusion feature.

[0017] The gated fusion features are fused with the spatiotemporal merged features after multi-scale convolution processing to obtain dynamic interaction features;

[0018] S5: The dynamic interaction features are processed by the prediction network to obtain the predicted BBCH time series data.

[0019] The S3, continuous time series features, extract features in the time dimension to generate seasonal adjustment vectors, specifically:

[0020] After averaging the features of the continuous time series in the time dimension, the features are extracted and linearly mapped to obtain the score weights of the four seasons. According to the score weights of the four seasons, the predefined four season templates are weighted combined to generate the seasonal adjustment vector.

[0021] Before extracting features in the time dimension, the S3 continuous time series feature further includes:

[0022] Based on the date information of the continuous time series, time embedding vectors at the daily, weekly, and monthly levels are constructed, and then fused with the continuous time series features through linear projection to obtain the spliced ​​time features.

[0023] S1 adjusts the daily temperature contribution value according to the day length, the preset photoperiod and the photoperiod sensitivity coefficient, specifically:

[0024] According to the formula: Calculate the photoperiod regulation factor;

[0025] Where ppfun represents the photoperiod regulation factor; ppsen represents the photoperiod sensitivity coefficient; CPP and L represent the preset photoperiod and day length, respectively;

[0026] The photoperiod adjustment factor was multiplied by the daily temperature contribution to obtain the adjusted daily temperature contribution.

[0027] The S1 calculates the daily temperature contribution value according to the formula:

[0028] accomplish;

[0029] Where DTC(T) is the daily temperature contribution; DTC max is the maximum value contributed by the preset temperature; T is the daily average temperature; T max 、T min Respectively, the preset maximum temperature and the preset minimum temperature; T opt1 、T opt2 They are the preset lower limit temperature and the preset upper limit temperature respectively.

[0030] The S2 growth period observation data includes the BBCH values ​​and their occurrence times at the sowing, seedling, tillering, jointing, heading, flowering, filling or maturity stages;

[0031] Geographic data, including the longitude, latitude, and altitude of the winter wheat site;

[0032] Soil data, including soil moisture and soil temperature.

[0033] The BBCH time series data predicted by S5 are the predicted continuous BBCH values ​​and growth stages.

[0034] After S5, the method further includes: providing recommendations using a language model based on the predicted BBCH time series data.

[0035] The S4 multi-scale convolution processing is specifically as follows:

[0036] The spatiotemporal merged features are processed by multiple sets of parallel channels of dilated convolution kernels with exponential expansion rates and then concatenated to generate multi-scale features;

[0037] Among them, each group of channels undergoes one-dimensional convolution, GELU nonlinear activation function processing, batch normalization, and adaptive average pooling in turn.

[0038] The present invention also provides a winter wheat growth period prediction and intelligent management system based on multi-source data, comprising:

[0039] Meteorological data processing module: used to obtain meteorological data from winter wheat sowing to the forecast stage and calculate daily temperature contribution values;

[0040] According to the day length, preset photoperiod and photoperiod sensitivity coefficient, the daily temperature contribution value is adjusted and accumulated day by day to obtain the daily accumulated temperature value;

[0041] Multi-source data processing module: used to obtain winter wheat growth period observation data, geographic data, remote sensing data, and soil data from sowing to the predicted stage, combined with daily accumulated temperature values, and aligned on a daily time scale to form a continuous time series;

[0042] Spatiotemporal modeling module: Continuous time series features are extracted in the time dimension to generate seasonal adjustment vectors, which are then fused with continuous time series features to obtain seasonal adjustment features.

[0043] In the spatial dimension, the continuous time series features use geographic data to construct spatial embedding vectors, which are then fused with seasonal adjustment features through dot product attention processing to obtain spatiotemporal merged features.

[0044] Dynamic Interaction Module: The spatiotemporal merge features are processed through bilinear interaction to generate a temporal interaction energy map. The absolute time difference matrix composed of all time step differences is calculated and combined with the preset temperature coefficient parameters. An exponential decay function is applied to obtain the time gating weight. The time gating weight is multiplied by the temporal interaction energy map and applied to the spatiotemporal merge features to generate a gated fusion feature.

[0045] The gated fusion features are fused with the spatiotemporal merged features after multi-scale convolution processing to obtain dynamic interaction features;

[0046] Prediction module: The dynamic interaction features are processed by the prediction network to obtain the predicted BBCH time series data.

[0047] Beneficial effects

[0048] The present invention utilizes multi-source heterogeneous data to generate continuous time series, and sequentially performs spatiotemporal modeling and dynamic feature interaction processing on the continuous time series features to predict the growing period of winter wheat. In the spatiotemporal modeling stage, the feature expression capability is enhanced from three aspects: time periodicity, seasonal gating, and geographic space dependence. In the dynamic feature interaction stage, the ability to characterize the time-varying interaction relationship between winter wheat time series variables is effectively improved through energy dynamic coupling. After completing the refined prediction of the daily growth process of winter wheat, the key growth stage turning points (such as tillering, jointing, heading, maturity, etc.) are extracted, and the prediction results are explained in natural language. At the same time, targeted and highly executable agricultural management suggestions are generated to facilitate the guidance of agricultural operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Schematic diagram of the process of step S3 in a method for predicting and intelligently managing the growth period of winter wheat based on multi-source data in an embodiment. DETAILED DESCRIPTION

[0050] The following examples are intended to illustrate the present invention rather than to further limit the present invention.

[0051] The present invention provides a winter wheat growth period prediction and intelligent management method based on multi-source data, comprising:

[0052] S1: Obtain meteorological data from winter wheat sowing to the forecast stage and calculate daily temperature contribution values;

[0053] After adjusting the daily temperature contribution value based on day length, preset photoperiod and photoperiod sensitivity coefficient, it is accumulated day by day to obtain the daily accumulated temperature value.

[0054] First, using meteorological data as the data source, knowledge of winter wheat physiological processes is introduced as a priori features in S1 to guide subsequent predictions.

[0055] Specifically, the daily temperature contribution value is calculated as follows:

[0056] accomplish;

[0057] Where DTC(T) is the daily temperature contribution; DTC max is the maximum value contributed by the preset temperature (such as 20); T is the daily average temperature; T max 、T min The preset maximum temperature (such as 40°C) and the preset minimum temperature (such as 0°C) are respectively; T opt1 、T opt2 They are respectively a preset lower limit temperature (such as 20°C) and a preset upper limit temperature (such as 30°C).

[0058] Considering that winter wheat is very sensitive to photoperiod from the seedling stage to the jointing stage, when the photoperiod is lower than the preset photoperiod (e.g. CPP of 16°C), the crop growth rate slows down, and the daily temperature contribution value needs to be adjusted according to the day length.

[0059] Furthermore, the daily temperature contribution value is adjusted according to the day length, the preset photoperiod and the photoperiod sensitivity coefficient, specifically:

[0060] According to the formula: Calculate the photoperiod regulation factor;

[0061] Where ppfun represents the photoperiod regulation factor; ppsen represents the photoperiod sensitivity coefficient; CPP and L represent the preset photoperiod and day length, respectively;

[0062] The photoperiod adjustment factor was multiplied by the daily temperature contribution to obtain the adjusted daily temperature contribution.

[0063] Among them, the length of day can be calculated based on the latitude and date of the winter wheat site.

[0064] S2: Obtain observation data, geographic data, remote sensing data, and soil data from the winter wheat sowing period to the predicted stage, combine them with daily accumulated temperature values, align them on a daily time scale, and form a continuous time series.

[0065] Among them, the observation data of growth period include BBCH values ​​and their occurrence time at sowing, seedling, tillering, jointing, heading, flowering, grain filling or maturity stages;

[0066] Geographic data, including the longitude, latitude, and altitude of the winter wheat site;

[0067] Remote sensing data, including Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI);

[0068] Soil data, including soil moisture and soil temperature.

[0069] Multi-source heterogeneous data requires preprocessing, including data standardization: unifying data from different sources into a standard format and dimension; outlier removal: using statistical methods to remove extreme abnormal data; missing value filling: using interpolation to fill missing data; and finally, alignment based on the daily time scale to form a continuous time series.

[0070] BBCH is an internationally accepted method for describing plant growth stages. It uses numerical codes to represent different stages of plant growth. For example, 1X represents the leafing stage of a seedling, where X represents the number of leaves; 2X represents the tillering stage. This coding system helps standardize plant growth observation and recording, facilitating cross-regional and cross-research collaboration.

[0071] S3: Extract features from the continuous time series features in the time dimension to generate seasonal adjustment vectors, which are then fused with the continuous time series features to obtain seasonal adjustment features.

[0072] In the spatial dimension, the continuous time series features use geographic data to construct spatial embedding vectors, which are then fused with seasonal adjustment features after dot product attention processing to obtain spatiotemporal merged features.

[0073] During the winter wheat growing period, important meteorological / remote sensing characteristics vary at different stages:

[0074] Early stage (such as sowing and seedling stage): air temperature and initial soil moisture are more critical;

[0075] Jointing and heading stage: sunshine duration and accumulated temperature are more critical;

[0076] Filling and maturity stages: Cumulative changes in precipitation and accumulated temperature are more important.

[0077] If only static feature weights are used, it is impossible to capture the changes in the contribution of dynamic time series features, resulting in decreased prediction accuracy and insufficient interpretability.

[0078] To this end, spatiotemporal modeling is performed in S3 to enhance feature expression capabilities from three aspects: temporal periodicity, seasonal gating, and geographic spatial dependence. Figure 1 shown.

[0079] Firstly, time embedding vectors at the daily, weekly and monthly levels are constructed based on the date information of the continuous time series. After linear projection and fusion with the continuous time series features, the spliced ​​time features are obtained to enhance the periodic features of the continuous time series.

[0080] Next, after averaging the continuous time series features (or spliced ​​time features) over time, features are extracted and linearly mapped to obtain score weights for the four seasons (spring, summer, autumn, and winter). Based on these score weights, the four predefined season templates are weighted and combined to generate a seasonal adjustment vector. This seasonal adjustment vector is fused with the continuous time series features (or spliced ​​time features) to generate the seasonal adjustment feature, thus achieving dynamic fusion of seasonal adjustment features and enhancing the ability to model seasonal patterns.

[0081] At the same time, geospatial information (longitude, latitude, and altitude) is integrated to construct a spatial embedding vector. This embedding vector is constructed using geographic data in the spatial dimension, and through dot-product attention processing, it reflects the spatial dependencies between different geographic locations. This is then fused with seasonal adjustment features to obtain a spatiotemporal merge feature.

[0082] The spatiotemporal merging features enable feature modeling that takes into account both the dynamics of time series and spatial heterogeneity, significantly improving the expressiveness and generalization performance in multi-source data-driven crop time series forecasting tasks.

[0083] In addition, before extracting features in the temporal dimension, the S3 continuous time series features also include mapping the continuous BBCH values ​​of the continuous time series to predefined growth stage labels (such as sowing period, tillering period, and jointing period) for supervised learning. This learning process captures the temporal variations in stage boundaries, captures the structural changes in BBCH values ​​across different growth stages, improves the sensitivity to identifying key transition stages, and enhances the ability to model the temporal structure of the winter wheat growth period, thereby significantly improving the prediction accuracy and generalization performance of the growth period.

[0084] S4: The spatiotemporal merged features are processed through bilinear interaction to generate a time interaction energy map. The absolute time difference matrix composed of all time step differences is calculated and combined with the preset temperature coefficient parameters. An exponential decay function is applied to obtain the time gating weight. The time gating weight is multiplied by the time interaction energy map and applied to the spatiotemporal merged features to generate a gated fusion feature.

[0085] The gated fusion features are fused with the spatiotemporal merged features after multi-scale convolution processing to obtain dynamic interaction features.

[0086] In addition to the spatiotemporal modeling of S3, it also includes the dynamic feature interaction processing of S4 to enhance the time-varying coupling modeling among multiple variables.

[0087] On the one hand, the spatiotemporal merging features are processed by bilinear interaction to generate a time interaction energy spectrum; the absolute time difference matrix composed of all time step differences is calculated, combined with the preset temperature coefficient parameters, and an exponential decay function is applied to obtain the time gating weight. The time gating weight is multiplied by the time interaction energy spectrum to generate a weighted energy map under temperature control, which is then applied to the spatiotemporal merging features to generate gated fusion features.

[0088] For example, the spatiotemporal merged features are expanded into tensors of the form [B, T, 1, D] and [B, 1, T, D], respectively. A bilinear interaction process is performed, where a linear mapping is applied to one of the tensors, followed by element-wise multiplication with the other tensor and summed along the feature dimension. This generates a temporal interaction energy map of the form [B, T, T], which quantifies the feature similarity and interaction strength between different time points. To introduce a temporal prior constraint on cross-timestep interactions, an absolute time difference matrix consisting of all timestep differences is calculated. A preset temperature coefficient parameter is then introduced as a conditioning factor. After multiplying this matrix with the absolute time difference matrix, an exponential decay function is applied to generate a temporal gating weight matrix that decays with increasing time difference, suppressing excessive interference between distant timesteps. The temporal gating weight matrix is ​​multiplied with the temporal interaction energy map to generate a temperature-controlled weighted energy map. This is then normalized to a cross-temporal attention distribution using a Softmax operation. Finally, a weighted summation is applied to the spatiotemporal merged features to generate a gated fusion feature.

[0089] The above operations complete the information aggregation and dynamic interaction of cross-time step features, thereby capturing the deep semantic dependencies between time steps in the time series.

[0090] On the one hand, the spatiotemporal merging features are processed by multi-scale convolution, specifically:

[0091] The spatiotemporal features are processed through multiple parallel channels of dilated convolution kernels with exponential expansion rates (e.g., four channels with expansion rates of 1, 2, 4, and 8, respectively) and then concatenated to generate multi-scale features. Each channel undergoes a one-dimensional convolution, a GELU nonlinear activation function, batch normalization, and adaptive average pooling. Adaptive average pooling maintains consistency in the temporal length of each channel's output, and then concatenates the multi-channel features along the dimensional axis to form a composite feature representation that combines local details with global trends.

[0092] Finally, the gated fusion features are fused with the spatiotemporal merged features after multi-scale convolution processing to obtain the dynamic interaction features.

[0093] Specifically:

[0094] The gated fusion features are concatenated with the spatiotemporal merged features processed by multi-scale convolution. After linear expansion, the feature distribution is stabilized through layer normalization. Gated linear units are used to split the features into two groups and perform element-wise multiplication, achieving compression and nonlinear mapping of deep features. Finally, a residual connection is used to add the processed results to the spatiotemporal merged features, achieving feature dimension alignment while ensuring gradient stability.

[0095] Through the synergistic effect of dilated convolution multi-scale perception, energy dynamic coupling modeling and gated fusion, the ability to characterize the time-varying interaction relationship between winter wheat time series variables is effectively improved.

[0096] S5: The dynamic interaction features are processed by a prediction network, such as a long short-term memory network (LSTM), to obtain the predicted BBCH time series data.

[0097] As a recurrent neural network structure, LSTM can effectively capture the dynamic evolution of patterns between previous and subsequent time steps in the temporal dimension. By gradually transmitting information and using gating mechanisms (input gate, forget gate, output gate), it extracts key temporal features from the input sequence.

[0098] Finally, the hidden state of the last time step in the LSTM output sequence (i.e., the global temporal features at the final moment) is selected as the summary of the sequence representation, which is considered to fully incorporate important dynamic information over the entire time period. This final representation is then input into two fully connected layers: one for predicting the continuous BBCH value (representing the numerical progression of the winter wheat growth period) and the other for outputting the growth stage classification (corresponding to the discrete growth stage label).

[0099] After obtaining the prediction results, the language model is used to give suggestions, thus building a "prediction-explanation-suggestion" linkage mechanism for actual agricultural management scenarios. Specifically:

[0100] Based on the predicted BBCH time series data, target information (such as the timing of growth periods, predicted trends, and plot climate) is extracted and encoded as structured prompts. This information is then fed into a language model designed for agricultural scenarios (such as a fine-tuned DeepSeek-V2 model). Enhanced with agricultural knowledge data, this language model intelligently analyzes current forecast results, identifies crop stages, and generates scientifically sound and actionable management recommendations.

[0101] For example, after the model predicts that the crop is currently entering the late tillering stage, the language model can generate information such as: "We are currently in the transition stage from the peak tillering stage to the early jointing stage. It is recommended to increase the supply of nitrogen fertilizer and keep the field moist to facilitate effective tillering formation", or "The heading time is expected to be delayed by 5 days this season. Please pay attention to the impact of cold air and carry out foliar fertilization and disease warnings in a timely manner."

[0102] Among them, in order for the language model to reasonably explain the prediction results, it is necessary to additionally build a winter wheat field knowledge base, covering agricultural knowledge text materials such as the climate requirements of different winter wheat growth stages, fertilizer and water management recommendations, and key points for pest and disease prevention and control.

[0103] This knowledge base includes the following:

[0104] The physiological and ecological characteristics and management key points of winter wheat at different stages; recommendations for zoning management under various domestic planting patterns; summary of national agricultural technology extension information and historical agricultural data.

[0105] The present invention utilizes multi-source heterogeneous data to generate continuous time series, and sequentially performs spatiotemporal modeling and dynamic feature interaction processing on the continuous time series features to predict the growing period of winter wheat. In the spatiotemporal modeling stage, the feature expression capability is enhanced from three aspects: time periodicity, seasonal gating, and geographic space dependence. In the dynamic feature interaction stage, the energy dynamic coupling modeling is used to effectively improve the characterization capability of the time-varying interaction relationship between the time series variables of winter wheat. After completing the refined prediction of the daily growth process of winter wheat, the key growth stage turning points (such as tillering, jointing, heading, maturity, etc.) are extracted, and the prediction results are explained in natural language. At the same time, targeted and highly executable agricultural management suggestions are generated to facilitate the guidance of agricultural operations.

[0106] The present invention also provides a winter wheat growth period prediction and intelligent management system based on multi-source data, comprising:

[0107] Meteorological data processing module: used to obtain meteorological data from winter wheat sowing to the forecast stage and calculate daily temperature contribution values;

[0108] According to the day length, preset photoperiod and photoperiod sensitivity coefficient, the daily temperature contribution value is adjusted and accumulated day by day to obtain the daily accumulated temperature value;

[0109] Multi-source data processing module: used to obtain winter wheat growth period observation data, geographic data, remote sensing data, and soil data from sowing to the predicted stage, combined with daily accumulated temperature values, and aligned on a daily time scale to form a continuous time series;

[0110] Spatiotemporal modeling module: Continuous time series features are extracted in the time dimension to generate seasonal adjustment vectors, which are then fused with continuous time series features to obtain seasonal adjustment features.

[0111] In the spatial dimension, the continuous time series features use geographic data to construct spatial embedding vectors, which are then fused with seasonal adjustment features through dot product attention processing to obtain spatiotemporal merged features.

[0112] Dynamic Interaction Module: The spatiotemporal merge features are processed through bilinear interaction to generate a temporal interaction energy map. The absolute time difference matrix composed of all time step differences is calculated and combined with the preset temperature coefficient parameters. An exponential decay function is applied to obtain the time gating weight. The time gating weight is multiplied by the temporal interaction energy map and applied to the spatiotemporal merge features to generate a gated fusion feature.

[0113] The gated fusion features are fused with the spatiotemporal merge features after multi-scale convolution processing to obtain dynamic interaction features; prediction module: the dynamic interaction features are processed by the prediction network to obtain the predicted BBCH time series data.

Claims

1. A method for predicting and intelligently managing the growth period of winter wheat based on multi-source data, characterized in that: include: S1: Obtain meteorological data from winter wheat sowing to the forecast stage and calculate daily temperature contribution values; According to the day length, preset photoperiod and photoperiod sensitivity coefficient, the daily temperature contribution value is adjusted and accumulated day by day to obtain the daily accumulated temperature value; S2: Obtain observation data, geographic data, remote sensing data, and soil data from the winter wheat sowing period to the predicted stage, combine them with daily accumulated temperature values, and align them on a daily time scale to form a continuous time series; S3: Extract features from the continuous time series features in the time dimension to generate seasonal adjustment vectors, which are then fused with the continuous time series features to obtain seasonal adjustment features. In the spatial dimension, the continuous time series features use geographic data to construct spatial embedding vectors, which are then fused with seasonal adjustment features through dot product attention processing to obtain spatiotemporal merged features. S4: The spatiotemporal merged features are processed through bilinear interaction to generate a time interaction energy map. The absolute time difference matrix composed of all time step differences is calculated and combined with the preset temperature coefficient parameters. An exponential decay function is applied to obtain the time gating weight. The time gating weight is multiplied by the time interaction energy map and applied to the spatiotemporal merged features to generate a gated fusion feature. The gated fusion features are fused with the spatiotemporal merged features after multi-scale convolution processing to obtain dynamic interaction features; S5: The dynamic interaction features are processed by the prediction network to obtain the predicted BBCH time series data.

2. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1, characterized in that: The S3, continuous time series features, extract features in the time dimension to generate seasonal adjustment vectors, specifically: After averaging the features of the continuous time series in the time dimension, the features are extracted and linearly mapped to obtain the score weights of the four seasons. According to the score weights of the four seasons, the predefined four season templates are weighted combined to generate the seasonal adjustment vector.

3. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1, characterized in that: Before extracting features in the time dimension, the S3 continuous time series feature further includes: Based on the date information of the continuous time series, time embedding vectors at the daily, weekly, and monthly levels are constructed, and then fused with the continuous time series features through linear projection to obtain the spliced ​​time features.

4. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1, characterized in that: S1 adjusts the daily temperature contribution value according to the day length, the preset photoperiod and the photoperiod sensitivity coefficient, specifically: According to the formula: Calculate the photoperiod regulation factor; Where, ppfun represents the photoperiod regulation factor; ppsen represents the photoperiod sensitivity coefficient; CPP and L represent preset photoperiod and day length, respectively; The photoperiod adjustment factor was multiplied by the daily temperature contribution to obtain the adjusted daily temperature contribution.

5. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1 is characterized in that: The S1 calculates the daily temperature contribution value according to the formula: accomplish; Where DTC(T) is the daily temperature contribution; DTC max is the maximum value contributed by the preset temperature; T is the daily average temperature; T max 、T min Respectively, the preset maximum temperature and the preset minimum temperature; T opt1 、T opt2 They are the preset lower limit temperature and the preset upper limit temperature respectively.

6. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1, characterized in that: The S2 growth period observation data includes the BBCH values ​​and their occurrence times at the sowing, seedling, tillering, jointing, heading, flowering, filling or maturity stages; Geographic data, including the longitude, latitude, and altitude of the winter wheat site; Soil data, including soil moisture and soil temperature.

7. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1, characterized in that: The BBCH time series data predicted by S5 are the predicted continuous BBCH values ​​and growth stages.

8. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1, characterized in that: After S5, the method further includes: providing suggestions using a language model based on the predicted BBCH time series data.

9. The winter wheat growth period prediction and intelligent management method based on multi-source data according to claim 1, characterized in that: The S4 multi-scale convolution processing is specifically as follows: The spatiotemporal merged features are processed by multiple sets of parallel channels of dilated convolution kernels with exponential expansion rates and then concatenated to generate multi-scale features; Among them, each group of channels undergoes one-dimensional convolution, GELU nonlinear activation function processing, batch normalization, and adaptive average pooling in turn.

10. A winter wheat growth period prediction and intelligent management system based on multi-source data, characterized in that: include: Meteorological data processing module: used to obtain meteorological data from winter wheat sowing to the forecast stage and calculate daily temperature contribution values; According to the day length, preset photoperiod and photoperiod sensitivity coefficient, the daily temperature contribution value is adjusted and accumulated day by day to obtain the daily accumulated temperature value; Multi-source data processing module: used to obtain winter wheat growth period observation data, geographic data, remote sensing data, and soil data from sowing to the predicted stage, combined with daily accumulated temperature values, and aligned on a daily time scale to form a continuous time series; Spatiotemporal modeling module: Continuous time series features are extracted in the time dimension to generate seasonal adjustment vectors, which are then fused with continuous time series features to obtain seasonal adjustment features. In the spatial dimension, the continuous time series features use geographic data to construct spatial embedding vectors, which are then fused with seasonal adjustment features through dot product attention processing to obtain spatiotemporal merged features. Dynamic Interaction Module: The spatiotemporal merge features are processed through bilinear interaction to generate a temporal interaction energy map. The absolute time difference matrix composed of all time step differences is calculated and combined with the preset temperature coefficient parameters. An exponential decay function is applied to obtain the time gating weight. The time gating weight is multiplied by the temporal interaction energy map and applied to the spatiotemporal merge features to generate a gated fusion feature. The gated fusion features are fused with the spatiotemporal merged features after multi-scale convolution processing to obtain dynamic interaction features; Prediction module: The dynamic interaction features are processed by the prediction network to obtain the predicted BBCH time series data.

Images