Forage grass yield prediction model construction method based on deep learning

Abstract

The invention relates to the technical field of deep learning, in particular to a forage grass yield prediction model construction method based on deep learning, which comprises the steps of constructing a multi-source data fusion module, establishing a cold start mechanism, designing a multi-modal deep learning prediction network, integrating a physical constraint mechanism and constructing a management decision support system. A seasonal attribution analysis function is realized; in the prior art, a simple data superposition or static weighted fusion scheme is generally adopted, and inherent defects of deficiency, different scales and heterogeneity of multi-source data are difficult to process, so that the fusion feature quality is poor; according to the method, firstly, a data blank is accurately filled through an intelligent algorithm based on space-time continuity, then heterogeneous data is unified to a standard grid by using a multi-scale pyramid engine, and finally, deep fusion is performed through an attention mechanism for dynamically calculating importance of each data source; the integrity, the consistency and the information density of the input data are remarkably improved, and a solid and reliable data foundation is laid for subsequent accurate prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a forage yield prediction model construction method based on deep learning. BACKGROUND

[0002] Forage yield prediction is an important basis for livestock production management and national food security strategic decision-making. Accurate prediction of forage yield is of great significance for ensuring the supply of livestock feed, stabilizing the grass product market, optimizing resource allocation, and responding to extreme climate risks. Traditional prediction methods mainly rely on time series extrapolation of historical statistical data, empirical models based on meteorological indicators, and regression analysis based on remote sensing vegetation indices. These methods can reflect the macro trend of yield to some extent.

[0003] However, existing prediction techniques have obvious limitations. First, they often rely on a single data source, such as using only remote sensing images or meteorological data, failing to fully utilize the complementary information between multi-source heterogeneous data, resulting in a bottleneck in prediction accuracy. Second, when facing newly planted areas with missing historical data, traditional models fail completely due to lack of training samples, i.e., there is a serious "cold start" problem. Third, mainstream data-driven prediction models, especially deep learning models, sometimes produce abnormal prediction values that violate plant physiology rules, have poor interpretability, and the prediction results cannot be effectively combined with specific agricultural production management measures, making it difficult to directly guide production practice. These shortcomings collectively limit the widespread application and value of forage yield prediction technology in practical business.

[0004] The present application aims to systematically solve the above problems. The present application proposes a forage yield prediction model construction method based on deep learning. The present application first integrates remote sensing, meteorological, soil, and historical yield data to form high-quality input; then initializes the new plot model through similar plot matching and transfer learning technology; further uses advanced deep learning models for accurate prediction and introduces physical mechanisms to ensure the rationality of the prediction results; finally, the prediction results are converted into operational field management recommendations and have the ability to attribute analysis of prediction bias, thereby realizing the closed loop from data to decision-making. SUMMARY

[0005] In order to overcome the problems proposed in the background art, the present application proposes a forage yield prediction model construction method based on deep learning.

[0006] The technical solution of the present application is: a forage yield prediction model construction method based on deep learning, comprising the following steps: S11: Construct a multi-source data fusion module integrating remote sensing data, meteorological data, soil data, and historical yield data. The multi-source data fusion module includes a data quality control submodule for handling data missing and incomplete cases; S12: Establish a cold start mechanism. For newly added plots lacking historical planting data, model initialization is achieved through similar plot matching and parameter migration learning. The cold start mechanism includes spatial proximity analysis and feature similarity calculation; S13: Design a multi-modal deep learning prediction network, including remote sensing data encoding branch, meteorological time series encoding branch, soil feature encoding branch, and multi-modal feature fusion module, outputting short-term monthly yield prediction and total yield prediction at the end of the season; S14: Integrate a physical constraint mechanism. Based on the forage grass growth physiological model and historical climate benchmarks, reasonable range constraints are imposed on model prediction results to prevent abnormal prediction values; S15: Construct a management decision support system to automatically generate irrigation, fertilization, and weeding management recommendations based on prediction results, soil parameters, and future weather forecasts; S16: Implement a season-end attribution analysis function. By comparing the differences between actual yield and predicted yield, and combining parameter changes during the growing season, key influencing factors of yield differences are identified.

[0007] As preferred, when constructing the multi-source data fusion module, the multi-source data fusion module specifically includes: A remote sensing data submodule for processing Sentinel-2, Landsat-8, and MODIS satellite images, extracting vegetation indices and biomass features; A meteorological data submodule integrating temperature, precipitation, sunshine duration, and humidity data to support time series feature extraction; A soil data submodule for integrating soil type, pH value, organic matter content, and water content information; A historical yield database for storing historical yield records at the plot level; A data quality control unit for filling gaps in remote sensing data using a spatiotemporal feature-based cloud gap filling algorithm, and using a time series interpolation method to complete meteorological data missing.

[0008] As preferred, when constructing the multi-source data fusion module, the following steps are specifically included: S21: Establish a multi-source data access interface to receive remote sensing data, meteorological data, soil data, and historical yield data; S22: Design a data quality control submodule that adopts differentiated processing strategies for different types of data missing; S23: Build a spatiotemporal alignment engine to unify multi-source heterogeneous data to the same spatiotemporal scale; S24: Implement feature-level data fusion to generate a fused feature tensor for the prediction model; S25: Establish a data quality evaluation system to quantify the credibility of each data source and the fusion effect.

[0009] As a preferred embodiment, the cold start mechanism includes the following steps: S31: Establish a plot feature space and extract geographic location, soil properties, topographic features, and climate zoning multi-dimensional feature vectors; S32: Design a composite similarity measurement function that considers both spatial proximity and feature similarity; S33: Implement a parameter transfer learning algorithm to initialize the model of a new plot based on the parameters of similar plots; S34: Build an incremental learning mechanism to dynamically adjust the transfer parameters using real-time data during the growing season.

[0010] As a preferred embodiment, when designing a composite similarity measurement function that considers both spatial proximity and feature similarity, the composite similarity measurement function adopts a space-feature dual constraint form, and the principle formula is: ; Where, is the geographic distance between plot A and B, is the spatial scale parameter, is the feature space similarity, is the temperature parameter, is the number of candidate similar plots.

[0011] As a preferred embodiment, when designing a multi-modal deep learning prediction network, the multi-modal deep learning prediction network adopts the following architecture: The remote sensing data encoding branch uses a multi-scale convolutional neural network to extract spatial features and process satellite image data; The meteorological time series encoding branch uses a gated recurrent unit network to capture temporal dependencies and process meteorological time series data; The soil feature encoding branch uses a fully connected network to process static soil attribute features; The multi-modal feature fusion module dynamically weights the contribution of each modal feature based on a cross-attention mechanism; The dual-output prediction head generates monthly yield sequence predictions and end-of-season total yield predictions, respectively.

[0012] As a preferred embodiment, the cross-attention mechanism of the multi-modal feature fusion module adopts the following attention weight calculation method: For the i-th modal query vector , the j-th modal key-value pair Calculate cross-attention: ; wherein the inter-modal prior matrix is calculated as: ; wherein is the key vector of the j-th modality, is the value vector of the j-th modality, , and is the learnable projection weight matrix, is the dimension of the key vector, is the dot product similarity between the query and the key, is the projected value vector, is the historical correlation coefficient between modalities i and j, is the modality feature distribution similarity, is the modality complementarity measure, , and are learnable weight parameters.

[0013] As preferred, when integrating the physical constraint mechanism, the following steps are included: S41: Establish a forage grass physiological growth model based on historical climate benchmarks to determine the appropriate range of key environmental factors at each growth stage; S42: Calculate the deviation of the current growing season environmental factors from the historical benchmarks to quantify the potential impact of climate anomalies on yield; S43: Design a physical constraint layer to constrain the deep learning model prediction results within the reasonable fluctuation range determined by the physiological model; S44: Implement a dynamic constraint adjustment mechanism to adaptively adjust the constraint strength according to the growth stage and crop type.

[0014] As preferred, when constructing the management decision support system, the management decision support system adopts a multi-objective optimization algorithm: with maximizing yield and minimizing cost as optimization objectives; considering water resource utilization efficiency, fertilizer utilization efficiency and economic benefit constraints; combining future weather forecast data to dynamically adjust management recommendations; outputting irrigation time and water quantity, fertilization type and quantity, and specific herbicide timing schemes.

[0015] As preferred, when implementing the end-of-season attribution analysis function, it specifically includes: S51: Establish a yield difference decomposition model to decompose the difference between actual yield and predicted yield into three components: environmental factor contribution, management measure contribution, and model error; S52: design a timing contribution tracking algorithm to quantify the influence of key factors at each growth stage on the difference in final yield; S53: realize multi-dimensional attribution analysis to comprehensively analyze the reasons for yield difference from the angles of climate anomaly, management effect and model bias; S54: generate an interpretable attribution report to provide improvement suggestions for next season planting decisions.

[0016] Advantages of the present application: 1. Compared with the simple data superposition or static weighted fusion scheme commonly used in the prior art, the inherent defects of multi-source data such as missing, different scales and heterogeneity are difficult to handle, resulting in poor fusion feature quality; the present application adopts a complete framework integrating advanced quality control, spatio-temporal alignment and attention weighted fusion; the scheme first accurately fills in the data gaps through intelligent algorithms based on spatio-temporal continuity, then uses a multi-scale pyramid engine to unify heterogeneous data such as remote sensing, weather, soil, etc. to standard grids, and finally performs deep fusion through an attention mechanism that dynamically calculates the importance of each data source; this scheme significantly improves the integrity, consistency and information density of the input data, laying a solid and reliable data foundation for subsequent accurate prediction; 2. Compared with the prior art prediction model which relies heavily on historical yield data of the plot itself for training, resulting in its inability to effectively predict newly added plots without any planting records; the present application designs a cold start mechanism combining similar plot matching and parameter transfer learning; the scheme calculates the composite similarity of the newly added plot and the historical plot in space and features, finds the best reference plot, and uses an elastic weight consolidation strategy to safely migrate the knowledge of the trained model to the new model, while dynamically optimizing with the season through incremental learning; this mechanism enables the model to quickly adapt to the new environment, effectively solving the initial data dependence problem during system expansion, and realizing rapid deployment and generalization of prediction ability; 3. Compared with the prior art which mostly uses a single model or simple splicing method to process multi-source data, it is difficult to capture the complex interaction between different modal data and the spatio-temporal dynamic characteristics; the present application designs a multi-modal deep learning network with exquisite structure; the network not only uses dedicated sub-networks to extract spatial features of remote sensing images, time series features of weather data and soil attribute features, but also innovatively introduces a fusion module based on cross-attention mechanism, which can intelligently evaluate and weight the contribution of different data sources to prediction; this scheme fully taps the complementary value of multi-source information, realizing more accurate short-term dynamic tracking and long-term trend prediction; 4、The existing pure data-driven deep learning model sometimes produces abnormal prediction values that violate the common sense of plant physiology, affecting the credibility and usability of the model in actual production; the present application creatively integrates a physical constraint layer; this scheme combines the physiological model of forage grass growth with the deep learning model, calculates the deviation of environmental factors from historical benchmarks, and uses the growth response function to constrain the model prediction value within a physiologically reasonable fluctuation range; this mechanism integrates domain knowledge into the black box model, ensuring the scientificity and rationality of the prediction results, greatly enhancing the reliability and practical value of the model output; 5、The existing prediction system often only provides yield results, failing to convert prediction information into specific production guidance; the present application constructs an intelligent management decision support system based on multi-objective optimization; this scheme takes yield maximization and resource cost minimization as the goal, integrates future weather, soil conditions and prediction results, and automatically generates a precision management scheme including irrigation, fertilization and weeding; this system realizes a closed loop from prediction to decision, directly converting data analysis results into executable production instructions, significantly improving the fine-grained level of agricultural management and economic benefits; 6、The existing model usually only gives the prediction value and lacks the ability to explain the difference in results, making it difficult for users to understand the reasons for the prediction deviation and improve management; the present application develops a deep end-of-season attribution analysis function; this scheme uses an improved attribution algorithm to decompose the difference between actual yield and prediction value into different environmental factors and management measures, and can trace the influence of each factor throughout the growing season; this function provides strong model interpretability, helping users accurately locate the problem, and provides a scientific basis for optimizing production strategies in the next season. BRIEF DESCRIPTION OF DRAWINGS

[0017] Fig. 1 The first three-dimensional structure schematic diagram of the forage grass yield prediction model construction method based on deep learning of the present application is shown. Fig. 2 The second three-dimensional structure schematic diagram of the forage grass yield prediction model construction method based on deep learning of the present application is shown. DETAILED DESCRIPTION

[0018] The present application will be further described below in conjunction with the drawings and examples.

[0019] Please refer to Figs. 1-2 The present application provides an embodiment: a forage grass yield prediction model construction method based on deep learning, comprising the following steps: Step one A multi-source data fusion module is constructed to integrate remote sensing data, meteorological data, soil data and historical yield data. The multi-source data fusion module includes a data quality control submodule for processing data missing and incomplete conditions. Specifically, the multi-source data fusion module includes: a remote sensing data submodule for processing Sentinel-2, Landsat-8 and MODIS satellite images, and extracting vegetation index and biomass features; a meteorological data submodule for integrating temperature, precipitation, sunshine duration and humidity data to support time series feature extraction; a soil data submodule for integrating soil type, pH value, organic matter content and water content information; a historical yield database for storing historical yield records at the plot level; a data quality control unit for processing remote sensing data missing by using a cloud gap filling algorithm based on spatio-temporal features, and using a time series interpolation method to complete meteorological data missing.

[0020] In this embodiment, when constructing the multi-source data fusion module, the following steps are specifically included: S21: Establish a multi-source data access interface to receive remote sensing data, meteorological data, soil data and historical yield data respectively; S22: Design a data quality control submodule to adopt differentiated processing strategies for different types of data missing; S23: Construct a spatio-temporal alignment engine to unify multi-source heterogeneous data to the same spatio-temporal scale; S24: Realize feature-level data fusion to generate a fusion feature tensor for use by the prediction model; S25: Establish a data quality evaluation system to quantify the reliability of each data source and the fusion effect.

[0021] In this embodiment, when designing the data quality control submodule, the data quality control submodule adopts a cloud gap filling algorithm based on spatio-temporal continuity, and the principle formula is: ; wherein, is a weight coefficient, and its calculation formula is: ; wherein, is the estimated value of the pixel to be filled at time t, is the value of the adjacent valid pixel, is the spatial distance, is the time distance, is the spatial scale parameter, is the time scale parameter, is the position of the pixel to be filled, ​the extracted multi-dimensional feature vector, is a feature similarity function.

[0022] Preferably, the feature similarity function adopts a modified cosine similarity, and the principle formula is: ; wherein, and are adjustment parameters for balancing the contributions of direction similarity and distance similarity, and represent two position feature vectors, represents an exponential function.

[0023] In the embodiment, when constructing the spatio-temporal alignment engine, the spatio-temporal alignment engine adopts a multi-scale pyramid alignment algorithm: establish a spatio-temporal pyramid structure, and align step by step from coarse to fine; spatial alignment adopts a feature point-based registration method, and the registration error function is: ; wherein, is the total alignment error, , and are weight coefficients, is a feature point matching error, is an intensity similarity error, is a smoothness constraint error; temporal alignment adopts a dynamic time warping algorithm to process sequence data of different time resolutions.

[0024] In the embodiment, the feature-level data fusion adopts an attention-weighted multi-modal fusion mechanism, specifically: for the feature vector of the i-th modality , the fusion weight is calculated as: ; wherein, is the attention weight of the i-th modality, is a context vector, is the number of modalities, represents a multi-layer perception; the final fusion feature is: ; wherein, is a linear transformation matrix of the i-th modality.

[0025] wherein, the data quality evaluation system includes a credibility propagation model: single data source credibility calculation: ; Fusion result uncertainty propagation: ; wherein, is the credibility of the kth data source, is the coverage rate, is the accuracy rate, is the timeliness, , and are weight coefficients, is the total uncertainty of the fusion result, is the fusion function, is the input value of the kth data source, is the uncertainty of the kth data source, is the correlation coefficient between the ith data source and the jth data source, is the uncertainty of the ith data source.

[0026] Specifically, the spatio-temporal interpolation of historical yield data adopts an adaptive Kriging method.

[0027] In the embodiment, the application realizes effective integration of remote sensing, meteorological, soil and historical yield data by constructing a multi-source data fusion module, using a cloud gap filling algorithm based on spatio-temporal continuity and an improved feature similarity calculation; solves the problems of scale unification and feature fusion of multi-source heterogeneous data through multi-scale pyramid spatio-temporal alignment and attention weighted multi-modal fusion mechanism; at the same time, a data quality evaluation system including credibility evaluation and uncertainty propagation is established, which significantly improves the reliability and accuracy of data fusion, provides high-quality, multi-dimensional and spatio-temporally consistent feature input for subsequent yield prediction models, and effectively overcomes the problems of data missing, scale inconsistency and poor fusion effect in traditional methods, and lays a solid data foundation for accurate prediction of forage grass yield.

[0028] Step two A cold start mechanism is established, and for newly added plots lacking historical planting data, model initialization is realized through similar plot matching and parameter transfer learning. The cold start mechanism includes spatial proximity analysis and feature similarity calculation, and specifically includes the following steps: S31: Establish a plot feature space and extract a multi-dimensional feature vector of geographical location, soil property, topographic feature and climate zoning; S32: Design a composite similarity measure function, considering spatial proximity and feature similarity; S33: Implement a parameter transfer learning algorithm to initialize a new plot model based on the model parameters of similar plots; S34: Build an incremental learning mechanism to dynamically adjust the transferred parameters using real-time data in the growing season.

[0029] In the design of the composite similarity measure function, considering the spatial proximity and feature similarity, the composite similarity measure function adopts a space-feature double constraint form, and the principle formula is: ; wherein, is the geographical distance between plots A and B, is the spatial scale parameter, is the feature space similarity, is the temperature parameter, is the number of candidate similar plots.

[0030] In this embodiment, the feature space similarity calculation adopts an adaptive weighted Mahalanobis distance: ; wherein, and are the feature vectors of plots A and B, is the feature covariance matrix, is the feature importance weight matrix, represents the matrix transposition operation.

[0031] Specifically, the parameter transfer learning adopts an elastic weight consolidation strategy, and the loss function is designed as: ; wherein, is the loss function of the elastic weight consolidation strategy, is the loss function of the new plot data, is the diagonal element of the Fisher information matrix, representing the parameter importance, is the i-th element of the current model parameter, is the original parameter value of the migration source, is the regularization intensity coefficient.

[0032] Specifically, the incremental learning mechanism adopts a sliding window online learning: Set the time window , update the model parameters at each time step t: ; wherein, is the model parameter at the current time step t, is the model parameter at the last time step, is the learning rate, is the data batch within the time window t, is the memory retention coefficient, is the cold start initial parameter.

[0033] In this embodiment, the application realizes accurate matching of new land plots and historical land plots by establishing a land plot feature space and multi-dimensional feature extraction, combining a spatial-feature double-constrained composite similarity measurement function; adopts a parameter migration learning algorithm of an elastic weight consolidation strategy to effectively balance old knowledge retention and new data adaptation when initializing a new land plot model, and prevents catastrophic forgetting; through an incremental learning mechanism of a sliding window online learning, dynamically adjusts model parameters using real-time data in the growing season, and realizes smooth transition from cold start to hot operation. The cold start mechanism significantly improves the prediction accuracy and convergence speed of the model on new land plots, solves the problem of strong dependence on historical data and slow adaptation to new land plots of traditional methods, and provides key technical support for rapid deployment and expansion of forage grass yield prediction systems.

[0034] Step three A multi-modal deep learning prediction network is designed, which includes a remote sensing data encoding branch, a meteorological time series encoding branch, a soil feature encoding branch, and a multi-modal feature fusion module, outputs short-term monthly yield prediction and total yield prediction at the end of the season; the multi-modal deep learning prediction network adopts the following architecture: The remote sensing data encoding branch uses a multi-scale convolutional neural network to extract spatial features and process satellite image data; The meteorological time series encoding branch uses a gated recurrent unit network to capture time dependence and process meteorological time series data; The soil feature encoding branch uses a fully connected network to process static soil attribute features; The multi-modal feature fusion module dynamically weights the contribution of each modal feature based on a cross-attention mechanism; The double-output prediction head generates monthly yield sequence prediction and total yield prediction at the end of the season.

[0035] As a preferred embodiment, the cross-attention mechanism of the multi-modal feature fusion module adopts the following attention weight calculation method: For the i-th modal query vector , the j-th modal key-value pair Calculate cross-attention: ; Where the inter-modal prior matrix is calculated as follows: ; Where is the key vector of the j-th modal, is the value vector of the j-th modal, , and are learnable projection weight matrices, is the dimension of the key vector, is the dot product similarity of the query and the key, is the projected value vector, is the historical correlation coefficient of modal i and j, is the modal feature distribution similarity, is the modal complementarity measure, , and are learnable weight parameters.

[0036] Specifically, the dual-output prediction head adopts a multi-task joint loss function: The total loss function is designed as: ; where, is the total loss function of the multi-task joint loss function, is the smooth L1 loss of monthly sequence prediction, is the mean square error loss of total output at the end of the quarter, is the consistency constraint loss, , and are task weight coefficients.

[0037] Specifically, the remote sensing data encoding branch adopts a multi-scale feature pyramid structure: Constructing a multi-scale representation from high-resolution detail features to low-resolution semantic features; Fusing feature maps of different scales through horizontal connection; Using a dilated convolution to expand the receptive field without increasing the number of parameters.

[0038] Specifically, the meteorological time series encoding branch integrates meteorological forecast uncertainty: Introducing an uncertainty gating mechanism in the GRU network; Adjusting the historical memory weight dynamically according to the weather forecast reliability; Outputting probabilistic time series features containing prediction intervals.

[0039] Specifically, the multi-modal feature fusion module contains a feature interaction enhancement mechanism: Calculating the inter-modal feature affinity through cross-modal attention; Using a graph neural network to model the complex inter-modal dependency; Achieving information complementarity and redundancy elimination at the feature level.

[0040] In this embodiment, the present application realizes efficient extraction of multi-source heterogeneous features by constructing a multi-modal deep learning prediction network, using multi-scale convolutional neural networks, gated recurrent units and fully connected networks to process remote sensing, meteorological and soil data respectively; innovatively designs a multi-modal fusion module based on cross-attention mechanism, introduces historical correlation, feature distribution similarity and modal complementarity to form a prior matrix, significantly improves the intelligence and accuracy of feature fusion; uses a double-output prediction head combined with a multi-task joint loss function to realize the collaborative prediction of short-term monthly yield sequence and total yield at the end of the season, and ensures the internal logical consistency of the prediction results of different time scales through consistency constraints. The network architecture fully utilizes the complementary advantages of multi-source data, effectively captures the spatio-temporal dynamic characteristics of forage grass growth, greatly improves the accuracy and reliability of yield prediction, and provides a powerful decision support tool for pasture production management.

[0041] Step four An integrated physical constraint mechanism is used to impose reasonable range constraints on the model prediction results based on the forage grass growth physiological model and historical climate benchmarks to prevent abnormal prediction values from being generated, including the following steps: S41: Establish a forage grass physiological growth model based on historical climate benchmarks to determine the suitable range of key environmental factors in each growth stage; S42: Calculate the deviation of the current growing season environmental factors from the historical benchmarks to quantify the potential impact of climate anomalies on yield; S43: Design a physical constraint layer to constrain the deep learning model prediction results within the reasonable fluctuation range determined by the physiological model; S44: Implement a dynamic constraint adjustment mechanism to adaptively adjust the constraint strength according to the growth stage and crop type.

[0042] In this embodiment, the physical constraint layer uses an adaptive constraint algorithm based on the growth response function, specifically: The constrained prediction value is calculated as: ; Where, is the growth response function of the ith environmental factor, defined as: ; Where, is the constrained prediction value, is the original prediction value of the deep learning model, is the deviation of the ith environmental factor from the historical benchmark, is the tolerance threshold of the ith environmental factor, is the constraint strength parameter, is the response curve shape parameter, is the indicator function, for fine-tuning the coefficients of the terms, for historical baseline yield.

[0043] Specifically, the environmental factor deviation The standardized relative deviation calculation is adopted, and the formula is: wherein, is the current observation value of the i-th environmental factor, is the historical baseline value of the i-th environmental factor, is the sensitivity coefficient of the i-th environmental factor to the forage grass, is the standard deviation of the historical data.

[0044] Specifically, the physiological growth model is established based on the light-temperature-water-soil production potential theory, and comprehensively considers photosynthesis, temperature effect, water balance and nutrient supply to determine the upper and lower limits of the theoretical yield.

[0045] Specifically, the dynamic constraint adjustment mechanism includes: The vegetative growth period is relatively loose, and a large fluctuation range is allowed; The reproductive growth period is strictly constrained to ensure stability during the key period of yield formation; The tolerance threshold of each environmental factor is adjusted according to the characteristics of the crop variety.

[0046] Specifically, the tolerance threshold The tolerance threshold is determined by quantile regression of historical yield and environmental factors, and the critical value of the environmental factor corresponding to the yield mutation point is taken as the physiological tolerance boundary.

[0047] In this embodiment, the present application establishes an intelligent physical constraint mechanism by constructing a forage grass physiological growth model based on historical climate benchmarks, combining standardized environmental factor deviation calculation and growth response function; the adaptive constraint algorithm is used to limit the deep learning prediction results within the physiological reasonable fluctuation range, and through the dynamic adjustment of the tolerance threshold and the constraint strength parameter, the effective filtering of abnormal prediction values is realized; at the same time, according to the characteristics of crop varieties and the differences of growth stages, the constraint strength is adaptively adjusted, which ensures the accuracy and applicability of the constraint mechanism. The physical constraint layer combines plant physiology knowledge and deep learning model organically, significantly improves the rationality and reliability of the prediction results, effectively prevents the absurd prediction of the model due to data deviation or insufficient training, and provides reliable physiological guarantee and business application value for forage yield prediction.

[0048] Step five A management decision support system is constructed to automatically generate irrigation, fertilization and weeding management suggestions according to the prediction results, soil parameters and future weather prediction. Specifically, the management decision support system adopts a multi-objective optimization algorithm: ​Maximizing yield and minimizing cost as optimization goals; Considering water resource utilization efficiency, fertilizer utilization efficiency and economic benefit constraints; Combining future weather forecast data, dynamically adjusting management recommendations; The output includes irrigation time and water quantity, fertilizer type and quantity, and specific herbicide application scheme.

[0049] In this embodiment, the present application constructs a management decision support system based on multi-objective optimization, takes yield maximization and cost minimization as core goals, comprehensively considers multiple constraints such as water resource utilization efficiency, fertilizer utilization efficiency and economic benefit, and dynamically generates precise irrigation, fertilization and weeding management schemes combined with future weather forecast data. Through intelligent algorithm optimization of resource allocation, the closed-loop application from prediction to decision is realized, scientific, economic and operable management recommendations are provided for forage grass production, the fine level of production management and resource utilization efficiency are significantly improved, and the production cost and environmental impact are effectively reduced, which brings significant economic and ecological benefits to herdsmen and enterprises.

[0050] Step six The end-of-season attribution analysis function is realized, the key influencing factors of yield difference are identified by comparing the difference between actual yield and predicted yield combined with parameter changes in the growing season, which specifically includes: S51: Establish a yield difference decomposition model to decompose the difference between actual yield and predicted yield into three components: environmental factor contribution, management measure contribution and model error; S52: Design a time-series contribution tracking algorithm to quantify the influence degree of key factors at each growth stage on the final yield difference; S53: Realize multi-dimensional attribution analysis to comprehensively analyze the reasons for yield difference from the angles of climate anomaly, management effect and model bias; S54: Generate an interpretable attribution report to provide improvement suggestions for next season planting decisions.

[0051] In this embodiment, the yield difference decomposition adopts an improved attribution algorithm based on Shapley value, specifically: The yield difference decomposition formula is:

[0052] Wherein, is the contribution degree of the ith environmental factor, and the calculation formula is: ; Wherein, is the difference between actual yield and predicted yield, is the contribution degree of the jth management measure, is the model residual term, for the feature subset, for the entire feature set, for the model prediction function, for the time weight function, reflecting the importance of factors at different growth stages.

[0053] Specifically, the time weight function weighted by growth stage sensitivity: ; wherein, is the current time point, is the time point of the kth key growth stage, is the sensitivity coefficient of factor i at stage k, is the total number of key stages of forage grass growth, is the time width parameter of the kth growth stage.

[0054] Specifically, the time sequence contribution tracking adopts a dynamic path integral method. Specifically, by analyzing the cumulative effect of changes in each factor during the entire growth season, the key time window and the dominant factor for yield formation are identified.

[0055] Specifically, the multi-dimensional attribution analysis includes: Climate dimension: analyzing the abnormal degree and timing of factors such as temperature, precipitation, and sunshine; Soil dimension: evaluating the impact of soil conditions such as changes in fertility and water status; Management dimension: quantifying the effects of measures such as irrigation, fertilization, and pest control; Model dimension: identifying systematic biases in the model and the need for parameter calibration.

[0056] Specifically, the attribution report generation adopts natural language generation technology: based on the results of attribution analysis, a comprehensive report containing main conclusions, improvement suggestions, and risk warnings is automatically generated.

[0057] In this embodiment, the present application establishes a yield difference decomposition model based on improved Shapley value, accurately decomposes the difference between actual yield and predicted yield into the contributions of environmental factors, management measures, and model errors, and uses a time weight function to quantify the dynamic influence of each factor at different growth stages by combining a time sequence contribution tracking algorithm; through multi-dimensional attribution analysis (climate, soil, management, and model dimensions) and a dynamic path integral method, the key time window and the dominant factor for yield formation are identified; finally, a natural language generation technology is used to automatically generate an interpretable attribution report. This function significantly improves the accuracy and interpretability of yield difference analysis, provides scientific decision-making basis for the next season's planting, helps users identify management deficiencies, optimize resource allocation, and guide model parameter calibration, effectively promoting the fine and intelligent management level of forage grass production.

[0058] Example 1: Applied to large-scale alfalfa planting ranch This example takes a 10,000-mu scale alfalfa planting ranch in northern China as the application scenario. The land of the ranch is contiguous, but there are differences in soil fertility inside, and plans to expand alfalfa planting on newly acquired adjacent land (no historical planting data).

[0059] The application process is as follows: Data fusion and cold start: The system first accesses the past three years of Sentinel-2 satellite images, daily weather data, detailed soil survey data, and the ranch's own recorded historical yield data at the plot level through the multi-source data fusion module.

[0060] The data quality control submodule automatically handles part of the missing data in the satellite image due to cloud cover, uses time series interpolation to complete the missing data of a meteorological station due to temporary failure, and uses the spatio-temporal alignment engine to unify all data to a 10-meter resolution, weekly time scale standard grid, generating a fused feature tensor.

[0061] For newly acquired land, the cold start mechanism takes effect immediately. In the plot feature space, the system matches the new land to two existing plots inside the ranch as high-similarity references based on soil type, terrain slope, and climate zoning. Subsequently, the parameter transfer learning algorithm transfers the model parameters trained based on the two reference plots and elastically initializes the prediction model for the new plot. After the start of the alfalfa growing season, the system uses the initial small amount of image and weather data to fine-tune the transferred model parameters through incremental learning, quickly adapting to the specific conditions of the new plot.

[0062] Yield prediction and physical constraints: The multi-modal deep learning prediction network begins to work. The remote sensing data encoding branch extracts the canopy coverage and growth status of alfalfa from the weekly updated satellite images; the weather time series encoding branch processes real-time and forecast temperature and precipitation sequences to identify drought or low temperature stress risks; the soil feature encoding branch inputs the water holding capacity and organic matter content of each plot.

[0063] These features are sent to the multi-modal feature fusion module, which dynamically calculates that soil moisture is the most critical factor affecting yield, followed by remote sensing growth information at the current growth stage. The network finally outputs the monthly yield prediction curve for the next three months and the total yield estimate at the end of the season.

[0064] The physical constraint mechanism works simultaneously. The built-in alfalfa physiology model judges that the current season's cumulative effective accumulated temperature is slightly lower than the historical average level, and according to the growth response function, its potential inhibitory effect on yield is quantified. Therefore, the initial data-driven optimistic prediction value of the model is moderately adjusted and constrained within a reasonable range determined by the physiological model, avoiding overestimation.

[0065] Decision support and attribution analysis: The management decision support system runs a multi-objective optimization algorithm based on the predicted yield potential, current soil moisture data, and future two-week weather forecasts. It generates a customized plan: it recommends a round of supplemental irrigation of 20 millimeters in 10 days for the 500 acres of the region that are slightly droughty; at the same time, given the current growth and forecasts, it recommends postponing the planned fertilizer plan by a week next month and adjusting the nitrogen fertilizer usage by 5% for optimal economic benefit.

[0066] After the end of the growing season, the end-of-season attribution analysis function is activated. The system compares the actual yield results with the model predictions and finds that the yield in the southeast region is slightly higher than the prediction. The attribution analysis points out through Shapley value decomposition that this is mainly due to the perfect match between an unexpected rainfall and irrigation in the key flowering period of alfalfa, with a water contribution of up to 30%. The report also points out that the model may slightly overestimate the sensitivity of the variety to early low temperatures. These analysis conclusions provide valuable quantitative basis for irrigation strategies and variety selection in the next year.

[0067] Example Two: Application to Regional Grassland Management and Policy Making This example takes the provincial agricultural management department as the application subject, aiming to make macro yield prediction and benefit evaluation of natural grasslands and artificial planted grasslands in the main pasture areas of the province.

[0068] The application process is as follows: Data fusion and cold start: The multi-source data fusion module integrates MODIS and Landsat-8 remote sensing data covering the whole province (taking into account coverage and resolution), gridded weather data from the meteorological bureau, provincial soil database, and satellite-reversed vegetation productivity data from previous years as historical yield reference.

[0069] In the face of a large number of small grasslands scattered in a vast area and lacking detailed historical records, the cold start mechanism shows great advantages. The system finds one or more similar templates for each blank grassland through spatial proximity and climate-soil characteristic similarity, realizes the rapid initialization and deployment of the prediction model, and makes the comprehensive prediction of the provincial range possible.

[0070] Yield prediction and physical constraints: The prediction network processes raster data of the whole province. The remote sensing branch uses a multi-scale convolutional network to capture features at different scales from pixels to regions; the meteorological branch analyzes the precipitation distribution and accumulated temperature patterns of the whole province; and the soil branch processes the spatial distribution of soil types.

[0071] In the multi-modal fusion, the prior matrix of the cross-attention mechanism gives a higher weight to the "meteorological-remote sensing" interaction, because on a large scale, climate conditions are the primary factor determining grassland productivity. The network finally outputs a prediction map of the grass yield distribution in the whole province and gives an estimate of the total yield of each administrative region.

[0072] The physical constraint layer plays the role of a security guard here. For some areas where the remote sensing signal is abnormal due to special terrain, the model may produce outliers. The physiological growth model sets absolute upper and lower limits for the predicted values based on the historical climate production potential of the region, effectively filtering these obviously unreasonable predictions and ensuring the reliability of the macro data provided to the government.

[0073] Decision support and attribution analysis: The management decision support system is upgraded from the field level to the regional level. Instead of outputting specific irrigation schemes, it performs disaster loss assessment and resource allocation simulation based on yield prediction results and meteorological drought indices. For example, the system predicts that the yield in a certain county in the north may decrease by 15% due to drought in summer, and accordingly issues an early warning to the management department, suggesting that the grass feed allocation plan be planned in advance or the animal husbandry disaster prevention and mitigation plan be started.

[0074] At the end of the season, the attribution analysis function analyzes the reasons for the spatial differentiation of yield from a macro perspective. Through time series contribution tracking, it is found that the yield in a certain area in the south is significantly better than in previous years, mainly due to the right timing of precipitation and the improved vegetation cover due to the implementation of the "Grain for Green" policy in the area. The positive contribution of management measures is quantitatively evaluated. This attribution report provides objective data support for evaluating the effectiveness of ecological policies and developing differentiated compensation and guiding policies.

[0075] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the application.

Claims

1. A method for constructing a forage yield prediction model based on deep learning, characterized in that: Includes the following steps: S11: Construct a multi-source data fusion module that integrates remote sensing data, meteorological data, soil data, and historical yield data. The multi-source data fusion module includes a data quality control sub-module to handle missing and incomplete data. S12: Establish a cold start mechanism. For newly added plots lacking historical planting data, model initialization is achieved through similar plot matching and parameter transfer learning. The cold start mechanism includes spatial proximity analysis and feature similarity calculation. S13: Design a multimodal deep learning prediction network, including remote sensing data coding branch, meteorological time series coding branch, soil feature coding branch and multimodal feature fusion module, to output short-term monthly yield forecast and quarter-end total yield forecast. S14: Integrating physical constraint mechanisms, based on forage growth physiological models and historical climate benchmarks, to impose reasonable range constraints on model prediction results and prevent abnormal prediction values ​​from being generated; S15: Build a management decision support system to automatically generate irrigation, fertilization and weeding management suggestions based on forecast results, soil parameters and future weather forecasts; S16: Implement end-of-season attribution analysis function. By comparing the difference between actual and predicted yields and combining the changes in parameters during the growing season, identify the key factors affecting yield differences.

2. The method for constructing a deep learning-based forage yield prediction model according to claim 1, characterized in that: When constructing a multi-source data fusion module, the multi-source data fusion module specifically includes: The remote sensing data submodule is used to process Sentinel-2, Landsat-8 and MODIS satellite imagery to extract vegetation indices and biomass characteristics; The meteorological data submodule integrates temperature, precipitation, sunshine duration and humidity data, and supports time-series feature extraction; The soil data submodule is used to integrate information on soil type, pH value, organic matter content, and water content; Historical production database is used to store historical production records at the plot level; The data quality control unit is used to process missing remote sensing data using a cloud gap filling algorithm based on spatiotemporal characteristics, and to complete missing meteorological data using a time-series interpolation method.

3. The method for constructing a deep learning-based forage yield prediction model according to claim 2, characterized in that: The construction of a multi-source data fusion module includes the following steps: S21: Establish a multi-source data access interface to receive remote sensing data, meteorological data, soil data, and historical yield data respectively; S22: Design a data quality control submodule to adopt differentiated processing strategies for different types of missing data; S23: Construct a spatiotemporal alignment engine to unify multi-source heterogeneous data to the same spatiotemporal scale; S24: Implement feature-level data fusion to generate a fused feature tensor for use by the prediction model; S25: Establish a data quality assessment system to quantify the credibility and integration effect of each data source.

4. The method for constructing a deep learning-based forage yield prediction model according to claim 3, characterized in that: Establishing a cold start mechanism includes the following steps: S31: Establish a plot feature space and extract multi-dimensional feature vectors of geographical location, soil properties, topographic features, and climate zoning; S32: Design a composite similarity metric function that considers both spatial proximity and feature similarity. S33: Implement the parameter transfer learning algorithm to initialize the new plot model based on the model parameters of similar plots; S34: Construct an incremental learning mechanism to dynamically adjust migration parameters using real-time data during the growing season.

5. The method for constructing a deep learning-based forage yield prediction model according to claim 4, characterized in that: When designing a composite similarity metric function that considers both spatial proximity and feature similarity, the composite similarity metric function adopts a space-feature dual-constraint form, and the principle formula is as follows: ； in, The geographical distance between plots A and B. For spatial scale parameters, For feature space similarity, For temperature parameters, This represents the number of candidate similar land parcels.

6. The method for constructing a deep learning-based forage yield prediction model according to claim 5, characterized in that: When designing a multimodal deep learning prediction network, the following architecture is adopted: The remote sensing data encoding branch uses a multi-scale convolutional neural network to extract spatial features and process satellite image data; The meteorological time-series coding branch uses a gated recurrent unit network to capture time dependencies and process meteorological time-series data. The soil feature coding branch uses a fully connected network to process static soil attribute features; The multimodal feature fusion module dynamically weights the contributions of each modality feature based on a cross-attention mechanism; The dual-output forecast head generates monthly production sequence forecasts and quarter-end total production forecasts, respectively.

7. The method for constructing a deep learning-based forage yield prediction model according to claim 6, characterized in that: The cross-attention mechanism of the multimodal feature fusion module adopts the following attention weight calculation method: For the query vector of the i-th modality key-value pairs with the j-th mode Calculate cross attention: ； Among them, the intermodal prior matrix The calculation formula is: ； in, Let j be the key vector of the j-th mode. Let j be the value vector of the j-th mode. , and The projected weight matrix is ​​a learnable matrix. Let be the dimension of the key vector. To query the dot product similarity with the key, This is the projected value vector. Let be the historical correlation coefficient between modes i and j. For modal feature distribution similarity, As a measure of modal complementarity, , and These are learnable weight parameters.

8. The method for constructing a deep learning-based forage yield prediction model according to claim 7, characterized in that: When integrating physical constraint mechanisms, the following steps are included: S41: Establish a forage physiological growth model based on historical climate benchmarks and determine the suitable range of key environmental factors for each growth stage; S42: Calculate the deviation of current growing season environmental factors from historical baselines to quantify the potential impact of climate anomalies on yield; S43: Design a physical constraint layer to constrain the prediction results of the deep learning model within the reasonable fluctuation range determined by the physiological model; S44: Implement a dynamic constraint adjustment mechanism to adaptively adjust the constraint intensity according to the growth stage and crop type.

9. The method for constructing a deep learning-based forage yield prediction model according to claim 8, characterized in that: When constructing a management decision support system, a multi-objective optimization algorithm is employed: The optimization objectives are to maximize output and minimize cost. Considering constraints such as water resource utilization efficiency, fertilizer utilization efficiency, and economic benefits; Management recommendations will be dynamically adjusted based on future weather forecast data. The output includes specific plans for irrigation time and water volume, fertilizer type and dosage, and weeding timing.

10. The method for constructing a deep learning-based forage yield prediction model according to claim 9, characterized in that: When implementing the quarter-end attribution analysis function, the specific steps include: S51: Establish a production difference decomposition model to decompose the difference between actual and predicted production into three components: environmental factor contribution, management measure contribution, and model error. S52: Design a time-series contribution tracking algorithm to quantify the impact of key factors at each growth stage on the final yield difference; S53: Implement multi-dimensional attribution analysis to comprehensively analyze the reasons for yield differences from the perspectives of climate anomalies, management effectiveness, and model bias. S54: Generate interpretable attribution reports to provide recommendations for improving planting decisions for the next season.

Images