Crop growth prediction method based on multi-source data fusion analysis

Abstract

The invention discloses a crop growth prediction method based on multi-source data fusion analysis, and the method comprises the steps: carrying out the standardization processing and missing value filling processing of multi-source data, constructing a time window, calculating the correlation between data sources, and generating a current data credibility vector based on historical credibility evaluation data; extracting features by adopting multi-scale decomposition, detecting abnormal points and analyzing abnormal degrees, and constructing an abnormal conduction network to evaluate conduction influence; performing feature importance evaluation and feature fusion; initial weight distribution is calculated, a weight prediction model is constructed, and consistency optimization is carried out on a weight matrix; performing hierarchical prediction by adopting a multi-level prediction model, and performing collaborative integration based on a prediction result and an optimized weight matrix; and calculating a prediction result deviation and carrying out optimization adjustment, and outputting a final prediction result. According to the invention, accurate prediction of the crop growth state is realized, and the reliability and practicability of the prediction result are improved.

Description

Technical Field

[0001] The invention belongs to the field of multi-source data fusion analysis, in particular to a crop growth prediction method based on multi-source data fusion analysis. Background Art

[0002] Crop growth prediction is an important part of modern precision agriculture and is of great significance for optimizing agricultural production management, improving agricultural production efficiency and ensuring food security. Accurately predicting crop growth conditions can help agricultural producers adjust planting strategies, optimize resource allocation, and formulate pest and disease control plans in a timely manner, thereby reducing production risks and improving crop yields and quality. Especially in the context of global climate change and frequent extreme weather events, crop growth prediction has become increasingly important for improving the risk resistance and sustainable development level of agricultural production.

[0003] At present, crop growth prediction mainly relies on remote sensing monitoring, ground sensor monitoring and historical agricultural data analysis. Remote sensing monitoring technology obtains crop growth information through satellites or drones, which can realize rapid monitoring of large areas of farmland; ground sensor networks obtain farmland microclimate and crop physiological characteristics data by deploying various environmental and physiological parameter sensors; historical agricultural data records the planting management measures and yield performance of previous years. The application of these technologies provides rich data support for crop growth prediction. Traditional prediction methods mainly use statistical models or simple machine learning algorithms, such as multivariate regression analysis, support vector machines, etc., to process and analyze data from a single source.

[0004] However, there are still some specific technical problems in the existing crop growth prediction methods: first, in the data preprocessing stage, there is a lack of systematic solutions for the spatiotemporal alignment and quality assessment of multi-source heterogeneous data, especially when dealing with the fusion of data with different spatiotemporal resolutions, it is difficult to accurately evaluate and control data quality; second, in terms of anomaly detection, existing methods often separate anomaly detection and impact assessment, and cannot effectively analyze the transmission relationship between abnormal conditions and different growth indicators, which makes it difficult to accurately assess the actual impact range and extent of abnormal events; third, in the process of feature extraction and fusion, there is a lack of dynamic evaluation mechanism for feature importance, which makes it difficult to adaptively adjust feature weights according to changes in crop growth stages, affecting the prediction model's accurate grasp of key growth characteristics; finally, in terms of prediction model construction, existing methods generally lack a collaborative optimization mechanism for multi-scale prediction results, making it difficult to ensure the consistency of prediction results at different time scales, and lack systematic evaluation and optimization methods for the reliability of prediction results. These technical problems have seriously affected the accuracy and practicality of crop growth prediction, and it is urgent to develop more advanced prediction methods. Summary of the invention

[0005] The purpose of the invention is to provide a crop growth prediction method based on multi-source data fusion analysis, in order to solve at least one technical problem existing in the prior art.

[0006] The technical solution is a crop growth prediction method based on multi-source data fusion analysis, which includes the following steps:

[0007] S1. Obtain the original remote sensing data set, ground sensor data set and historical agricultural data set, perform standardization and missing value filling to obtain a complete data matrix; based on the complete data matrix, construct a time window and calculate the correlation between data sources to generate a data source time series correlation tensor; calculate the current data credibility vector based on the pre-stored historical credibility assessment data and the data source time series correlation tensor;

[0008] S2. Based on the complete data matrix, a multi-scale decomposition method is used to extract features and obtain a multi-scale feature tensor. Based on the multi-scale feature tensor and the data source time series correlation tensor, abnormal points are detected and the degree of abnormality is analyzed to generate an abnormal degree matrix. Based on the abnormal degree matrix and the data source time series correlation tensor, an abnormal conduction network is constructed and the conduction impact is evaluated, and a conduction impact matrix is ​​output.

[0009] S3. Based on the complete data matrix and the multi-scale feature tensor, a multi-scale feature extraction method is used to generate a multi-scale feature set; based on the multi-scale feature set, feature importance evaluation is performed to obtain a feature importance matrix; based on the feature importance matrix and the conduction influence matrix, feature fusion is performed to obtain a fused feature matrix;

[0010] S4. Based on the current data credibility vector and the conduction influence matrix, the initial weight distribution is calculated to obtain the initial weight vector; based on the initial weight vector and the fusion feature matrix, a weight prediction model is constructed to generate a dynamic weight matrix; the dynamic weight matrix is ​​optimized for consistency to obtain an optimized weight matrix;

[0011] S5. Based on the fusion feature matrix, a multi-level prediction model is used to perform hierarchical prediction to obtain a multi-scale prediction result set; based on the multi-scale prediction result set and the optimized weight matrix, the prediction results are collaboratively integrated to output the collaborative prediction results;

[0012] S6. Based on the multi-scale prediction result set and the collaborative prediction result, calculate the deviation between the prediction results of different scales to obtain a consistency evaluation report; based on the consistency evaluation report, optimize and adjust the collaborative prediction result to output the final prediction result.

[0013] Beneficial effects: The present invention establishes a reliable data foundation, realizes early identification of growth abnormalities and impact assessment; realizes adaptive adjustment of the prediction model by capturing the key characteristics of crop growth; ensures the accuracy and reliability of the prediction results through multi-level prediction and result optimization; realizes the intelligent fusion of multi-source heterogeneous data and establishes a dynamic evaluation mechanism for data credibility; improves the accuracy and practicability of crop growth prediction, and can provide reliable data support for agricultural production decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the present invention.

[0015] Figure 2 This is a flow chart of step S1 of the present invention.

[0016] Figure 3 This is a flow chart of step S2 of the present invention.

[0017] Figure 4 This is a flow chart of step S3 of the present invention.

[0018] Figure 5 This is a flow chart of step S4 of the present invention.

[0019] Figure 6 This is a flow chart of step S5 of the present invention.

[0020] Figure 7 This is a flow chart of step S6 of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 As shown, this application proposes a crop growth prediction method based on multi-source data fusion analysis, which includes the following steps:

[0022] S1. Obtain the original remote sensing data set, ground sensor data set and historical agricultural data set, perform standardization and missing value filling to obtain a complete data matrix; based on the complete data matrix, construct a time window and calculate the correlation between data sources to generate a data source time series correlation tensor; calculate the current data credibility vector based on the pre-stored historical credibility assessment data and the data source time series correlation tensor;

[0023] S2. Based on the complete data matrix, a multi-scale decomposition method is used to extract features and obtain a multi-scale feature tensor. Based on the multi-scale feature tensor and the data source time series correlation tensor, abnormal points are detected and the degree of abnormality is analyzed to generate an abnormal degree matrix. Based on the abnormal degree matrix and the data source time series correlation tensor, an abnormal conduction network is constructed and the conduction impact is evaluated, and a conduction impact matrix is ​​output.

[0024] S3. Based on the complete data matrix and the multi-scale feature tensor, a multi-scale feature extraction method is used to generate a multi-scale feature set; based on the multi-scale feature set, feature importance evaluation is performed to obtain a feature importance matrix; based on the feature importance matrix and the conduction influence matrix, feature fusion is performed to obtain a fused feature matrix;

[0025] S4. Based on the current data credibility vector and the conduction influence matrix, the initial weight distribution is calculated to obtain the initial weight vector; based on the initial weight vector and the fusion feature matrix, a weight prediction model is constructed to generate a dynamic weight matrix; the dynamic weight matrix is ​​optimized for consistency to obtain an optimized weight matrix;

[0026] S5. Based on the fusion feature matrix, a multi-level prediction model is used to perform hierarchical prediction to obtain a multi-scale prediction result set; based on the multi-scale prediction result set and the optimized weight matrix, the prediction results are collaboratively integrated to output the collaborative prediction results;

[0027] S6. Based on the multi-scale prediction result set and the collaborative prediction result, calculate the deviation between the prediction results of different scales to obtain a consistency evaluation report; based on the consistency evaluation report, optimize and adjust the collaborative prediction result to output the final prediction result.

[0028] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0029] S11, receiving the original remote sensing data set, the ground sensor data set and the historical agricultural data set, constructing a three-dimensional index structure containing spatial position and time information, and forming a spatiotemporal index matrix; based on the spatiotemporal index matrix, standardizing the numerical data in the original remote sensing data set, the ground sensor data set and the historical agricultural data set to generate a standardized data matrix; based on the data change trend of adjacent time points in the standardized data matrix, using the interpolation method to fill in the missing values ​​to generate a complete data matrix;

[0030] S12. Based on the complete data matrix, a dynamic sliding time window is constructed to obtain a time window sequence; within each time window sequence, the correlation coefficients between different data sources are calculated to generate a correlation coefficient matrix; the correlation coefficient matrix is ​​combined in the time dimension to form a data source time series correlation tensor;

[0031] S13. Read the pre-stored historical credibility assessment data, calculate the transition probability between data source states based on the historical credibility assessment data, and generate a state transition matrix; evaluate the state of each data source in the current time window according to the state transition matrix and the data source time series correlation tensor to obtain the evaluated data source state; divide the data source into different credible states based on the evaluated data source state and the preset time series correlation significance level, and finally obtain the current data credibility vector.

[0032] In one embodiment of the present application, the time series correlation coefficient TC(i, j, t) = Σ[(Xi(t)-μi(t))·(Xj(t)-μj(t))] / {[Σ(Xi(t)-μi(t)) 2 ]·[Σ(Xj(t)-μj(t)) 2 ]} 0.5 ·exp(-|Δt| / τ); dynamic weight matrix W(t) = softmax(TC(t)·R(t)); where Xi(t) and Xj(t) are the time series of the two data sources; μi(t) and μj(t) are local means; Δt is the time delay; τ is the time decay parameter; R(t) is the reliability matrix; and softmax is the normalization function.

[0033] State transition probability P(Si→Sj) = N(Si, Sj) / N(Si); credibility state vector CS(t) = softmax[P·CS(t-1)]; transition entropy TE(i, j) = -∑p(si, sj)log[p(sj|si) / p(sj)]; evaluation credibility RC(i) = CS(i)·[1-H(Si)]·Q(i); where N(Si, Sj) is the number of state transitions; N(Si) is the number of occurrences of state i; p(si, sj) is the joint probability; H(Si) is the state entropy; Q(i) is the quality factor.

[0034] This embodiment realizes the efficient fusion and credibility assessment of multi-source heterogeneous data required for crop growth prediction by constructing a multi-level data processing architecture. Specifically, firstly, the integrity and consistency of the data are ensured by standardization processing and missing value filling, which is crucial to the accuracy of crop growth prediction, because crop growth data often has data missing problems caused by acquisition equipment failure, weather influence, etc. Secondly, by constructing the time window and calculating the correlation between the data sources, the data source time series correlation tensor is generated, which can effectively capture the dynamic correlation between different data sources (such as remote sensing data, ground sensor data, historical agricultural data), especially the feature correlation at different growth stages of crops. The generation of a complete data matrix mainly solves the problem of missing data and ensures the integrity of the data; the setting of the time window is to capture the time series characteristics and dynamic change patterns of the data; in time series prediction, the missing data is processed first, and then time windows of different lengths are set to analyze the trend; data completion ensures the integrity of the time window analysis, and step-by-step processing can better control the quality of each step, which is convenient for adjusting and optimizing the parameters of each step. Finally, by combining historical credibility evaluation data to calculate the current data credibility vector, this credibility evaluation mechanism based on historical data can effectively identify and filter out unreliable data sources, such as sensor data affected by extreme weather or remote sensing image data of poor quality, thereby providing a high-quality data basis for subsequent prediction analysis. This embodiment forms an adaptive data quality control system that can dynamically adjust the credibility weight of a data source according to its historical performance and current status, thereby improving the input data quality of the crop growth prediction model.

[0035] According to one aspect of the present application, step S11 is further:

[0036] S111, receiving the original remote sensing data set, the ground sensor data set and the historical farming data set, extracting the timestamp information and the spatial coordinate information of each data set, and obtaining the time series vector and the spatial coordinate vector; constructing a time index tree based on the time series vector to obtain the time index matrix; constructing a spatial index tree based on the spatial coordinate vector to generate the spatial index matrix; combining the time index matrix and the spatial index matrix to obtain the spatiotemporal index matrix;

[0037] S112, based on the spatiotemporal index matrix, normalizing the numerical attributes of the original remote sensing data set, the ground sensor data set, and the historical agricultural data set to obtain a normalized data set; performing noise smoothing on the normalized data set to generate a smoothed data set; performing outlier detection based on the smoothed data set to obtain a cleaned data set; reorganizing the cleaned data set according to the spatiotemporal index matrix to output a standardized data matrix;

[0038] S113, based on the time span of missing data in the standardized data matrix, classify the missing positions to obtain a missing type matrix; based on the missing type matrix, select a corresponding interpolation strategy to generate an interpolation strategy vector; according to the interpolation strategy vector, divide the missing data into different processing groups to obtain a missing processing grouping matrix;

[0039] S114. Based on the missing processing grouping matrix and the standardized data matrix, different interpolation strategies are used to fill the missing values ​​to obtain a filled data set; data smoothing is performed on the filled data set to generate a smoothed filled matrix; the smoothed filled matrix is ​​merged with the non-missing data in the standardized data matrix to output a complete data matrix.

[0040] In one embodiment of the present application, the normalization method is: standardized value X'(i, t) = [X(i, t) - μ(t)] / [σ(t) + λ·v(t)]; smoothing value S(i, t) = α·X'(i, t) + (1-α)·MA(i, t); dynamic noise threshold NT(t) = β·MAD(t)·log[1+var(t)]; wherein X(i, t) is the original data; μ(t) is the mean in the time window; σ(t) is the standard deviation; v(t) is the coefficient of variation; λ is the stability parameter, which is taken as 0.1; MA(i, t) is the moving average; α is the smoothing coefficient, which is adaptively adjusted according to data volatility; MAD(t) is the median absolute deviation; var(t) is the variance in the window; β is the noise control parameter, which is taken as 0.8.

[0041] Adaptive interpolation strategy: missing type score ST(i) = w1·L(i) + w2·D(i) + w3·C(i); interpolation weight W(j) = exp(-d(i, j) / h)·Q(j); predicted value P(i) = ΣW(j)·V(j) / ΣW(j); where L(i) is the missing duration score; D(i) is the data density score; C(i) is the change trend score; w1, w2, and w3 are weight coefficients obtained through historical data training; d(i, j) is the spatiotemporal distance; h is the bandwidth parameter; Q(j) is the quality score of the reference point; and V(j) is the observation value of the reference point.

[0042] Data filling optimization: local similarity LS(i, j) = cos(Xi, Xj)·exp(-d(i, j) / h); filling weight FW(j) = LS(i, j)·Q(j)·T(j); predicted value PV(i) = ∑[FW(j)·V(j)] / ∑FW(j); smoothing coefficient α(t) = 1 / [1+exp(-k·var(t))]; where Xi, Xj are feature vectors; d(i, j) is the distance; Q(j) is the data quality; T(j) is the time weight; var(t) is the local variance.

[0043] This embodiment realizes the efficient organization and preprocessing of multi-source crop growth data by constructing a spatiotemporal index system and a data standardization processing flow. Specifically, first, a spatiotemporal index matrix is ​​constructed by extracting timestamp information and spatial coordinate information. This index structure can quickly locate and access crop growth data at a specific spatiotemporal position, and is particularly suitable for the organization and management of large-scale distributed monitoring data. Then, by normalizing and smoothing the numerical attributes, the comparability and consistency of data from different sources are ensured, and the data scale and noise differences caused by different sensors and different acquisition devices can be effectively processed. Finally, the missing data is repaired through a classification filling strategy to form a complete data matrix, which can adaptively select the optimal interpolation strategy according to the time span and spatial distribution characteristics of the missing data, such as using local interpolation for short-term missing and model prediction for long-term missing, thereby maintaining the spatiotemporal continuity and physical rationality of the data to the greatest extent. This embodiment forms an efficient data preprocessing system, which lays a solid data foundation for subsequent analysis and prediction.

[0044] According to one aspect of the present application, step S13 is further:

[0045] S131, reading pre-stored historical credibility assessment data, extracting data source state change characteristics, and obtaining a state characteristic matrix; constructing a state transition relationship based on the state characteristic matrix, and generating a state transition matrix.

[0046] S132, receiving the state transfer matrix and the data source timing correlation tensor, extracting the correlation features of the current time window, and obtaining a correlation feature vector; evaluating the data source state according to the correlation feature vector, and generating a state evaluation vector.

[0047] S133, receiving the state evaluation vector, combining it with a preset credibility evaluation rule, obtaining an initial credibility vector; dynamically adjusting the initial credibility vector to generate an adjusted credibility vector; and outputting the current data credibility vector.

[0048] This embodiment realizes the dynamic evaluation and control of the quality of crop growth monitoring data by establishing a data credibility assessment mechanism. First, a state transition matrix is ​​constructed based on historical credibility assessment data. This state transition model based on historical data can accurately characterize the dynamic change law of the credibility state of different data sources, and is particularly suitable for dealing with long-term data quality assessment problems in crop growth monitoring. Then, by combining the time series correlation characteristics for state assessment, an accurate judgment of the current data source state is achieved, and abnormal associations between data sources can be discovered in a timely manner, such as detecting sensor failures or acquisition anomalies. Finally, the current data credibility vector is generated through a dynamic adjustment mechanism, and its credibility rating can be dynamically adjusted according to the real-time performance of the data source, ensuring the timeliness and accuracy of the data quality assessment results. This embodiment constructs a complete data quality control system, which can provide reliable quality assurance for subsequent data fusion and predictive analysis.

[0049] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0050] S21. Based on the complete data matrix, the wavelet transform method is used to perform multi-level decomposition on the data to obtain a multi-scale feature tensor; based on the multi-scale feature tensor, the local density analysis method is used to identify abnormal points and generate an abnormal point set; according to the local density deviation degree of each abnormal point in the abnormal point set, the abnormal degree matrix is ​​calculated;

[0051] S22. Based on the data source time series correlation tensor, a directed weighted graph is constructed to obtain an initial conduction graph; based on the abnormal point set and the abnormal degree matrix, the conduction weights between the nodes in the initial conduction graph are calculated to obtain a weighted conduction graph; based on the weighted conduction graph, an abnormal conduction adjacency matrix is ​​generated;

[0052] S23. Based on the abnormal conduction adjacency matrix and the current data credibility vector, the conduction importance of each abnormal point is calculated to obtain the conduction importance vector; based on the abnormal conduction adjacency matrix, the conduction distance between each node is calculated; based on the conduction importance vector and the conduction distance, a conduction influence attenuation function is constructed to generate a conduction attenuation matrix; the conduction attenuation matrix is ​​combined with the abnormality degree matrix to output the conduction influence matrix.

[0053] In one embodiment of the present application, the transmission weight is calculated as follows: Transmission weight TW(i, j) = exp(-d(i, j) / σ)·[ρ(i)·ρ(j)] α·cos(θij); weighted conduction graph weight WG(i, j, t) = TW(i, j)·AS(i)·AS(j)·TD(t); adjacency matrix A(i, j) = {WG(i, j, t), if WG(i, j, t) > θt; 0, otherwise}; where d(i, j) is the distance between nodes; σ is the scale parameter; ρ(i) and ρ(j) are the node importances; α is the importance adjustment parameter; θij is the conduction direction angle; AS(i) and AS(j) are the abnormality scores; TD(t) is the time decay factor; θt is the dynamic threshold.

[0054] This embodiment realizes the accurate identification and impact assessment of abnormal patterns in crop growth data by implementing multi-scale anomaly detection and conduction analysis. First, the wavelet transform method is used to decompose the data at multiple levels to obtain a multi-scale feature tensor, which can simultaneously capture the changing characteristics of crop growth data at different time scales, from fast-changing daytime fluctuations to slowly changing seasonal trends. The local density analysis method is used to identify abnormal points and calculate the abnormal degree matrix, which is particularly suitable for processing complex abnormal patterns in crop growth data, such as growth abnormalities caused by sudden pests and diseases or fluctuations in physiological indicators caused by extreme weather. On this basis, an abnormal conduction network is constructed and the conduction impact is evaluated, which can accurately track the diffusion path of abnormal conditions between different growth indicators, such as how nutrient deficiency gradually affects multiple physiological indicators such as leaf area index and photosynthesis efficiency. By calculating the conduction importance and constructing the conduction impact attenuation function, a quantitative assessment of the scope and degree of abnormal impact is achieved, which is of great significance for timely discovery and intervention of potential risks in the process of crop growth. Among them, anomaly detection focuses on single-point or local abnormal patterns; anomaly conduction analysis focuses on the correlation between anomalies and the spread of impact; the parallel processing between the two improves computing efficiency, can capture different types of anomalies, and the two can also verify and complement each other.

[0055] According to one aspect of the present application, step S21 is further:

[0056] S211, based on the complete data matrix, construct a multi-layer wavelet transform coefficient matrix to obtain a wavelet coefficient set; based on the wavelet coefficient set, perform coefficient reconstruction to generate a reconstruction coefficient matrix; perform a scale fusion operation on the reconstruction coefficient matrix to obtain a fusion coefficient sequence; reorganize the fusion coefficient sequence into a tensor form to obtain a multi-scale feature tensor;

[0057] S212. Based on the multi-scale feature tensor, calculate the local density distribution features to obtain a density distribution matrix; based on the density distribution matrix, construct a density distance map to generate a distance relationship matrix; according to the distance relationship matrix, calculate the local anomaly factor to obtain an anomaly factor vector; according to the anomaly factor vector, determine the abnormal data point and output the anomaly point set;

[0058] S213. Based on the outlier point set and the multi-scale feature tensor, the local density deviation value of each outlier point is calculated to obtain a density deviation vector; based on the density deviation vector, an outlier degree assessment model is constructed to generate an assessment result vector; the assessment result vector is converted into a matrix form to obtain an outlier degree matrix.

[0059] In one embodiment of the present application, wavelet transform: wavelet coefficient C(a, b) = 1 / sqrt(a)·∫x(t)·ψ*((tb) / a)dt; reconstruction coefficient RC(j) = Σ[C(ai, bj)·Ki]; fusion coefficient FC(t) = Σ[RC(j)·W(j, t)]; wherein a is a scale parameter; b is a translation parameter; x(t) is an input signal; ψ(t) is a wavelet basis function; Ki is a reconstruction kernel function; W(j, t) is a time-varying weight coefficient, which is determined by the wavelet energy distribution.

[0060] Local density analysis method: local density ρ(i) = Σ(χ(d(i, j) - dc)); local distance Δ(i) = min(d(i, j)), for ρ(j) > ρ(i); anomaly factor γ(i) = ρ(i) · Δ(i); where d(i, j) is the distance between sample points i and j; dc is the cutoff distance, taking the 2% quantile of all distances; χ(x) is the cutoff function, when x < 0, χ(x) = 1, otherwise χ(x) = 0; ρ(i) represents the local density of sample point i, that is, the number of sample points whose distance to sample point i is less than dc; Δ(i) represents the distance from sample point i to the nearest sample point with a greater density than it; γ(i) is the anomaly factor used to identify anomalies.

[0061] Density distance map construction method: Improved density distance value DM(i, j) = exp(-ρ(i)·ρ(j) / σ 2)·d(i, j)·w(t); adaptive threshold AT(t) = μ(t) + k·σ(t); density distance matrix R(i, j) = {1, if DM(i, j) < AT(t); 0, otherwise}; where ρ(i) and ρ(j) are the local densities of sample points i and j; σ is the density scale parameter, which is determined by the variance of the density distribution of the sample points; d(i, j) is the Euclidean distance between sample points; w(t) is the time series weight coefficient, which decays exponentially with time; μ(t) is the mean of the density distance at the current moment; σ(t) is the standard deviation of the density distance at the current moment; k is the threshold adjustment coefficient, which defaults to 2.5.

[0062] This embodiment realizes accurate identification and quantitative evaluation of abnormal patterns in crop growth data by implementing multi-scale decomposition and local density analysis. Specifically, the multi-layer wavelet transform is first used to decompose and reconstruct the data, which can decompose the crop growth data into different frequency components, effectively separating normal growth trends, seasonal changes and abnormal fluctuations. Then, by calculating the local density distribution characteristics and constructing a density distance map, the accurate identification of abnormal points is realized, which is particularly suitable for processing complex abnormal patterns in crop growth data, such as gradual abnormalities and sudden abnormalities. Finally, by evaluating the degree of local density deviation of the abnormal points, a quantitative description of the degree of abnormality is achieved, which can accurately measure the severity of the abnormality and provide a decision-making basis for subsequent early warning and intervention. This embodiment forms a multi-level abnormal analysis system that can comprehensively evaluate abnormal conditions during crop growth.

[0063] According to one aspect of the present application, step S23 is further:

[0064] S231. Receive the abnormal conduction adjacency matrix and the current data credibility vector, calculate the node importance ranking according to the conduction network structure, and obtain the node ranking vector; construct the importance update rule based on the node ranking vector and generate the importance update matrix; iteratively calculate the importance update matrix and output the conduction importance vector.

[0065] S232. Receive the conduction importance vector, calculate the conduction distance between the abnormal points, and obtain the conduction distance matrix; construct the conduction attenuation function based on the conduction distance matrix, and generate the attenuation function parameter vector; calculate the conduction attenuation value according to the attenuation function parameter vector, and output the conduction attenuation matrix.

[0066] S233. Receive the conduction attenuation matrix and the abnormality degree matrix, construct a conduction impact calculation model, and obtain an impact calculation parameter vector; perform conduction impact calculation based on the impact calculation parameter vector to generate an impact assessment vector; convert the impact assessment vector into a matrix form and output a conduction impact matrix.

[0067] In one embodiment of the present application, the conduction attenuation function is: conduction distance D(i, j) = [∑(wk·|xik-xjk| p )] 1 / p ; Attenuation function AF(d) = exp(-d / σ)·[1+α·log(1+β·d)]; Conduction intensity TI(i, j) = AF(D(i, j))·IS(i)·IS(j); where wk is the feature weight; p is the distance norm; σ is the attenuation parameter; α, β are adjustment coefficients; IS is the node importance, exp is the exponential function, xik is the value of the i-th node on the k-th feature, and d is the distance between two nodes.

[0068] Calculation of transmission impact: direct impact DI(i, j) = TI(i, j)·AS(i); indirect impact II(i, j) = ∑[TI(i, k)·TI(k, j)·AS(k)]; total impact TI(i, j) = DI(i, j) + λ·II(i, j); impact matrix IM(i, j) = TI(i, j)·exp(-t / τ); where TI is the transmission intensity; AS is the anomaly score; λ is the indirect impact weight; t is the time interval; τ is the time attenuation parameter.

[0069] This embodiment realizes the network analysis and quantitative evaluation of abnormal effects on crop growth by establishing an abnormal conduction evaluation mechanism. First, the importance of nodes is calculated based on the abnormal conduction network structure, and the physiological indicators or environmental factors that play a key role in the abnormal conduction process can be identified. Then, by constructing a conduction attenuation function, a quantitative description of the abnormal impact propagation process is realized. This attenuation model that considers spatial distance and time delay can accurately characterize the diffusion law of abnormal impacts, such as the spatial spread of pests and diseases or the step-by-step conduction of nutritional deficiencies. Finally, the overall impact degree is evaluated through the conduction impact calculation model, which takes into account both direct and indirect impacts, and can comprehensively evaluate the overall impact of abnormal events on crop growth. This embodiment constructs a complete impact evaluation system, which provides a scientific basis for crop growth management.

[0070] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0071] S31. Based on the complete data matrix and the multi-scale feature tensor, a multi-scale convolution kernel group is constructed to obtain a convolution kernel matrix; based on the convolution kernel matrix, feature extraction is performed to obtain a feature component set; the feature component set is combined to generate a multi-scale feature set;

[0072] S32. Based on the multi-scale feature set, a contribution analysis method is used to calculate the importance of each feature to obtain an initial importance vector; based on the initial importance vector and the conduction influence matrix, the feature importance is corrected to obtain a corrected importance vector; the corrected importance vector is converted into a matrix form, and a feature importance matrix is ​​output;

[0073] S33. Based on the multi-scale feature set and the feature importance matrix, a feature fusion weight function is constructed to obtain a fusion weight vector; based on the fusion weight vector, the features are weightedly combined to generate an initial fusion matrix; based on the initial fusion matrix and the conduction influence matrix, optimization and adjustment are performed to obtain a fusion feature matrix.

[0074] In one embodiment of the present application, the multi-scale convolution kernel: convolution kernel function K(x, y, s) = [1 / (2πσs 2 )]·exp[-(x 2 +y 2 ) / (2σs 2 )]·G(s); scale parameter σs = σ0·β s ; Feature response F(x, y, s) = ∬I(u, v)·K(xu, yv, s)dudv; where s is the scale index; σ0 is the reference scale; β is the scale factor; G(s) is the scale normalization function; I(u, v) is the input feature map; x, y are the spatial coordinates.

[0075] Multi-scale feature extraction method: Spatial feature extraction: SP(x, y, s) = ∬I(u, v) G(xu, yv, σs)dudv; Gaussian scaling kernel G(x, y, σ) = [1 / (2πσ 2 )]·exp[-(x 2 +y 2 ) / (2σ 2 )]; scale parameter σs = σ0·2 s / n ; Differential feature DF(x, y, s) = SP(x, y, s+1) - SP(x, y, s); Directional gradient DG(x, y, s) = [Gx(x, y, s) 2 +Gy(x,y,s) 2 ] 0.5 ; Feature response map FR(x, y, s) = DF(x, y, s)·DG(x, y, s)·w(s); where I(u, v) is the input image; σ0 is the reference scale; s is the scale index; n is the number of scale divisions; Gx and Gy are the gradients in the x and y directions; w(s) is the scale weight.

[0076] Time series feature extraction: time domain decomposition TD(t, s) = ∑[x(t)·ψ((tb) / 2 s )]; frequency domain transform FT(ω, s) = ∫TD(t, s) · exp(-jωt)dt; energy spectrum ES(s) = |FT(ω, s)| 2 ; Time-frequency feature TF(t, s) = TD(t, s)·ES(s)·exp(-s / τ); where x(t) is the time series; ψ(t) is the wavelet basis function; b is the translation parameter; s is the decomposition scale; τ is the time attenuation parameter.

[0077] Feature fusion: local feature LF(i, s) = max{FR(x, y, s)}; global feature GF(i, s) = mean{TF(t, s)}; combined feature CF(i) = ∑[LF(i, s)·α(s) + GF(i, s)·β(s)]; feature matrix FM(i, j) = CF(i)·RF(j); where α(s) and β(s) are scale fusion weights; RF(j) is the feature correlation factor.

[0078] Dynamic feature selection: information gain IG(f) = H(C) - H(C|f); feature importance FI(i) = IG(f)·[1 + log(1+var(f))]; selection threshold ST(t) = μ(FI) + k·σ(FI); final feature set FS = {f | FI(f) > ST(t)}; where H(C) is category entropy; H(C|f) is conditional entropy; var(f) is feature variance; k is threshold adjustment parameter.

[0079] Feature optimization: Objective function OF(w) = ∑[L(yi, f(xi, w)) + λ·||w|| 2 + γ·TR(w)]; temporal regularization term TR(w) = ∑|w(t)-w(t-1)|; gradient update w(t+1) = w(t) - η·ΞOF / Ξw; where L is the loss function; λ and γ are regularization parameters; η is the learning rate; w is the feature weight, and Ξ is the partial derivative.

[0080] This embodiment realizes the comprehensive capture and optimal combination of crop growth data features through multi-scale feature extraction and fusion mechanism. Specifically, firstly, a multi-scale convolution kernel group is used for feature extraction, which can simultaneously consider the changing characteristics of crop growth data at different spatiotemporal scales, from the microscopic leaf growth state to the macroscopic field group performance. Through the feature importance evaluation mechanism, the features of different scales and sources are quantitatively evaluated, and the most valuable feature combination for crop growth prediction can be accurately identified, such as the most indicative morphological features, spectral features or meteorological features are respectively selected at different growth stages. Finally, through the feature fusion process, the evaluated features are optimized and combined, which not only considers the individual importance of the features, but also considers the interaction relationship between the features through the conduction influence matrix, thereby constructing a comprehensive feature system that can reflect the current growth state of crops and predict future development trends. This embodiment improves the characterization ability of the growth state of crops and provides a more reliable feature basis for subsequent prediction analysis.

[0081] According to one aspect of the present application, step S32 is further:

[0082] S321, based on the multi-scale feature set, extract feature contribution information to obtain a contribution vector; based on the contribution vector, calculate the feature combination effect to generate a combination effect matrix; normalize the combination effect matrix to obtain an initial importance vector;

[0083] S322, constructing a feature correction model based on the initial importance vector and the conduction influence matrix to obtain a correction parameter vector; adjusting the feature importance based on the correction parameter vector to generate an adjustment coefficient matrix; updating the feature importance according to the adjustment coefficient matrix and outputting a correction importance vector;

[0084] S323. Based on the modified importance vector, construct an importance stability evaluation index to obtain a stability index vector; based on the stability index vector, optimize the modified importance vector to generate an optimized importance vector; convert the optimized importance vector into a matrix form to obtain a feature importance matrix.

[0085] In one embodiment of the present application, the feature contribution: information gain IG(f) = H(Y) - H(Y|f); conditional entropy H(Y|f) = -Σp(f)·Σp(y|f)·log[p(y|f)]; combined effect CE(fi, fj) = IG(fi, fj) - [IG(fi) + IG(fj)]; feature importance FI(i) = IG(fi) + λ·ΣCE(fi, fj)·w(t); wherein H(Y) is the system entropy; p(f) is the feature probability distribution; p(y|f) is the conditional probability; IG(fi, fj) is the joint information gain of the feature pair; λ is the combined effect weight; w(t) is the timing adjustment factor.

[0086] This embodiment realizes the accurate quantification and optimization selection of crop growth characteristics by constructing a feature importance evaluation and dynamic adjustment mechanism. First, by extracting feature contribution information and calculating feature combination effects, the most valuable feature combination for crop growth prediction can be accurately identified, which is particularly suitable for processing feature screening problems in multi-source data fusion scenarios. For example, the most indicative feature combination can be automatically identified at different growth stages, such as focusing on leaf area index and chlorophyll content in the vegetative growth period, while more attention is paid to ear characteristics and grain filling indicators in the reproductive growth period. Then, the importance is dynamically adjusted through the feature correction model. This correction mechanism that considers the conduction effect can timely reflect the temporal changes in feature importance, such as automatically increasing the weight of water-related features under drought stress, and increasing the importance of the spot index when pests and diseases occur. Finally, by constructing an importance stability evaluation index for optimization, the reliability and stability of the feature importance evaluation results are ensured. This optimization mechanism can effectively balance the influence of short-term fluctuations and long-term trends, and provide a more reliable basis for feature selection.

[0087] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0088] S41. Based on the current data credibility vector, calculate the basic weight to obtain a basic weight vector; based on the basic weight vector and the conduction influence matrix, adjust the weight to generate an adjusted weight vector; normalize the adjusted weight vector and output an initial weight vector;

[0089] S42, based on the initial weight vector and the fusion feature matrix, construct a sequence prediction model to obtain a weight prediction model; use the weight prediction model to predict the weight change trend and generate a prediction weight sequence; convert the prediction weight sequence into a matrix form to obtain a dynamic weight matrix;

[0090] S43. Based on the dynamic weight matrix, a set of weight constraints is constructed to obtain a constraint matrix; based on the constraint matrix, the weight distribution is optimized to generate an optimized weight sequence; the optimized weight sequence is converted into a matrix form to obtain an optimized weight matrix.

[0091] In one embodiment of the present application, weight prediction: time series feature vector TV(t) = [H(t), V(t), M(t), S(t)]; prediction model PM(t+1) = LSTM[TV(t), TV(t-1), ..., TV(tk)]; prediction weight PW(t+1) = softmax(PM(t+1))·AF(t); wherein H(t) is the historical trend feature; V(t) is the volatility feature; M(t) is the momentum feature; S(t) is the seasonal feature; k is the lookback window size; LSTM is the long short-term memory network; AF(t) is the attention factor.

[0092] In another embodiment of the present application, the weight prediction model: weight change ΔW(t) = α·[W(t)-W(t-1)] + β·∑[W(t)-W(tk)]·exp(-k / τ) + γ·▽L(t); prediction weight WP(t+1) = W(t) +ΔW(t)·[1+η·S(t)]; stability coefficient S(t) = 1 / [1+exp(κ·var(W(t)))]; final weight WF(t+1)= softmax[WP(t+1)·Q(t)]; wherein α is the momentum coefficient; β is the historical influence coefficient; γ is the gradient coefficient; τ is the time decay parameter; k is the lookback window size; η is the stability adjustment coefficient; κ is the variance sensitivity; Q(t) is the quality adjustment factor; L(t) is the loss function; var(W(t)) is the weight variance.

[0093] This embodiment realizes adaptive weight adjustment of different data sources and features in crop growth prediction by establishing a dynamic weight allocation and optimization mechanism. First, the initial weight allocation is calculated based on data credibility and conduction influence, and the influence of different data sources in the prediction model can be reasonably allocated according to their historical performance and current status, which is particularly suitable for processing the fusion problem of multi-source heterogeneous data in crop growth monitoring. Then, by constructing a weight prediction model, a dynamic prediction of the weight change trend is realized, and the dynamic changes of the importance of various influencing factors in the crop growth process can be captured in time, such as automatically adjusting the weights of environmental factors such as temperature, moisture, and light at different growth stages. Finally, through weight consistency optimization, the rationality and stability of weight allocation are ensured, and the adverse effects of drastic fluctuations in weights on the prediction results can be effectively avoided, while ensuring the continuity and smoothness of weight allocation at different growth stages. This embodiment forms an adaptive weight optimization system that can automatically adjust the influence of each prediction factor according to the dynamic characteristics of crop growth.

[0094] According to one aspect of the present application, step S42 is further:

[0095] S421. Receive the initial weight vector and the fusion feature matrix, build a weighted time series feature extraction model, and obtain a time series feature vector; establish a sequence prediction rule based on the time series feature vector and generate a prediction rule matrix; build a prediction model based on the prediction rule matrix and output a weighted prediction model.

[0096] S422, receiving a weight prediction model, extracting the changing trend of each weight component, and obtaining a trend feature matrix; performing weight sequence prediction based on the trend feature matrix, and generating a prediction sequence set; performing combination optimization on the prediction sequence set, and outputting a prediction weight sequence.

[0097] S423, receiving the prediction weight sequence, calculating the deviation between the predicted value and the actual value, and obtaining the deviation evaluation vector; constructing the correction rule based on the deviation evaluation vector, and generating the correction parameter matrix; adjusting the prediction sequence according to the correction parameter matrix, converting the adjusted sequence into a matrix form, and outputting the dynamic weight matrix.

[0098] In one embodiment of the present application, the sequence prediction model: feature vector X(t) = [H(t), S(t), T(t), V(t)]; sequence prediction value P(t+1) = W1·LSTM[X(t)] + W2·GRU[X(t)] + W3·TCN[X(t)]; attention weight A(t) = softmax[Q(t)·K(t) / sqrt(d)]·V(t); final prediction F(t+1) = P(t+1)·A(t)·[1+γ·M(t)]; wherein H(t) is the historical trend feature; S(t) is the seasonal feature; T(t) is the periodic feature; V(t) is the fluctuation feature; W1, W2, W3 are model fusion weights; Q(t), K(t), V(t) are attention calculation parameters; d is the feature dimension; γ is the momentum factor; M(t) is the memory item.

[0099] Weighted time series feature extraction: autocorrelation coefficient AC(k) = cov(w(t), w(tk)) / var(w(t)); trend strength TS(t) = ∑[w(t)-MA(t)]·exp(-k / τ); periodic indicator PI(t) = max{|FFT(w(t))|}; feature vector FV(t) = [AC(1:k), TS(t), PI(t), V(t)]; where k is the lag order; MA is the moving average; τ is the time constant; V(t) is the volatility.

[0100] This embodiment realizes the intelligent adjustment and precise control of the weights of the crop growth prediction model by establishing a weight sequence prediction and dynamic optimization mechanism. Specifically, a weight time series feature extraction model is first constructed, which can accurately capture the dynamic law of weight changes, and is particularly suitable for processing the periodic changes and trend changes of the weights of different influencing factors in the process of crop growth. For example, it can accurately predict the weight change trend of environmental factors such as light, temperature, and moisture at different growth stages. Then, by analyzing the change trend of each weight component for sequence prediction, the turning point and anomaly of the weight change can be discovered in time, such as adjusting the weights of relevant environmental factors in advance before the extreme weather arrives. Finally, by calculating the prediction deviation and constructing the correction rule, the dynamic optimization of the weight prediction result is realized, and the prediction model can be continuously adjusted and improved according to the feedback of the prediction effect to ensure the accuracy and reliability of the weight prediction.

[0101] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0102] S51. Based on the fusion feature matrix, a prediction model group including different spatiotemporal scales is constructed to obtain a prediction model set; the prediction model set is used to perform hierarchical prediction on the feature data to generate a prediction result sequence; the prediction result sequence is organized according to the spatiotemporal scale to obtain a multi-scale prediction result set;

[0103] S52. Based on the multi-scale prediction result set, a scale constraint relationship model is constructed to obtain a scale constraint matrix; based on the scale constraint matrix and the optimized weight matrix, the prediction results are weightedly fused to generate a fused prediction sequence; the fused prediction sequence is processed for consistency and the collaborative prediction result is output.

[0104] In one embodiment of the present application, multi-scale prediction is fused: scale correlation SC(i, j) = exp(-|si-sj| / τ)·corr(Pi, Pj); fusion weight FW(i) = softmax[SC(i, ·)·PR(i)]; final prediction FP(t) = ΣFW(i)·Pi(t)·Q(i, t); wherein si, sj are prediction scales; τ is a scale attenuation parameter; corr(Pi, Pj) is a prediction result correlation coefficient; PR(i) is prediction reliability; Q(i, t) is a quality adjustment factor; Pi(t) is a prediction result of each scale.

[0105] This embodiment realizes accurate prediction and dynamic adjustment of crop growth trends by constructing a multi-level prediction model and a collaborative prediction mechanism. First, a multi-level prediction model is used for hierarchical prediction. This hierarchical prediction strategy can simultaneously consider short-term fluctuations and long-term trends of crop growth, from changes in physiological indicators at the daily scale to the growth and development process at the seasonal scale. By establishing a constraint relationship model between scales, the consistency of prediction results at different time scales is ensured. This constraint mechanism is particularly suitable for processing multi-scale dynamic characteristics in the process of crop growth, such as ensuring the coordination and unification of short-term prediction results and long-term growth trends. Finally, the collaborative integration of prediction results based on the optimized weight matrix realizes the optimization and fusion of the results of different prediction models, and can make full use of the advantages of different models, such as combining the stability of statistical models and the nonlinear fitting ability of machine learning models, so as to obtain more accurate and reliable prediction results. This embodiment forms a multi-level, multi-scale prediction system that can adapt to complex changes in the process of crop growth.

[0106] According to one aspect of the present application, step S52 is further:

[0107] S521, receiving a set of multi-scale prediction results, extracting the prediction relationship between different scales, and obtaining a scale relationship vector; constructing constraint rules based on the scale relationship vector and generating a constraint rule matrix; establishing a scale constraint model according to the constraint rule matrix and outputting the scale constraint matrix.

[0108] S522, receiving a multi-scale prediction result set and an optimized weight matrix, constructing a result fusion model, and obtaining a fusion parameter vector; calculating a weighted fusion result based on the fusion parameter vector, and generating an initial fusion sequence; performing boundary constraint processing on the initial fusion sequence, and outputting a fusion prediction sequence.

[0109] S523, receiving the fused prediction sequence and the scale constraint matrix, calculating the consistency index of the prediction result, and obtaining the consistency vector; constructing the collaborative optimization rule based on the consistency vector, and generating the optimization adjustment matrix; collaboratively adjusting the prediction result according to the optimization adjustment matrix, and outputting the collaborative prediction result.

[0110] In one embodiment of the present application, the scale constraint relationship is: scale consistency CI(i, j) = 1 - ||Pi-Pj|| / max(||Pi||, ||Pj||); constraint strength CS(i, j) = CI(i, j)·exp(-|si-sj| / τ); constraint matrix CM(i, j) = {CS(i, j), if CS(i, j) > θ; 0, otherwise}; wherein Pi, Pj are prediction results of different scales; si, sj are scale parameters; τ is the scale attenuation coefficient; θ is the constraint threshold.

[0111] This embodiment achieves precise control and dynamic adjustment of crop growth prediction by establishing a collaborative optimization mechanism for multi-scale prediction results. First, by extracting the prediction relationship between different scales to construct constraint rules, the consistency of prediction results at different time scales can be ensured, such as ensuring the coordination and unification of daily scale prediction results with weekly and monthly scale prediction trends. Then, by weightedly combining the multi-scale prediction results through the fusion model, the advantages of different prediction models can be fully utilized, such as combining the nonlinear fitting ability of the machine learning model and the stability of the statistical model. Finally, by constructing collaborative optimization rules to adjust the prediction results, not only the prediction accuracy is considered, but also the practicality and interpretability of the prediction results are considered, which can provide a more reliable prediction basis for agricultural production decisions.

[0112] like Figure 7 As shown, according to one aspect of the present application, step S6 is further:

[0113] S61. Based on the multi-scale prediction result set and the collaborative prediction result, a prediction result deviation calculation model is constructed to obtain a deviation matrix; based on the deviation matrix, a consistency evaluation index system is constructed to generate an evaluation index vector; based on the evaluation index vector, a consistency evaluation report is generated;

[0114] S62. Based on the consistency assessment report and collaborative prediction results, construct a result optimization objective function to obtain an optimization target matrix; based on the optimization target matrix, optimize and adjust the collaborative prediction results to generate an optimized prediction sequence; perform final processing on the optimized prediction sequence and output the final prediction result.

[0115] In one embodiment of the present application, the prediction result deviation calculation model is: deviation score D(t) = {[y(t)-y*(t)] 2 ·exp(-t / τ) + α·| y* (t)-MA(t)|·σ(t) + β·[y* (t)- y* (t-1)] 2}·Q(t); time series weight W(t) = softmax[D(t)·exp(-λ·t)]; weighted deviation WD(t) = ∑[W(t)·D(t)]; where MA(t) is the moving average; σ(t) is the local standard deviation; Q(t) is the data quality coefficient; τ is the time decay parameter; α, β are balance coefficients; λ is the time series decay rate, y(t) is the actual observation value at time t, and y*(t) is the predicted value at time t.

[0116] Result optimization objective function: Optimization objective F(t) = w1·∑[y(t)- y*(t)] 2 / N + w2·∑| y*(t)- y*(t-1)| / N + w3·∑| y*(t)-MP(t)| / N + w4·TV(y*) + w5·R(y*); Constraints C1:y*min ≤ y* (t) ≤ y*max; C2: | y*(t)- y*(t-1)| ≤ Δ; C3: |∑wi-1| ≤ ε; where y(t) is the actual value; y*(t) is the predicted value; N is the number of samples; MP(t) is the multi-model ensemble predicted value; TV(y*) is the total variation regularization term; R(y*) is the smoothness regularization term; w1 to w5 are adaptive weight coefficients; Δ is the maximum change threshold; ε is the weight error threshold, y*min is the minimum predicted value, and y*max is the maximum predicted value.

[0117] This embodiment achieves quality control and precision improvement of crop growth prediction results by establishing a prediction result evaluation and optimization mechanism. Specifically, by first calculating the deviations between prediction results of different scales and generating a consistency evaluation report, the reliability of the prediction results can be comprehensively tested, especially when dealing with prediction deviations at different growth stages of crops. Then, an optimization objective function is constructed based on the consistency evaluation report to achieve directional optimization of the prediction results, and the optimization target can be automatically adjusted according to the characteristics of different growth stages, such as paying more attention to the accuracy of the prediction in the critical growth period and paying more attention to the smoothness of the prediction in the stable growth period. Finally, through the result optimization and adjustment process, the reliability and practicality of the final prediction results are ensured, taking into account not only the prediction accuracy, but also the practical application value of the prediction results, such as ensuring that the prediction results can provide effective guidance for agricultural activity decision-making. This embodiment forms a closed-loop quality control system that can continuously improve the quality of the prediction results.

[0118] According to one aspect of the present application, step S62 is further:

[0119] S621. Receive the consistency assessment report and collaborative prediction results, extract the optimization target features, and obtain the target feature vector; construct the optimization constraints based on the target feature vector and generate the constraint matrix; establish the target optimization model according to the constraint matrix and output the optimization target matrix.

[0120] S622, receiving the optimization target matrix and the collaborative prediction results, constructing a multi-objective optimization model, and obtaining an optimization parameter vector; performing optimization calculation based on the optimization parameter vector to generate an initial optimization sequence; smoothing the initial optimization sequence, and outputting an optimization prediction sequence.

[0121] S623, receiving the optimized prediction sequence, calculating the stability index of the prediction result, and obtaining the stability vector; constructing the correction rule based on the stability vector, and generating the correction parameter matrix; making the final adjustment to the optimized sequence according to the correction parameter matrix, and outputting the final prediction result.

[0122] In one embodiment of the present application, the optimization target is constructed as follows: accuracy target AC(t) = -∑[y(t)-y*(t)] 2 / n; smoothness target SM(t) = -∑| y*(t)- y*(t-1)|; consistency target CO(t) = -∑| y*(t)-MP(t)|; combined target OF(t) = w1·AC(t) + w2·SM(t) + w3·CO(t); where y(t) is the actual value; y*(t) is the predicted value; MP(t) is the multi-scale average prediction; w1, w2, w3 are the target weights.

[0123] Stability assessment: Stability index SI(t) = 1 - {Σ[FP(t)-MA(t)] 2 / (n·var(t))}; correction parameter MP(t) = exp(-SI(t) / α)·[1+log(1+σ(t))]; final prediction value FV(t) = FP(t)·[1+β·MP(t)·sign(ΔFP(t))]; where MA(t) is the moving average; n is the number of samples; var(t) is the prediction variance; α is the stability adjustment parameter; σ(t) is the prediction standard deviation; β is the correction strength coefficient; ΔFP(t) is the prediction change.

[0124] This embodiment achieves quality control and precision improvement of crop growth prediction results by constructing a prediction result optimization and stability evaluation mechanism. First, by extracting optimization target features and constructing constraint conditions, it is possible to simultaneously consider multiple aspects such as prediction accuracy, prediction stability, and actual application needs, which is particularly suitable for dealing with complex optimization problems in crop growth prediction. Then, the prediction results are optimized and adjusted through a multi-objective optimization model, which can ensure the smoothness and continuity of the prediction results while ensuring the accuracy of the prediction, and avoid the adverse effects of drastic fluctuations in the prediction results on agricultural production decisions. Finally, by calculating the stability index and constructing correction rules, the final adjustment of the optimization results is achieved, which can further improve the reliability and practicality of the prediction results, and provide more stable and reliable decision support for agricultural production management. This embodiment forms a complete quality control system that can continuously improve the practical value of the prediction results.

[0125] In another embodiment of the present application, step S1 can also be: receiving the original remote sensing data set, the ground sensor data set and the historical agricultural data set, constructing a time window and performing data preprocessing to generate a time window sequence matrix; performing missing value filling according to the time window sequence matrix to obtain a complete data matrix; based on the complete data matrix and the pre-stored historical credibility assessment data, calculating the correlation and credibility between data sources, and outputting the data source time series correlation tensor and the current data credibility vector. Specifically:

[0126] S1a, receiving the original remote sensing data set, ground sensor data set and historical agricultural data set, constructing the initial time window, and generating the initial time window sequence; performing data standardization according to the initial time window sequence to obtain the standardized data sequence; organizing the standardized data sequence by time window, and outputting the time window sequence matrix.

[0127] S1b, receiving the time window sequence matrix, calculating the statistical characteristics of each time window, and generating a window statistical characteristic matrix; interpolating missing values ​​based on the window statistical characteristic matrix to obtain a completed data sequence; performing noise filtering on the completed data sequence, and outputting a complete data matrix.

[0128] S1c, receiving the complete data matrix and historical credibility assessment data, calculating the data source state transition probability, and obtaining the state transition matrix; calculating the time series correlation based on the state transition matrix, and generating the correlation feature tensor; evaluating the data credibility according to the correlation feature tensor, and outputting the data source time series correlation tensor and the current data credibility vector.

[0129] This embodiment generates a complete data matrix by standardizing and filling missing values ​​of the original remote sensing data set, ground sensor data set and historical agricultural data set, ensuring the integrity and consistency of the data, improving the data quality, and providing a reliable data basis for subsequent analysis. By constructing a time window and calculating the correlation between data sources, a data source time series correlation tensor is generated, which can accurately capture the time dependency between data sources and provide an important reference for data analysis and prediction. By dynamically evaluating the credibility of data, anomalies and inconsistencies in data can be discovered and processed in a timely manner, improving the accuracy and reliability of data analysis. This embodiment can effectively integrate remote sensing data, ground sensor data and historical agricultural data, provide comprehensive crop growth information, and help improve the accuracy and reliability of crop growth prediction. By calculating the state transition probability and time series correlation of the data source, a correlation feature tensor is generated, and the credibility of the data is evaluated based on the correlation feature tensor, which provides scientific support data for crop growth prediction and agricultural decision-making, and improves the accuracy and effectiveness of prediction and decision-making.

[0130] In another embodiment of the present application, step S2 may also be: receiving a complete data matrix, extracting features using a multi-scale decomposition method, and obtaining a multi-scale feature tensor; detecting abnormal points based on the multi-scale feature tensor, and generating an abnormal point feature set; constructing an abnormal conduction network based on the abnormal point feature set and analyzing the conduction impact, and outputting a conduction impact matrix. Specifically:

[0131] S2a, receiving the complete data matrix, performing multi-layer wavelet transform, and obtaining the wavelet coefficient matrix; reconstructing and fusing the wavelet coefficient matrix to generate a reconstructed feature sequence; organizing the reconstructed feature sequence into a tensor form, and outputting a multi-scale feature tensor.

[0132] S2b, receiving multi-scale feature tensors, calculating local density distribution, and obtaining a density feature matrix; identifying outliers based on the density feature matrix, generating an initial outlier set; performing feature extraction on the initial outlier set, and outputting an outlier feature set.

[0133] S2c, receiving the feature set of abnormal points, constructing the conduction network topology, and generating the conduction network matrix; calculating the conduction path and strength based on the conduction network matrix, and obtaining the conduction feature vector; evaluating the impact range and degree according to the conduction feature vector, and outputting the conduction impact matrix.

[0134] This embodiment uses multi-layer wavelet transform to perform multi-scale decomposition on the complete data matrix, which can extract features at different scales and generate multi-scale feature tensors, so as to capture multi-level information in the data and improve the accuracy and comprehensiveness of feature extraction. The local density analysis method is used to identify abnormal points and generate abnormal point feature sets, which can accurately detect abnormal points in the data and analyze their abnormal degree, providing a reliable basis for subsequent abnormal conduction analysis. By constructing the conduction network topology, generating the conduction network matrix, and calculating the conduction path and strength, the conduction relationship between the abnormal points can be fully analyzed, which can reveal the mutual influence and conduction mechanism between the abnormal points and provide a scientific basis for abnormal management and control. Based on the conduction feature vector, the influence range and degree of the abnormal point are evaluated, and the conduction influence matrix is ​​output, which can quantify the conduction influence of the abnormal point, help identify key abnormal points and high-risk areas, and improve the effectiveness of abnormal management. This embodiment helps to fully understand and evaluate the abnormal situation in the data and provide strong support for scientific decision-making.

[0135] In another embodiment of the present application, step S111 can also be: receiving the original remote sensing data set, the ground sensor data set and the historical agricultural data set, extracting the spatiotemporal features to construct a time index tree and a space index tree, and generating a spatiotemporal feature vector; constructing a multidimensional index structure based on the spatiotemporal feature vector to obtain an index structure matrix; converting the index structure matrix into a standardized format, and outputting the spatiotemporal index matrix and the basic feature vector.

[0136] Step S31 may also be: receiving a complete data matrix, a multi-scale feature tensor and a basic feature vector, fusing the spatiotemporal feature information to construct a multi-scale convolution kernel, and obtaining an enhanced convolution kernel matrix; extracting features using the enhanced convolution kernel matrix to obtain a feature component set; and combining the feature component set to generate a multi-scale feature set.

[0137] This embodiment generates a spatiotemporal feature vector by extracting spatiotemporal features and constructing a time index tree and a space index tree, which can more accurately capture the spatiotemporal information in the data, improve the accuracy of feature extraction, and provide a reliable basis for subsequent data analysis and processing. By fusing spatiotemporal feature information to construct a multi-scale convolution kernel and obtain an enhanced convolution kernel matrix, it is possible to more comprehensively extract multi-scale features in the data, which helps to improve the performance and prediction accuracy of the model. The enhanced convolution kernel matrix is ​​used to extract features, obtain a set of feature components, and combine them to generate a multi-scale feature set, which can capture multi-level information in the data and provide rich feature support for subsequent analysis and prediction. This embodiment helps to further improve the accuracy and reliability of crop growth prediction.

[0138] In another embodiment of the present application, step S132 may also be: receiving the state transfer matrix and the data source timing correlation tensor, extracting the timing correlation characteristics, and obtaining the correlation feature vector; evaluating the data source state based on the correlation feature vector, and generating a state evaluation vector; combining the correlation feature vector and the state evaluation vector, and outputting the weight reference matrix.

[0139] Step S41 can also be: receiving the current data credibility vector and the weight reference matrix, calculating the basic weight of the data source, and obtaining the basic weight vector; adjusting the weight based on the basic weight vector and the conduction influence matrix to generate an adjusted weight vector; performing normalization processing on the adjusted weight vector and outputting the initial weight vector.

[0140] This embodiment receives the state transfer matrix and the data source timing correlation tensor, extracts the timing correlation features and generates a correlation feature vector, which can more accurately evaluate the state of the data source, improve the accuracy of the data source state evaluation, and provide a reliable basis for subsequent weight calculations. By calculating the basic weight of the data source, and adjusting the weight in combination with the conduction influence matrix to generate an adjusted weight vector, it can dynamically reflect the changes in the data source and improve the scientificity and rationality of the weight distribution. This embodiment helps to fully understand and evaluate the actual situation of the data source and provide strong support for scientific decision-making; at the same time, through accurate state evaluation and dynamic weight adjustment, it can provide decision makers with scientific decision support data, help decision makers formulate more effective management and planning strategies, and improve the ability to deal with complex environments and uncertainties.

[0141] In another embodiment of the present application, step S213 can also be: receiving a set of abnormal points and a multi-scale feature tensor, calculating a local density deviation value, and obtaining a density deviation vector; constructing an abnormality assessment model based on the density deviation vector and generating an abnormality assessment matrix; converting the abnormality assessment matrix into a standard format, and outputting an abnormality degree matrix and a density feature matrix.

[0142] The method also includes step S214, receiving a multi-scale feature tensor and a density feature matrix, calculating a comprehensive density distribution feature, and obtaining a comprehensive density distribution matrix; based on the comprehensive density distribution matrix, constructing an improved density distance map, and generating an improved distance relationship matrix; according to the improved distance relationship matrix, calculating a weighted local anomaly factor, and obtaining an optimized anomaly factor vector; based on the optimized anomaly factor vector, determining abnormal data points, and outputting an optimized anomaly point set.

[0143] This embodiment calculates the local density deviation value and constructs an anomaly assessment model to generate an anomaly assessment matrix, which can more accurately identify and evaluate anomalies in the data, improve the accuracy of anomaly detection, and provide a reliable basis for subsequent anomaly management and control. By calculating the comprehensive density distribution characteristics and constructing an improved density distance map, an improved distance relationship matrix is ​​generated, which can comprehensively analyze the density distribution characteristics in the data, help to better understand the structure and distribution of the data, and improve the comprehensiveness of anomaly detection. The optimized set of anomaly points output by this embodiment provides high-quality data support for anomaly management and control, and helps to improve the efficiency and effectiveness of anomaly handling.

[0144] The present invention realizes the full-process intelligent processing of crop growth prediction by constructing a complete multi-source data fusion analysis framework. First, a reliable data foundation is established through multi-source data preprocessing and credibility assessment; then, through anomaly detection and conduction analysis, early identification and impact assessment of growth anomalies are achieved; then, through multi-scale feature extraction and fusion, the key characteristics of crop growth are captured; then, through dynamic weight optimization, adaptive adjustment of the prediction model is achieved; finally, through multi-level prediction and result optimization, the accuracy and reliability of the prediction results are ensured. The present invention forms a closed-loop intelligent prediction system, which improves the accuracy and practicality of crop growth prediction and can provide reliable data support for agricultural production decision-making; it realizes the intelligent fusion of multi-source heterogeneous data, establishes a dynamic evaluation mechanism for data credibility, constructs a networked analysis model for abnormal conduction, develops a multi-scale fusion method for feature extraction, and designs a collaborative optimization strategy for the prediction model, which can effectively solve the complex problems in crop growth prediction.

[0145] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A crop growth prediction method based on multi-source data fusion analysis, characterized in that: The steps include: S1, obtain the original remote sensing data set, ground sensor data set and historical agricultural data set, perform standardization and missing value filling to obtain a complete data matrix; Based on the complete data matrix, a time window is constructed and the correlation between data sources is calculated to generate a data source time series correlation tensor; The current data credibility vector is calculated based on the pre-stored historical credibility evaluation data and the data source time series correlation tensor; S2. Based on the complete data matrix, a multi-scale decomposition method is used to extract features and obtain a multi-scale feature tensor. Based on the multi-scale feature tensor and the data source time series correlation tensor, abnormal points are detected and the degree of abnormality is analyzed to generate an abnormal degree matrix. Based on the abnormal degree matrix and the data source time series correlation tensor, an abnormal conduction network is constructed and the conduction impact is evaluated, and a conduction impact matrix is ​​output. S3, based on the complete data matrix and the multi-scale feature tensor, a multi-scale feature extraction method is used to generate a multi-scale feature set; Based on the multi-scale feature set, feature importance evaluation is performed to obtain the feature importance matrix; According to the feature importance matrix and the conduction influence matrix, feature fusion is performed to obtain a fused feature matrix; S4. Calculate the initial weight distribution based on the current data credibility vector and the conduction influence matrix to obtain an initial weight vector; Based on the initial weight vector and the fusion feature matrix, a weight prediction model is constructed to generate a dynamic weight matrix; Perform consistency optimization on the dynamic weight matrix to obtain the optimized weight matrix; S5. Based on the fusion feature matrix, a multi-level prediction model is used to perform hierarchical prediction to obtain a multi-scale prediction result set; based on the multi-scale prediction result set and the optimized weight matrix, the prediction results are collaboratively integrated to output the collaborative prediction results; S6. Based on the multi-scale prediction result set and the collaborative prediction result, calculate the deviation between the prediction results of different scales to obtain a consistency evaluation report; based on the consistency evaluation report, optimize and adjust the collaborative prediction result to output the final prediction result.

2. The crop growth prediction method based on multi-source data fusion analysis according to claim 1 is characterized in that: Step S1 is further as follows: S11, receiving the original remote sensing data set, the ground sensor data set and the historical agricultural data set, constructing a three-dimensional index structure containing spatial position and time information, and forming a spatiotemporal index matrix; According to the spatiotemporal index matrix, the numerical data in the original remote sensing data set, ground sensor data set and historical agricultural data set are standardized to generate a standardized data matrix; according to the data change trend of adjacent time points in the standardized data matrix, the missing values ​​are filled by interpolation method to generate a complete data matrix; S12. Based on the complete data matrix, a dynamic sliding time window is constructed to obtain a time window sequence; In each time window sequence, the correlation coefficients between different data sources are calculated to generate a correlation coefficient matrix; Combine the correlation coefficient matrices in the time dimension to form a data source time series correlation tensor; S13, reading pre-stored historical credibility evaluation data, and based on the historical credibility evaluation data, calculating the transition probability between data source states, and generating a state transition matrix; According to the state transfer matrix and the data source timing correlation tensor, the state of each data source in the current time window is evaluated to obtain the evaluated data source state; Based on the evaluation of the data source status and the preset time series correlation significance level, the data source is divided into different trustworthy states, and finally the current data credibility vector is obtained.

3. The crop growth prediction method based on multi-source data fusion analysis according to claim 2 is characterized in that: Step S2 is further as follows: S21. Based on the complete data matrix, the wavelet transform method is used to perform multi-level decomposition on the data to obtain a multi-scale feature tensor; based on the multi-scale feature tensor, the local density analysis method is used to identify abnormal points and generate an abnormal point set; according to the local density deviation degree of each abnormal point in the abnormal point set, the abnormal degree matrix is ​​calculated; S22. Based on the data source time series correlation tensor, a directed weighted graph is constructed to obtain an initial conduction graph; based on the abnormal point set and the abnormal degree matrix, the conduction weights between the nodes in the initial conduction graph are calculated to obtain a weighted conduction graph; based on the weighted conduction graph, an abnormal conduction adjacency matrix is ​​generated; S23, based on the abnormal conduction adjacency matrix and the current data credibility vector, calculate the conduction importance of each abnormal point to obtain a conduction importance vector; based on the abnormal conduction adjacency matrix, calculate the conduction distance between each node; Based on the conduction importance vector and the conduction distance, a conduction influence attenuation function is constructed to generate a conduction attenuation matrix; the conduction attenuation matrix is ​​combined with the abnormality degree matrix to output the conduction influence matrix.

4. The crop growth prediction method based on multi-source data fusion analysis according to claim 3 is characterized in that: Step S3 is further as follows: S31. Based on the complete data matrix and the multi-scale feature tensor, a multi-scale convolution kernel group is constructed to obtain a convolution kernel matrix; based on the convolution kernel matrix, feature extraction is performed to obtain a feature component set; Combine feature component sets to generate multi-scale feature sets; S32. Based on the multi-scale feature set, a contribution analysis method is used to calculate the importance of each feature to obtain an initial importance vector; based on the initial importance vector and the conduction influence matrix, the feature importance is corrected to obtain a corrected importance vector; the corrected importance vector is converted into a matrix form, and a feature importance matrix is ​​output; S33, constructing a feature fusion weight function based on the multi-scale feature set and the feature importance matrix to obtain a fusion weight vector; Based on the fusion weight vector, the features are weighted and combined to generate the initial fusion matrix; based on the initial fusion matrix and the conduction influence matrix, optimization and adjustment are performed to obtain the fusion feature matrix.

5. The crop growth prediction method based on multi-source data fusion analysis according to claim 4 is characterized in that: Step S4 is further as follows: S41. Based on the current data credibility vector, calculate the basic weight to obtain a basic weight vector; based on the basic weight vector and the conduction influence matrix, adjust the weight to generate an adjusted weight vector; normalize the adjusted weight vector and output an initial weight vector; S42, based on the initial weight vector and the fusion feature matrix, construct a sequence prediction model to obtain a weight prediction model; use the weight prediction model to predict the weight change trend and generate a prediction weight sequence; convert the prediction weight sequence into a matrix form to obtain a dynamic weight matrix; S43, constructing a weight constraint condition set based on the dynamic weight matrix to obtain a constraint condition matrix; Optimize the weight distribution based on the constraint matrix and generate an optimized weight sequence; The optimized weight sequence is converted into a matrix form to obtain the optimized weight matrix.

6. The crop growth prediction method based on multi-source data fusion analysis according to claim 5 is characterized in that: Step S5 is further as follows: S51. Based on the fusion feature matrix, a prediction model group including different spatiotemporal scales is constructed to obtain a prediction model set; the prediction model set is used to perform hierarchical prediction on the feature data to generate a prediction result sequence; the prediction result sequence is organized according to the spatiotemporal scale to obtain a multi-scale prediction result set; S52, based on the multi-scale prediction result set, constructing an inter-scale constraint relationship model to obtain a scale constraint matrix; Based on the scale constraint matrix and the optimized weight matrix, the prediction results are weightedly fused to generate a fused prediction sequence; the fused prediction sequence is processed for consistency and the collaborative prediction results are output.

7. The crop growth prediction method based on multi-source data fusion analysis according to claim 6 is characterized in that: Step S6 is further as follows: S61. Based on the multi-scale prediction result set and the collaborative prediction result, a prediction result deviation calculation model is constructed to obtain a deviation matrix; based on the deviation matrix, a consistency evaluation index system is constructed to generate an evaluation index vector; Generate a consistency assessment report based on the assessment indicator vector; S62. Based on the consistency assessment report and the collaborative prediction results, a result optimization objective function is constructed to obtain an optimization objective matrix; Based on the optimization target matrix, the collaborative prediction results are optimized and adjusted to generate an optimized prediction sequence; the optimized prediction sequence is finally processed to output the final prediction results.

8. The crop growth prediction method based on multi-source data fusion analysis according to claim 7 is characterized in that: Step S11 is further as follows: S111, receiving the original remote sensing data set, the ground sensor data set and the historical farming data set, extracting the timestamp information and the spatial coordinate information of each data set, obtaining the time series vector and the spatial coordinate vector; constructing a time index tree based on the time series vector, and obtaining the time index matrix; According to the spatial coordinate vector, a spatial index tree is constructed to generate a spatial index matrix; Combine the time index matrix and the space index matrix to obtain the space-time index matrix; S112, based on the spatiotemporal index matrix, normalizing the numerical attributes of the original remote sensing dataset, the ground sensor dataset, and the historical farming dataset to obtain a normalized dataset; Perform noise smoothing on the normalized data set to generate a smoothed data set; perform outlier detection based on the smoothed data set to obtain a cleaned data set; Reorganize the cleaned data set according to the spatiotemporal index matrix and output the standardized data matrix; S113, based on the time span of missing data in the standardized data matrix, classify the missing positions to obtain a missing type matrix; based on the missing type matrix, select a corresponding interpolation strategy to generate an interpolation strategy vector; according to the interpolation strategy vector, divide the missing data into different processing groups to obtain a missing processing grouping matrix; S114. Based on the missing processing grouping matrix and the standardized data matrix, different interpolation strategies are used to fill the missing values ​​to obtain a filled data set; data smoothing is performed on the filled data set to generate a smoothed filled matrix; the smoothed filled matrix is ​​merged with the non-missing data in the standardized data matrix to output a complete data matrix.

9. The crop growth prediction method based on multi-source data fusion analysis according to claim 7 is characterized in that: Step S21 is further as follows: S211, based on the complete data matrix, construct a multi-layer wavelet transform coefficient matrix to obtain a wavelet coefficient set; based on the wavelet coefficient set, perform coefficient reconstruction to generate a reconstruction coefficient matrix; perform a scale fusion operation on the reconstruction coefficient matrix to obtain a fusion coefficient sequence; reorganize the fusion coefficient sequence into a tensor form to obtain a multi-scale feature tensor; S212. Based on the multi-scale feature tensor, calculate the local density distribution features to obtain a density distribution matrix; based on the density distribution matrix, construct a density distance map to generate a distance relationship matrix; according to the distance relationship matrix, calculate the local anomaly factor to obtain an anomaly factor vector; according to the anomaly factor vector, determine the abnormal data point and output the anomaly point set; S213. Based on the outlier point set and the multi-scale feature tensor, the local density deviation value of each outlier point is calculated to obtain a density deviation vector; based on the density deviation vector, an outlier degree assessment model is constructed to generate an assessment result vector; the assessment result vector is converted into a matrix form to obtain an outlier degree matrix.

10. The crop growth prediction method based on multi-source data fusion analysis according to claim 7, characterized in that: Step S32 is further as follows: S321, based on the multi-scale feature set, extract feature contribution information to obtain a contribution vector; based on the contribution vector, calculate the feature combination effect to generate a combination effect matrix; normalize the combination effect matrix to obtain an initial importance vector; S322, constructing a feature correction model based on the initial importance vector and the conduction influence matrix to obtain a correction parameter vector; adjusting the feature importance based on the correction parameter vector to generate an adjustment coefficient matrix; updating the feature importance according to the adjustment coefficient matrix and outputting a correction importance vector; S323. Based on the modified importance vector, construct an importance stability evaluation index to obtain a stability index vector; based on the stability index vector, optimize the modified importance vector to generate an optimized importance vector; convert the optimized importance vector into a matrix form to obtain a feature importance matrix.

Images