Crop evapotranspiration prediction method and system fusing time sequence and deep learning

Abstract

The invention relates to the technical field of agriculture, and particularly discloses a crop evapotranspiration prediction method and system fusing a time sequence and deep learning, and the method comprises the following steps: S1, data preprocessing; s2, performing depth time feature mining on the preprocessed data; s3, different branches respectively process features of different time scales and spatial scales, and the features extracted by each branch are fused at a subsequent level; s4, training and optimizing the deep learning model; s5, constructing a multi-dimensional prediction result evaluation system, and guiding the improvement and optimization of each step; the system comprises a data preprocessing module, a time sequence feature mining module, a deep learning model construction module, a model training and optimizing module and a prediction result evaluation and feedback module. Through cooperative work of multiple modules, the advantages of time sequence analysis and deep learning are fully played, the precision and adaptability of crop evapotranspiration prediction are improved, and a scientific basis is provided for agricultural water resource management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural technology, in particular to a crop evapotranspiration prediction method and system fusing time series and deep learning. BACKGROUND

[0002] In the field of agriculture, accurate prediction of crop evapotranspiration plays a crucial role in improving agricultural water resource utilization efficiency, ensuring agricultural product quality and safety, optimizing agricultural production layout, promoting sustainable development of agriculture, and enhancing agricultural climate adaptation ability. It is an indispensable key link for modern agricultural fine management. Traditional evapotranspiration prediction methods mainly rely on physical models and empirical formulas. Physical models usually require a large amount of meteorological data and soil parameters, and have poor adaptability in different regions. Empirical formulas have limited applicability in different crop types and regions.

[0003] In recent years, time series analysis and deep learning technology have been widely applied in many fields, but in the prediction of crop evapotranspiration, how to fuse these two technologies to fully exert their respective advantages is still a challenging problem. Existing fusion methods have shortcomings in data processing, feature extraction, model construction and prediction result evaluation, such as insufficient data processing leading to inaccurate feature extraction, poor model adaptability, and difficulty in evaluating the credibility of prediction results. SUMMARY

[0004] The purpose of the present application is to solve the shortcomings of existing technology in data processing, feature extraction, model construction and prediction result evaluation, and to propose a crop evapotranspiration prediction method and system fusing time series and deep learning, which fully utilizes the advantages of time series analysis and deep learning, improves the accuracy and adaptability of crop evapotranspiration prediction, and provides scientific basis for agricultural water resource management.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme: A crop evapotranspiration prediction method fusing time series and deep learning, comprising the following specific steps: S1: Data preprocessing: cleaning, normalizing, removing outliers, filling missing values, and unifying data scales of collected historical meteorological data, soil data and crop growth data, and extracting initial features through feature engineering, including but not limited to calculating the moving average of meteorological data, the gradient change of soil data and the growth rate of crop growth data, to enhance data quality and extract potential useful features, and prepare for subsequent model processing; S2: Time series feature mining: Utilize improved time series analysis algorithms, such as the enhanced version of wavelet transform-based time series decomposition method, to conduct deep time feature mining on the preprocessed data, which can more meticulously depict the time sequence characteristics of the data and provide more rich time dimension feature input for the deep learning model; S3: Deep learning model construction: Design a deep learning model with a multi-scale feature fusion architecture, namely a multi-branch LSTM-CNN hybrid neural network structure, different branches handle features of different time scales and spatial scales, and the features extracted by each branch are fused in subsequent layers, enabling the model to more effectively learn the complex spatio-temporal correlation in the data. Compared with traditional single-structure deep learning models, this architecture can more accurately depict the complex mechanism of crop evapotranspiration influenced by multiple factors, and improve prediction accuracy; S4: Model training and optimization: Train and optimize the deep learning model using adaptive learning rate optimization algorithms combined with Bayesian optimization strategies; S5: Prediction result evaluation and feedback: Build a multi-dimensional prediction result evaluation system, consider multiple indicators, compare the differences between the prediction results and the measured data, analyze the sensitivity of the crop growth model to evapotranspiration changes, determine the main sources of model prediction bias, such as weather data noise, crop growth stage division error, or model internal feature extraction defects, and feedback the evaluation results to S1-S3 to guide targeted improvement and optimization of each step, such as adjusting data preprocessing algorithms, improving time series decomposition strategies, or modifying deep learning model structures, to continuously iterate and upgrade the prediction model, constantly improve the accuracy, reliability, and practicality of the prediction results, form a complete, logically coherent, and self-optimizing crop evapotranspiration prediction system, effectively solve the problems of insufficient data processing, inaccurate feature extraction, poor model adaptability, and difficulty in evaluating the credibility of prediction results in existing fusion time series and deep learning crop evapotranspiration prediction methods.

[0006] As a further technical solution of the present application, in S2, deep time feature mining includes extracting long-term trends, seasonal fluctuations, and periodic patterns in the data, capturing subtle change patterns at different time scales, such as short-term fluctuation characteristics, and constructing a more comprehensive and accurate time series feature matrix.

[0007] As a further technical solution of the present invention, in S2, the improved time series analysis algorithm introduces a multi-resolution analysis framework and adaptive threshold processing. When performing wavelet decomposition, it automatically adjusts the number of decomposition layers and threshold parameters according to the energy distribution of different frequency components, so as to more accurately separate the signal features at different time scales. This reduces the feature loss or noise residue problems caused by fixed parameter settings in traditional wavelet decomposition, and provides a cleaner and more representative time series feature input for deep learning models, which helps to improve the learning effect and prediction accuracy of models on complex agricultural time series data.

[0008] As a further technical solution of the present invention, in S3, the multi-branch LSTM-CNN hybrid neural network structure includes a deep LSTM branch that specifically processes long-term trend features, a shallow LSTM branch that captures short-term fluctuation features, and a CNN branch that extracts spatial correlation features.

[0009] As a further technical solution of the present invention, in S3, the deep learning model of the multi-scale feature fusion architecture adopts an attention mechanism to enhance the feature fusion effect at each feature fusion point. By calculating the importance weight of the features extracted from different branches, the model can automatically focus on the feature combination that is more critical to the prediction of evapotranspiration and dynamically adjust the contribution of each branch feature to the final prediction result. Compared with the simple feature splicing fusion method, it can more effectively integrate multi-scale feature information and improve the feature expression ability and prediction performance of the model when processing complex agricultural scenario data.

[0010] As a further technical solution of the present invention, in S4, a relatively large but reasonable learning rate is used in the initial stage to enable the model to quickly converge to the vicinity of the optimal solution region. Subsequently, Bayesian optimization dynamically adjusts the hyperparameters based on the error feedback during the model training process, including fine adjustment of the learning rate and optimization of the regularization parameter, in order to avoid overfitting and further explore the performance potential of the model. Compared with conventional fixed learning rate or single optimization algorithm, it can more efficiently improve the model's generalization ability under complex agricultural environment data and ensure the stable prediction performance of the model on data from different regions, different crop types and different growth stages.

[0011] As a further technical solution of the present invention, in S4, the adaptive learning rate optimization algorithm combined with the Bayesian optimization strategy specifically includes: In the early stages of model training, a larger initial learning rate is used to quickly bring the model parameters closer to the optimal solution region of the loss function. When the slope of the training loss curve is lower than a preset threshold, the Bayesian optimization process is triggered. Based on the current hyperparameter settings of the model, a probabilistic model between hyperparameters and model validation error is constructed under the Gaussian process prior distribution. New hyperparameter candidate points are obtained by sampling, and their corresponding expected improvement values ​​are calculated. The hyperparameter combination with the largest expected improvement value is selected to update the model. Repeat this process until the model's error on the validation set meets the convergence condition or reaches the maximum number of iterations. This ensures the efficiency of model training while fully exploring the model's performance potential, and guarantees the model's generalization ability and prediction stability in complex and ever-changing agricultural data environments.

[0012] As a further technical solution of the present invention, in S5, multiple indicators are comprehensively considered, including prediction error indicators (such as root mean square error, mean absolute error, etc.), uncertainty quantification indicators, and consistency evaluation indicators between prediction results and actual crop physiological responses.

[0013] As a further technical solution of the present invention, in S5, the multi-dimensional prediction result evaluation system introduces a rationality evaluation rule based on crop physiological and ecological knowledge. For example, based on the evapotranspiration variation pattern of different growth stages of crops, the physical response relationship between meteorological conditions and evapotranspiration, and the principle of soil moisture balance, the evapotranspiration time series predicted by the model is logically tested. When the prediction result violates the common sense of crop physiological and ecological knowledge, such as predicting an abnormally high evapotranspiration during the crop dormancy period, the evaluation index will be deducted accordingly and a detailed error source analysis will be triggered. By comparing with historical data and visualizing the output of the intermediate feature layer of the model, the steps or data links that lead to unreasonable prediction results can be accurately located, providing a clear direction and basis for subsequent model feedback optimization, enabling the prediction model to continuously improve in a direction that is more in line with actual agricultural production, and improving the credibility and usability of the prediction results in practical applications.

[0014] A crop evapotranspiration prediction system integrating time series analysis and deep learning is provided to implement a crop evapotranspiration prediction method that integrates time series analysis and deep learning. The system includes the following modules: Data preprocessing module: used to clean and normalize the collected historical meteorological, soil and crop growth data, and at the same time perform feature engineering to extract initial features; Time series feature mining module: Utilizes improved time series analysis algorithms to perform in-depth time feature mining on preprocessed data, and constructs a comprehensive and accurate time series feature matrix; Deep learning model building module: Used to design deep learning models with multi-scale feature fusion architecture. Different branches process features at different time and spatial scales, and the features extracted by each branch are fused in subsequent layers. Model training and optimization module: Adaptive learning rate optimization algorithm combined with Bayesian optimization strategy is used to train and optimize deep learning models; Prediction Result Evaluation and Feedback Module: This module is used to construct a multi-dimensional prediction result evaluation system, identify the main sources of model prediction bias, and feed the evaluation results back to the data preprocessing module, time series feature mining module, and deep learning model building module to guide each module in targeted improvements and optimizations.

[0015] The beneficial effects of this invention are as follows: 1. Improved prediction accuracy: Through improved time series analysis algorithms and deep learning model architecture, the temporal dependencies and spatial correlation features in the data can be more comprehensively mined, resulting in more accurate prediction results compared to traditional methods.

[0016] 2. Enhanced model adaptability: Applicable to evapotranspiration prediction in different regions, crop types and growth stages, it can automatically learn based on the data characteristics of different regions and has wide applicability.

[0017] 3. Improve data utilization: Make full use of multi-source information such as historical meteorological data, soil data, and crop growth data to explore the potential value in the data and provide richer basis for forecasting.

[0018] 4. Intelligent Automation: It realizes full-process automation from data preprocessing to prediction result evaluation, reduces the workload of manual feature engineering, improves prediction efficiency, and facilitates practical application and promotion. Attached Figure Description

[0019] Figure 1 This is a flowchart of a crop evapotranspiration prediction method that integrates time series and deep learning proposed in this invention; Figure 2 This is a block diagram of a crop evapotranspiration prediction system that integrates time series and deep learning proposed in this invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Please see the appendix Figure 1 A method for predicting crop evapotranspiration that integrates time series analysis and deep learning includes the following specific steps: S1: Data preprocessing: Clean and normalize the collected historical meteorological data, soil data and crop growth data, remove outliers, fill missing values, unify the data scale, and perform feature engineering to extract initial features, including but not limited to calculating the moving average of meteorological data, the gradient change of soil data and the growth rate of crop growth data, in order to enhance data quality and extract potentially useful features to prepare for subsequent model processing. The feature engineering process for extracting initial features specifically includes: S11: For the temperature series in meteorological data, the moving average method is used to extract the moving average temperature characteristics. The calculation formula is as follows: ,in: For the first Temperature observations at each moment, The sliding window size (take an odd number, determined based on the fluctuation cycle of temperature data and forecasting needs, generally 3-7 days). For the first The moving average temperature value at each time point; S12: For the soil moisture content sequence in the soil data, calculate the gradient change characteristics of soil moisture content within adjacent soil depth intervals. The gradient is calculated using the three-point difference method, and the calculation formula is as follows: ,in: For the first Soil moisture content at each soil depth The interval between adjacent soil depths (determined based on the actual interval of soil sampling, generally 1000 m / s) ), For the first Soil moisture content gradient at various soil depths; S13: For the plant height sequence in crop growth data, calculate the growth rate characteristics. The growth rate is calculated using the difference method, and the calculation formula is as follows: ,in: For the first Plant height at each observation time, The time interval between adjacent observation times (determined based on the actual crop growth observation frequency, generally 1-2 days). For the first The crop height growth rate at each observation time.

[0022] S2: Time Series Feature Mining: Utilizing improved time series analysis algorithms, such as the wavelet transform-based enhanced time series decomposition method, deep time feature mining is performed on the preprocessed data. This includes extracting long-term trends, seasonal fluctuations, and periodic patterns from the data, capturing subtle changes at different time scales, such as short-term fluctuation features, and constructing a more comprehensive and accurate time series feature matrix. This allows for a more detailed characterization of the time series properties of the data, providing richer time-dimensional feature inputs for deep learning models. The improved time series analysis algorithm introduces a multi-resolution analysis framework and adaptive threshold processing. When performing wavelet decomposition, it automatically adjusts the number of decomposition layers and threshold parameters according to the energy distribution of different frequency components. This allows for more accurate separation of signal features at different time scales, reducing feature loss or noise residue caused by fixed parameter settings in traditional wavelet decomposition. It provides purer and more representative time series feature inputs for deep learning models, which helps improve the learning effect and prediction accuracy of models on complex agricultural time series data. S21: The preprocessed data sequence As input, where For the number of data points, Indicates the first The observation values ​​at each time point; S22: Selecting appropriate wavelet basis functions and scale parameters Translation parameters The wavelet basis functions satisfy the orthogonality and normalization conditions, and their expressions are as follows: Commonly used ones include Daubechies wavelets ( The order of the wavelet is typically 2-10. Higher orders require smoother signals but also increase computational complexity. Other wavelet basis functions and their orders can be selected based on data characteristics and prediction requirements. S23: Perform multi-scale wavelet decomposition on the data sequence. Decomposed into approximate and detailed parts at different scales, for the first Wavelet decomposition of scale, approximation coefficients and detail coefficient The calculation formulas are as follows: in, and The scaling function and wavelet function are respectively the scale function and wavelet function at the th Scale, location The discrete sampled values ​​at each point are obtained by scaling and shifting the wavelet basis functions. For location indexes, the value range is determined based on data length and scale parameters; decompose scale. Generally, the time span of the data and the shortest periodic characteristics to be captured are determined. For example, for daily observation data, if it is desired to capture monthly, quarterly, and annual periodic characteristics, then... Set to 3-5 floors; S24: Utilizing an improved thresholding method for detail coefficients For denoising and feature enhancement, traditional hard and soft thresholding methods suffer from pseudo-Gibbs phenomenon and blurred detail features. An improved method based on Garrote thresholding is adopted, with the calculation formula as follows: in, This is a threshold parameter, generally ranging from 1 to 3. The specific value can be determined through cross-validation based on the noise level and feature strength of the data. This is a sign function used to preserve the sign information of the detail coefficients. This method, while removing noise, can better preserve abrupt changes and detailed features in the signal. S25: Perform wavelet reconstruction, recombine the processed approximation coefficients and detail coefficients to obtain feature sequences at different scales. The reconstruction formula is: in, The reconstructed feature sequence contains temporal feature information at different scales; S26: Extract key features from feature sequences at different scales, construct a time series feature matrix, and perform specific steps for each scale. The feature sequence is analyzed, and its statistical characteristics such as mean, variance, skewness, and kurtosis are calculated, as well as extreme value characteristics such as maximum, minimum, range, and percentile. Simultaneously, waveform features are extracted from the feature sequence, such as rising slope, falling slope, and peak-to-trough interval. These features are then arranged and combined according to scale and feature type to form a dataset of size [size missing]. The feature matrix of , where The number of feature dimensions extracted. This represents the number of samples.

[0023] This feature matrix serves as input to the deep learning model, providing a comprehensive and accurate description of time-series features. This helps the deep learning model learn the temporal patterns in the data more effectively and improves the accuracy of crop evapotranspiration prediction.

[0024] S3: Deep Learning Model Construction: Design a deep learning model with a multi-scale feature fusion architecture, namely a multi-branch LSTM-CNN hybrid neural network structure, including a deep LSTM branch for processing long-term trend features, a shallow LSTM branch for capturing short-term fluctuation features, and a CNN branch for extracting spatial correlation features. Different branches process features at different time and spatial scales, and the features extracted by each branch are fused in subsequent layers, enabling the model to learn the complex spatiotemporal relationships in the data more effectively. Compared with traditional single-structure deep learning models, this architecture can more accurately characterize the complex mechanism of crop evapotranspiration being affected by multiple factors, thus improving prediction accuracy. The deep learning model with a multi-scale feature fusion architecture uses an attention mechanism at each feature fusion point to enhance the feature fusion effect. By calculating the importance weights of features extracted from different branches, the model can automatically focus on feature combinations that are more critical to evapotranspiration prediction and dynamically adjust the contribution of each branch feature to the final prediction result. Compared with a simple feature splicing fusion method, it can more effectively integrate multi-scale feature information and improve the model's feature expression ability and prediction performance when processing complex agricultural scenario data. S31: Long Short-Term Memory Branch (LSTM Branch): S311: Long Short-Term Memory (LSTM) Layer 1: Focuses on extracting long-term trend features; LSTM units control information flow through forget gates, input gates, and output gates, and their calculation formula is as follows: in: These are the outputs of the forget gate, input gate, and output gate, respectively. Candidate cell state; This is the unit state; It is in a hidden state; It is the sigmoid activation function; This indicates element-wise multiplication; These are the weight matrices for the forget gate, input gate, candidate cell state, and output gate, respectively. For the corresponding bias term; For the first Input at each time step; This is the hidden state from the previous moment; LSTM layer 1 is stacked There are LSTM units, and the number of hidden units in each unit is . It is used to capture time dependencies over longer periods, such as annual cycle features; S312: Long Short-Term Memory (LSTM) Layer 2: Focuses on extracting short-term fluctuation features; similar to LSTM Layer 1, but with a different number of hidden units. Furthermore, by adjusting the time step or using a smaller sliding window input, it focuses on capturing fluctuation features on short-term time scales, such as weekly or monthly cycle features; the output of LSTM layer 2 is: ,in, The input data is cropped using a short time window. The cropping method is determined based on the time span of the short-term features. For example, for daily data, the short-term window can be 7-14 days. S32: Convolutional Neural Network Branch (CNN Branch): S321: One-dimensional convolutional layer (Conv1D): used to extract local spatiotemporal features; the kernel size is... Step size is The number of input channels is The number of output channels is The formula for calculating convolution operations is: ,in: For the first Feature maps of each output channel; For the first The first output channel Each convolutional kernel weight; Input data at time step The eigenvector at that location; For the first The bias term for each output channel; For activation functions, such as ReLU; Local spatiotemporal features at different scales are extracted using multiple convolutional kernels, and the output feature map of the convolutional layer is as follows: ,in This represents the time step after convolution; S322: Pooling Layer: Used to reduce feature dimensionality and enhance feature translation invariance; employs max pooling or average pooling operations, with a pooling window size of [size missing]. Step size is The formula for calculating the output of the max pooling layer is: , in: For the pooling process Feature vectors at each time step; The feature vectors within the pooling window; The pooling layer outputs a feature map. ,in This represents the time step after pooling; S323: Fully Connected Layer: Maps the pooled features to a higher-level abstract space, extracting globally correlated features; the number of neurons in the fully connected layer is... The calculation formula is as follows: ,in: The output feature vector of the CNN branch; This is the weight matrix of the fully connected layer; For bias terms; S33: Feature Fusion Layer: This layer fuses the features extracted by the LSTM and CNN branches using feature concatenation, combining the outputs of the LSTM branch... and the output of the CNN branch They are concatenated according to the feature dimensions to form a joint feature vector. : , where [;] represents a concatenation operation on the feature dimension; The feature fusion layer integrates features from different time and spatial scales to provide a comprehensive feature representation for the subsequent decision-making layer. S34: Decision Layer: Predicts based on the fused feature vectors; the decision layer consists of a fully connected layer that uses a linear activation function to output the predicted crop evapotranspiration value. The calculation formula is as follows: ,in: This is the weight matrix of the output layer; This is a bias term.

[0025] S4: Model Training and Optimization: An adaptive learning rate optimization algorithm combined with a Bayesian optimization strategy is used to train and optimize the deep learning model. In the initial stage, a relatively large but reasonable learning rate is used to enable the model to converge quickly to the vicinity of the optimal solution region. Subsequently, Bayesian optimization dynamically adjusts the hyperparameters based on the error feedback during the model training process, including fine adjustment of the learning rate and optimization of regularization parameters, in order to avoid overfitting and further explore the performance potential of the model. Compared with conventional fixed learning rate or single optimization algorithm, it can more efficiently improve the model's generalization ability under complex agricultural environment data and ensure the stable predictive performance of the model on data from different regions, different crop types and different growth stages. In the early stages of model training, a large initial learning rate is used to quickly bring the model parameters closer to the optimal solution region of the loss function. When the slope of the training loss curve falls below a preset threshold, a Bayesian optimization process is triggered. Based on the current hyperparameter settings, a probabilistic model between hyperparameters and model validation error is constructed under a Gaussian process prior distribution. New hyperparameter candidate points are obtained through sampling, and their corresponding expected improvement values ​​are calculated. The hyperparameter combination with the largest expected improvement value is selected to update the model. This process is repeated until the model error on the validation set meets the convergence condition or reaches the maximum number of iterations. This ensures model training efficiency while fully exploring the model's performance potential, ensuring the model's generalization ability and predictive stability in complex and ever-changing agricultural data environments. S41: Initialize model parameters: Randomly initialize the parameters of the deep learning model, such as the weight matrix of the LSTM branch. and bias terms Convolutional kernel weights of CNN branches Bias terms The weight matrix of the fully connected layer Bias terms and the weight matrix of the decision-making layer Bias terms Initialize hyperparameters, including the initial learning rate. (Generally set to 0.01), the exponential decay rate of the first and second moments in the Adam optimization algorithm. (Usually set to 0.9) (Generally set to 0.999), numerical stability parameter (Generally set as) The number of iterations for Bayesian optimization (Determined based on computing resources and time costs, generally set to) ), parameters of the acquisition function (as expected improved) (usually set to 0.01), etc. S42: Adaptive Learning Rate Optimization Training: The model is trained using the Adam optimization algorithm, and the model parameters are updated using the following formula: in: and These are the first-order moment estimate and the second-order moment estimate, respectively; and These represent the first-order moment estimate (i.e., the exponentially weighted average of the gradient) and the second-order moment estimate (i.e., the exponentially weighted average of the squared gradient), respectively, and have been biased corrected. For gradient; For the first The learning rate for each training cycle; These are model parameters; In the initial stage of training, use a larger learning rate. This allows the model to converge quickly to the vicinity of a relatively optimal solution region; S43: Bayesian Optimization Preparation: During model training, record the training loss and validation loss; when the validation loss is continuous... No longer significantly decreased within a training period (i.e., the decrease was below the preset threshold) If the value is typically set to 0.001, the Bayesian optimization process is triggered; the hyperparameter space for Bayesian optimization is constructed, including the learning rate. (The search scope is generally set to) Regularization parameters (The search scope is generally set to) )wait; S44: Bayesian Optimization Loop: For the first Bayesian optimization loop... The next iteration; S441: A surrogate model between hyperparameters and validation loss is constructed using Gaussian process regression. The kernel function of the Gaussian process uses the radial basis function (RBF), and its expression is as follows: ,in: and For hyperparameter vectors; The variance of the signal; For the first The length scale parameter of the hyperparameter; S442: Optimize the acquisition function (Expected Improvement, EI) to select new hyperparameter candidates; the formula for calculating EI is: in: and These are the mean and standard deviation predicted by the surrogate model, respectively. This represents the current minimum verification loss; and These are the cumulative distribution function and probability density function of the standard normal distribution, respectively. S443: Evaluate new hyperparameter candidates, retrain the model using the selected hyperparameters, and record the new validation loss; S444: Update the Gaussian process regression model by adding new hyperparameter candidate points and their corresponding validation loss to the training data; S45: Stopping condition judgment: When the Bayesian optimization reaches the preset number of iterations. Or verify the loss in continuous In the Bayesian optimization iterations, the decrease no longer becomes significant (the decrease is below a preset threshold). When the value is typically set to 0.0001, the optimization process stops.

[0026] S5: Prediction Result Evaluation and Feedback: Construct a multi-dimensional prediction result evaluation system, comprehensively considering prediction error indicators (such as root mean square error, mean absolute error, etc.), uncertainty quantification indicators, and consistency evaluation indicators between prediction results and actual crop physiological responses. By comparing and analyzing the differences between prediction results and measured data, and combining the sensitivity analysis of crop growth models to changes in evapotranspiration, determine the main sources of model prediction bias, such as meteorological data noise, crop growth stage segmentation errors, or defects in feature extraction within the model. Feedback the evaluation results to S1-S3 to guide targeted improvements and optimizations in each step, such as adjusting data preprocessing algorithms, improving time series decomposition strategies, or modifying the deep learning model structure, to achieve continuous iterative upgrades of the prediction model, thereby continuously improving the accuracy, reliability, and practicality of prediction results. This forms a complete, logically coherent crop evapotranspiration prediction system with self-optimization capabilities, effectively solving practical problems in existing crop evapotranspiration prediction methods that integrate time series and deep learning, such as insufficient data processing, inaccurate feature extraction, poor model adaptability, and difficulty in evaluating the credibility of prediction results. In the multi-dimensional prediction result evaluation system, a rationality evaluation rule based on crop physiological and ecological knowledge is introduced. For example, based on the evapotranspiration variation pattern of crops at different growth stages, the physical response relationship between meteorological conditions and evapotranspiration, and the principle of soil moisture balance, the evapotranspiration time series predicted by the model is logically tested. When the prediction result violates the common sense of crop physiological and ecological knowledge, such as predicting abnormally high evapotranspiration during the crop dormancy period, the evaluation index will be deducted accordingly and a detailed error source analysis will be triggered. Through comparison with historical data and visualization of the output of the intermediate feature layer of the model, the steps or data links that lead to unreasonable prediction results can be accurately located, providing a clear direction and basis for subsequent model feedback optimization. This enables the prediction model to continuously improve in a direction that is more in line with the actual agricultural production situation, thereby improving the credibility and usability of the prediction results in practical applications. S51: Calculation of Prediction Error Index: S511: Root Mean Square Error (RMSE): Measures the square root of the mean of the squared errors between predicted and measured values. The formula is: ,in: For the first Measured crop evapotranspiration of a sample For the corresponding predicted value, The number of samples; S512: Mean Absolute Error (MAE): Measures the mean of the absolute error between the predicted and measured values. The formula is as follows: ; S513: Coefficient of Determination ( The formula for measuring the goodness of fit between predicted and measured values ​​is: ,in This is the average of the measured values; S52: Calculation of uncertainty quantification index: S521: Predicted Interval Coverage Probability (PICP): Measures the probability that a predicted interval contains the observed value. The calculation formula is as follows: ,in: As an indicator function, when the measured value Falling within the prediction range The value is 1 if the time condition is met, and 0 otherwise. and These are the lower and upper limits of the prediction interval, respectively. S522: Average Width of Forecast Interval (PINAW): Measures the average width of the forecast interval, calculated using the following formula: ; S53: Indicators for assessing the consistency between predicted results and actual crop physiological responses: S531: Reasonableness Assessment Based on Crop Growth Stages: Establish assessment rules based on the variation patterns of evapotranspiration at different crop growth stages; for example, evapotranspiration is typically lower during the sowing and maturity stages, and higher during the vigorous growth period; calculate a reasonableness score by comparing the predicted results with these patterns. If the prediction result conforms to the pattern, a full score of 1 is awarded; otherwise, points are deducted based on the degree of deviation. S532: Sensitivity Assessment Based on Meteorological Conditions: The reasonableness of the forecast results is assessed based on the physical response relationship between evapotranspiration and meteorological factors (such as temperature, humidity, wind speed, and solar radiation). For example, evapotranspiration typically increases when temperature rises and humidity decreases. A sensitivity score is obtained by calculating the consistency between the forecast results and these response relationships. The calculation formula is: ,in: For the number of meteorological factors, To predict the effect of evapotranspiration on the first The rate of change of each meteorological factor To measure the evapotranspiration rate for the first The rate of change of each meteorological factor is used to measure consistency by calculating the relative error between the two factors; S533: Assessment Based on Soil Moisture Balance: Based on the principle of soil moisture balance, assess the consistency between predicted evapotranspiration and changes in soil moisture content; the calculation formula is: ,in: For the first Changes in soil moisture content at each time step For precipitation, This represents the amount of soil moisture seepage. The predicted evapotranspiration is given. When the predicted evapotranspiration is consistent with the soil moisture balance, a score of 1 is given; otherwise, points are deducted based on the magnitude of the error. S534: Consistency Comprehensive Evaluation Index: The consistency comprehensive evaluation index is obtained by weighted summation of the above consistency evaluation indicators. The calculation formula is: ,in: These are the weighting coefficients, and Adjustments will be made based on the actual assessment priorities. S54: Source Analysis of Model Prediction Bias S541: Compare and analyze the differences between the predicted results and the measured data: Draw time series graphs, scatter plots and error distribution graphs of the predicted values ​​and measured values ​​to visually observe the magnitude and distribution characteristics of the prediction deviation; calculate the statistical indicators of the prediction error at different time scales (such as day, month, quarter) and analyze the changing patterns of the error; S542: Sensitivity Analysis: Using a crop growth model, analyze the sensitivity of evapotranspiration to various input factors (meteorological data, soil data, and crop growth data). By changing the value of each input factor while keeping other factors constant, calculate the rate of change of evapotranspiration to determine the model's sensitivity to each input factor. For example, for temperature in meteorological data, set different temperature variation ranges, calculate the corresponding changes in evapotranspiration, and obtain the temperature sensitivity coefficient. The calculation formula is: ,in: This represents the change in evapotranspiration. This represents the change in temperature; similarly, the sensitivity coefficients to other meteorological factors such as humidity, wind speed, and solar radiation, as well as factors such as soil moisture content, soil texture, and crop growth stage, can be calculated. S543: Sources of Deviation: Based on the results of prediction error indices, uncertainty quantification indices, and consistency assessment indices, combined with sensitivity analysis, determine the main sources of model prediction bias. For example, if RMSE and MAE are large, and sensitivity analysis indicates that the model is highly sensitive to solar radiation in meteorological data, it may indicate problems with the accuracy of solar radiation data or feature extraction. If PICP is low and PINAW is large, it may indicate that the model underestimates the uncertainty of evapotranspiration, requiring optimization of the model's uncertainty quantification part. If the consistency assessment indices... Lower, and A low score may indicate that the model's response to meteorological conditions is not accurately modeled, and the model structure or feature engineering needs to be adjusted. S55: Evaluation Result Feedback and Model Optimization S551: Feedback to S1: If the evaluation results indicate that there are problems with the data quality, such as a lot of noise or outliers, guide step S1 to adjust the data cleaning algorithm, such as using a more suitable data interpolation method to fill in missing values ​​and optimizing the outlier detection and processing strategy. S552: Feedback to S2: When the prediction error index shows that the model is insufficient in capturing certain time scale features, adjust the parameters of the time series decomposition algorithm, such as changing the wavelet basis function, decomposition scale, etc., or optimize the feature extraction method, increase or modify the feature dimension to better characterize the time series characteristics in the data. S553: ​​Feedback to S3: Based on the consistency assessment of the prediction results and the actual crop physiological response, as well as sensitivity analysis, optimize the architecture and hyperparameters of the deep learning model; for example, if the model does not accurately predict the change pattern of evapotranspiration during the crop growth stage, feature inputs related to the crop growth state can be added, and the structure of each branch of the model can be adjusted, such as increasing the number of LSTM units to enhance the ability to capture long-term trends, or optimizing the size of the CNN convolution kernel to better extract spatial correlation features.

[0027] A crop evapotranspiration prediction system integrating time series analysis and deep learning is provided to implement a crop evapotranspiration prediction method that integrates time series analysis and deep learning. The system includes the following modules: Data preprocessing module: used to clean and normalize the collected historical meteorological, soil and crop growth data, and at the same time perform feature engineering to extract initial features; Time series feature mining module: Utilizes improved time series analysis algorithms to perform in-depth time feature mining on preprocessed data, and constructs a comprehensive and accurate time series feature matrix; Deep learning model building module: Used to design deep learning models with multi-scale feature fusion architecture. Different branches process features at different time and spatial scales, and the features extracted by each branch are fused in subsequent layers. Model training and optimization module: Adaptive learning rate optimization algorithm combined with Bayesian optimization strategy is used to train and optimize deep learning models; Prediction Result Evaluation and Feedback Module: This module is used to construct a multi-dimensional prediction result evaluation system, identify the main sources of model prediction bias, and feed the evaluation results back to the data preprocessing module, time series feature mining module, and deep learning model building module to guide each module in targeted improvements and optimizations.

[0028] As can be seen from the above description, the embodiments of the present invention achieve the following technical effects: improved prediction accuracy: through the improved time series analysis algorithm and deep learning model architecture, the temporal dependency and spatial correlation features in the data can be more comprehensively mined, and the prediction results are more accurate compared with traditional methods.

[0029] Enhanced model adaptability: Applicable to evapotranspiration prediction in different regions, crop types and growth stages, it can automatically learn based on the data characteristics of different regions and has wide applicability.

[0030] Improve data utilization: Make full use of multi-source information such as historical meteorological data, soil data, and crop growth data to explore the potential value in the data and provide richer basis for forecasting.

[0031] Intelligent automation: It realizes full-process automation from data preprocessing to prediction result evaluation, reduces the workload of manual feature engineering, improves prediction efficiency, and facilitates practical application and promotion.

[0032] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0033] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this specification. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A crop evapotranspiration prediction method fusing time series and deep learning, characterized in that, The method comprises the following specific steps: S1: data preprocessing: cleaning, normalizing and extracting initial features through feature engineering on collected historical meteorological data, soil data and crop growth data; S2: time series feature mining: using an improved time series analysis algorithm to deeply mine time features from the preprocessed data; S3: deep learning model construction: designing a deep learning model with a multi-scale feature fusion architecture, i.e. a multi-branch LSTM-CNN hybrid neural network structure, different branches processing features of different time scales and spatial scales, and the features extracted by each branch being fused in subsequent levels; S4: model training and optimization: training and optimizing the deep learning model using an adaptive learning rate optimization algorithm combined with a Bayesian optimization strategy; S5: prediction result evaluation and feedback: building a multi-dimensional prediction result evaluation system, considering multiple indicators, comparing the differences between the prediction results and the measured data, analyzing the sensitivity of the crop growth model to evapotranspiration changes, determining the main source of model prediction deviation, and feeding back the evaluation results to S1-S3 to guide targeted improvement and optimization of each step.

2. The crop evapotranspiration prediction method of claim 1, wherein In S2, the deep time feature mining includes extracting long-term trends, seasonal fluctuations and periodic patterns in the data, capturing subtle change rules at different time scales, and constructing a more comprehensive and accurate time series feature matrix.

3. The crop evapotranspiration forecasting method of claim 2, wherein, In S2, the improved time series analysis algorithm automatically adjusts the decomposition layer number and threshold value according to the energy distribution of different frequency components during wavelet decomposition by introducing a multi-resolution analysis framework and adaptive threshold processing, more accurately separating signal features at different time scales.

4. The crop evapotranspiration prediction method of claim 1, wherein, In S3, the multi-branch LSTM-CNN hybrid neural network structure includes a deep LSTM branch that specifically processes long-term trend features, a shallow LSTM branch that captures short-term fluctuation features, and a CNN branch that extracts spatial correlation features.

5. The crop evapotranspiration forecasting method of claim 4, wherein, In S3, the deep learning model with a multi-scale feature fusion architecture uses an attention mechanism at each feature fusion point to enhance the feature fusion effect, calculates the importance weights of the features extracted by different branches, and enables the model to automatically focus on feature combinations that are more critical to evapotranspiration prediction, dynamically adjusting the contribution of each branch feature to the final prediction result.

6. The crop evapotranspiration forecasting method of claim 1, wherein, In S4, a larger but reasonable learning rate is used in the initial stage to quickly converge the model to the vicinity of the optimal solution region, and then the Bayesian optimization dynamically adjusts the hyperparameters according to the error feedback during model training to avoid overfitting and further tap the performance potential of the model.

7. The crop evapotranspiration forecasting method of claim 6, wherein, In S4, the adaptive learning rate optimization algorithm combined with the Bayesian optimization strategy specifically includes: In the early stage of model training, a larger initial learning rate is used to quickly move the model parameters to the vicinity of the optimal solution region of the loss function; When the change slope of the training loss curve is lower than a preset threshold, a Bayesian optimization process is triggered to construct a probability model between the hyperparameters and the model validation error under a Gaussian process prior distribution based on the current hyperparameter settings of the model, to obtain a new hyperparameter candidate point by sampling and calculate its corresponding expected improvement value, and to select the hyperparameter combination with the maximum expected improvement value to update the model. This process is repeated until the error of the model on the validation set meets the convergence condition or reaches the maximum number of iterations. 8.The crop evapotranspiration prediction method of fusing time series and deep learning according to claim 1, wherein, In S5, the comprehensive consideration of multiple indicators includes prediction error indicators, uncertainty quantification indicators, and prediction result consistency evaluation indicators with actual crop physiological responses.

9. The crop evapotranspiration forecasting method of claim 8, wherein, In S5, in the multi-dimensional prediction result evaluation system, reasonable evaluation rules based on crop physiological and ecological knowledge are introduced to perform logical testing on the evapotranspiration time series predicted by the model. When the prediction result violates the common sense of crop physiology and ecology, the steps or data links that lead to unreasonable prediction results are accurately located through comparison with historical data and visualization of the intermediate feature layer output of the model, providing a clear direction and basis for subsequent model feedback optimization.

10. A crop evapotranspiration prediction system fusing time series and deep learning, characterized in that, A crop evapotranspiration prediction method combining time series and deep learning, for implementing any of claims 1-9, comprises the following modules: A data preprocessing module for cleaning and normalizing the collected historical meteorological, soil, and crop growth data, and extracting initial features through feature engineering; A time series feature mining module for performing deep time feature mining on the preprocessed data using an improved time series analysis algorithm to construct a comprehensive and accurate time series feature matrix; A deep learning model construction module for designing a deep learning model with a multi-scale feature fusion architecture, different branches processing features of different time scales and spatial scales, and the features extracted by each branch being fused in subsequent levels; A model training and optimization module for training and optimizing the deep learning model using an adaptive learning rate optimization algorithm combined with a Bayesian optimization strategy; A prediction result evaluation and feedback module for constructing a multi-dimensional prediction result evaluation system to determine the main source of model prediction bias and feeding back the evaluation results to the data preprocessing module, time series feature mining module, and deep learning model construction module to guide the targeted improvement and optimization of each module.

Images