Method, device and medium for predicting transpiration of medicinal crops

Abstract

The present invention discloses a method, device, and medium for predicting transpiration of medicinal crops, including the following specific steps: collecting transpiration-related environmental data, preprocessing the data, and determining key influencing factors of transpiration; constructing a CNN-W model and a CNN-S model based on the preprocessed data, and determining a feature matrix in combination with the key influencing factors of transpiration, wherein the CNN-W model is used to capture the time dependency of meteorological environmental data, and the CNN-S model is used to capture the time dependency of soil environmental data; constructing a global encoder and a local encoder, extracting time series features based on the feature matrix; fusing the extracted time series features, and predicting transpiration based on the fused features. By fusing and outputting the time series features and subsequently extracting features based on time-dependent covariates, the autocorrelation of the time series data is effectively captured, and the dependency between transpiration and environmental influencing factors is explored, thereby improving the accuracy and stability of transpiration prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction, and in particular to a method, device and medium for predicting the transpiration amount of medicinal crops. Background Art

[0002] Artificial intelligence technology has made significant progress in agriculture, encompassing applications such as crop disease and pest detection and crop maturity testing, making modern agriculture more efficient. As a specialty cash crop, medicinal crop cultivation is a part of agriculture and should draw on the experience of intelligent agriculture to address challenges encountered during its cultivation and promote the high-quality development of the Traditional Chinese Medicine (TCM) industry.

[0003] my country has a long history of cultivating traditional Chinese medicine (TCM). However, in recent years, some farmers, driven by the pursuit of yield and economic returns, have blindly applied water and fertilizer to their crops, resulting in poor quality and significant waste of cultivation resources. Transpiration of medicinal crops is fundamental to determining field water requirements during cultivation, formulating irrigation strategies, and achieving precision irrigation. With the trend toward a water-saving society, the active adoption of precision irrigation technologies is crucial for conserving water resources, reducing production costs, and increasing economic returns for farmers.

[0004] Patent "CN106651012A A crop transpiration prediction method based on an improved extreme learning machine" provides a crop transpiration prediction method, which collects soil environmental data and meteorological environmental data of farmland and normalizes them to obtain a training set, trains the extreme learning machine network and improves the extreme learning machine, and finally inputs the normalized data collected again into the improved extreme learning machine to predict crop transpiration.

[0005] The patent "CN114117905A A greenhouse crop irrigation method based on deep neural network" constructs a crop transpiration rate prediction model based on deep neural network, and then inputs the crop planting date, current date, current time, and greenhouse environmental parameter prediction values ​​into the crop transpiration rate prediction model to obtain the crop transpiration rate prediction value, and integrates the time within the interval of the next irrigation time to obtain the crop transpiration amount prediction value.

[0006] Existing technologies for calculating medicinal crop transpiration require the use of multiple sensors to measure environmental factors, such as meteorological and soil factors. Historical transpiration measurements and some environmental factors form a time series with a chronological order, and the current data value is correlated with the previous data value. Existing transpiration prediction methods are computationally complex and fail to fully exploit the long-term dependencies between transpiration and environmental factors, making them unsuitable for accurate long-term prediction of medicinal crop transpiration. Summary of the Invention

[0007] The technical problem to be solved by the present invention is that the existing transpiration prediction method is complex in calculation and cannot fully explore the long-term dependence between transpiration and environmental influencing factors. It is not suitable for long-term accurate prediction of transpiration of medicinal crops. The purpose is to provide a method, device and medium for predicting transpiration of medicinal crops. By screening environmental time series data, extracting global information and local information of time series data, effectively capturing the autocorrelation of time series data and the dependence between time series data, designing a residual network to extract features of the encoder fusion output and subsequent dependent time covariates, and mapping the output features to probability predictions, finally obtaining the predicted value of transpiration of medicinal crops, effectively capturing the autocorrelation of time series data, and exploring the dependence between transpiration and environmental influencing factors, the accuracy and stability of transpiration prediction are improved.

[0008] The present invention is achieved through the following technical solutions:

[0009] The first aspect of the present invention provides a method for predicting the transpiration rate of medicinal crops, comprising the following specific steps:

[0010] Collect environmental data related to transpiration, pre-process the data, and determine the key factors affecting transpiration;

[0011] Based on the preprocessed data, a CNN-W model and a CNN-S model are constructed, and a characteristic matrix is ​​determined by combining key influencing factors of transpiration. The CNN-W model is used to capture the time dependency of meteorological environmental data, and the CNN-S model is used to capture the time dependency of soil environmental data.

[0012] Construct a global encoder and a local encoder to extract temporal features based on the feature matrix;

[0013] The extracted time series features are fused and transpiration is predicted based on the fused features.

[0014] The present invention determines the key influencing factors of transpiration by collecting environmental data related to transpiration, constructs CNN-W model and CNN-S model to determine the feature matrix, constructs a global encoder to extract the global information of time series data, constructs a local encoder to extract the local information of time series data, effectively captures the autocorrelation of time series data and the dependency between time series data, fuses and outputs the time series features and subsequently extracts features of time-dependent covariates, and maps the output features to probability prediction, ultimately obtaining the predicted value of transpiration of medicinal crops. It effectively captures the autocorrelation of time series data and explores the dependency between transpiration and environmental influencing factors, thereby improving the accuracy and stability of transpiration prediction.

[0015] Furthermore, the data preprocessing specifically includes:

[0016] Calculate the mean and variance of environmental data related to transpiration;

[0017] Based on the mean and variance, the Laida criterion is used to screen abnormal data and correct the abnormal data to obtain the corrected environmental data;

[0018] Normalizing the corrected environmental data to obtain normalized environmental data;

[0019] The normalized environmental data were calculated using the Penman formula to obtain the reference crop transpiration;

[0020] Based on the transpiration of reference crops and the crop coefficient corresponding to the planting of medicinal crops, the key influencing factors of transpiration are determined.

[0021] Furthermore, the determining of the characteristic matrix specifically includes:

[0022] Based on the normalized environmental data, the CNN-W model and the CNN-S model are constructed;

[0023] The key factors affecting transpiration were input into the CNN-W model and the CNN-S model as time-dependent covariates, and the temporal correlation of the key factors affecting transpiration was captured through one-dimensional convolution.

[0024] The temporal correlation of the key factors affecting transpiration is combined with the output of the fully connected layer through a convolution operation to obtain a feature matrix.

[0025] Furthermore, the extracting of time series features based on the feature matrix specifically includes:

[0026] The global encoder module is used to extract temporal global information, and the local encoder module is used to extract temporal local information.

[0027] Furthermore, the extracting of temporal global information using the global encoder module specifically includes:

[0028] Constructing a temporal convolutional network, wherein the residual block of the temporal convolutional network includes two layers of multi-scale extended causal convolution;

[0029] Add layer normalization and LeakyReLU activation function to each layer of the residual block of the temporal convolutional network, use 1×1 convolution to adjust the network input tensor, and determine the expansion factor;

[0030] The extended causal convolution is determined based on the size of different expansion factors, and the extended causal convolution operation with different receptive field ranges is used in each layer to determine the comprehensive global feature information and obtain the output feature Y1.

[0031] Furthermore, the extracting of temporal local information by using a local encoder module specifically includes:

[0032] Perform convolution operation on the feature matrix to obtain feature X′, and perform self-calibration operation on feature X′ to obtain calibration feature X′1;

[0033] Map the calibration feature X′1 to the original feature space to obtain the output feature S1;

[0034] Perform a self-calibration operation on the output feature S1 to obtain the output feature S2;

[0035] The output feature S1 and the output feature S2 are connected in series to obtain the connection feature S, and the feature S is convolved to obtain the output feature Y2 containing the local identification feature.

[0036] Furthermore, the fusing of the extracted time series features specifically includes:

[0037] The global encoder output feature Y1 and the local encoder output feature Y2 are connected in series to output the fused feature information;

[0038] The output fused feature information includes local feature information and global feature information of the time series data.

[0039] Furthermore, the transpiration rate prediction specifically includes:

[0040] Construct a Resnet-V network to capture the information of input features, combine it with the encoder fusion output at time t, and obtain the potential output feature δ t+ω

[0041] A fully connected layer is used to transform the potential output feature δ t+ω Mapped to the probability prediction, the transpiration of medicinal crops was obtained.

[0042] A second aspect of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for predicting the transpiration of medicinal crops is implemented.

[0043] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting the transpiration of medicinal crops.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] The present invention uses principal component analysis to screen environmental time series data, designs a multi-scale extended causal convolutional network to extract global information from time series data, improves the self-calibration network to extract local information from time series data, effectively captures the autocorrelation of time series data and the dependency between time series data, designs a residual network to extract features from the encoder fusion output and subsequent dependent time covariates, and maps the output features to probability predictions, ultimately obtaining predicted values ​​for the transpiration of medicinal crops. This effectively captures the autocorrelation of time series data and explores the dependency between transpiration and environmental influencing factors, thereby improving the accuracy and stability of transpiration prediction.

[0046] The solution proposed in this invention uses the principal component analysis method to reduce the number of input parameters of the prediction model and improve the model efficiency; the convolutional network effectively mines the dependency between transpiration and environmental influencing factors, and effectively captures the autocorrelation of time series data, thereby improving the accuracy and stability of long-term transpiration prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings:

[0048] Figure 1 is an overall flow chart of the invention in an embodiment of the present invention;

[0049] Figure 2 It is composed of the CNN-W model in the embodiment of the present invention;

[0050] Figure 3 It is composed of the CNN-S model in the embodiment of the present invention;

[0051] Figure 4 is the convolutional neural network model structure in an embodiment of the present invention;

[0052] Figure 5 It is the temporal convolutional network structure in the embodiment of the present invention;

[0053] Figure 6 It is a multi-scale extended causal convolution structure in an embodiment of the present invention;

[0054] Figure 7 : This is the mSC network structure in the embodiment of the present invention;

[0055] Figure 8 is the mSC in the embodiment of the present invention j (j=1, 2) network structure;

[0056] Figure 9 This is the Resnet-V network structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0058] Example 1

[0059] like Figure 1 As shown, the first aspect of this embodiment provides a method for predicting the transpiration rate of medicinal crops, including the following specific steps:

[0060] Collect environmental data related to transpiration, pre-process the data, and determine the key factors affecting transpiration;

[0061] Based on the preprocessed data, the CNN-W model and CNN-S model were constructed. The characteristic matrix was determined by combining the key influencing factors of transpiration. The CNN-W model was used to capture the time dependency of meteorological environmental data, and the CNN-S model was used to capture the time dependency of soil environmental data.

[0062] Construct a global encoder and a local encoder to extract temporal features based on the feature matrix;

[0063] The extracted time series features are fused and transpiration is predicted based on the fused features.

[0064] This embodiment determines the key influencing factors of transpiration by collecting environmental data related to transpiration, constructs CNN-W model and CNN-S model to determine the feature matrix, constructs a global encoder to extract the global information of time series data, constructs a local encoder to extract the local information of time series data, effectively captures the autocorrelation of time series data and the dependency between time series data, fuses and outputs the time series features and subsequently extracts features of time-dependent covariates, and maps the output features to probability prediction, ultimately obtaining the predicted value of transpiration of medicinal crops. It effectively captures the autocorrelation of time series data and explores the dependency between transpiration and environmental influencing factors, thereby improving the accuracy and stability of transpiration prediction.

[0065] The collection of transpiration-related environmental data specifically includes: using a multi-element louvered box to collect meteorological environmental data for field medicinal crop planting, and using a soil moisture sensor to collect soil environmental data for field medicinal crop planting. The collected environmental data include daily average soil temperature and humidity, soil pH value, rainfall, daily average air temperature and humidity, daily average total solar radiation, wind speed at a height of 2m, and atmospheric pressure.

[0066] Data preprocessing specifically includes:

[0067] Step 1: Screen abnormal environmental data and make corrections. (i) Processing is performed to calculate the mean and variance of environmental data. The formula is:

[0068]

[0069]

[0070] The Laida criterion is used to filter abnormal data. The formula is:

[0071]

[0072] Substitute the mean and variance of the environmental data. The data that meets the above formula is abnormal data. The abnormal data is replaced with the average value of the data on both sides of the abnormal data value for correction. The formula is:

[0073]

[0074] Where i is the category of environmental data, N is the total amount of environmental data in category i, is the nth environmental data in category i, is the average value of environmental data in category i, σ i is the standard deviation of environmental data in category i, ε is the predefined threshold, is the nth corrected environmental data in category i, and the corrected environmental data is recorded as X′ (i) .

[0075] Step 2: Corrected environmental data X′ (i) Normalization is performed to make all data in the same dimension to speed up model training. The formula is:

[0076]

[0077] Among them, i is the category of environmental data, is the nth environmental data in category i, is the minimum value of environmental data in category i, is the maximum value of environmental data in category i, is the nth normalized environmental data in category i, and the normalized environmental data is recorded as X *(i) .

[0078] Step 3: Screening of environmental factors affecting transpiration. Using normalized environmental data X according to the Penman formula *(i) Calculate the reference crop transpiration using the formula:

[0079]

[0080] Where ET0 is the reference crop transpiration; Δ is the slope of the saturated water vapor pressure temperature curve; R n is the average net radiation on the surface of medicinal crops; H is the soil heat flux; T hr is the average air temperature; u2 is the average wind speed at 2 meters; γ is the hygrometer constant; e s is the saturated water vapor pressure; e a is the actual water vapor pressure.

[0081] Multiply the reference crop transpiration ET0 by the crop coefficient K corresponding to the medicinal crop planted c , the transpiration rate ET of medicinal crops can be obtained c , the formula is:

[0082] ET c =ET0×K c

[0083] Environmental factors influencing transpiration were screened. The 10 most common environmental factors affecting transpiration are daily average air temperature and humidity, wind speed at 2m height, sunshine hours, rainfall, atmospheric pressure, light intensity, daily average soil temperature and humidity, and daily average net solar radiation. SPSS software was used for principal component analysis to decouple environmental factors affecting transpiration, reduce the dimensionality of environmental data, and reduce the amount of model calculations. Calculation of environmental factors and transpiration ET of medicinal crops. c The correlation table is shown in Table 1:

[0084] Table 1 Environmental factors and transpiration ET of medicinal crops c Correlation

[0085]

[0086] Among them, ET c is the transpiration of medicinal crops, X *(1) is the daily average air temperature, X *(2) is the average daily humidity of the air, X *(3) is the wind speed at a height of 2m, X *(4) is the sunshine hours, X *(5) is the rainfall, X *(6) is the atmospheric pressure, X *(7) is the light intensity, X *(8) is the daily average soil temperature, X *(9) is the daily average soil moisture, X *(10) is the daily average net solar radiation.

[0087] Environmental factors with a correlation greater than 0.5 with the transpiration of medicinal crops were selected, including: air temperature, air humidity, wind speed, light intensity, soil moisture, and net solar radiation. The correlations with the transpiration of medicinal crops were 0.62, -0.61, 0.64, 0.65, -0.60, and 0.59, respectively. Among them, air temperature, wind speed, light intensity, and net solar radiation were positively correlated with the transpiration of medicinal crops, while air humidity and soil moisture were negatively correlated with transpiration. Therefore, the above environmental factors were screened as key influencing factors of the transpiration of medicinal crops.

[0088] Building a prediction model specifically includes:

[0089] Step 1: Extract environmental data features. Based on the filtered normalized environmental data X *(i) , a CNN-W model is constructed to capture the time dependency of meteorological environment data, and a CNN-S model is constructed for the time dependency of soil environment data. Among them, meteorological environment influencing factors include air temperature, air humidity, wind speed, light intensity, and net solar radiation, and soil environment influencing factors include soil humidity;

[0090] The CNN-W model consists of a convolutional layer and an average pooling layer, and its structure is as follows: Figure 2 As shown;

[0091] The CNN-S model consists of a convolutional layer and an average pooling layer, and its structure is as follows: Figure 3 shown.

[0092] The key factors affecting transpiration are input into the feature extraction network as time-dependent covariates. The CNN-W model and CNN-S model capture the temporal correlation of the factors through one-dimensional convolution and use the fully connected layer to combine the high-level features of the factors. The output of the fully connected layer is combined with the transpiration of medicinal crops y t The next step input X of the prediction model is obtained through convolution operation i , the expression is:

[0093] X i =K0(D w,s (W,S)+y t )

[0094] Among them, X i is the feature matrix of {I×N}, i∈(1,I) is the dimension of time series data, I=7 is consistent with the total number of time series categories of model input, N is the length of time series data in category i, K0 is the convolution operation, D w,s is the CNN-W and CNN-S feature extraction models, W is the meteorological environment data, S is the soil environment data, y t is the transpiration of medicinal crops, y t With ET c Indicates that the data content is the same.

[0095] Step 2: Convolutional neural network model construction. Time series modeling can better capture the potential characteristics of time series data. Global and local encoders are designed to extract the characteristic information of transpiration time series data. The model structure is as follows: Figure 4 The specific steps are as follows:

[0096] a. Model input feature X i They are used for the global encoder module to extract global temporal information and the local encoder module to extract local temporal information.

[0097] b. Design a global encoder to extract global information from time series data. The specific steps are as follows:

[0098] 1) The global encoder is composed of a stack of multi-scale temporal convolutional networks (mTCNs) to capture the dependencies between key factors affecting transpiration and historical transpiration measurements;

[0099] 2) The residual block of the temporal convolutional network includes two layers of multi-scale extended causal convolution, and adds layer normalization and LeakyReLU activation function in each layer to better select features. The network uses 1×1 convolution to adjust the shape of the input tensor to ensure that it matches the output tensor shape. The temporal convolutional network structure is as follows Figure 5 As shown:

[0100] 3) Design extended convolutions based on different expansion factor sizes. The multi-scale extended causal convolution sets the maximum expansion factors of each layer of the network from bottom to top to d = 1, 2, 4, and 8 respectively. Each layer of the network needs to calculate all extended convolutions that are less than the maximum expansion factor of this layer to prevent the low-scale feature information in the high-level network from being discarded. The network adds a global average pooling layer to reduce the dimension of the feature map and further extract feature information. Using convolution operations with different receptive field ranges at each layer can obtain rich and comprehensive global feature information, and can also reveal the overall trend of transpiration time series data. The multi-scale extended causal convolution structure is as follows: Figure 6 As shown:

[0101] 4) The global feature information of the time series data is extracted through the global encoder to obtain the output feature Y1, which is used as the feature input of the subsequent decoder. The multi-scale extended causal convolution can extract feature information of different scales, improve the network's feature information learning ability, and improve the prediction accuracy and performance of the prediction model.

[0102] c. Design a local encoder to extract local information of time series data. The specific steps are as follows:

[0103] 1) Use the modified Self-Calibrated (mSC) module to extract and enhance local information, capture the autocorrelation of time series data, and improve the network layer and convolution operations to extract richer local features of transpiration-related time series data. The mSC network structure is as follows: Figure 7 As shown:

[0104] 2) Design a self-calibration block to form a residual self-calibration convolution instead of the traditional self-calibration convolution structure, so that the model can adaptively learn local feature information. j (j=1, 2) The network structure is as follows Figure 8 As shown:

[0105] 3) Input feature X i Through convolution operation The feature X′ is obtained. The potential space after the convolution operation can guide the original feature transformation to generate a more expressive feature representation. The convolution operation The formula is:

[0106]

[0107] Among them, X′ is the convolution output feature; LeakyReLU is the activation function, which speeds up the network calculation speed through linear operations and avoids the fixed feature problem caused by the original self-calibration convolution using the ReLU activation function.

[0108] 4) In the self-calibration operation, each spatial position adaptively considers the surrounding information as an embedding in the latent space, which serves as a guide to the original space. In addition, the self-calibration operation only considers the appropriate context range of each spatial position to avoid information contamination from irrelevant areas. The formula for the self-calibration operation on feature X′ is:

[0109]

[0110] Among them, X′1 is the calibration feature, and the deconvolution operation Indicates feature transformation of feature X′; convolution operation Used to form attention weights for calibration; sigmoid is the activation function, · is the matrix dot product, and * is the convolution operation.

[0111] 5) Map the calibrated feature X′1 to the original feature space to obtain the output feature S1, the formula is:

[0112]

[0113] 6) Self-calibration convolution extracts more local discriminative features of time series data by expanding the field of view within the convolution layer. The residual network is designed to capture multi-level features and can extract rich hierarchical complementary information. The formula for the self-calibration operation of feature S1 is:

[0114]

[0115]

[0116]

[0117] in, represents the convolution operation, represents the deconvolution operation, X′2 is the calibration feature, S2 represents the output feature; the input feature X″ is converted into The original spatial features are obtained to preserve the original spatial context.

[0118] 7) Features S1 and S2 are concatenated to obtain the connected feature S, which collects different types of context information. The convolution operation K0 is performed on the feature S to obtain the output feature Y2 containing the local discriminant feature, which is used as the feature input of the subsequent decoder. The formula is:

[0119] S=concat(S1,S2)

[0120] Y2=S*K0

[0121] d. Design a decoder to map the output features to probability predictions. The specific steps are as follows:

[0122] 1) The global encoder output feature Y1 and the local encoder output feature Y2 are connected in series to obtain the encoder fusion output, which contains the local feature information and global feature information of the time series data;

[0123] 2) The decoder input consists of the time-dependent covariates after time t The Resnet-V network is designed to capture the information of input features by fusion with the encoder output. The Resnet-V network uses the nonlinear residual function R(·) to calculate the covariate For processing, R(·) consists of two fully connected layers, LeakyReLU activation function, and layer normalization, and ω is the time interval. The network uses 1×1 convolution to adjust the shape of the encoder fusion output tensor to ensure that it matches the shape of the residual function R(·) output tensor. The Resnet-V network structure is as follows Figure 9 shown.

[0124] The latent output features δ of the Resnet-V network t+ω Expressed as:

[0125]

[0126] in, is the encoder fusion output at time t, δ t+ω is the potential output feature of the decoder at time t+ω.

[0127] 3) Use the fully connected layer to transform the potential output feature δ t+ω Mapped to the probability prediction to obtain the final transpiration prediction value.

[0128] Step 3: Convolutional neural network model training. Given a Gaussian distribution, use maximum likelihood estimation to obtain the network parameter output (μ, σ), and construct a negative log-likelihood function as the loss function to optimize the model. The expression is as follows:

[0129]

[0130] Among them, l is the likelihood function, μ is the expectation, σ is the standard deviation, and y is the target value of transpiration.

[0131] (4) Prediction of transpiration of medicinal crops

[0132] Given that the distribution is Gaussian, the probability prediction can be described as:

[0133] P(y t+ω )~G(μ t+ω ,σ t+ω )

[0134] Where P is the probability distribution of transpiration, y t+ω is the transpiration at time t+ω, μ t+ω y t+ω The expectation of t+ω is the standard deviation.

[0135] A second aspect of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for predicting the transpiration of medicinal crops is implemented.

[0136] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for predicting the transpiration of medicinal crops.

[0137] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting transpiration of medicinal crops, characterized in that: The specific steps include: Collect environmental data related to transpiration, pre-process the data, and determine the key factors affecting transpiration; Based on the preprocessed data, a CNN-W model and a CNN-S model are constructed, and a characteristic matrix is ​​determined by combining key influencing factors of transpiration. The CNN-W model is used to capture the time dependency of meteorological environmental data, and the CNN-S model is used to capture the time dependency of soil environmental data. The determining feature matrix specifically includes: Based on the normalized environmental data, the CNN-W model and the CNN-S model are constructed; The key factors affecting transpiration were input into the CNN-W model and the CNN-S model as time-dependent covariates, and the temporal correlation of the key factors affecting transpiration was captured through one-dimensional convolution. The time correlation of the key factors affecting transpiration is combined with the output of the fully connected layer through convolution operation to obtain the feature matrix; Construct a global encoder and a local encoder to extract temporal features based on the feature matrix. Specifically, the global encoder module is used to extract temporal global information, and the local encoder module is used to extract temporal local information. The use of the global encoder module to extract the temporal global information specifically includes: Constructing a temporal convolutional network, wherein the residual block of the temporal convolutional network includes two layers of multi-scale extended causal convolution; Add layer normalization and LeakyReLU activation function to each layer of the residual block of the temporal convolutional network, use 1×1 convolution to adjust the network input tensor, and determine the expansion factor; The extended causal convolution is determined based on the different expansion factor sizes, and the extended causal convolution operation with different receptive field ranges is used in each layer to determine the comprehensive global feature information and obtain the output feature ; The extracting of temporal local information by using the local encoder module specifically includes: Perform convolution operation on the feature matrix to obtain the feature , for features Perform self-calibration operation to obtain calibration characteristics ; The calibration feature Mapping to the original feature space to obtain output features ; Output features Perform self-calibration operation to obtain output characteristics ; The output features and output features Perform concatenation operations to obtain connection features , for features Perform convolution operation to obtain output features containing local discriminant features ; The extracted time series features are fused and transpiration is predicted based on the fused features.

2. The method for predicting transpiration of medicinal crops according to claim 1, wherein: The data preprocessing specifically includes: Calculate the mean and variance of environmental data related to transpiration; Based on the mean and variance, the Laida criterion is used to screen abnormal data and correct the abnormal data to obtain the corrected environmental data; Normalizing the corrected environmental data to obtain normalized environmental data; The normalized environmental data were calculated using the Penman formula to obtain the reference crop transpiration; Based on the transpiration of reference crops and the crop coefficient corresponding to the planting of medicinal crops, the key influencing factors of transpiration are determined.

3. The method for predicting transpiration of medicinal crops according to claim 1, wherein: The fusing of the extracted time series features specifically includes: Output features to the global encoder And the local encoder output features Perform a series operation and output the fused feature information; The output fused feature information includes local feature information and global feature information of the time series data.

4. The method for predicting transpiration of medicinal crops according to claim 1, wherein: The transpiration rate prediction specifically includes: Construct Resnet-V network to capture input feature information, combined with The moment encoder fuses the output to obtain the potential output features ; Use the fully connected layer to transform the potential output features Mapped to the probability prediction, the transpiration of medicinal crops was obtained.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for predicting the transpiration of medicinal crops according to any one of claims 1 to 4 is implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for predicting the transpiration of medicinal crops according to any one of claims 1 to 4 is implemented.

Images