A method for predicting survival rate of pot seedling transplanting culture

Abstract

The present invention proposes a method for predicting the survival rate of pot seedling transplanting and cultivation, and relates to the field of machine learning. The pot seedling transplanting and cultivation survival rate prediction process proposed by the present invention includes constructing a pot seedling transplanting and cultivation data set, decomposing the time series into trend, seasonality and residual components through STL sequence decomposition, so as to extract long-term trend, periodic fluctuation and random disturbance features respectively, using sine and cosine functions to generate position coding to add relative order information to the components, and then constructing multiple interactive learning modules to deeply extract the unique features of each component, and fuse the common features of the three components through the information sharing module to enhance feature interaction, while introducing residual connections to alleviate the gradient dissipation problem, and finally, inputting the comprehensive features of the three components into the fully connected layer to generate the prediction results of the survival rate of pot seedling transplanting and cultivation, and the method can effectively improve the accuracy of the prediction of the survival rate of pot seedling transplanting and cultivation.

Description

Technical Field

[0001] The invention belongs to the technical field of machine learning, and in particular relates to a method for predicting the survival rate of pot seedling transplanting culture. Background Art

[0002] Pot seedling transplanting is a cultivation method in which seedlings are transplanted from seedling pots to the field or production environment. The survival rate of pot seedling transplanting directly affects the growth and development of crops and the final yield. A low survival rate may lead to an extended production cycle, waste of resources and economic losses. Therefore, predicting the survival rate of pot seedling transplanting is of great significance for optimizing planting management, improving resource utilization and ensuring stable and high yields. It can help farmers and related personnel take intervention measures in advance to reduce risks that may arise during the cultivation process.

[0003] The main methods for predicting the survival rate of pot seedling transplanting culture usually include empirical judgment and statistical analysis. The empirical judgment method relies on the experience of farmers or experts and is subjective and limited. The statistical analysis method makes predictions by analyzing historical data or a single variable and is difficult to deal with the interaction of multiple complex factors. These traditional methods can usually only analyze simpler features and lack the ability to deeply mine multidimensional data, making it difficult to comprehensively and accurately predict the survival rate of pot seedling transplanting culture.

[0004] The survival rate of seedling transplanting in pots is affected by many factors, including environmental conditions, plant characteristics and fertilizer application amount. There are often complex nonlinear relationships between these factors, making it difficult for traditional methods to accurately predict their survival rates. Machine learning can effectively process multi-dimensional and diversified data, automatically discover complex associations hidden in the data, and identify features that have a significant impact on survival rates through feature extraction algorithms, providing farmers and managers with timely decision-making support and helping them quickly adjust their planting strategies based on prediction results. Summary of the invention

[0005] The present invention provides a method for predicting the survival rate of seedling transplanting culture in pots, which aims to decompose a time series into a trend component, a seasonal component and a residual component through STL sequence decomposition, and then each component is processed by multiple interactive learning modules, in which the features of each component are extracted separately, and interactive learning with the features of other components is realized, so as to generate the final features of each component, and finally, the final features of each component are spliced ​​and integrated and input into the fully connected layer to accurately predict the survival rate of seedling transplanting culture in pots.

[0006] The technical method adopted by the present invention to achieve the above-mentioned purpose specifically comprises the following steps:

[0007] S1. Collect pot seedling transplanting and cultivation data, including daily average temperature, day and night temperature difference, light intensity, wind speed, soil moisture, soil pH value, seedling height, seedling leaf color, and transplanting time, and pre-process the collected data to ensure data quality and consistency;

[0008] S2. Using the STL sequence decomposition method, the time series is decomposed into trend component, seasonal component and residual component through local weighted regression. The trend component represents the long-term trend, the seasonal component represents the periodic fluctuation, and the residual component represents the random disturbance or abnormal fluctuation.

[0009] S3, using sine and cosine functions to alternately generate codes at different positions, so that the model can recognize the relative order of time steps and concatenate them with the three components respectively;

[0010] S4, constructing interactive learning modules, inputting each component into multiple interactive learning modules respectively, extracting component features in each interactive learning module, and ensuring information flow between components through feature interaction, specifically including the following steps:

[0011] S41. Construct a component feature extraction module, input the three components of trend, seasonality and residual into two fully connected layers respectively, perform nonlinear mapping on each component through ReLU activation function, and extract its feature representation to provide rich feature information for subsequent model learning;

[0012] S42, constructing an information sharing module, first extracting the shared features of the three components to capture the common information between the components, and then combining the shared features with the features of the three components respectively, so that each component not only contains its own features, but also pays attention to the features of other components, thereby enhancing feature interaction and fusion;

[0013] S43, introduce residual connection, by adding the final generated component features to their original input features to form a residual connection, so as to alleviate the gradient dissipation problem of deep networks caused by stacking multiple interactive learning modules, thereby improving the network's expressiveness and model performance;

[0014] S5. The final features of the three components of trend, season and residual are concatenated, and the fully connected layer is used to further calculate the prediction results of the survival rate of pot seedling transplanting culture.

[0015] Preferably, in step S1, the pot seedling transplanting and cultivation data are collected, including the average daily temperature, the temperature difference between day and night, the light intensity, the wind speed, the soil moisture, the soil pH value, the seedling height, the seedling leaf color, and the transplanting time, and it is checked whether there are missing values ​​in the data. If there are missing values, they are filled by interpolation to maintain the time series and trend of the data.

[0016] Preferably, in step S2, the time series data is input Where T is the time window size, n is the number of data types, is the time series data of the ith data in T consecutive time steps, is the monitoring value of the i-th data at the t-th moment. For each time series of data, the STL sequence decomposition algorithm is used to decompose it into trend component, seasonal component and residual component. The specific calculation process is as follows:

[0017] E T，i , S T，i , R T，i =STL(x T，i );

[0018] E T，i =[e 1，i , e 2，i ,...,e T，i ];

[0019] S T，i =[s 1，i ,s 2，i ,...,s T，i ];

[0020] R T，i =[r 1，i , r 2，i , ..., r T，i ];

[0021] In the formula are the trend component, seasonal component and residual component obtained by decomposing the i-th data through STL sequence, are the trend value, seasonal value and residual value obtained by STL sequence decomposition of the ith data at time t, respectively. STL(·) is the STL sequence decomposition algorithm.

[0022] Preferably, STL sequence decomposition decomposes time series data into trend, seasonal and residual components, where the trend component can reveal long-term change trends, the seasonal component reflects periodic fluctuations, and the residual component represents random disturbances or anomalies. STL sequence decomposition not only helps the model to achieve targeted learning on different components, but also improves the accuracy and stability of predictions, while enhancing the interpretability of the model.

[0023] Preferably, in step S3, sine and cosine functions are used to generate a periodic position encoding vector for the time series so as to capture the position features of different time steps. The specific calculation process is as follows:

[0024]

[0025] In the formula is the encoding value of the position encoding vector in even dimensions, is the encoding value of the position encoding vector in odd dimensions, t∈[1, T] is the time step, d is the total dimension of the position encoding vector, k∈[0, d / 2] is the index in the position encoding vector dimension, sin(·) is the sine function, cos(·) is the cosine function, and the position encoding vector of the tth time step can be obtained by concatenating all dimensions of the position encoding vector of the tth time step. Then it is spliced ​​into the trend component, seasonal component and residual component of time step t, so as to introduce location information for each component. The specific calculation process is as follows:

[0026]

[0027] Where PE=[PE 1 , PE 2 , ..., PE T ] is the total position encoding vector of T consecutive time steps, are the concatenation results of position coding, trend component, seasonal component and residual component, respectively, where h = T + d, is the concatenation operation in the time step dimension.

[0028] Preferably, the position information of the time step is introduced into the trend, seasonality and residual components through position encoding, which enhances the model's perception of the time step sequence. The alternation of sine and cosine functions to generate codes at different positions can preserve the sequential relationship of the time series, enabling the model to better understand the laws of periodic changes, thereby improving the accuracy of modeling time series. At the same time, the periodic characteristics of position encoding are also helpful in processing sequence data with long-term dependencies.

[0029] Preferably, in step S4, N interactive learning modules are constructed, firstly, the features of the input trend component, seasonal component and residual component are deeply extracted, and then the information flow between the components is ensured through the feature interaction mechanism, so as to capture the relationship between the components.

[0030] Preferably, in the steps S4 and S41, a component feature extraction module is constructed. First, the three component features of trend, seasonality and residual are respectively input into the fully connected layer to extract preliminary features, and activated using the ReLU activation function. The specific calculation process is as follows:

[0031]

[0032] In the formula The features of the three components of trend, seasonality and residual are obtained by fully connected layers and ReLU function activation. is the trainable weight matrix, is a trainable bias parameter, ReLU(·) is a ReLU activation function. In order to enhance the generalization ability of the model, the Dropout regularization technique is applied to randomly shut down some neurons during the training process, and then the regularized component features are passed through the fully connected layer and the ReLU activation function again to further extract deep features. The specific calculation process is as follows:

[0033]

[0034] In the formula The three component features of trend, seasonality and residual are obtained by passing the fully connected layer and ReLU function again to obtain the feature vectors. is the trainable weight matrix, is a trainable bias parameter, Dropout(·) is the Dropout regularization operation, and ReLU(·) is the ReLU activation function.

[0035] Preferably, the fully connected layer and ReLU activation function are used to extract the deep features of the three components, so that the model can obtain a deep understanding of the input data at different levels. At the same time, the Dropout technology is applied to enhance the generalization ability of the model and ensure the robustness of the model in complex tasks.

[0036] Preferably, in the steps S4 and S42, an information sharing module is constructed to extract shared features of the three components. The shared features are used to capture the common information of the three components to enhance the model's learning ability for global features. The specific calculation process is as follows:

[0037]

[0038] In the formula is the shared eigenvector of the three components, The shared eigenvectors generated are The eigenvectors of the trend, seasonality, and residual components are added to each other to enhance the information flow between the components. The specific calculation process is as follows:

[0039]

[0040] In the formula The feature vectors are obtained by adding the trend, seasonality and residual components to the shared features.

[0041] Preferably, shared features are extracted through the information sharing module to achieve feature sharing among trend, seasonality and residual components. The shared features reflect the commonalities of the three components and help capture global patterns in time series. Combining the shared features with each component can enhance the generalization ability of the model while retaining the unique characteristics of each component, thereby improving the accuracy of the prediction.

[0042] Preferably, residual connection is introduced in steps S4 and S43, and the three component features are added to their original input component features respectively to obtain the final feature vectors of the three components. The specific calculation process is as follows:

[0043]

[0044] In the formula The three component features of trend, seasonality and residual of the i-th data are connected by residual to obtain the feature vector. are the initial trend, seasonality and residual component features input to the interactive learning module respectively. In the next interactive learning module, It will be used as input feature, namely E′ T,i , S′ T,i , R′ T，i The updated value of .

[0045] Preferably, superimposing multiple layers of interactive learning modules enhances the feature expression ability of the components. By introducing residual connections, not only the feature information of the original input is retained, but also the gradient vanishing problem caused by the multiple layers of interactive learning modules can be effectively alleviated, thereby maintaining the integrity of the features in the deep structure of the model.

[0046] Preferably, in step S5, the trend, seasonality and residual component features of the same data are first concatenated to form a complete feature representation of the data, and then the complete feature vectors of all data are added to generate a comprehensive feature vector of all data features. Finally, the comprehensive feature vector is input into the fully connected layer to calculate the survival rate prediction value of pot seedling transplanting culture. The specific calculation process is as follows:

[0047]

[0048] P = Sigmoid (WF + b);

[0049] In the formula is the comprehensive feature vector of all data features, For splicing operation, is the predicted value of the final survival rate of pot seedling transplanting culture, Sigmoid(·) is the sigmoid function, is the trainable weight matrix, is a trainable bias parameter.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention subdivides the time series into trend, seasonal and residual components through STL sequence decomposition, ensuring that the model can capture the long-term trend, periodic fluctuations and random disturbances of the data, and uses position encoding to enhance the model's sequential perception of time steps. Subsequently, the feature representation of each component is extracted in the interactive learning module, and shared features are introduced to realize the information flow between multiple components, capturing the commonalities and mutual relationships of each component, and alleviating the gradient dissipation problem through residual connections. Finally, by splicing the features of the three components and inputting them into the fully connected layer, an accurate prediction of the survival rate of pot seedling transplanting culture is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a step-by-step diagram of a method for predicting the survival rate of pot seedling transplanting culture.

[0052] Figure 2 This is the interactive learning module diagram.

[0053] Figure 3 This is the component feature extraction module diagram.

[0054] Figure 4 This is the information sharing module diagram.

[0055] Figure 5 This is a diagram showing the prediction effect of survival rate of pot seedling transplanting culture. DETAILED DESCRIPTION

[0056] The present invention proposes a method for predicting the survival rate of seedling transplanting and cultivation in pots. The method divides the time series into trend, seasonal and residual components through STL sequence decomposition, so that the model can separate the long-term trend, periodic fluctuation and random disturbance in the data. Then, the position coding generated by sine and cosine functions is adopted, and the coding is embedded into each component to enhance the time sequence perception ability of the model. Subsequently, the feature representation of each component is extracted in multiple interactive learning modules, and the information flow and commonality capture between multiple components are realized through shared feature interaction. At the same time, residual connection is introduced to alleviate the problem of gradient dissipation. Finally, the features of the three components are spliced ​​and input into the fully connected layer to realize the survival rate prediction of seedling transplanting and cultivation in pots. The technical scheme in the embodiment of the present invention is described in detail and completely below, which specifically includes the following steps: Figure 1 shown.

[0057] S1. Collect 62 days of pot seedling transplanting and cultivation data, including daily average temperature, day and night temperature difference, light intensity, wind speed, soil moisture, soil pH value, seedling height, seedling leaf color and transplanting time. After the data collection is completed, the missing values ​​are filled using the interpolation method to ensure the integrity of the time series data. Then the processed data is divided into training set and test set in a ratio of 8:2. For the input data of the model, the sliding window method is used to construct the feature sequence, and the window size is set to 7 days, so that the model can learn and capture the dynamic changes of the data in the time dimension, and provide support for subsequent predictions.

[0058] S2. Use the STL sequence decomposition method to decompose time series data into three components: trend, seasonality, and residual. The trend component represents the long-term trend of the data, the seasonal component reveals the periodic fluctuations of the data, and the residual component is used to represent random disturbances or abnormal fluctuations. Through STL sequence decomposition, the different component characteristics in the data can be effectively separated, providing a clearer feature basis for subsequent modeling and analysis.

[0059] Furthermore, in step S2, the time series data is input Where 7 is the time window size, n=9 is the number of data types, is the time series data of the i-th data in 7 consecutive time steps, is the monitoring value of the i-th data at the t-th moment. The STL sequence decomposition algorithm is used to decompose the time series of each data into trend component, seasonal component and residual component. The specific calculation process is as follows:

[0060] E 7，i , S 7，i , R 7，i =STL(x 7，i );

[0061] E 7，i =[e 1，i , e 2，i , ..., e 7，i ];

[0062] S 7，i =[s 1，i , S 2，i , ..., S 7，i ];

[0063] R 7，i =[r 1，i , r 2,i ,...,r 7，i ];

[0064] In the formula are the trend component, seasonal component and residual component obtained by STL sequence decomposition of the ith data, respectively. STL(·) is the STL sequence decomposition algorithm. They are the trend value, seasonal value and residual value obtained by STL sequence decomposition of the i-th data at the tth moment.

[0065] S3. Sine and cosine functions are used alternately to generate position codes, providing unique coding values ​​for each time step, so that the model can identify the relative order of time steps. These position codes are then concatenated with the trend component, seasonal component and residual component respectively, so that the model can consider the time position information while processing each component.

[0066] Furthermore, in step S3, a periodic position code is generated for each time step to capture the relative order characteristics of the time step, wherein the even dimension and odd dimension of the position code vector are calculated by sine and cosine functions respectively, and the specific calculation process is as follows:

[0067]

[0068] In the formula are the encoding values ​​of the position encoding vector in even and odd dimensions respectively, t∈[1,7] is the time step, d=10 is the total dimension of the position encoding vector, k∈[0,5] is the dimension index of the position encoding vector, sin(·) is the sine function, cos(·) is the cosine function, and the position encoding vector is obtained by concatenating all the position encoding values ​​of the tth time step in the order of the dimensions. Next, the position encoding vector PE t Spliced ​​to the tth time step of the trend component, seasonal component and residual component to introduce location information into each component. The specific calculation process is as follows:

[0069]

[0070]

[0071] Where PE=[PE 1 , PE 2 , ..., PE 7 ] is the total position encoding vector of 7 consecutive time steps, are the concatenation results of the position code and the trend component, seasonal component and residual component respectively, where the dimension 17 is the sum of the time window size and the position code vector dimension. is the concatenation operation in the time step dimension.

[0072] S4. Construct 5 interactive learning modules to deeply extract the features of the input trend component, seasonal component and residual component, and establish information flow channels between the components through the feature interaction mechanism. Specifically, the following steps are included: S41. Construct a component feature extraction module, and input the input trend, seasonality and residual components into the fully connected layer and ReLU activation function respectively to achieve preliminary feature extraction, and then apply the Dropout regularization technology to randomly close some neurons during the training process. Finally, extract its feature representation again through the fully connected layer and ReLU activation function to provide rich feature information for subsequent model learning.

[0073] Furthermore, in step S4, the interactive learning module is as follows: Figure 2 As shown, a component feature extraction module is constructed in S41, such as Figure 3 As shown, the component feature E′ 7，i , S′ 7，i , R′ 7，i Input the fully connected layer respectively and activate it through the ReLU activation function to extract preliminary features. The specific calculation process is as follows:

[0074]

[0075] In the formula are the initial feature vectors obtained after the features of the three components of trend, seasonality and residual are activated by the fully connected layer and the ReLU function. ReLU(·) is the ReLU activation function. is the trainable weight matrix, is a trainable bias parameter, and then the Dropout regularization technique is applied to randomly close some neurons during the training process to enhance the generalization ability of the model. The regularized component features are passed through the fully connected layer and ReLU activation function again to further extract deep features. The specific calculation process is as follows:

[0076]

[0077] In the formula The three component features of trend, seasonality and residual are obtained by passing the fully connected layer and ReLU function again to obtain the feature vectors. is the trainable weight matrix, is a trainable bias parameter, Dropout(·) is the Dropout regularization operation, and ReLU(·) is the ReLU activation function.

[0078] S42. Construct an information sharing module. First, generate shared features of the three components through the Hadamard product. Then, combine the shared features with the three component features respectively, so that each component not only contains its own features, but also pays attention to the features of other components.

[0079] Furthermore, in step S4, the interactive learning module is as follows: Figure 2 As shown, in S42, an information sharing module is constructed, such as Figure 4 As shown, the shared eigenvectors of the three components are first calculated by the Hadamard product to capture the common information of the three components. The specific calculation process is as follows:

[0080]

[0081] In the formula is the shared eigenvector of the three components, is the Hadamard product operation, that is, the corresponding elements are multiplied bit by bit, and the generated shared eigenvector The eigenvectors of the trend, seasonality, and residual components are added to enhance the information flow between the components. The specific calculation process is as follows:

[0082]

[0083]

[0084] In the formula The trend, seasonality and residual components are integrated after the shared features.

[0085] S43. Introduce the residual connection mechanism to add the final generated component features to their original input component features to form a residual connection. This residual structure helps to alleviate the gradient dissipation problem that may occur in deep networks when multiple interactive learning modules are stacked.

[0086] Furthermore, in step S4, the interactive learning module is as follows: Figure 2 As shown, a residual connection is introduced in S43, and the three component features generated in the previous step are added to the three component features of the corresponding original input to obtain the final feature vectors of the three components. The specific calculation process is as follows:

[0087]

[0088] In the formula The final feature vector is obtained by residual connection of the three component features of trend, seasonality and residual of the i-th data. are the initial trend, seasonality and residual component features input to the interactive learning module respectively. In the next interactive learning module, will be E′T,i , S′ T,i , R′ T，i Enter the updated value of .

[0089] S5. The final feature vectors of the three components of trend, seasonality and residual are concatenated to form a comprehensive feature representation. Then, the concatenated comprehensive feature vector is input into the fully connected layer. After calculation, the survival rate prediction result of pot seedling transplanting culture is obtained.

[0090] Furthermore, in step S5, the three component feature vectors of the trend, seasonality and residual of the same data are first concatenated to form a complete feature representation of the data, and then the complete feature vectors of all data are added to generate a comprehensive feature vector containing all data. The specific calculation process is as follows:

[0091]

[0092] In the formula is the comprehensive feature vector of all data features, For the splicing operation, finally, the comprehensive feature vector is input into the fully connected layer and activated by the sigmoid function to obtain the predicted value of the survival rate of pot seedling transplanting culture. The specific calculation process is as follows:

[0093] P = Sigmoid (WF + b);

[0094] In the formula is the predicted value of the final survival rate of pot seedling transplanting culture, Sigmoid(·) is the sigmoid function, is the trainable weight matrix, is a trainable bias parameter.

[0095] Furthermore, this method uses Python 3.8 programming language and is based on the deep learning framework PyTorch for model development. During the training process, NVIDIA RTX 3090 GPU is used to provide computing support, and the initial learning rate is set to 1×10 -3 , the batch size is 12, the Dropout regularization probability is set to 0.3, and the loss function is the mean square error MSE.

[0096] Furthermore, the method predicts the effect as Figure 5 As shown in the figure, the ordinate is the survival rate of pot seedling transplanting culture (%), and the abscissa is time (days). The black solid line represents the predicted value of the survival rate of pot seedling transplanting culture, and the gray dotted line represents the actual survival rate of pot seedling transplanting culture. By comparing and observing, it can be found that the prediction results of the model are basically consistent with the true value curve trend, indicating that the method has a high degree of fit in the prediction of pot seedling transplanting culture survival rate, which verifies the effectiveness and reliability of the method.

[0097] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the survival rate of pot seedling transplanting culture, characterized in that: The following steps are involved: S1. Collect pot seedling transplanting and cultivation data, including daily average temperature, day and night temperature difference, light intensity, wind speed, soil moisture, soil pH value, seedling height, seedling leaf color, and transplanting time, check the collected data, use interpolation method to fill in missing values, and ensure the integrity of the data; S2. Use the STL sequence decomposition method to decompose the time series data into trend component, seasonal component and residual component to effectively separate the long-term trend, cyclical fluctuation and random disturbance in the data, providing a clearer feature basis for subsequent modeling and analysis; S3, using sine and cosine functions to alternately generate codes for different time steps to provide information about the relative order of the time steps, and concatenating the generated position codes with the trend, seasonality, and residual components, respectively, to facilitate the model to understand the sequential characteristics of the time series; S4. Construct an interactive learning module to deeply extract the features of each component and perform interactive learning to realize information sharing and fusion, which specifically includes the following steps: S41. Build a component feature extraction module, input the trend, seasonality and residual components into two fully connected layers respectively, and combine the ReLU activation function for nonlinear mapping to extract the unique feature representation of each component. At the same time, the Dropout regularization technique is applied during the training process to randomly close some neurons to enhance the generalization ability of the model. S42, constructing an information sharing module to extract shared features of the three components to capture common information between the components. By combining the shared features with the features of each component, each component contains information about its own features and other components, thereby enhancing the interaction and fusion between the component features. S43, introduce residual connection, add the final generated component features to their original input features, alleviate the problem of gradient dissipation in deep networks, enhance the network's expressiveness, and improve model performance; S5. The final features of the three components of trend, season and residual are concatenated to form a comprehensive feature vector, and then the comprehensive feature vector is input into the fully connected layer and the sigmoid function to calculate the prediction result of the survival rate of pot seedling transplanting culture.

2. The method for predicting the survival rate of pot seedling transplanting culture according to claim 1, characterized in that: In step S2, first input the time series data with a time window size of T Where n is the number of data types, is the time series data of the ith data in T consecutive time steps, is the monitoring value of the i-th data at the t-th time. Then, the time series of each data is decomposed into trend component, seasonal component and residual component by STL sequence decomposition algorithm. The specific calculation process is as follows: T，i , S T，i , R T，i =STL(x T，i ); Where E T，i =[e 1，i , e 2，i ..., e T，i ]、S T，i =[s 1，i ,s 2，i , ..., s T，i ]、 are the trend component, seasonal component and residual component obtained by decomposing the i-th data through STL sequence, e t，i 、s t，i , are the trend value, seasonal value and residual value obtained by STL sequence decomposition of the ith data at time t, respectively. STL(·) is the STL sequence decomposition algorithm.

3. A method for predicting the survival rate of pot seedling transplanting culture according to claim 2, characterized in that: In step S3, sine and cosine functions are used to generate periodic position encoding vectors for the time series to obtain position features at different time steps. The specific calculation process is as follows: Where PE t,2k , are the encoding values ​​of the position encoding vector in even and odd dimensions respectively, t∈[1,T] is the time step index, d is the total dimension of the position encoding vector, k∈[0,d / 2] is the index in the position encoding vector dimension, sin(·) and cos(·) are the sine function and cosine function respectively, and the position encoding value of the tth time step is concatenated into a complete position encoding vector according to the dimension. It is then spliced ​​to the tth time step in the trend, seasonality, and residual components so that each component contains location information. The specific calculation process is as follows: Where FE = [PE1, PE2, ..., PE T ] is the total position encoding vector of T consecutive time steps, E′ T,i , S′ T，i , are the concatenation results of position coding, trend component, seasonal component and residual component, respectively, where h = T + d, is the concatenation operation in the time step dimension.

4. A method for predicting the survival rate of pot seedling transplanting culture according to claim 3, characterized in that: In the step S4, the interactive learning module is constructed. In the specific step S41, a component feature extraction module is constructed. First, the three component features of trend, seasonality and residual are respectively input into the fully connected layer, and the ReLU activation function is applied for nonlinear activation. The specific calculation process is as follows: In the formula They are the feature vectors obtained after the three components of trend, seasonality and residual are activated by the fully connected layer and ReLU function. is the trainable weight matrix, is a trainable bias parameter, ReLU(·) is a ReLU activation function, and then the obtained component feature vector is regularized by Dropout to randomly close some neurons during the training process to enhance the generalization ability of the model. The regularized component features are passed through the fully connected layer and ReLU activation function again to further extract deep features. The specific calculation process is as follows: In the formula The three component features of trend, seasonality and residual are obtained by passing the fully connected layer and ReLU function again to obtain the feature vectors. is the trainable weight matrix, is a trainable bias parameter, Dropout(·) is the Dropout regularization operation, and ReLU(·) is the ReLU activation function.

5. A method for predicting the survival rate of pot seedling transplanting culture according to claim 4, characterized in that: In the step S4, constructing an interactive learning module, in the specific step S42, an information sharing module is constructed to capture the common information of the three components to generate a shared feature vector to enhance the model's learning ability for global features, and then the generated shared feature vector is added to the feature vectors of the trend, seasonality and residual components respectively to enhance the information flow between the components. The specific calculation process is as follows: In the formula is the shared eigenvector of the three components, For Hadamard, The feature vectors are obtained by adding the trend, seasonality and residual components to the shared features.

6. A method for predicting the survival rate of pot seedling transplanting culture according to claim 5, characterized in that: In the step S4, constructing the interactive learning module, in the specific step S43, a residual connection is introduced to add the feature vectors of the three components to their original input component features respectively to obtain the updated final feature vectors of the three components. The specific calculation process is as follows: In the formula The three component features of the trend, seasonality and residual of the i-th data are connected by residuals to obtain the feature vector, E′ T，i , S ′T，i , are the initial trend component features, seasonal component features and residual component features input to the interactive learning module respectively. In the next interactive learning module, Will replace E' T，i , S′ T，i , R′ T，i Become the component features of the input.

7. The method for predicting the survival rate of pot seedling transplanting culture according to claim 1, characterized in that: In order to predict the survival rate of pot seedlings transplanted and cultivated, the data of pot seedlings transplanted and cultivated were collected, including the average daily temperature, the temperature difference between day and night, the light intensity, the wind speed, the soil moisture, the soil pH value, the seedling height, the seedling leaf color and the transplanting time. The missing values ​​in the data were filled by the interpolation method to ensure the integrity and consistency of the time series data. Then the preprocessed data set was divided into the training set and the validation set for the training and validation of the model, providing a reliable data basis for the prediction of the survival rate of pot seedlings transplanted and cultivated.

Images