CO 2 Prediction method and device for tomato net primary productivity driven by balance

Abstract

The present invention provides a method and device for predicting the net primary productivity of tomatoes driven by CO2 balance. The method includes: performing data fusion and key feature extraction on multi-modal data of a greenhouse to obtain key features; inputting the key features and physical parameters into a CO2 balance calculation model to obtain the CO2 cumulative consumption of the greenhouse; and inputting the CO2 cumulative consumption into a net primary productivity estimation model to obtain the net primary productivity of tomatoes. The method and device for predicting the net primary productivity of tomatoes driven by CO2 balance provided by the present invention, by taking the CO2 balance of the greenhouse as the main driving factor for predicting the net primary productivity, realize the real-time calculation and monitoring of the CO2 cumulative consumption in the greenhouse. Using multi-modal data fusion and deep learning technologies, a CO2 balance calculation model and a net primary productivity estimation model are constructed in sequence, realizing an "end-to-end" process from data collection to data analysis, and improving the accuracy of net primary productivity prediction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to a method and device for predicting the net primary productivity of tomatoes driven by CO 2 balance. Background Art

[0002] The net primary productivity (NPP) of tomatoes in a greenhouse refers to the organic carbon fixed by photosynthesis of tomatoes minus the organic carbon consumed by respiration within a unit time (a specific time period such as this year or since planting), and the common unit is g / m 2 / yr, which reflects the growth rate and biomass accumulation of tomatoes. The prediction of NPP of tomatoes in a greenhouse is of great significance and value for evaluating the growth status of tomatoes, predicting the yield and quality of tomatoes, guiding water and fertilizer management and pest control in the greenhouse, etc. Theoretically, CO 2 is an essential substance for tomato photosynthesis, and the CO 2 balance has an important impact on the prediction of NPP of tomatoes in a greenhouse and is also a product of tomato respiration.

[0003] Existing technologies often only use single or limited data sources, collect single-modal or few-modal data inside and outside the greenhouse, lack the use of multi-modal data of tomatoes in the greenhouse, resulting in low data quality and integrity, and it is difficult to fully reflect the growth characteristics and status of tomatoes in the greenhouse, affecting the accuracy and reliability of prediction. Summary of the Invention

[0004] The present invention provides a method and device for predicting the net primary productivity of tomatoes driven by CO 2 balance, so as to improve the prediction accuracy of the net primary productivity of tomatoes in a greenhouse.

[0005] The present invention provides a method for predicting the net primary productivity of tomatoes driven by CO 2 balance, including:[[]]

[0006] Collecting multi-modal data of the greenhouse, where the multi-modal data includes one or more of the CO 2 concentration change data in the greenhouse, temperature data in the greenhouse, humidity data in the greenhouse, light intensity data in the greenhouse, wind speed data outside the greenhouse, growth status data of crops in the greenhouse, remote sensing data of the greenhouse, and control operation data of the greenhouse;

[0007] Performing data fusion on the multi-modal data, and extracting key features from the fused data based on a multi-modal variational autoencoder to obtain the key features of the multi-modal data;

[0008] Input the key features and the physical parameters of the greenhouse into a pre-constructed CO 2 equilibrium calculation model to obtain the CO 2 cumulative consumption in the greenhouse output by the CO 2 equilibrium calculation model. The CO 2 equilibrium calculation model is constructed based on a deep kernel learning neural network improved by a kernel function combined with a deep Gaussian process;

[0009] Input the CO 2 cumulative consumption into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model. The net primary productivity estimation model is constructed based on a dense connection network combined with an attention mechanism network.

[0010] According to a method for predicting the net primary productivity of tomatoes driven by CO 2 equilibrium provided by the present invention, the step of inputting the key features and the physical parameters of the greenhouse into a pre-constructed CO 2 equilibrium calculation model to obtain the CO 2 cumulative consumption in the greenhouse output by the CO 2 equilibrium calculation model includes:

[0011] Input the key features and the physical parameters of the greenhouse into a deep kernel learning neural network for deep feature extraction to obtain the deep features of the key features;

[0012] Based on a kernel function, map the deep features to a high-dimensional feature space to obtain high-dimensional deep features;

[0013] Based on a deep Gaussian process, perform CO 2 cumulative consumption prediction on the high-dimensional deep features to obtain the CO 2 cumulative consumption in the greenhouse.

[0014] According to a method for predicting the net primary productivity of tomatoes driven by CO 2 equilibrium provided by the present invention, the step of inputting the CO 2 cumulative consumption into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model includes:

[0015] Input the CO 2 cumulative consumption into a dense connection network for feature extraction to obtain the first residual feature;

[0016] Input the CO 2 cumulative consumption into an attention mechanism network for feature extraction to obtain the second residual feature;

[0017] Fuse the first residual feature and the second residual feature to obtain a fused feature, and perform a net primary productivity prediction on the fused feature to obtain the net primary productivity of tomatoes in the greenhouse.

[0018] According to a method for predicting the net primary productivity of tomatoes driven by a CO 2 balance, extracting key features of the fused data based on a multi-modal variational autoencoder to obtain key features of the multi-modal data, including:

[0019] Based on an encoder, map the fused data to a distribution of latent variables, and based on a decoder, construct a multi-modal data matrix from the distribution of latent variables. The multi-modal variational autoencoder is composed of an encoder and a decoder;

[0020] Based on the multi-modal data matrix, determine the key features of the multi-modal data.

[0021] According to a method for predicting the net primary productivity of tomatoes driven by a CO 2 balance, the training process of the CO 2 balance calculation model includes:

[0022] Based on Hamiltonian Monte Carlo sampling of stochastic gradient Hamiltonian Monte Carlo, adaptively adjust the parameters of the initial CO 2 balance calculation model to obtain a CO 2 balance calculation model with preliminary parameter adjustment;

[0023] Based on the noise of Langevin dynamics of stochastic gradient Langevin dynamics, adaptively adjust the parameters of the CO 2 balance calculation model with preliminary parameter adjustment to obtain the CO 2 balance calculation model.

[0024] According to a method for predicting the net primary productivity of tomatoes driven by a CO 2 balance, the training process of the net primary productivity estimation model includes:

[0025] Based on the second moment of the gradient of adaptive gradient clipping, adaptively adjust the clipping threshold of the gradient of the initial net primary productivity estimation model to obtain a net primary productivity estimation model with preliminary parameter adjustment;

[0026] Based on the first moment and the second moment of the gradient of adaptive momentum estimation, adaptively adjust the learning rate and momentum of the net primary productivity estimation model with preliminary parameter adjustment to obtain the net primary productivity estimation model.

[0027] The present invention also provides a device for predicting the net primary productivity of tomatoes driven by CO balance, comprising: 2 A data acquisition module, configured to acquire multi-modal data of a greenhouse, where the multi-modal data includes one or more of the CO concentration change data in the greenhouse, the temperature data in the greenhouse, the humidity data in the greenhouse, the light intensity data in the greenhouse, the wind speed data outside the greenhouse, the growth state data of crops in the greenhouse, the remote sensing data of the greenhouse, and the control operation data of the greenhouse;

[0028] A feature extraction module, configured to perform data fusion on the multi-modal data, and perform key feature extraction on the fused data based on a multi-modal variational auto-encoder to obtain the key features of the multi-modal data; 2 A parameter estimation module, configured to input the key features and the physical parameters of the greenhouse into a pre-constructed CO balance calculation model to obtain the cumulative consumption amount of CO in the greenhouse output by the CO balance calculation model, where the CO balance calculation model is constructed by combining a deep kernel learning neural network improved based on a kernel function and a deep Gaussian process;

[0029] A prediction module, configured to input the cumulative consumption amount of CO into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, where the net primary productivity estimation model is constructed by combining a densely connected network and an attention mechanism network.

[0030] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the computer program, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above.

[0031] The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above.

[0032] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the computer program, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above.

[0033] The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above.

[0034] The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above. 2 The present invention also provides a computer program product, comprising a computer program, where when the computer program is executed by a processor, it implements the method for predicting the net primary productivity of tomatoes driven by CO balance as described in any one of the above.

[0035] The CO 2 balance-driven prediction method and device for the net primary productivity of tomatoes, by taking the CO 2 balance in the greenhouse as the main driving factor for predicting the net primary productivity, realizes the real-time calculation and monitoring of the cumulative consumption of CO 2 in the greenhouse. At the same time, by using multi-modal data fusion and deep learning technologies, a CO 2 balance calculation model and a net primary productivity estimation model are constructed in sequence, realizing an "end-to-end" process from data collection to data analysis, and improving the prediction accuracy of the net primary productivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly describe the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0037] Figure 1 is a schematic flowchart of the method for predicting the net primary productivity of tomatoes driven by the CO 2 balance provided by the present invention;

[0038] Figure 2 is a schematic structural diagram of the device for predicting the net primary productivity of tomatoes driven by the CO 2 balance provided by the present invention;

[0039] Figure 3 is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0041] The net primary productivity of tomatoes in the greenhouse refers to the organic carbon fixed by photosynthesis minus the organic carbon consumed by respiration by tomatoes within a unit time (a specific time period such as this year or since planting) (the common unit is g / m 2 / yr), which reflects the growth rate and biomass accumulation of tomatoes. The prediction of tomato NPP in greenhouses is of great significance and value for evaluating the growth status of tomatoes, predicting tomato yield and quality, guiding water and fertilizer management and pest control in greenhouses, etc. Theoretically, CO 2 is an essential substance for tomato photosynthesis, and the CO 2 balance has an important impact on the prediction of tomato NPP in greenhouses and is also the product of tomato respiration. At present, the prediction methods of tomato NPP in greenhouses mainly include two categories: model-based methods and data-based methods. Whether it is model-based methods or data-based methods, key factors need to be considered, that is, the CO 2 balance. Among them, the CO 2 balance refers to the balance relationship between the input and output of CO 2 in the greenhouse, which reflects the CO 2 concentration change and demand in the greenhouse. However, the CO 2 concentration in the greenhouse is affected by various factors such as greenhouse ventilation, light, temperature, humidity, soil, plants, etc., and is a complex dynamic non-linear change process.

[0042] The main tomato NPP prediction methods in related methods are as follows:

[0043] Model-based methods: mainly include mechanical models, empirical models and mechanism models. Mechanical models are simple models established according to the physical laws in greenhouses (such as greenhouse energy balance models, CO2 balance models, etc.). Empirical models are statistical models established according to experimental data in greenhouses (such as linear regression models, non-linear regression models, etc.). Mechanism models are complex models established according to the biochemical processes in greenhouses (such as photosynthesis models, respiration models, growth models, etc.). The advantage of model-based methods is that they can better reflect the physical and biological processes in greenhouses and are conducive to explaining the causal relationship between CO2 balance and tomato NPP in greenhouses. However, a large number of environmental factors and physiological parameters are required as inputs, which are often difficult to obtain or have large uncertainties, resulting in the accuracy and stability of the models being affected. In addition, the model structure and parameters are often fixed and difficult to adapt to the dynamic changes and different planting conditions in greenhouses.

[0044] Data-based methods: include traditional machine learning methods and deep learning methods. Traditional machine learning methods use temperature, humidity, CO 2Sensor data such as concentration, through feature extraction and feature selection, establish a data-driven NPP prediction model (support vector machine SVM, artificial neural network ANN, random forest RF, etc.). Deep learning methods use multi-modal data such as spectra, images, and videos in the greenhouse to directly learn high-level features from the data and establish an "end-to-end" NPP prediction model (convolutional neural network CNN, recurrent neural network RNN, variational autoencoder VAE, etc.). The advantage of the above methods is to automatically learn and optimize using a large amount of data, adapt to the complex and non-linear data distribution in the greenhouse, and improve the accuracy and efficiency of NPP prediction. However, a large amount of labeled data is required as training data, which is difficult to obtain or has large errors, resulting in the generalization ability and credibility of the model being affected. In addition, the model structure and parameters are often black boxes, making it difficult to explain the internal relationship between CO 2 balance and the evolution process between tomato NPP.

[0045] The tomato NPP prediction technology in the greenhouse in the related methods mainly has the following deficiencies:

[0046] Insufficient data dimension. Existing technologies often only use single or limited data sources, collect single-modal or few-modal data inside and outside the greenhouse, and lack the use of multi-modal data of tomatoes in the greenhouse, resulting in low data quality and integrity, making it difficult to fully reflect the growth characteristics and states of tomatoes in the greenhouse, and affecting the accuracy and reliability of prediction.

[0047] Insufficient model ability. Existing technologies often adopt simple or fixed model structures. For example, statistical regression models, machine learning models, etc., usually take the NPP of tomatoes in the greenhouse as the output of the model and the environmental factors inside and outside the greenhouse as the input of the model for prediction. Ignoring the growth dynamics and non-linear characteristics of tomatoes in the greenhouse, it is difficult to adapt to the complex and changing environmental conditions in the greenhouse, and it is also difficult to effectively use the CO 2 balance to reflect the NPP level of tomatoes in the greenhouse, resulting in weak expression ability and computing ability of the model, and affecting the accuracy and stability of prediction.

[0048] Insufficient visualization / interpretability. Existing technologies often lack effective visualization methods, making it difficult to intuitively and clearly display the multi-modal data and prediction results of tomatoes in the greenhouse, and it is also difficult to effectively and meaningfully analyze the characteristics and relationships of tomatoes in the greenhouse, resulting in low data utilization rate and value, and affecting the interpretability and credibility of prediction results.

[0049] Aiming at the defects in the related methods, the present invention proposes a method for predicting tomato net primary productivity driven by CO 2 balance, Figure 1 For the CO provided by the present invention2 Schematic flow chart of a method for predicting the net primary productivity of tomatoes driven by carbon balance. Refer to Figure 1 , the present invention provides a CO 2 The method for predicting the net primary productivity of tomatoes driven by carbon balance may include:

[0050] Step 110, collecting multimodal data of the greenhouse, where the multimodal data includes one or more of the CO 2 concentration change data in the greenhouse, temperature data in the greenhouse, humidity data in the greenhouse, light intensity data in the greenhouse, wind speed data outside the greenhouse, growth state data of crops in the greenhouse, remote sensing data of the greenhouse, and control operation data of the greenhouse;

[0051] Step 120, performing data fusion on the multimodal data, and extracting key features from the fused data based on a multimodal variational autoencoder to obtain the key features of the multimodal data;

[0052] Step 130, inputting the key features and the physical parameters of the greenhouse into a pre-constructed CO 2 balance calculation model to obtain the cumulative consumption of CO 2 in the greenhouse output by the CO 2 balance calculation model, and the CO 2 balance calculation model is constructed based on a deep kernel learning neural network improved by a kernel function combined with a deep Gaussian process;

[0053] Step 140, inputting the cumulative consumption of CO 2 into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, and the net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network.

[0054] The CO provided by the present invention 2The execution entity of the balance-driven tomato net primary productivity prediction method can be an electronic device, a component in the electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a handheld computer, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., and the non-mobile electronic device can be a server, a Network Attached Storage (NAS), or a personal computer (PC), etc. The present invention does not make specific limitations.

[0055] The following takes a computer executing the CO provided by the present invention 2 balance-driven tomato net primary productivity prediction method as an example to elaborate on the technical solution of the present invention in detail.

[0056] In step 110, multi-modal data of the greenhouse is collected, and the multi-modal data includes one or more of the CO 2 concentration change data in the greenhouse, the temperature data in the greenhouse, the humidity data in the greenhouse, the light intensity data in the greenhouse, the wind speed data outside the greenhouse, the growth state data of the crops in the greenhouse, the remote sensing data of the greenhouse, and the control operation data of the greenhouse.

[0057] Specifically, sensors and cameras inside and outside the intelligent greenhouse, as well as unmanned aerial vehicles (UAVs) or satellites, can be used to collect the multi-modal data inside and outside the greenhouse. The collection includes the CO 2 concentration change, temperature, humidity, light, wind speed, growth state of the crops, leaf area of the crops, color of the crops, etc. inside the greenhouse, as well as high-resolution remote sensing data of the greenhouse (including surface temperature, albedo, etc. of the pixel where the greenhouse is located), and relevant control operation data (including ventilation, CO 2 supplementation, etc.).

[0058] In order to obtain the multi-modal data inside and outside the greenhouse, sensors and cameras inside and outside the greenhouse, as well as unmanned aerial vehicles (UAVs) or satellites, are used. Specifically, the following devices and methods are used:

[0059] Use the CO 2 sensor inside the greenhouse to measure the CO 2 concentration change in the greenhouse, with the unit of ppm, the highest sampling frequency being once per minute, and the data format being timestamp and CO 2A pair of concentration values, denoted as C(t), where t is time and C(t) is the CO 2 concentration.

[0060] Measure the temperature and humidity inside the greenhouse using the temperature and humidity sensors in the greenhouse. The units are °C and %RH respectively. The highest sampling frequency is once per minute. The data format is a triple of timestamp and temperature and humidity values, denoted as T(t) and H(t), where t is time, T(t) is the temperature, and H(t) is the humidity.

[0061] Measure the light intensity inside the greenhouse using a light sensor. The sampling frequency is at most once per minute. The data format is a pair of timestamp and light intensity value, denoted as L(t), where t is time and L(t) is the light intensity.

[0062] Measure the wind speed outside the greenhouse using a wind speed sensor. The unit is m / s. The highest sampling frequency is once per minute. The data format is a pair of timestamp and wind speed value, denoted as W(t), where t is time and W(t) is the wind speed.

[0063] Use a camera to photograph the growth status of the crops in the greenhouse, and collect the height, stem diameter, leaf area, leaf color, number of flowers and fruits, etc. of the crops in combination with manual records. The sampling frequency is once per day. The data format is a pair of timestamp and image data, denoted as S(t), where t is time and S(t) is the image data.

[0064] Use a drone (UAV) or satellite to photograph the high-resolution remote sensing data of the greenhouse, including the surface temperature of the greenhouse, the albedo of the greenhouse, etc. The sampling frequency is once per week. The data format is a pair of timestamp and image data, denoted as R(t), where t is time and R(t) is the image data.

[0065] Use the control system of the greenhouse to record the relevant control operation data of the greenhouse, including ventilation, CO 2 supplementation, irrigation, fertilization, etc. The sampling frequency is at the time of each operation. The data format is a triple of timestamp and operation type and parameter value, denoted as O(t), where t is time and O(t) is the operation type and parameter value.

[0066] In step 120, perform data fusion on the multi-modal data, and extract key features from the fused data based on the multi-modal variational autoencoder to obtain the key features of the multi-modal data.

[0067] Specifically, in order to preprocess and analyze the data, use a multi-source data fusion deep learning model based on the multi-modal variational autoencoder (MVAE) to extract high-level features and patterns of the data, as well as the interactions and correlations between the data.

[0068] Specifically, the following steps and methods are used:

[0069] For numerical data such as CO concentration changes, temperature, humidity, light intensity, wind speed, etc. inside the greenhouse, methods such as data cleaning, data normalization, and data interpolation are used to remove outliers, missing values, noise, etc., so that the data conforms to a normal distribution, has no missing values, no noise, etc., and is expressed as 2 and so on. etc.

[0070] For image data such as the growth status of crops inside the greenhouse and high-resolution remote sensing data of the greenhouse, methods such as image enhancement and image feature extraction are used to increase the contrast and clarity of the images, segment the regions of interest in the images, and extract features such as the color, texture, and shape of the images, and are expressed as and so on.

[0071] For categorical data such as related control operations of the greenhouse, methods such as data encoding, data mapping, and data embedding are used to convert the data into numerical data, so that the data has computability, comparability, embeddability, etc., and is expressed as etc.

[0072] All the preprocessed data is aligned and combined into a multi-modal data matrix according to the timestamp, and is expressed as where t is the time and X(t) is the multi-modal data matrix.

[0073] A multi-source data fusion deep learning model using a multi-modal variational autoencoder (MVAE) is used to perform data fusion on the multi-modal data matrix, extract high-level features and patterns of the data, as well as the interactions and correlations between the data.

[0074] Specifically, the following structure and formula are used:

[0075] The multi-modal variational autoencoder (MVAE) consists of an encoder and a decoder. The encoder maps the multi-modal data matrix to the distribution of latent variables, and the decoder reconstructs the distribution of latent variables into the multi-modal data matrix, while minimizing the reconstruction error and the difference between the distribution of latent variables and the prior distribution, and is expressed as:

[0076] Encoder:

[0077] Decoder:

[0078] Objective function:

[0079] Among them, M is the number of modalities, z is the latent variable, φ and θ are the parameters of the model, p(z) is the prior distribution of the latent variable, usually a standard normal distribution, and KL is the Kullback-Leibler divergence used to measure the similarity between two distributions.

[0080] The encoder consists of multiple sub-encoders. Each sub-encoder is responsible for processing the data of a modality and uses structures such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to map the data of each modality into a Gaussian distribution, expressed as:

[0081]

[0082] Among them, μ m and are the mean and variance of the Gaussian distribution, determined by the output of the sub-encoder, m is the index of the modality, and z is the latent variable.

[0083] The decoder consists of multiple sub-decoders. Each sub-decoder is responsible for reconstructing the data of a modality and uses structures such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to reconstruct the distribution of the latent variable into the data of each modality, expressed as:

[0084]

[0085] Among them, f m (z) is the output of the sub-decoder, is the variance of the reconstruction error, determined by the parameters of the sub-decoder, m is the index of the modality, and z is the latent variable.

[0086] Using the method of variational inference, optimize the objective function of the multi-modal variational autoencoder (MVAE), minimizing the reconstruction error and the difference between the distribution of the latent variable and the prior distribution, expressed as:

[0087] Optimization objective:

[0088] Optimization method: Stochastic Gradient Descent (SGD) or other optimization algorithms

[0089] Among them, φ and θ are the parameters of the multi-modal variational autoencoder (MVAE), X(t) is the multi-modal data matrix, and L(φ, θ; X(t)) is the objective function.

[0090] Use the output of the multi-modal variational autoencoder (MVAE), that is, the distribution of the latent variable, as the high-level features and patterns of the data, as well as the interactions and correlations between the data, expressed as:

[0091] Z(t) = q φ (z || X(t));

[0092] where t is time, Z(t) is the distribution of latent variables, and q φ (z||X(t)) is the output of the encoder of a multi-modal variational autoencoder (MVAE).

[0093] In step 130, the key features and the physical parameters of the greenhouse are input into a pre-constructed CO 2 balance calculation model to obtain the CO 2 cumulative consumption in the greenhouse output by the CO 2 balance calculation model. The CO 2 balance calculation model is constructed based on a deep kernel learning neural network improved by a kernel function combined with a deep Gaussian process.

[0094] Using the key features extracted from multi-modal data inside and outside the greenhouse, physical parameters (including greenhouse volume, surface area, thermal conductivity coefficient of greenhouse materials, etc.) and other relevant control operations (including ventilation, CO 2 supplementation, etc.) as inputs, and taking the CO 2 cumulative consumption of the greenhouse as the output, a CO 2 balance calculation model is established. By combining the deep neural network of deep kernel learning (DKL) and the kernel function, the processing of the nonlinearity and uncertainty of data is enhanced, and the calculation accuracy and reliability of CO 2 balance are improved. At the same time, by using the feature extraction of the deep neural network and the feature mapping of the kernel function, the expression ability and generalization ability of the model are improved.

[0095] By combining multi-layer Gaussian processes of deep Gaussian process (DGP), the processing of the nonlinearity and uncertainty of data is enhanced, and the calculation accuracy and reliability of CO 2 balance are improved. At the same time, by using the Bayesian inference and variational inference of the Gaussian process, the inference ability and approximation ability of the model are improved. The function of this step is to use the multi-modal data inside and outside the greenhouse, as well as other relevant physical parameters, and a more advanced machine learning model to improve the calculation accuracy and reliability of CO 2 balance, and provide a CO 2 balance calculation model for the training and testing of the subsequent CO 2 balance calculation model.

[0096] In step 140, the CO 2 cumulative consumption is input into a pre-constructed net primary productivity estimation model to obtain the tomato net primary productivity in the greenhouse output by the net primary productivity estimation model. The net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network.

[0097] Build a net primary productivity estimation model using a Dense Convolutional Network (DenseNet) and an Attention Network to estimate the NPP in a greenhouse, improving the estimation accuracy and feature selection ability of the model.

[0098] Utilize the features of all previous layers of each layer of the Dense Convolutional Network (DenseNet) to enhance the feature transfer and feature reuse of the network, improving the efficiency and performance of the network. Use the attention weights of the Attention Network to enhance the feature selection and feature attention of the network, improving the flexibility and robustness of the network.

[0099] Utilize CO 2 Balance the results of the calculation model, as well as multi-modal data inside and outside the greenhouse, and advanced deep learning models to improve the estimation accuracy and generalization ability of net primary productivity.

[0100] It can be understood that taking the CO 2 balance inside the greenhouse as the main driving factor for NPP prediction, realizing the real-time calculation and monitoring of the cumulative consumption amount of CO 2 inside the greenhouse, as well as the analysis and evaluation of its impact on NPP. Fully consider the impact of the CO 2 concentration inside the greenhouse on the photosynthesis and respiration of tomatoes, as well as the impact of various environmental factors inside the greenhouse on the CO 2 concentration, improving the accuracy and reliability of NPP prediction. Compared with the model-based methods in related methods, it does not require a large number of environmental factors and physiological parameters as inputs, reducing the complexity and uncertainty of the model; compared with the data-based methods in related methods, it does not require a large number of labeled data as training data, reducing the error and unbelievability of the model.

[0101] Based on multi-modal data fusion and deep learning technologies, successively build a CO 2 balance calculation model and a net primary productivity estimation model to achieve an "end-to-end" process from data collection to data analysis. Based on the Multi-modal Variational Autoencoder (MVAE) for data fusion, extract the complementary and correlated information between multi-modal data, enhancing the expression ability and information content of the data. The method of the present invention uses Deep Kernel Learning (DKL) and Deep Gaussian Process (DGP) for CO 2 balance calculation, enhancing the non-linear processing ability and probabilistic reasoning ability of the model. At the same time, use the Dense Convolutional Network (DenseNet) and the Attention Network to estimate the net primary productivity, improving the estimation accuracy and feature selection ability of the model.

[0102] The CO 2A method for predicting the net primary productivity of tomatoes driven by balance, which realizes the real-time calculation and monitoring of the cumulative consumption of CO₂ in the greenhouse by taking the CO₂ balance in the greenhouse as the main driving factor for predicting the net primary productivity. At the same time, by using multi-modal data fusion and deep learning technologies, a CO₂ balance calculation model and a net primary productivity estimation model are constructed in sequence to achieve an "end-to-end" process from data collection to data analysis, improving the accuracy of predicting the net primary productivity. 2 balance as the main driving factor for predicting the net primary productivity, realizing the real-time calculation and monitoring of the cumulative consumption of CO₂ in the greenhouse. 2 cumulative consumption. Meanwhile, using multi-modal data fusion and deep learning technologies, a CO₂ 2 balance calculation model and a net primary productivity estimation model are constructed in sequence to achieve an "end-to-end" process from data collection to data analysis, improving the accuracy of predicting the net primary productivity.

[0103] In one embodiment, the key features and the physical parameters of the greenhouse are input into a pre-constructed CO₂ 2 balance calculation model to obtain the CO₂ 2 cumulative consumption output by the CO₂ 2 balance calculation model in the greenhouse, including: inputting the key features and the physical parameters of the greenhouse into a deep kernel learning neural network for deep feature extraction to obtain the deep features of the key features; based on the kernel function, mapping the deep features to a high-dimensional feature space to obtain high-dimensional deep features; based on a deep Gaussian process, performing CO₂ 2 cumulative consumption prediction on the high-dimensional deep features to obtain the CO₂ 2 cumulative consumption in the greenhouse.

[0104] Optionally, the specific implementation process can be:

[0105] Using physical parameters such as the volume of the greenhouse, the surface area, and the thermal conductivity of the material of the greenhouse, represented as V, A, k, etc., for calculating the heat balance and CO₂ 2 balance.

[0106] Using the output of a multi-modal variational autoencoder (MVAE), that is, the distribution of latent variables, as the high-level features and laws of the data, as well as the interaction and correlation between the data, represented as Z(t), where t is time and Z(t) is the distribution of latent variables, for providing the fusion and representation of multi-modal data inside and outside the greenhouse.

[0107] Using the combination of a deep neural network and a kernel function in deep kernel learning (DKL) to enhance the processing of the nonlinearity and uncertainty of the data, improving the calculation accuracy and reliability of the CO₂ 2 balance, and at the same time using the feature extraction of the deep neural network and the feature mapping of the kernel function to improve the expression ability and generalization ability of the model.

[0108] Specifically, the following structure and formula are used:

[0109] Deep Kernel Learning (DKL) consists of a deep neural network and a kernel function. The deep neural network maps the high-level features and patterns of the data, as well as the interactions and correlations between the data, into a low-dimensional feature space, and the kernel function maps the low-dimensional feature space into a high-dimensional feature space, expressed as:

[0110] f(X(t)) = k(g(X(t)));

[0111] Where f(X(t)) is the output of Deep Kernel Learning (DKL), k(·) is the kernel function, g(X(t)) is the output of the deep neural network, and X(t) is the high-level features and patterns of the data, as well as the interactions and correlations between the data.

[0112] The deep neural network consists of multiple fully connected layers, convolutional layers, recurrent layers, etc., and uses techniques such as activation functions, normalization layers, dropout layers, etc. to improve the nonlinearity, stability, robustness, etc. of the network, expressed as:

[0113] g(X(t)) = F(X(t); θ);

[0114] Where F is the function of the deep neural network, θ is the parameter of the deep neural network, and X(t) is the high-level features and patterns of the data, as well as the interactions and correlations between the data.

[0115] The kernel function is a Gaussian kernel function, Laplace kernel function, polynomial kernel function, etc. Using the kernel trick, it maps the low-dimensional feature space into a high-dimensional feature space to improve the feature expression ability and discrimination ability, expressed as:

[0116]

[0117] Where exp(·) is the exponential function, ||·|| is the norm, μ and σ 2 are the parameters of the Gaussian kernel function, and g(X(t)) is the output of the deep neural network.

[0118] Using the combination of multiple-layer Gaussian processes of Deep Gaussian Process (DGP) to enhance the processing of the nonlinearity and uncertainty of the data, and improve the computational accuracy and reliability of CO 2 At the same time, using the Bayesian inference and variational inference of the Gaussian process to improve the inference ability and approximation ability of the model.

[0119] Specifically, the following structure and formula are used:

[0120] Deep Gaussian Process (DGP) consists of multiple Gaussian processes. Each Gaussian process takes the output of the Gaussian process in the previous layer as input and outputs a Gaussian process, expressed as:

[0121] f l(X(t)) = GP(m l (X(t)), k l (X(t), X(t)));

[0122] where, f l (X(t)) is the output of the Gaussian process of the l-th layer, GP(·, ·) is the distribution of the Gaussian process, m l (X(t)) is the mean function of the Gaussian process of the l-th layer, k l (X(t), X(t)) is the covariance function of the Gaussian process of the l-th layer, X(t) is the high-level features and patterns of the data, as well as the interactions and correlations between the data, or the output of the Gaussian process of the previous layer.

[0123] The mean function and covariance function of the Gaussian process are determined by the output of deep kernel learning (DKL). Using the non-linearity and uncertainty of deep kernel learning (DKL), the expressive power and generalization ability of the Gaussian process are improved, which is expressed as:

[0124] m l (X(t)) = f(X(t));

[0125] k l (X(t), X(t)) = f(X(t))f(X(t)) T ;

[0126] where, f(X(t)) is the output of deep kernel learning (DKL), and f(X(t))f(X(t)) T is the outer product of the output of deep kernel learning (DKL), and X(t) is the high-level features and patterns of the data, as well as the interactions and correlations between the data.

[0127] Using the method of Bayesian inference, the parameters of the deep Gaussian process (DGP) are optimized to maximize the posterior probability, improving the inference ability and approximation ability of the model, which is expressed as:

[0128] Optimization objective:

[0129] Optimization method: Variational inference or other inference algorithms

[0130] where, φ and θ are the parameters of the deep Gaussian process (DGP), and Y(t) is the cumulative consumption of CO 2 in the greenhouse, and p(φ, θ|Y(t)) is the posterior probability.

[0131] Using the output of the deep Gaussian process (DGP), that is, the output of the Gaussian process of the last layer, as the prediction of the cumulative consumption of CO 2 in the greenhouse, which is expressed as:

[0132] Y(t) = f L (X(t));

[0133] where t is time, Y(t) is the cumulative consumption of CO₂ in the greenhouse, 2 and f L (X(t)) is the output of the Gaussian process of the last layer.

[0134] Optionally, the specific values of the core parameters of the constructed CO₂ balance calculation model can be as shown in Table 1: 2

[0135] Table 1 CO₂ Balance Calculation Model Core Parameter Value Table 2

[0136]

[0137]

[0138] In one embodiment, the training process of the CO₂ balance calculation model includes: Hamiltonian Monte Carlo sampling based on Stochastic Gradient Hamiltonian Monte Carlo (SGHMC), adaptively adjusting the parameters of the initial CO₂ balance calculation model to obtain the CO₂ balance calculation model after preliminary parameter adjustment; based on the noise of Langevin dynamics of Stochastic Gradient Langevin Dynamics (SGLD), adaptively adjusting the parameters of the CO₂ balance calculation model after preliminary parameter adjustment to obtain the CO₂ balance calculation model. 2 2 2 2 2

[0139] Optionally, the specific training process can be as follows:

[0140] Use the Hamiltonian Monte Carlo sampling of Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) to adaptively adjust the parameters of the model to maximize the posterior probability and improve the effect and efficiency of Bayesian inference of the model. Use the noise of Langevin dynamics of Stochastic Gradient Langevin Dynamics (SGLD) to adaptively adjust the parameters of the model to minimize the variational lower bound and improve the effect and efficiency of variational inference of the model. Evaluate the performance of the CO₂ balance calculation model, such as mean squared error (MSE) or correlation coefficient, etc. The role of this step is to use advanced optimization algorithms to adjust the parameters of the CO₂ balance calculation model to minimize the calculation error of the CO₂ balance, and to evaluate the performance of the CO₂ balance calculation model, providing the training and test results of the model for the confidence and error analysis of the subsequent CO₂ balance calculation model. 2 2 2 2 2

[0141] ​​​​​​​​​​​​Specifically, the following steps and methods are used:

[0142] Use sampling of Hamiltonian Monte Carlo with Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) to adaptively adjust the parameters of the model, maximize the posterior probability, and improve the effect and efficiency of Bayesian inference of the model.

[0143] Specifically, the following formulas and parameters are used:

[0144] Sampling of Hamiltonian Monte Carlo with Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) is a gradient-based Markov chain Monte Carlo (MCMC) sampling method. It uses simulation of Hamiltonian dynamics to effectively explore the posterior distribution in high-dimensional space and uses stochastic gradients to adapt to large-scale data, expressed as:

[0145] Hamiltonian dynamics:

[0146] Stochastic gradient:

[0147] where θ t is the parameter of the model, v t is the auxiliary variable, ∈ t is the step size, U(θ t ) is the potential energy function, C is the friction coefficient, η t is the noise of the standard normal distribution, N is the total amount of data, n is the number of data for each sampling, y i is the observed value of the data, p(θ t ) is the prior distribution of the model, p(y i ∣θ t ) is the likelihood function of the model.

[0148] Sampling of Hamiltonian Monte Carlo with Stochastic Gradient Hamiltonian Monte Carlo (SGHMC) uses the following parameter settings to adapt to the data and model in the embodiments of the present invention:

[0149] The step size ∈ t is 0.01, and a fixed step size is used to maintain the stability of sampling.

[0150] The friction coefficient C is 0.05, and a smaller friction coefficient is used to reduce the bias of sampling.

[0151] The number of data for each sampling n is 100, and a smaller number of data is used to improve the efficiency of sampling.

[0152] The number of iterations of sampling is 1000, and a larger number of iterations is used to improve the convergence of sampling.

[0153] The noise of Langevin dynamics using Stochastic Gradient Langevin Dynamics (SGLD) adaptively adjusts the parameters of the model, minimizes the variational lower bound, and improves the effect and efficiency of variational inference of the model.

[0154] Specifically, the following formula and parameters are used:

[0155] The noise of Langevin dynamics of Stochastic Gradient Langevin Dynamics (SGLD) is a gradient-based stochastic optimization algorithm that effectively optimizes the objective function in high-dimensional space using the simulation of Langevin dynamics, expressed as:

[0156]

[0157] where θ t is the parameter of the model, ∈ t is the step size, L(θ t ) is the objective function, η t is the noise of the standard normal distribution.

[0158] The noise of Langevin dynamics of Stochastic Gradient Langevin Dynamics (SGLD) uses the following parameter settings to adapt to the data and model in the embodiments of the present invention:

[0159] The step size ∈ t is 0.01, and a fixed step size is used to maintain the stability of the optimization.

[0160] The objective function L(θ t ) is the variational lower bound, and the variational lower bound is used to approximate the posterior distribution, expressed as:

[0161]

[0162] where q φ (z∣X(t)) is the variational distribution, p θ (Y(t),z∣X(t)) is the joint distribution, p(z∣X(t)) is the posterior distribution, φ and θ are the parameters of the model, Y(t) is the cumulative consumption of CO in the greenhouse 2 , z is the latent variable, X(t) is the high-level features and patterns of the data, as well as the interactions and correlations between the data, and KL is the Kullback-Leibler divergence used to measure the similarity of two distributions.

[0163] Evaluate the performance of the CO 2 balance calculation model, such as mean square error (MSE) or correlation coefficient, etc.

[0164] Specifically, the following metrics and methods are used:

[0165] The mean squared error (MSE) is an index for measuring the difference between the predicted value and the true value, and is expressed as the mean of the squared errors:

[0166]

[0167] where T is the total number of time points, Y(t) is the true value of the cumulative consumption of CO₂ in the greenhouse, 2 is the predicted value of the cumulative consumption of CO₂ in the greenhouse, is the square error. 2 is the predicted value of the cumulative consumption of CO₂ in the greenhouse, is the square error.

[0168] The correlation coefficient is an index for measuring the correlation between the predicted value and the true value, and is expressed as the ratio of the covariance and standard deviation of the predicted value and the true value:

[0169]

[0170] where, is the covariance of the predicted value and the true value, σ Y and are the standard deviations of the predicted value and the true value, Y and are the vectors of the predicted value and the true value.

[0171] Optionally, confidence and error analysis can be performed on the CO₂ balance calculation model. 2 Using the posterior distribution of the Gaussian process of Bayesian optimization (BO), adaptively adjust the hyperparameters of the model to maximize the performance of the model, improve the confidence and error analysis of the model, and at the same time use the sampling and prediction of the Gaussian process to improve the calculation of the confidence interval and error range of the model. Using the posterior distribution of the weights of the neural network of Bayesian neural network (BNN), adaptively adjust the structure of the model to maximize the performance of the model, improve the confidence and error analysis of the model, and at the same time use the distribution and variance of the output of the neural network to improve the calculation of the confidence interval and error range of the model. The role of this step is to use advanced confidence and error analysis techniques to enhance the credibility of the results of the CO₂ balance calculation model and the analysis of errors, and provide the calculation results of CO₂ balance for subsequent NPP estimation.

[0172] Specifically, the following steps and methods are used: 2 Using the posterior distribution of the Gaussian process of Bayesian optimization (BO), adaptively adjust the hyperparameters of the model to maximize the performance of the model, improve the confidence and error analysis of the model. 2 Specifically, the following structure and formula are used:

[0173] Specifically, the following steps and methods are used:

[0174] Using the posterior distribution of the Gaussian process of Bayesian optimization (BO), adaptively adjust the hyperparameters of the model to maximize the performance of the model, improve the confidence and error analysis of the model.

[0175] Specifically, the following structure and formula are used:

[0176] Bayesian optimization (BO) is a global optimization method based on Bayesian inference. It effectively finds the optimal hyperparameters in high-dimensional space using the posterior distribution of the Gaussian process. At the same time, it uses an acquisition function to balance exploration and exploitation, expressed as:

[0177] Gaussian process: f(θ) ~ GP(m(θ), k(θ, θ));

[0178] Acquisition function: α(θ) = E[f(θ)|D t ;

[0179] Optimization objective:

[0180] where f(θ) is the objective function, θ is the hyperparameter, m(θ) is the mean function of the Gaussian process, k(θ, θ) is the covariance function of the Gaussian process, α(θ) is the acquisition function, D t is the observed dataset, and t is the number of iterations.

[0181] The mean function and covariance function of the Gaussian process are determined by the output of the deep Gaussian process (DGP). Using the non-linearity and uncertainty of the deep Gaussian process (DGP) improves the expressive power and generalization ability of the Gaussian process, expressed as:

[0182] m(θ) = f L (X(θ));

[0183] k(θ, θ) = f L (X(θ))f L (X(θ)) T ;

[0184] where f L (X(θ)) is the output of the Gaussian process of the last layer, and f L (X(θ))f L (X(θ)) T is the outer product of the output of the Gaussian process of the last layer. X(θ) is the high-level features and patterns of the data, as well as the interactions and correlations between the data, or the output of the Gaussian process of the previous layer.

[0185] The acquisition function is expected improvement (EI) or upper confidence bound (UCB) or probability of improvement (PoI), etc. Using different strategies to balance exploration and exploitation, expressed as:

[0186] Expected improvement: α(θ) = E[max(f(θ) - f(θ * ), 0)];

[0187] Upper confidence bound: α(θ) = μ(θ) + βσ(θ);

[0188] Probability improvement:

[0189] Among them, f(θ) is the objective function, θ is the hyperparameter, f(θ * ) is the maximum value of the objective function, μ(θ) and σ(θ) are the mean and standard deviation of the posterior distribution of the Gaussian process, β is the confidence coefficient, and Φ(·) is the cumulative distribution function of the standard normal distribution.

[0190] Use the optimizer of Bayesian optimization (BO) to perform optimization and find the optimal hyperparameters, which is expressed as:

[0191]

[0192] Among them, θ * is the optimal hyperparameter, α(θ) is the acquisition function, and θ is the hyperparameter.

[0193] Use the optimal hyperparameters to train the CO 2 balanced calculation model, and use the posterior distribution of the weights of the neural network of Bayesian neural network (BNN) to adaptively adjust the structure of the model, maximize the performance of the model, and improve the confidence and error analysis of the model.

[0194] Specifically, use the following structure and formula:

[0195] Bayesian neural network (BNN) is a neural network based on Bayesian inference. It uses the posterior distribution of the weights of the neural network to effectively find the optimal structure of the neural network in high-dimensional space. At the same time, it uses variational inference to effectively approximate the posterior distribution in large-scale data, which is expressed as:

[0196] Neural network: f(X(t)) = F(X(t); W);

[0197] Posterior distribution: p(W||Y(t)) ∝ p(Y(t)|W)p(W);

[0198] Variational distribution: q φ (W) ≈ p(W||Y(t));

[0199] Variational lower bound:

[0200] Among them, f(X(t)) is the output of the neural network, F(·; W) is the function of the neural network, W

[0201] is the weight of the neural network, p(W|Y(t)) is the posterior distribution of the weights of the neural network, p(Y(t)|W) is the likelihood function of the neural network, p(W) is the prior distribution of the weights of the neural network, q φ(W) is the variational distribution of the weights of the neural network, φ is the parameter of the variational distribution, Y(t) is the cumulative consumption of CO₂ in the greenhouse, and KL is the Kullback-Leibler divergence that measures the similarity between two distributions. 2 The cumulative consumption, and KL is the Kullback-Leibler divergence that measures the similarity between two distributions.

[0202] The neural network is composed of multiple fully connected layers, convolutional layers, recurrent layers, etc., and uses techniques such as activation functions, normalization layers, dropout layers, etc. to improve the nonlinearity, stability, robustness, etc. of the network, which is expressed as:

[0203] f(X(t)) = F(X(t); W);

[0204] Among them, F(·; W) is the function of the neural network, W is the weight of the neural network, and X(t) is the high-level features and laws of the data, as well as the interactions and correlations between the data.

[0205] The posterior distribution of the weights of the neural network is determined by the method of Bayesian inference. Using Bayes' theorem, the prior distribution and the likelihood function are combined to obtain the posterior distribution, which is expressed as:

[0206] p(W|Y(t)) ∝ p(Y(t)|W)p(W);

[0207] Among them, p(W|Y(t)) is the posterior distribution of the weights of the neural network, p(Y(t)|W) is the likelihood function of the neural network, p(W) is the prior distribution of the weights of the neural network, Y(t) is the cumulative consumption of CO₂ in the greenhouse, and W is the weight of the neural network. 2 The cumulative consumption, and W is the weight of the neural network.

[0208] The variational distribution of the weights of the neural network is approximated by the method of variational inference. Using the variational lower bound, the log-likelihood of the posterior distribution and the Kullback-Leibler divergence of the variational distribution are separated to obtain the variational distribution, which is expressed as:

[0209] q φ (W) ≈ p(W|Y(t));

[0210] Among them, q φ (W) is the variational distribution of the weights of the neural network, p(W|Y(t)) is the posterior distribution of the weights of the neural network, φ is the parameter of the variational distribution, Y(t) is the cumulative consumption of CO₂ in the greenhouse, and W is the weight of the neural network. 2 The cumulative consumption, and W is the weight of the neural network.

[0211] Using the method of variational inference, optimize the parameters of the variational distribution of the weights of the neural network, minimize the variational lower bound, and improve the effect and efficiency of variational inference of the model, which is expressed as:

[0212] Optimization objective:

[0213] Optimization method: Stochastic Gradient Descent (SGD) or other optimization algorithms;

[0214] Where φ is the parameter of the variational distribution of the weights of the neural network, and L(φ) is the variational lower bound, expressed as:

[0215]

[0216] Where q φ (W) is the variational distribution of the weights of the neural network, p(Y(t)|W) is the likelihood function of the neural network, p(W) is the prior distribution of the weights of the neural network, Y(t) is the cumulative consumption of CO in the greenhouse, W is the weight of the neural network, and KL is the Kullback-Leibler divergence. 2

[0217] Use the output of the variational distribution of the weights of the neural network, that is, the approximation of the posterior distribution of the weights of the neural network, as the optimal value of the weights of the neural network, expressed as:

[0218] W * = q φ (W);

[0219] Where W * is the optimal value of the weights of the neural network, q φ (W) is the variational distribution of the weights of the neural network, and φ is the parameter of the variational distribution.

[0220] Use the optimal weights of the neural network to test the CO 2 balance calculation model, and use the distribution and variance of the output of the neural network to improve the calculation of the confidence interval and error range of the model.

[0221] Specifically, use the following formula and method:

[0222] The distribution and variance of the output of the neural network are determined by the approximation of the posterior distribution of the weights of the neural network. Using the mean and variance of the approximation of the posterior distribution of the weights of the neural network, the distribution and variance of the output of the neural network are obtained and expressed as:

[0223] Distribution: p(Y(t)||X(t)) = N(f(X(t)), σ 2 );

[0224] Variance:

[0225] Where p(Y(t)|X(t)) is the distribution of the output of the neural network, N(·,·) is the normal distribution, f(X(t)) is the output of the neural network, X(t) is the high-level features and patterns of the data, as well as the interactions and correlations between the data, and σ2 is the variance of the output of the neural network, is the variance of the variational distribution of the weights of the neural network, q φ (W) is the variational distribution of the weights of the neural network, and φ is the parameter of the variational distribution.

[0226] Using the distribution and variance of the output of the neural network, calculate the confidence interval and error range of the output of the neural network. Using the properties of the normal distribution, obtain the confidence interval and error range of the output of the neural network, expressed as:

[0227] Confidence interval: [μ - ασ, μ + ασ];

[0228] Error range: 2ασ;

[0229] where μ is the mean of the output of the neural network, σ is the standard deviation of the output of the neural network, and α is the confidence level, generally taking values such as 1.96 or 2.58.

[0230] In one embodiment, the CO 2 cumulative consumption is input into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, including: the CO 2 cumulative consumption is input into a dense connection network for feature extraction to obtain a first residual feature; the CO 2 cumulative consumption is input into an attention mechanism network for feature extraction to obtain a second residual feature; the first residual feature and the second residual feature are fused to obtain a fused feature, and the fused feature is used for net primary productivity prediction to obtain the net primary productivity of tomatoes in the greenhouse.

[0231] Optionally, utilize the features of all the previous layers of each layer of the dense connection network (DenseNet) to enhance the feature transfer and feature reuse of the network, and improve the efficiency and performance of the network. Utilize the attention weights of the attention mechanism network (AttentionNetwork) to enhance the feature selection and feature attention of the network, and improve the flexibility and robustness of the network. The role of this step is to utilize CO 2 to balance the results of the calculation model, as well as multi-modal data inside and outside the greenhouse, and advanced deep learning models, to improve the estimation accuracy and generalization ability of NPP, and provide a net primary productivity estimation model for the training and testing of subsequent net primary productivity estimation models.

[0232] Specifically, use the following structure and formula:

[0233] DenseNet (Dense Convolutional Network) is an improved deep neural network based on the Deep Residual Network (ResNet). It uses dense connections to connect the features of all previous layers of each layer, which is expressed as:

[0234] x l =H l ([x 0 ,x 1 ,…,x l-1 );

[0235] Among them, x l is the output of the l-th layer, H l (·) is the function of the l-th layer, and [x 0 ,x 1 ,…,x l-1 is the concatenation of the outputs of all previous layers of the l-th layer, and l is the number of layers.

[0236] The function of each layer of DenseNet consists of a Batch Normalization (BN) layer, a Rectified Linear Unit (ReLU) layer, and a Convolution (Conv) layer. Using the techniques of a standard Convolutional Neural Network (CNN) can improve the nonlinearity, stability, robustness, etc. of the network, which is expressed as:

[0237] H l (·)=Conv(ReLU(BN(·)));

[0238] Among them, BN(·) is the function of the Batch Normalization (BN) layer, ReLU(·) is the function of the Rectified Linear Unit (ReLU) layer, and Conv(·) is the function of the Convolution (Conv) layer.

[0239] The Convolution (Conv) layer of each layer of DenseNet uses a fixed kernel size and stride to keep the size of the feature map unchanged, which is expressed as:

[0240] Conv(·)=Conv k,s (·);

[0241] Among them, Conv k,s (·) is the function of the Convolution (Conv) layer, k is the kernel size, s is the stride, and k = 3, s = 1, etc. are taken.

[0242] The number of channels of the output of each layer of DenseNet is controlled by a fixed growth rate to control the number of network parameters and computational complexity, which is expressed as:

[0243] C l =C 0 +k l ;

[0244] Among them, C l is the number of channels of the output of the l-th layer, and C 0 is the number of channels of the output of the 0-th layer. k l is the growth rate of the l-th layer, and generally k l = 12, 24, 48, etc.

[0245] The features of the output of each layer of the Dense Connection Network (DenseNet) are used as the feature transfer and feature reuse of the network, improving the efficiency and performance of the network, which is expressed as:

[0246] F l = x l ;

[0247] Among them, F l is the feature of the output of the l-th layer, x l is the output of the l-th layer, and l is the number of layers.

[0248] The Attention Mechanism Network is a deep neural network based on Self-Attention. Using attention weights, it sums up the features of all previous layers of each layer, which is expressed as:

[0249]

[0250] Among them, y l is the output of the l-th layer, a li is the attention weight between the l-th layer and the i-th layer, and F i is the feature of the output of the i-th layer. l and i are the numbers of layers.

[0251] The attention weight of each layer of the Attention Mechanism Network is composed of a Multi-Head Self-Attention layer, a Feed-Forward layer, a Residual Connection layer, and a Layer Normalization layer. Using standard Natural Language Processing (NLP) techniques, it improves the feature selection and feature attention of the network, which is expressed as:

[0252] a li = LN(RC(FF(MHSA(F l , F i ))));

[0253] Among them, MHSA(·,·) is the function of the Multi-Head Self-Attention layer, FF(·) is the function of the Feed-Forward layer, RC(·) is the function of the Residual Connection layer, LN(·) is the function of the Layer Normalization layer, F l ,F i is the feature of the output of the l-th layer and the i-th layer, where l and i are the layer numbers.

[0254] The feature of the output of each layer of the Attention Network, as the feature selection and feature attention of the network, improves the flexibility and robustness of the network, and is expressed as:

[0255] G l = y l ;

[0256] Among them, G l is the feature of the output of the l-th layer, and y l is the output of the l-th layer, where l is the layer number.

[0257] Using the outputs of the DenseNet and the Attention Network, that is, the features of the output of each layer, as the high-level features and laws of the data, as well as the interaction and correlation between the data, are expressed as:

[0258] X(t) = [F 0 , F 1 , …, F L , G 0 , G 1 , …, G L ;

[0259] Among them, t is the time, X(t) is the high-level features and laws of the data, as well as the interaction and correlation between the data, F 0 , F 1 , …, F L are the features of the output of each layer of the DenseNet, G 0 , G 1 , …, G L are the features of the output of each layer of the Attention Network, and L is the number of layers.

[0260] Using the output of the net primary productivity estimation model, that is, the output of the last layer, as the estimated value of NPP, is expressed as:

[0261] N(t) = f(X(t));

[0262] Where t is time, N(t) is the estimated value of NPP, f(f(t)) is the output of the last layer, and X(t) is the high-level features and patterns of the data.

[0263] Optionally, the core parameters of the constructed net primary productivity estimation model can be as shown in Table 2.

[0264] Table 2 Core Parameter Table of Net Primary Productivity Estimation Model

[0265]

[0266]

[0267] In one embodiment, the training process of the net primary productivity estimation model includes: adaptively adjusting the clipping threshold of the gradient of the initial net primary productivity estimation model based on the second moment of the gradient of adaptive gradient clipping (AdaGradClip) to obtain the net primary productivity estimation model after preliminary parameter adjustment; adaptively adjusting the learning rate and momentum of the net primary productivity estimation model after preliminary parameter adjustment based on the first moment and second moment of the gradient of adaptive momentum estimation (AdaM) to obtain the net primary productivity estimation model.

[0268] Using the second moment of the gradient of adaptive gradient clipping (AdaGradClip) to adaptively adjust the clipping threshold of the gradient to prevent gradient explosion or gradient disappearance, and improve the stability and convergence of optimization. Using the first moment and second moment of the gradient of adaptive momentum estimation (AdaM) to adaptively adjust the learning rate and momentum to accelerate the speed and effect of optimization. Evaluate the performance of the net primary productivity estimation model, such as mean square error (MSE) or correlation coefficient, etc. The role of this step is to use the optimization algorithm to adjust the parameters of the net primary productivity estimation model to minimize the estimation error of NPP, and to evaluate the performance of the net primary productivity estimation model, providing the model training and test results for the subsequent feature extraction and feature visualization of the net primary productivity estimation model.

[0269] Specifically, the following steps and methods are used:

[0270] Using the second moment of the gradient of adaptive gradient clipping (AdaGradClip) to adaptively adjust the clipping threshold of the gradient to prevent gradient explosion or gradient disappearance, and improve the stability and convergence of optimization.

[0271] Specifically, the following formula and parameters are used:

[0272] Adaptive Gradient Clipping (AdaGradClip) is an improved optimization algorithm based on Adaptive Gradient (AdaGrad). It uses the second moment of the gradient to adaptively adjust the gradient clipping threshold, which is expressed as:

[0273] Second moment of the gradient:

[0274] Gradient clipping threshold:

[0275] Gradient clipping:

[0276] where r t is the second moment of the gradient at the t-th iteration, r t-1 is the second moment of the gradient at the (t - 1)-th iteration, is the gradient at the t-th iteration, δ t is the gradient clipping threshold at the t-th iteration, η is the initial learning rate, ∈ is the smoothing term, θ t is the parameter at the t-th iteration, and J(θ t ) is the loss function at the t-th iteration.

[0277] The second moment of the gradient and the gradient clipping threshold of Adaptive Gradient Clipping (AdaGradClip) use the following parameter settings to adapt to the data and model in the embodiments of the present invention:

[0278] The initial learning rate η is 0.01. Using a smaller initial learning rate maintains the stability of optimization.

[0279] The smoothing term ∈ is 10 -8 , and using a smaller smoothing term prevents the denominator from being zero.

[0280] Using the first moment and second moment of the gradient of Adaptive Momentum Estimation (AdaM) to adaptively adjust the learning rate and momentum, accelerating the speed and effect of optimization.

[0281] Specifically, the following formula and parameters are used:

[0282] Adaptive Momentum Estimation (AdaM) is an optimization algorithm based on the combination of Adaptive Gradient (AdaGrad) and Momentum. It uses the first moment and second moment of the gradient to adaptively adjust the learning rate and momentum, which is expressed as:

[0283] First moment of the gradient:

[0284] Second moment of the gradient:

[0285] Bias correction:

[0286] Parameter update:

[0287] Among them, m t and v t are the first and second moments of the gradient at the t-th iteration, m t-1 and v t-1 are the first and second moments of the gradient at the (t - 1)-th iteration, is the gradient at the t-th iteration, β 1 and β 2 are the decay rates of the first and second moments of the gradient, and are the bias corrections of the first and second moments of the gradient at the t-th iteration, η is the initial learning rate, ∈ is the smoothing term, θ t is the parameter at the t-th iteration, J(θ t ) is the loss function at the t-th iteration.

[0288] For the first and second moments of the gradient of Adaptive Moment Estimation (AdaM), the following parameter settings are used to adapt to the data and model in the embodiments of the present invention:

[0289] The initial learning rate η is 0.01, and a smaller initial learning rate is used to maintain the stability of optimization.

[0290] The smoothing term ∈ is 10 -8 , and a smaller smoothing term is used to prevent the denominator from being zero.

[0291] The decay rate β 1 of the first moment of the gradient is 0.9, and a larger decay rate is used to maintain the inertia of the first moment of the gradient.

[0292] The decay rate β 2 of the second moment of the gradient is 0.999, and a larger decay rate is used to maintain the stability of the second moment of the gradient.

[0293] Use the optimizers of Adaptive Gradient Clipping (AdaGradClip) and Adaptive Moment Estimation (AdaM) for optimization to find the optimal parameters, expressed as:

[0294]

[0295] Among them, θ * is the optimal parameter, J(θ) is the loss function, and θ is the parameter.

[0296] Use the optimal parameters to train and test the net primary productivity estimation model, and use indicators such as mean square error (MSE) or correlation coefficient (introduced in the previous steps) to evaluate the performance of the net primary productivity estimation model.

[0297] Optionally, feature extraction and feature visualization can be performed on the net primary productivity estimation model.

[0298] Using the graph structure of the graph convolutional neural network (GCN), enhance the extraction and representation of the topological and geometric features of the data, and improve the expression ability and discrimination ability of the features. Using the tensor structure of tensor decomposition (TD), enhance the decomposition and reconstruction of the multi-dimensional and multi-modal features of the data, and improve the compression ability and interpretability of the features. At the same time, perform an interpretable and visual display of the results of the net primary productivity estimation model, including using heat maps, scatter plots, bar charts, etc. The role of this step is to use the technologies of feature extraction and feature visualization to enhance the understanding and mining of the internal laws of the data, as well as the interpretability and visualization of the model results, and provide data analysis and display for the subsequent application and evaluation of the model.

[0299] Specifically, the following steps and methods are used:

[0300] Using the graph structure of the graph convolutional neural network (GCN), enhance the extraction and representation of the topological and geometric features of the data, and improve the expression ability and discrimination ability of the features.

[0301] Specifically, the following structure and formula are used:

[0302] The graph convolutional neural network (GCN) is a deep neural network based on the graph structure. Using graph convolution, aggregate the information of the node and its neighbors, and represent the node as a feature vector, expressed as:

[0303]

[0304] where, H (l) is the output of the l-th layer, σ(·) is the activation function, is the degree matrix of the graph, is the adjacency matrix of the graph, W (l) is the weight matrix of the l-th layer, and l is the number of layers

[0305] The graph convolutional neural network (GCN) uses the following parameter settings to adapt to the data and model in the embodiments of the present invention:

[0306] The activation function σ(·) is the rectified linear unit (ReLU). Using the non-linearity, stability, robustness, etc. of the rectified linear unit (ReLU), expressed as σ(x) = max(0, x), where x is the input and σ(x) is the output.

[0307] The number of layers l is 2. Using two layers of graph convolutional neural network (GCN) to balance the complexity and performance of the network, expressed as l = 2, where l is the number of layers.

[0308] The output dimension of the first layer is 64, and a 64-dimensional feature vector is used to represent the nodes of the graph, denoted as d 1 = 64, where d 1 is the output dimension of the first layer.

[0309] The output dimension of the second layer is 1, and a 1-dimensional feature vector is used to represent the NPP estimation value of the nodes, denoted as d 2 = 1, where d 2 is the output dimension of the second layer.

[0310] Use the tensor structure of tensor decomposition (TD) to enhance the decomposition and reconstruction of multi-dimensional and multi-modal features of data, and improve the compression ability and interpretability of features.

[0311] Specifically, use the following structure and formula:

[0312] Tensor decomposition (TD) is a dimensionality reduction method based on the tensor structure. It decomposes the multi-dimensional and multi-modal features of data into several low-dimensional core tensors and factor matrices, denoted as:

[0313]

[0314] where X is the original data tensor, R is the rank of the tensor, λ r is the weight of the r-th core tensor, G r is the r-th core tensor, A (n) is the n-th factor matrix, N is the order of the tensor, × n is the n-th mode multiplication of the tensor.

[0315] Tensor decomposition (TD) uses the following parameter settings to adapt to the data and model in the embodiments of the present invention:

[0316] The data tensor X is a third-order tensor, and a three-dimensional data structure is used to represent the multi-modal features of the data, denoted as , where I is the size of the first dimension, representing the time dimension of the data, J is the size of the second dimension, representing the spatial dimension of the data, and K is the size of the third dimension, representing the modal dimension of the data.

[0317] The rank R of the tensor is 10, and 10 core tensors and factor matrices are used to approximate the original data tensor, denoted as R = 10, where R is the rank of the tensor.

[0318] The core tensor G r is a third-order tensor, and a three-dimensional data structure is used to represent the low-dimensional features of the data, denoted as where I r is the size of the first dimension, representing the low-dimensional features of the time dimension of the data, J ris the size of the second dimension, representing the low-dimensional features of the spatial dimension of the data, K r is the size of the third dimension, representing the low-dimensional features of the modal dimension of the data, and r is the serial number of the core tensor.

[0319] The factor matrix A(n) is a two-dimensional matrix, using a two-dimensional data structure to represent the feature transformation of each dimension of the data, denoted as where S n is the size of the nth dimension, representing the original features of the nth dimension of the data, R is the rank of the tensor, representing the low-dimensional features of the nth dimension of the data, and n is the serial number of the dimension.

[0320] Using the results of tensor decomposition (TD), namely the core tensor and the factor matrix, as the decomposition and reconstruction of the multi-dimensional and multi-modal features of the data, to improve the compression ability and interpretability of the features, denoted as where X is the original data tensor, is the approximate data tensor, R is the rank of the tensor, λ r is the weight of the rth core tensor, G r is the rth core tensor, A (n) is the nth factor matrix, and n is the serial number of the dimension.

[0321] Using the result of the net primary productivity estimation model, that is, the output of the last layer, as the estimated value of NPP, denoted as N(t) = f(X(t)), where t is the time, N(t) is the estimated value of NPP, f(X(t)) is the output of the last layer, and X(t) is the high-level features and patterns of the data, as well as the interactions and correlations between the data.

[0322] Using the result of the net primary productivity estimation model, that is, the estimated value of NPP, for interpretable and visual display, including using heatmaps, scatter plots, bar charts, etc.

[0323] Specifically, use the following methods and graphs:

[0324] Heatmap, reflecting the distribution and changes of the estimated value of NPP in the spatial dimension, using the depth of color to represent the magnitude and trend of the estimated value of NPP.

[0325] Scatter plot, showing the relationship and similarity between the estimated value and the true value of NPP in the time dimension, using the position and shape of the points to represent the differences and consistencies between the estimated value and the true value of NPP.

[0326] Bar chart, presenting the comparison and evaluation between the estimated value and the true value of NPP in the modal dimension, using the height and color of the bars to represent the error and accuracy between the estimated value and the true value of NPP.

[0327] In one embodiment, key features of the fused data are extracted based on a multi-modal variational autoencoder to obtain the key features of the multi-modal data, including: mapping the fused data to a distribution of latent variables based on an encoder, and constructing a multi-modal data matrix from the distribution of latent variables based on a decoder, where the multi-modal variational autoencoder consists of an encoder and a decoder; determining the key features of the multi-modal data based on the multi-modal data matrix.

[0328] A multi-modal variational autoencoder (MVAE) consists of an encoder and a decoder. The encoder maps a multi-modal data matrix to a distribution of latent variables, and the decoder reconstructs the distribution of latent variables into a multi-modal data matrix while minimizing the reconstruction error and the difference between the distribution of latent variables and the prior distribution, expressed as:

[0329] Encoder:

[0330] Decoder: Objective function:

[0331] where M is the number of modalities, z is the latent variable, φ and θ are the parameters of the model, p(z) is the prior distribution of the latent variable usually a standard normal distribution, and KL is the Kullback-Leibler divergence used to measure the similarity of two distributions.

[0332] The encoder consists of multiple sub-encoders. Each sub-encoder is responsible for processing the data of a modality and uses structures such as a convolutional neural network (CNN) or a recurrent neural network (RNN) to map the data of each modality to a Gaussian distribution, expressed as:

[0333]

[0334] where μ m and are the mean and variance of the Gaussian distribution, determined by the output of the sub-encoder, m is the index of the modality, and z is the latent variable.

[0335] The decoder consists of multiple sub-decoders. Each sub-decoder is responsible for reconstructing the data of a modality and uses structures such as a convolutional neural network (CNN) or a recurrent neural network (RNN) to reconstruct the distribution of latent variables into the data of each modality, expressed as:

[0336]

[0337] where f m (z) is the output of the sub-decoder, is the variance of the reconstruction error, determined by the parameters of the sub-decoder, m is the index of the modality, and z is the latent variable.

[0338] Using the method of variational inference, optimize the objective function of the multi-modal variational autoencoder (MVAE), minimize the reconstruction error and the difference between the distribution of the latent variables and the prior distribution, expressed as:

[0339] Optimization objective:

[0340] Optimization method: Stochastic Gradient Descent (SGD) or other optimization algorithms

[0341] Among them, φ and θ are the parameters of the multi-modal variational autoencoder (MVAE), X(t) is the multi-modal data matrix, and L(φ, θ; X(t)) is the objective function.

[0342] Use the output of the multi-modal variational autoencoder (MVAE), that is, the distribution of the latent variables, as the high-level features and laws of the data, as well as the interactions and correlations between the data, expressed as:

[0343] Z(t) = q φ (z || X(t));

[0344] Among them, t is the time, Z(t) is the distribution of the latent variables, and q φ (z || X(t)) is the output of the encoder of the multi-modal variational autoencoder (MVAE).

[0345] Figure 2 The structural schematic diagram of the tomato net primary productivity prediction device driven by CO 2 balance provided by the present invention is shown as Figure 2 shown, and the device includes:

[0346] A data acquisition module 210 for acquiring multi-modal data of the greenhouse, where the multi-modal data includes one or more of the CO 2 concentration change data in the greenhouse, the temperature data in the greenhouse, the humidity data in the greenhouse, the light intensity data in the greenhouse, the wind speed data outside the greenhouse, the growth state data of the crops in the greenhouse, the remote sensing data of the greenhouse, and the control operation data of the greenhouse;

[0347] A feature extraction module 220 for performing data fusion on the multi-modal data and extracting key features from the fused data based on a multi-modal variational autoencoder to obtain the key features of the multi-modal data;

[0348] A parameter estimation module 230 for inputting the key features and the physical parameters of the greenhouse into a pre-constructed CO 2 balance calculation model to obtain the CO 2 cumulative consumption amount of CO in the greenhouse output by the balance calculation model, the CO 2 ​2 The equilibrium calculation model is constructed based on a deep kernel learning neural network with improved kernel functions combined with a deep Gaussian process;

[0349] The prediction module 240 is used to 2 The cumulative consumption is input into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, wherein the net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network.

[0350] The CO provided in the embodiment of the present invention 2 The equilibrium-driven prediction device for tomato net primary productivity was used to calculate the CO 2 Balance is the main driver of net primary productivity projections, achieving greenhouse CO 2 Real-time calculation and monitoring of cumulative consumption. At the same time, multimodal data fusion and deep learning technology are used to build CO 2 The balance calculation model and net primary productivity estimation model realize an "end-to-end" process from data collection to data analysis, improving the accuracy of net primary productivity prediction.

[0351] In one embodiment, the parameter estimation module 230 is specifically used for:

[0352] The key characteristics and physical parameters of the greenhouse are input into a pre-built CO 2 The equilibrium calculation model was used to obtain the CO 2 The CO in the greenhouse output by the balance calculation model 2 Cumulative consumption, including:

[0353] Inputting the key features and the physical parameters of the greenhouse into a deep kernel learning neural network to extract deep features, thereby obtaining deep features of the key features;

[0354] Based on the kernel function, the depth feature is mapped to a high-dimensional feature space to obtain a high-dimensional depth feature;

[0355] The high-dimensional deep features are CO-enhanced based on the deep Gaussian process. 2 Cumulative consumption prediction, the CO 2 Cumulative consumption.

[0356] In one embodiment, the prediction module 240 is specifically used to:

[0357] The CO 2 The accumulated consumption is input into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, including:

[0358] Input the cumulative consumption of the CO 2 into a densely connected network for feature extraction to obtain a first residual feature;

[0359] Input the cumulative consumption of the CO 2 into an attention mechanism network for feature extraction to obtain a second residual feature;

[0360] Fuse the first residual feature and the second residual feature to obtain a fused feature, and perform a net primary productivity prediction on the fused feature to obtain the net primary productivity of tomatoes in the greenhouse.

[0361] In one embodiment, the feature extraction module 220 is specifically configured to:

[0362] The key feature extraction of the fused data based on the multimodal variational autoencoder to obtain the key features of the multimodal data includes:

[0363] Map the fused data to a distribution of latent variables based on an encoder, and construct a multimodal data matrix from the distribution of latent variables based on a decoder. The multimodal variational autoencoder is composed of an encoder and a decoder;

[0364] Determine the key features of the multimodal data based on the multimodal data matrix.

[0365] In one embodiment, the parameter estimation module 230 is further specifically configured to:

[0366] CO 2 The training process of the CO balance calculation model includes:

[0367] Based on Hamiltonian Monte Carlo sampling of stochastic gradient Hamiltonian Monte Carlo, adaptively adjust the parameters of the initial CO balance calculation model to obtain a CO balance calculation model with preliminary parameter adjustment; 2 Based on the noise of Langevin dynamics of stochastic gradient Langevin dynamics, adaptively adjust the parameters of the CO balance calculation model with preliminary parameter adjustment to obtain the CO balance calculation model. 2 balance calculation model;

[0368] Based on the noise of Langevin dynamics of stochastic gradient Langevin dynamics, adaptively adjust the parameters of the CO balance calculation model with preliminary parameter adjustment to obtain the CO balance calculation model. 2 balance calculation model; 2 balance calculation model.

[0369] In one embodiment, the prediction module 240 is further specifically configured to:

[0370] The training process of the net primary productivity estimation model includes:

[0371] Based on the second moment of the gradient with adaptive gradient clipping, adaptively adjust the clipping threshold of the gradient of the initial net primary productivity estimation model to obtain the net primary productivity estimation model after preliminary parameter adjustment;

[0372] Based on the first moment and the second moment of the gradient with adaptive momentum estimation, adaptively adjust the learning rate and momentum of the net primary productivity estimation model after the preliminary parameter adjustment to obtain the net primary productivity estimation model.

[0373] Figure 3 Illustrated is a schematic diagram of the physical structure of an electronic device, as Figure 3 shown. The electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340. Among them, the processor 310, the communication interface 320, and the memory 330 complete communication with each other through the communication bus 340. The processor 310 may call the logical instructions in the memory 330 to execute CO 2 A tomato net primary productivity prediction method driven by carbon balance, the method includes:

[0374] Collect multi-modal data of the greenhouse, and the multi-modal data includes one or more of the CO 2 concentration change data in the greenhouse, the temperature data in the greenhouse, the humidity data in the greenhouse, the light intensity data in the greenhouse, the wind speed data outside the greenhouse, the growth state data of the crops in the greenhouse, the remote sensing data of the greenhouse, and the control operation data of the greenhouse;

[0375] Perform data fusion on the multi-modal data, and extract key features from the fused data based on a multi-modal variational autoencoder to obtain the key features of the multi-modal data;

[0376] Input the key features and the physical parameters of the greenhouse into a pre-constructed CO 2 balance calculation model to obtain the CO 2 cumulative consumption in the greenhouse output by the CO 2 balance calculation model, and the CO 2 balance calculation model is constructed based on a deep kernel learning neural network improved by a kernel function combined with a deep Gaussian process;

[0377] Input the CO 2 cumulative consumption into a pre-constructed net primary productivity estimation model to obtain the tomato net primary productivity in the greenhouse output by the net primary productivity estimation model. The net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network.

[0378] In addition, when the logical instructions in the above-mentioned memory 330 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0379] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the CO provided by the above-mentioned various methods. 2 A method for predicting the net primary productivity of tomatoes with balanced driving, the method comprising:

[0380] Collecting multi-modal data of the greenhouse, the multi-modal data including one or more of the CO concentration change data in the greenhouse, the temperature data in the greenhouse, the humidity data in the greenhouse, the light intensity data in the greenhouse, the wind speed data outside the greenhouse, the growth state data of the crops in the greenhouse, the remote sensing data of the greenhouse, and the control operation data of the greenhouse; 2 Fusing the multi-modal data, and extracting key features from the fused data based on a multi-modal variational autoencoder to obtain the key features of the multi-modal data;

[0381] Inputting the key features and the physical parameters of the greenhouse into a pre-constructed CO balance calculation model to obtain the cumulative consumption amount of CO in the greenhouse output by the CO balance calculation model. The CO balance calculation model is constructed by combining a deep kernel learning neural network improved by a kernel function and a deep Gaussian process;

[0382] Inputting the key features and the physical parameters of the greenhouse into a pre-constructed CO 2 balance calculation model to obtain the CO 2 cumulative consumption amount output by the CO balance calculation model of the CO in the greenhouse 2 The CO 2 balance calculation model is constructed by combining a deep kernel learning neural network improved by a kernel function and a deep Gaussian process;

[0383] Inputting the CO 2The cumulative consumption is input into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, wherein the net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network.

[0384] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the above-mentioned CO 2 A balance-driven method for predicting net primary productivity of tomato, comprising:

[0385] Collect multimodal data of the greenhouse, the multimodal data including CO 2 One or more of concentration change data, temperature data in the greenhouse, humidity data in the greenhouse, light intensity data in the greenhouse, wind speed data outside the greenhouse, growth status data of crops in the greenhouse, remote sensing data of the greenhouse, and control operation data of the greenhouse;

[0386] Performing data fusion on the multimodal data, and extracting key features of the fused data based on a multimodal variational autoencoder to obtain key features of the multimodal data;

[0387] The key characteristics and the physical parameters of the greenhouse are input into a pre-built CO 2 The equilibrium calculation model was used to obtain the CO 2 The CO in the greenhouse output by the balance calculation model 2 Cumulative consumption, the CO 2 The equilibrium calculation model is constructed based on a deep kernel learning neural network with improved kernel functions combined with a deep Gaussian process;

[0388] The CO 2 The cumulative consumption is input into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, wherein the net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network.

[0389] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0390] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0391] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting net primary productivity of tomatoes driven by CO2 balance, characterized in that: The method comprises: Collecting multimodal data of the greenhouse, the multimodal data including one or more of CO2 concentration change data in the greenhouse, temperature data in the greenhouse, humidity data in the greenhouse, light intensity data in the greenhouse, wind speed data outside the greenhouse, growth status data of crops in the greenhouse, remote sensing data of the greenhouse, and control operation data of the greenhouse; Performing data fusion on the multimodal data, and extracting key features of the fused data based on a multimodal variational autoencoder to obtain key features of the multimodal data; The key features and the physical parameters of the greenhouse are input into a pre-constructed CO2 balance calculation model to obtain the cumulative CO2 consumption in the greenhouse output by the CO2 balance calculation model, wherein the CO2 balance calculation model is constructed based on a deep kernel learning neural network improved by a kernel function combined with a deep Gaussian process, and the training process of the CO2 balance calculation model includes: adaptively adjusting the parameters of the initial CO2 balance calculation model based on Hamiltonian Monte Carlo sampling of stochastic gradient Hamiltonian Monte Carlo to obtain the CO2 balance calculation model after the preliminary parameter adjustment; adaptively adjusting the parameters of the CO2 balance calculation model after the preliminary parameter adjustment based on the noise of Langevin dynamics of stochastic gradient Langevin dynamics to obtain the CO2 balance calculation model; The accumulated CO2 consumption is input into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, wherein the net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network, and the training process of the net primary productivity estimation model includes: adaptively adjusting the clipping threshold of the gradient of the initial net primary productivity estimation model based on the second-order moment of the gradient of the adaptive gradient clipping to obtain the net primary productivity estimation model after the preliminary parameter adjustment; adaptively adjusting the learning rate and momentum of the net primary productivity estimation model after the preliminary parameter adjustment based on the first-order moment and the second-order moment of the gradient of the adaptive momentum estimation to obtain the net primary productivity estimation model; The step of inputting the key features and the physical parameters of the greenhouse into a pre-built CO2 balance calculation model to obtain the cumulative CO2 consumption in the greenhouse output by the CO2 balance calculation model includes: Inputting the key features and the physical parameters of the greenhouse into a deep kernel learning neural network to extract deep features, thereby obtaining deep features of the key features; Based on the kernel function, the depth feature is mapped to a high-dimensional feature space to obtain a high-dimensional depth feature; The cumulative consumption of CO2 is predicted based on the deep Gaussian process on the high-dimensional deep features to obtain the cumulative consumption of CO2 in the greenhouse.

2. The method for predicting tomato net primary productivity driven by CO2 balance according to claim 1, characterized in that: The step of inputting the accumulated CO2 consumption into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model comprises: Inputting the accumulated CO2 consumption into a densely connected network for feature extraction to obtain a first residual feature; Inputting the accumulated CO2 consumption into the attention mechanism network for feature extraction to obtain a second residual feature; The first residual feature and the second residual feature are fused to obtain a fused feature, and the net primary productivity of tomatoes in the greenhouse is predicted using the fused feature.

3. The method for predicting tomato net primary productivity driven by CO2 balance according to claim 1, characterized in that: The key features of the fused data are extracted based on the multimodal variational autoencoder to obtain the key features of the multimodal data, including: Mapping the fused data to a distribution of latent variables based on an encoder, and constructing the distribution of latent variables into a multimodal data matrix based on a decoder, wherein the multimodal variational autoencoder is composed of an encoder and a decoder; Based on the multimodal data matrix, key features of the multimodal data are determined.

4. A CO2 balance-driven tomato net primary productivity prediction device, characterized in that: include: A data acquisition module, used to collect multimodal data of the greenhouse, wherein the multimodal data includes one or more of CO2 concentration change data in the greenhouse, temperature data in the greenhouse, humidity data in the greenhouse, light intensity data in the greenhouse, wind speed data outside the greenhouse, growth status data of crops in the greenhouse, remote sensing data of the greenhouse, and control operation data of the greenhouse; A feature extraction module, used to perform data fusion on the multimodal data, and extract key features of the fused data based on a multimodal variational autoencoder to obtain key features of the multimodal data; A parameter estimation module is used to input the key features and the physical parameters of the greenhouse into a pre-constructed CO2 balance calculation model to obtain the cumulative CO2 consumption in the greenhouse output by the CO2 balance calculation model, wherein the CO2 balance calculation model is constructed based on a deep kernel learning neural network improved by a kernel function combined with a deep Gaussian process, and the training process of the CO2 balance calculation model includes: adaptively adjusting the parameters of the initial CO2 balance calculation model based on Hamiltonian Monte Carlo sampling of stochastic gradient Hamiltonian Monte Carlo to obtain the CO2 balance calculation model after the preliminary parameter adjustment; adaptively adjusting the parameters of the CO2 balance calculation model after the preliminary parameter adjustment based on the noise of Langevin dynamics of stochastic gradient Langevin dynamics to obtain the CO2 balance calculation model; A prediction module is used to input the CO2 cumulative consumption into a pre-constructed net primary productivity estimation model to obtain the net primary productivity of tomatoes in the greenhouse output by the net primary productivity estimation model, wherein the net primary productivity estimation model is constructed based on a densely connected network combined with an attention mechanism network, and the training process of the net primary productivity estimation model includes: adaptively adjusting the clipping threshold of the gradient of the initial net primary productivity estimation model based on the second-order moment of the gradient of the adaptive gradient clipping to obtain the net primary productivity estimation model after the preliminary parameter adjustment; adaptively adjusting the learning rate and momentum of the net primary productivity estimation model after the preliminary parameter adjustment based on the first-order moment and second-order moment of the gradient of the adaptive momentum estimation to obtain the net primary productivity estimation model; The step of inputting the key features and the physical parameters of the greenhouse into a pre-built CO2 balance calculation model to obtain the cumulative CO2 consumption in the greenhouse output by the CO2 balance calculation model includes: Inputting the key features and the physical parameters of the greenhouse into a deep kernel learning neural network to extract deep features, thereby obtaining deep features of the key features; Based on the kernel function, the depth feature is mapped to a high-dimensional feature space to obtain a high-dimensional depth feature; The cumulative consumption of CO2 is predicted based on the deep Gaussian process on the high-dimensional deep features to obtain the cumulative consumption of CO2 in the greenhouse.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the CO2 balance-driven tomato net primary productivity prediction method as described in any one of claims 1 to 3 is implemented.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting tomato net primary productivity driven by CO2 balance as claimed in any one of claims 1 to 3 is implemented.

7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for predicting tomato net primary productivity driven by CO2 balance as claimed in any one of claims 1 to 3 is implemented.

Images