An intelligent prediction method for rice phenological periods based on environmental data and deep learning

Abstract

An intelligent prediction method for rice phenological periods based on environmental data and deep learning, belonging to the field of smart agriculture technology. To achieve accurate prediction of rice phenological periods, the present invention includes obtaining environmental data and phenological period data throughout the year of rice growth, and constructing a phenological model dataset; performing data preprocessing on the phenological model dataset for subsequent model training and prediction; building a rice phenological period prediction model based on the Encoder-Decoder architecture of the Seq2Seq model, introducing a Bayesian neural network for uncertainty assessment, and formulating a determination criterion for updating the phenological model; training the rice phenological period prediction model according to 8 phenological periods divided by the rice growth cycle; performing real-time prediction and update based on the daily updated environmental data. The present invention can process and analyze daily meteorological data in real time, and dynamically update the prediction results of phenological periods, improving the efficiency and benefits of agricultural production.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart agriculture, and particularly relates to an intelligent prediction method for rice phenological periods based on environmental data and deep learning. Background Art

[0002] As one of the most important food crops in the world, the growth process of rice undergoes a series of phenological stages. The time nodes of these phenological periods can provide a basis for guiding production and managing disaster risks. Therefore, accurately predicting each phenological period is of great significance for the precise management and yield increase of rice.

[0003] Traditional phenological period prediction methods mainly rely on manual observation or camera-based monitoring technology. The manual observation method requires agricultural technicians to regularly observe the growth of rice in the field and record the start and end times of each phenological period. Although this method is simple, it highly depends on the experience of personnel, is prone to subjective errors, and has limited observation frequency, making it difficult to capture some important minor changes in the rice growth process. In addition, manual observation is inefficient in large-scale farmland operations and cannot achieve refined management. The phenological period monitoring technology based on cameras automatically detects the growth of rice through image processing and machine vision methods. Although it solves the limitations of manual observation to a certain extent, the installation and maintenance costs of cameras are relatively high, and their monitoring range is limited and cannot cover large areas of farmland. Especially under adverse weather conditions, the image quality collected by cameras will significantly decline, thus affecting the recognition accuracy of phenological periods. In addition, image processing methods are easily affected by background lighting, camera angles, and crop planting densities, further reducing the prediction accuracy.

[0004] In response to the above problems, in recent years, some rice phenological period prediction methods based on mathematical statistical models and process models have emerged. These methods predict the growth stages of rice by fitting environmental data and growth models. However, traditional statistical models are limited by the accuracy of modeling and cannot fully capture the non-linear relationship between complex environmental factors and the rice growth process. Process models, on the other hand, rely on detailed biological mechanisms, have a complex modeling process, and require a large amount of experimental data for calibration, which makes their adaptability poor between different regions and years and difficult to achieve large-scale application. Summary of the Invention

[0005] The problem to be solved by the present invention is to accurately predict the rice phenological periods through environmental data, and propose an intelligent prediction method for rice phenological periods based on environmental data and deep learning.

[0006] To achieve the above object, the present invention is realized through the following technical solutions:

[0007] An intelligent prediction method for rice phenological periods based on environmental data and deep learning, comprising the following steps:

[0008] S1. Obtain the environmental data and phenological period data of rice growth throughout the year, and construct a phenological model data set;

[0009] S2. Perform data preprocessing on the phenological model data set for subsequent model training and prediction;

[0010] S3. Build a rice phenological period prediction model based on the Encoder-Decoder architecture of the Seq2Seq model, introduce a Bayesian neural network for uncertainty assessment, and formulate a criterion for updating the phenological model;

[0011] S4. Train the rice phenological period prediction model according to the 8 phenological periods divided by the rice growth cycle;

[0012] S5. Perform real-time prediction and update based on the daily updated environmental data.

[0013] Furthermore, the specific implementation method of step S1 includes the following steps:

[0014] S1.1. Set the acquisition of environmental data and phenological period data of rice growth throughout the year, including the maximum temperature, minimum temperature, average daily temperature, daily precipitation, daily sunshine duration, daily humidity, and daily wind speed;

[0015] S1.2. Set the entire growth cycle of rice to include 8 main phenological periods: the seedling stage is from sowing to tillering, the tillering stage is from tillering to jointing, the jointing stage is from jointing to booting, the booting stage is from booting to heading, the heading stage is from heading to milk ripening, the milk ripening stage is from milk ripening to dough ripening, the dough ripening stage is from dough ripening to full ripening, and the full ripening stage is from full ripening to harvesting;

[0016] S1.3. For the key time nodes of the set phenological periods, including 9 nodes: sowing, tillering, jointing, booting, heading, milk ripening, dough ripening, full ripening, and harvesting;

[0017] S1.4. Respectively collect the environmental data and phenological period data of rice growth throughout the year for each phenological period, mark the phenological node time by establishing a timestamp date_ori, match and mark the environmental data with the corresponding rice phenological periods, and construct a phenological model data set.

[0018] Furthermore, the specific implementation method of step S2 includes the following steps:

[0019] S2.1. Organize the data collected in step S1 into time series data in units of days, and perform normalization processing on each data sample;

[0020] S2.2. Divide the annual environmental data according to the growth cycle of rice, and divide it into the seedling stage, tillering stage, jointing stage, booting stage, heading stage, milk ripening stage, dough ripening stage, and full ripening stage. Among them, the seedling stage is marked as 1, the tillering stage is marked as 2, the jointing stage is marked as 3, the booting stage is marked as 4, the heading stage is marked as 5, the milk ripening stage is marked as 6, the dough ripening stage is marked as 7, the full ripening stage is marked as 8, and other non-rice planting periods are marked as 0 for the training of the phenological period model; during the model training process, according to the phenological period markings of the annual data, a prediction model for 9 phenological nodes is trained.

[0021] Further, the specific implementation method of step S3 includes the following steps:

[0022] S3.1. Construct an Encoder using two bidirectional LSTMs as the model, and set the input feature X of the encoder as a three-dimensional tensor, where B represents the batch size, T represents the time series length of 20 days, and F represents the number of daily features, that is, 7 environmental features;

[0023] S3.2. First, use a fully connected layer to expand the input features, and set the weight matrix as bias term The shape of the output tensor is batch_size, T, 256, and the expression of the output tensor X' is:

[0024] X′ = ReLU(XW1 + b1)

[0025] where ReLU is the activation function;

[0026] S3.3. Input the output into a two-layer stacked bidirectional long short-term memory network Bi-LSTM structure. The first layer of bidirectional LSTM is used to process the input data and output the forward and backward hidden states, with dimensions of 128 respectively, and the total dimension is 256. The output shape is The expression is:

[0027]

[0028] where, respectively represent the forward and backward hidden states of the first time step t in the Bi-LSTM. biLSTM1 represents the first layer of bidirectional long short-term memory network, and X' t represents the input feature at time t;

[0029] The second layer of bidirectional LSTM receives the output of the first layer of bidirectional LSTM and further processes the data. The finally encoded time series features are output as The output shape is [batch_size, T, 256], and the expression is:

[0030]

[0031] Among them, respectively represent the forward and backward hidden states of the second-layer Bi-LSTM at time t, biLSTM2 represents the second-layer bidirectional long short-term memory network, and H represents the output of the first-layer bidirectional LSTM;

[0032] S3.4. The decoder Decoder receives the features output by the encoder and decodes them, decoding the encoded temporal features into future phenological prediction values. The decoder includes two layers of convolutional neural network CNN and a fully connected layer. The first layer of CNN uses a convolutional kernel to convert the input features from [batch_size, T, 16, 16] to [batch_size, T, 8, 8]. The expression is:

[0033] C1 = ReLU(Convld(H, W2))

[0034] Among them, C1 represents the output of the first layer of convolution, ReLU represents the activation function, H represents the input features of the encoder, and W2 represents the weight matrix of the first layer of convolutional kernel;

[0035] The second layer of CNN uses a convolutional kernel to further convert the features to batch_size, T, 4, 4. The decoded feature output is The expression is:

[0036] C2 = ReLU(Convld(C1, W3))

[0037] S3.5. Flatten the output C2 of the decoder into a vector form to obtain a shape of batch_size, T×16. After flattening the output of the convolutional network into a two-dimensional vector, a fully connected layer is introduced for final prediction processing, which is used to predict 9 time nodes of the rice growth cycle;

[0038] S3.6. Use the Bayesian layer to introduce uncertainty evaluation and calculate the confidence interval. Set the weight W4 and bias b4 as random variables that satisfy the probability distribution. The expression is:

[0039]

[0040] y = C2W4 + b4

[0041] Among them, W4 and b4 are represented as random variables in the Bayesian layer, representing the weight and bias of the model respectively, μ represents the mean, and σ 2 represents the variance, y represents the output of the Bayesian layer, and C2 represents the output features of the second convolutional layer in the decoder;

[0042] The output shape of the Bayesian layer is batch_size, 1, corresponding to the predicted value of each sample;

[0043] During the training process, the loss function is modified to a Bayesian loss function that includes uncertainty and is optimized by maximizing the variational lower bound ELBO. The expression is:

[0044]

[0045] where q(w) is the estimated weight distribution, p(w) is the prior distribution of the weights, and D is the training data;

[0046] S3.7. When making predictions, use Monte Carlo sampling (MC) to run the model and generate multiple different predicted values.

[0047] By sampling 100 times, 100 predicted values are obtained, and the mean and standard deviation of the predictions are obtained; the confidence interval of each predicted value is calculated through the standard deviation;

[0048] First, determine a confidence level of 95%, then calculate the corresponding z value, and then use the following formula to calculate the upper and lower bounds of each predicted value:

[0049] CI = μ ± z·σ

[0050] where μ is the mean of the predicted value, σ is the standard deviation, and z is the corresponding z value. For a 95% confidence interval, z is 1.96;

[0051] S3.8. Set a decision rule to ensure the accuracy of the prediction and the determination of whether to proceed to the next phenological period. When any two of the following conditions are met, stop updating the data, end the current phenological period prediction, and start the next phenological period prediction:

[0052] S3.8.1. Stability of the predicted value: Set as the predicted value on the t-th day. If the predicted values for k consecutive days fluctuate less, it indicates that the prediction of the model has become stable. The specific determination condition is:

[0053]

[0054] where ∈1 is the set threshold, taken as 1 day, i.e., ∈1 = 1;

[0055] S3.8.2. Convergence of the confidence interval: Set if the width of the confidence interval CI w is less than the threshold ∈2 for k consecutive days, indicating that the prediction of the model is stable enough. The determination formula is:

[0056] CI w (t) = CIupper (t)-CI lower (t)

[0057] CI w (t) ≤ ∈2 & CI w (t - 1) ≤ ∈2 & CI w (t - 2) ≤ ∈2 &

[0058] Among them, ∈2 is the threshold of the confidence interval width, set to 2 days;

[0059] S3.8.3. Proximity between the actual date and the predicted date: When the predicted phenological period is approaching, when the difference between the predicted date d pred and the current date d current is less than the set threshold, this rule is satisfied:

[0060] |d pred - d current | ≤ ∈3

[0061] Among them, ∈3 is the set threshold, taken as 10 days.

[0062] Furthermore, for the training of the independent prediction model described in step S4, the collected environmental data corresponds to 8 predicted phenological periods, that is, 9 time nodes. Training is carried out for different time nodes to predict each phenological period respectively.

[0063] Furthermore, step S5 first obtains the environmental data from January 1st to the current date, and introduces the weather forecast data for the next few days. At the same time, the sequence padding technology is adopted to add padding value 0 at the end of the input sequence to fill the input sequence to the fixed sequence length of the model, which is 365;

[0064] During the prediction process, the uncertainty of the model output is evaluated to determine the credibility of the prediction result; the prediction result is updated every day. When the judgment criterion is met, the sowing time node is determined; continue to predict. When the judgment criterion is met, the tillering time node is determined, and then the prediction of the rice seedling stage is completed, and the prediction of the tillering stage is started. And so on, gradually complete the prediction of subsequent jointing, booting, heading, milk ripening, dough ripening, full ripening and harvesting time nodes, and finally realize the comprehensive prediction of the entire phenological period of rice.

[0065] Advantages of the present invention:

[0066] For the intelligent prediction method of rice phenological period based on environmental data and deep learning described in the present invention, compared with the traditional manual observation and phenological period monitoring method based on cameras, the present invention does not rely on expensive physical devices. The cost is reduced through the prediction of environmental data, and it can be widely applied to the rice growth cycle management in different regions and different years.

[0067] An intelligent prediction method for rice phenological periods based on environmental data and deep learning described in the present invention utilizes the Seq2Seq model architecture in deep learning, combines bidirectional long short-term memory networks (Bi-LSTM) and convolutional neural networks (CNN), and can effectively capture the complex characteristics of meteorological data during the rice growth process. Through this model architecture, nine key time nodes of eight main phenological periods of rice can be accurately predicted.

[0068] An intelligent prediction method for rice phenological periods based on environmental data and deep learning described in the present invention can provide a credibility assessment for each prediction result by introducing a Bayesian neural network. By calculating the confidence interval, the system can quantify the accuracy of the prediction, provide a reliability reference for the prediction result to the user, and thus ensure the stability of the prediction and the scientific nature of decision-making.

[0069] An intelligent prediction method for rice phenological periods based on environmental data and deep learning described in the present invention can process and analyze daily meteorological data in real time and dynamically update the prediction results of phenological periods, providing efficient decision-making support for the precision agriculture management of rice. By predicting in advance the time nodes of each phenological period of rice, agricultural producers can scientifically and reasonably arrange planting and management plans according to the requirements of different growth stages, improving the efficiency and benefits of agricultural production. Description of the Drawings

[0070] Figure 1 is a flowchart of an intelligent prediction method for rice phenological periods based on environmental data and deep learning described in the present invention;

[0071] Figure 2 is a schematic diagram of rice phenological nodes and phenological periods of the present invention;

[0072] Figure 3 is a working flowchart of the intelligent prediction method for rice phenological periods of the present invention;

[0073] Figure 4 is an Encoder-Decoder model architecture diagram of the present invention;

[0074] Figure 5 is a schematic diagram of data and model update of the present invention. Detailed Embodiments

[0075] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the specific embodiments described are only a part of the embodiments of the present invention, rather than all of the specific embodiments. Usually, the components of the specific embodiments of the present invention described and shown in the accompanying drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0076] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents the selected specific embodiments of the present invention. Based on the specific embodiments of the present invention, all other specific embodiments obtained by those skilled in the art without creative efforts fall within the scope of protection of the present invention.

[0077] In order to further understand the content, features and effects of the present invention, the following specific embodiments are exemplified and are accompanied by Figure 1 -Accompanying Figure 5 The details are as follows:

[0078] Example 1:

[0079] An intelligent prediction method for rice phenological periods based on environmental data and deep learning, comprising the following steps:

[0080] S1. Obtain the environmental data and phenological period data of the rice growth throughout the year, and construct a phenological model data set;

[0081] Further, the specific implementation method of step S1 includes the following steps:

[0082] S1.1. Set that the environmental data and phenological period data obtained throughout the year of rice growth include the maximum temperature, minimum temperature, average daily temperature, daily precipitation, daily sunshine duration, daily humidity, and daily wind speed;

[0083] S1.2. Set that the entire growth cycle of rice includes 8 main phenological periods: the seedling stage is from sowing to tillering, the tillering stage is from tillering to jointing, the jointing stage is from jointing to booting, the booting stage is from booting to heading, the heading stage is from heading to milk ripening, the milk ripening stage is from milk ripening to dough ripening, the dough ripening stage is from dough ripening to full ripening, and the full ripening stage is from full ripening to harvesting;

[0084] S1.3. For the key time nodes of the set phenological periods, including 9 nodes: sowing, tillering, jointing, booting, heading, milk ripening, dough ripening, full ripening, and harvesting;

[0085] S1.4. Collect the environmental data and phenological data of rice growth throughout the year for each phenological period, mark the phenological node time by establishing a timestamp date_ori, match and mark the environmental data with the corresponding rice phenological periods, and construct a phenological model dataset.

[0086] S2. Preprocess the data according to the phenological model dataset for subsequent model training and prediction;

[0087] Furthermore, the specific implementation method of step S2 includes the following steps:

[0088] S2.1. Organize the data collected in step S1 into time series data in days and perform normalization processing on each data sample;

[0089] S2.2. Divide the annual environmental data according to the growth cycle of rice, and divide it into the seedling stage, tillering stage, jointing stage, booting stage, heading stage, milk ripening stage, dough ripening stage, and full ripening stage. Among them, the seedling stage is marked as 1, the tillering stage is marked as 2, the jointing stage is marked as 3, the booting stage is marked as 4, the heading stage is marked as 5, the milk ripening stage is marked as 6, the dough ripening stage is marked as 7, the full ripening stage is marked as 8, and other non-rice planting periods are marked as 0 for the training of the phenological period model; during the model training process, according to the phenological period marking of the annual data, train the prediction models of 9 phenological nodes.

[0090] S3. Build a rice phenological period prediction model based on the Encoder-Decoder architecture of the Seq2Seq model, introduce a Bayesian neural network for uncertainty assessment, and formulate a criterion for determining the update of the phenological model;

[0091] Furthermore, the specific implementation method of step S3 includes the following steps:

[0092] S3.1. Construct two bidirectional LSTMs as the Encoder of the model, and set the input feature X of the encoder as a three-dimensional tensor, where B represents the batch size, T represents the time series length of 20 days, and F represents the number of daily features, that is, 7 environmental features;

[0093] S3.2. First, use a fully connected layer to expand the input features, and set the weight matrix as Bias term The shape of the output tensor is batch_size, T, 256, and the expression of the output tensor X' is:

[0094] X′ = ReLU(XW1 + b1)

[0095] where ReLU is the activation function;

[0096] S3.3. Input the output into a two-layer stacked bidirectional long short-term memory network (Bi-LSTM) structure. The first layer of Bi-LSTM is used to process the input data, outputting forward and backward hidden states with dimensions of 128 respectively, and a total dimension of 256. The output shape is The expression is:

[0097]

[0098] Where, respectively represent the forward and backward hidden states at the first time step t in the Bi-LSTM. biLSTM1 represents the first layer of the bidirectional long short-term memory network, and X' t represents the input feature at time step t;

[0099] The second layer of Bi-LSTM receives the output of the first layer of Bi-LSTM to further process the data. The finally encoded time series features are output as The output shape is [batch_size, T, 256]. The expression is:

[0100]

[0101] Where, respectively represent the forward and backward hidden states of the second layer of Bi-LSTM at time step t. biLSTM2 represents the second layer of the bidirectional long short-term memory network, and H represents the output of the first layer of Bi-LSTM;

[0102] S3.4. The decoder (Decoder) receives the features output by the encoder and decodes them, decoding the encoded time series features into future phenological period prediction values. The decoder includes two layers of convolutional neural network (CNN) and a fully connected layer. The first layer of CNN uses a convolutional kernel to convert the input features from [batch_size, T, 16, 16] to [batch_size, T, 8, 8]. The expression is:

[0103] C1 = ReLU(Convld(H, W2))

[0104] Where, C1 represents the output of the first layer of convolution, ReLU represents the activation function, H represents the input features of the encoder, and W2 represents the weight matrix of the first layer of convolutional kernel;

[0105] The second layer of CNN uses a convolutional kernel to further convert the features to batch_size, T, 4, 4. The decoded features are output as The expression is:

[0106] C2 = ReLU(Convld(C1, W3))

[0107] S3.5. Flatten the output C2 of the decoder into a vector form to obtain a shape of batch_size, T×16. After flattening the output of the convolutional network into a two-dimensional vector, a fully connected layer is introduced for the final prediction process to predict 9 time nodes of the rice growth cycle;

[0108] S3.6. Use the Bayesian layer to introduce uncertainty assessment and calculate the confidence interval. Set the weight W4 and bias b4 as random variables that satisfy the probability distribution. The expression is:

[0109]

[0110] y = C2W4 + b4

[0111] where W4 and b4 are represented as random variables in the Bayesian layer, representing the weight and bias of the model respectively, μ represents the mean, σ 2 represents the variance, y represents the output of the Bayesian layer, and C2 represents the output features of the second convolutional layer in the decoder;

[0112] The output shape of the Bayesian layer is obtained as batch_size, 1, corresponding to the predicted value of each sample;

[0113] During the training process, modify the loss function to the Bayesian loss function containing uncertainty, and optimize it by maximizing the variational lower bound ELBO. The expression is:

[0114]

[0115] where q(w) is the estimated weight distribution, p(w) is the prior distribution of the weight, and D is the training data;

[0116] S3.7. When making predictions, use Monte Carlo sampling (MonteCarlo, MC) to run the model to generate multiple different predicted values

[0117] Obtain 100 predicted values by sampling 100 times to get the mean and standard deviation of the prediction; calculate the confidence interval of each predicted value through the standard deviation;

[0118] First, determine a confidence level of 95%, then calculate the corresponding z value, and then use the following formula to calculate the upper and lower bounds of each predicted value:

[0119] CI = μ ± z·σ

[0120] where μ is the mean of the predicted value, σ is the standard deviation, and z is the corresponding z value. For a 95% confidence interval, z is 1.96;

[0121] S3.8. Set judgment rules to ensure the accuracy of prediction and the judgment of whether to enter the next phenological period. When any two of the following conditions are met, stop updating data, end the prediction of the current phenological period, and start predicting the next phenological period:

[0122] S3.8.1. Stability of predicted values: Set as the predicted value on the t-th day. If the predicted values for k consecutive days fluctuate little, it indicates that the prediction of the model has tended to be stable. The specific judgment conditions are:

[0123]

[0124] where ∈1 is the set threshold, taken as 1 day, i.e., ∈1 = 1;

[0125] S3.8.2. Convergence of confidence intervals: Set if the width of the confidence interval CI w is less than the threshold ∈2 for k consecutive days, indicating that the prediction of the model is stable enough. The judgment formula is:

[0126] CI w (t) = CI upper (t) - CI lower (t)

[0127] CI w (t) ≤ ∈2 & CI w (t) - 1) ≤ ∈2 & CI w (t - 2) ≤ ∈2 &

[0128] where ∈2 is the threshold of the width of the confidence interval, set to 2 days;

[0129] S3.8.3. Proximity of actual date to predicted date: When the predicted phenological period is approaching, when the difference between the predicted date d pred and the current date d current is less than the set threshold, this rule is satisfied:

[0130] |d pred - d current | ≤ ∈3

[0131] where ∈3 is the set threshold, taken as 10 days.

[0132] S4. Train a rice phenological period prediction model based on the 8 phenological periods divided according to the rice growth cycle;

[0133] Furthermore, for the independent prediction model trained in step S4, the environmental data collected corresponds to the 8 predicted phenological periods, that is, 9 time nodes. Train for different time nodes and predict each phenological period respectively.

[0134] S5. Perform real-time prediction and update based on the daily updated environmental data;

[0135] Further, in step S5, first obtain the environmental data from January 1st to the current date, introduce the weather forecast data for the next few days, and at the same time adopt the sequence padding technology to add padding value 0 at the end of the input sequence to fill the input sequence to the fixed sequence length of 365 of the model;

[0136] During the prediction process, evaluate the uncertainty of the model output to determine the credibility of the prediction result; update the prediction result every day, and when the judgment criterion is met, determine the sowing time node; continue to predict, when the judgment criterion is met, determine the tillering time node, and then complete the prediction of the rice seedling stage and start the prediction of the tillering stage. And so on, gradually complete the prediction of subsequent jointing, booting, heading, milk ripening, dough ripening, full ripening and harvesting time nodes, and finally achieve a comprehensive prediction of the entire phenological period of rice.

[0137] Further, this embodiment aims to solve the problem that the current traditional methods are restricted by time, technology and region and cannot achieve inclusive and simplified services for the overall population. It adopts the Encoder-Decoder architecture of the Seq2Seq model, which can effectively process time series data with mismatched input and output lengths. For data with different lengths of rice growth cycles in different years and regions, first perform data preprocessing, input the environmental data into the Encoder, and then input the encoded data into the Decoder for prediction. At the same time, it combines the bidirectional long short-term memory network (Bi-LSTM) and the convolutional neural network (CNN). Bi-LSTM can effectively capture the long-term dependence relationships in the time series, while CNN improves the prediction accuracy by extracting local features from the environmental data. In addition, the Bayesian neural network (BNN) is introduced for uncertainty evaluation. BNN can evaluate the prediction credibility of the model by giving the confidence interval of the prediction result, and quantify the uncertainty of the prediction through the Monte Carlo sampling technology to ensure the stability and reliability of the prediction. This innovative phenological period prediction method does not need to rely on the traditional camera monitoring system, which not only reduces the cost, but also provides an efficient and scientific basis for guiding and recommending subsequent agricultural production tasks, promoting the development of intelligent agriculture.

[0138] It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without more limitations, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, article or device comprising the said element.

[0139] Although the present application has been described above with reference to specific embodiments, various improvements can be made thereto and components thereof can be replaced with equivalents without departing from the scope of the present application. In particular, as long as there is no structural conflict, the various features in the specific embodiments disclosed in the present application can be combined with each other in any way, and the reason for not exhaustively describing the situations of these combinations in this specification is only for the sake of saving space and resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An intelligent prediction method for rice phenological periods based on environmental data and deep learning, characterized in that, It includes the following steps: S1. Obtain the environmental data and phenological data of rice growth throughout the year, and construct a phenological model dataset; The specific implementation method of step S1 includes the following steps: S1.

1. Set that the environmental data and phenological data obtained for rice growth throughout the year include the maximum temperature, minimum temperature, average daily temperature, daily precipitation, daily sunshine duration, daily humidity, and daily wind speed; S1.

2. Set the entire growth cycle of rice to include 8 main phenological periods: the seedling stage is from sowing to tillering, the tillering stage is from tillering to jointing, the jointing stage is from jointing to booting, the booting stage is from booting to heading, the heading stage is from heading to milk ripening, the milk ripening stage is from milk ripening to dough ripening, the dough ripening stage is from dough ripening to full ripening, and the full ripening stage is from full ripening to harvesting; S1.

3. For the key time nodes of the set phenological periods, including 9 nodes: sowing, tillering, jointing, booting, heading, milk ripening, dough ripening, full ripening, and harvesting; S1.

4. Collect the environmental data and phenological data of rice growth throughout the year for each phenological period respectively. Mark the phenological node time by establishing a timestamp date_ori, match and mark the environmental data with the corresponding rice phenological periods, and construct a phenological model dataset; S2. Perform data preprocessing according to the phenological model dataset for subsequent model training and prediction; S3. Build a rice phenological period prediction model based on the Encoder-Decoder architecture of the Seq2Seq model, introduce a Bayesian neural network for uncertainty assessment, and formulate a criterion for judging the update of the phenological model; S4. Train the rice phenological period prediction model according to the 8 phenological periods divided by the rice growth cycle; S5. Perform real-time prediction and update based on the daily updated environmental data.

2. The intelligent prediction method for rice phenological period based on environmental data and deep learning according to claim 1, wherein, The specific implementation method of step S2 includes the following steps: S2.

1. Organize the data collected in step S1 into time series data in units of days, and perform normalization processing on each data sample; S2.

2. Divide the annual environmental data according to the rice growth cycle, and divide it into the seedling stage, tillering stage, jointing stage, booting stage, heading stage, milk ripening stage, dough ripening stage, and full ripening stage. Among them, the seedling stage is marked as 1, the tillering stage is marked as 2, the jointing stage is marked as 3, the booting stage is marked as 4, the heading stage is marked as 5, the milk ripening stage is marked as 6, the dough ripening stage is marked as 7, the full ripening stage is marked as 8, and other non-rice planting periods are marked as 0 for the training of the phenological period model; during the model training process, train the prediction models of 9 phenological nodes according to the phenological period markings of the annual data.

3. An intelligent prediction method for rice phenological periods based on environmental data and deep learning according to claim 2, characterized in that, The specific implementation method of step S3 includes the following steps: S3.

1. Construct an Encoder using two bidirectional LSTMs as the model, and set the input feature X of the encoder as a three-dimensional tensor, where B represents the batch size, T represents the time series length of 20 days, and F represents the number of daily features, i.e., 7 environmental features; S3.

2. First, use the fully connected layer to expand the input features, and set the weight matrix as bias term The shape of the output tensor is batch_size, T, 256, and the expression of the output tensor X′ is: x′ = ReLU(XW1 + b1) where ReLU is the activation function; S3.

3. Input the output into a two-layer stacked bidirectional long short-term memory network (Bi-LSTM) structure. The first layer of Bi-LSTM is used to process the input data, outputting forward and backward hidden states with dimensions of 128 respectively, and the total dimension is 256. The output shape is The expression is: Among them, respectively represent the forward and backward hidden states at the first time step t in the Bi-LSTM, biLSTM1 represents the first layer of bidirectional long short-term memory network, and X′ t represents the input feature at time t; The second - layer bidirectional LSTM receives the output of the first - layer bidirectional LSTM to further process the data, and finally the encoded time - series features are output as The output shape is [batch_size, T, 256], and the expression is: Among them, respectively represent the forward and backward hidden states of the second-layer Bi-LSTM at time t, biLSTM2 represents the second-layer bidirectional long short-term memory network, and H represents the output of the first-layer bidirectional LSTM; S3.

4. The decoder receives the features output by the encoder and decodes them, decoding the encoded temporal features into future phenological prediction values. The decoder consists of two layers of convolutional neural network (CNN) and a fully connected layer. The first layer of CNN uses a convolutional kernel to convert the input features from [batch_size, T, 16, 16] to [batch_size, T, 8, 8]. The expression is: C1 = ReLU(Convld(H, W2)) where C1 represents the output of the first layer of convolution, ReLU represents the activation function, H represents the input features of the encoder, and W2 represents the weight matrix of the first layer of convolution kernel; The second-layer CNN uses convolutional kernels Further transform the features into batch_size, T, 4, 4, and the decoded feature output is The expression is: C2 = ReLU(Convld(C1, W3)) S3.

5. Flatten the output C2 of the decoder into a vector form to obtain a shape of batch_size, T×16. After flattening the output of the convolutional network into a two-dimensional vector, a fully connected layer is introduced for the final prediction process to predict 9 time nodes of the rice growth cycle. S3.

6. Use a Bayesian layer to introduce uncertainty assessment and calculate the confidence interval. Set the weight W4 and bias b4 as random variables that satisfy a probability distribution, and the expression is: y = C2W4 + b4 Among them, W4 and b4 are represented as random variables in the Bayesian layer, representing the weight and bias of the model respectively, μ represents the mean, σ 2 represents the variance, y represents the output of the Bayesian layer, and C2 represents the output features of the second convolutional layer in the decoder; Obtain the output shape of the Bayesian layer as batch_size, 1, corresponding to the predicted value of each sample. During the training process, modify the loss function to a Bayesian loss function that includes uncertainty, and optimize it by maximizing the variational lower bound ELBO. The expression is: where q(w) is the estimated weight distribution, p(w) is the prior distribution of the weights, and D is the training data. S3.

7. When making predictions, use Monte Carlo sampling (MC) to run the model and generate multiple different predicted values Obtain 100 predicted values by sampling 100 times to get the mean and standard deviation of the predictions; calculate the confidence interval for each predicted value through the standard deviation. First, determine a confidence level of 95%, then calculate the corresponding z value, and then use the following formula to calculate the upper and lower bounds of each predicted value: CI = μ ± z·σ where μ is the mean of the predicted value, σ is the standard deviation, and z is the corresponding z value. For a 95% confidence interval, z is 1.

96. S3.

8. Set a decision rule to ensure the accuracy of the prediction and the determination of whether to proceed to the next phenological period. When any two of the following conditions are met, stop updating the data, end the current phenological period prediction, and start the next phenological period prediction: S3.8.

1. Stability of predicted values: Set as the predicted value for the t-th day. If the predicted values for k consecutive days fluctuate slightly, it indicates that the prediction of the model has tended to be stable. The specific determination condition is as follows: where ∈1 is the set threshold, taken as 1 day, i.e., ∈1 = 1; S3.8.

2. Confidence Interval Convergence: Set that if the confidence interval width CI w is less than the threshold ∈2 for k consecutive days, it indicates that the prediction of the model is stable enough, and the judgment formula is: CI w CI(t) = CI upper CI(t) - CI lower CI(t) CI w CI(t) ≤ ∈2& w CI(t - 1) ≤ ∈2& w CI(t - 2) ≤ ∈2& where ∈2 is the threshold for the width of the confidence interval, set to 2 days; S3.8.

3. Proximity between the actual date and the predicted date: When the predicted phenological period is approaching, if the difference between the predicted date d pred and the current date d current is less than the set threshold, then this rule is satisfied: |d pred -d current |≤∈3 where ∈3 is the set threshold, taken as 10 days.

4. The intelligent prediction method for rice phenological period based on environmental data and deep learning according to claim 3, characterized in that, For the training of the independent prediction model described in step S4, the collected environmental data corresponds to 8 phenological periods of the prediction, i.e., 9 time nodes, and training is carried out for different time nodes to predict each phenological period respectively.

5. The intelligent prediction method for rice phenological periods based on environmental data and deep learning according to claim 4, characterized in that Step S5 first obtains the environmental data from January 1st to the current date, and introduces the weather forecast data for the next few days. At the same time, the sequence padding Padding technology is adopted to add a padding value of 0 at the end of the input sequence to pad the input sequence to the fixed sequence length of the model, which is 365. During the prediction process, conduct uncertainty assessment on the model output to determine the credibility of the prediction result; update the prediction result every day. When the decision criteria are met, determine the sowing time node; continue to predict. When the decision criteria are met, determine the tillering time node, and then complete the prediction of the rice seedling stage and start the prediction of the tillering stage, and so on, gradually complete the prediction of the subsequent jointing, booting, heading, milk ripening, dough ripening, full ripening, and harvesting time nodes, and finally achieve a comprehensive prediction of the entire phenological period of the rice.

Images