Method and system for processing accumulated temperature and phenological period information based on the Internet of Things

Abstract

The present application discloses a method and system for processing accumulated temperature and phenological period information based on the Internet of Things. The method comprises: training an initialized prediction model based on a first training data set to obtain a first model, wherein the first training data set is training data for a first crop; determining a characteristic difference between the first crop and a second crop, and determining a parameter type corresponding to the characteristic difference; determining an adjustment direction of the parameter type according to the characteristic difference and the parameter type; mapping the parameter type that needs to be adjusted in the first model to the weights of at least part of the parameters in the network, and adjusting them according to the adjustment direction to obtain a second model; and training the second model using a second training data set until a training condition is met, wherein the second training data set is training data for a second crop.

Description

Technical Field

[0001] The present application relates to agricultural Internet of Things and artificial intelligence technologies, and in particular to an Internet of Things-based method and system for processing accumulated temperature and phenological period information. Background Art

[0002] In the agricultural Internet of Things system, there are many detectable variables, and the accumulation of these variables will have a considerable impact on the target results, such as in the phenological period prediction task. The traditional estimation model, which estimates through some main parameters, has indeed achieved certain results and has advantages such as simple calculation models. However, in fact, there are many factors that affect the target results. Therefore, if it is necessary to further explore deep hidden information, a more complex deep neural network solution is needed. The inventor believes that although the above model has made some progress in increasing the types of input data, it also ignores the impact of the fluctuation characteristics of some important cumulative factors in the accumulation process on the results.

[0003] At the same time, for model training, the similarity of predictions for the same type of targets on the same task is low if the corresponding model for each type is trained from scratch. Summary of the invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention provides a method and system for processing accumulated temperature and phenological period information based on the Internet of Things.

[0005] On the one hand, the embodiment of the present application provides a method for processing accumulated temperature and phenological period information based on the Internet of Things, including:

[0006] Constructing a prediction model for crop phenological period, the prediction model comprising at least one temporal feature extraction subnetwork for extracting temporal features of key parameters and a global feature extraction subnetwork for extracting global features, and a classification network for classifying fusion features of temporal features and global features, wherein the key parameters at least include effective accumulated temperature information;

[0007] Training the initialized prediction model based on a first training data set to obtain a first model, wherein the first training data set is training data for a first crop;

[0008] Determine a characteristic difference between the first crop and the second crop, and determine a parameter type corresponding to the characteristic difference;

[0009] Determining an adjustment direction of the parameter type according to the characteristic difference and the parameter type;

[0010] Mapping the parameter types that need to be adjusted in the first model to weights of at least some parameters in the network, and adjusting them according to the adjustment direction to obtain a second model;

[0011] The second model is trained using a second training data set until a training condition is met, wherein the second training data set is training data for a second crop.

[0012] In some embodiments, the parameters of the prediction model include at least multiple information of: planting time, effective accumulated temperature, land parameters, fertilization parameters, precipitation, sunshine time, remote sensing data, and pest parameters;

[0013] Among them, the important parameters include at least effective accumulated temperature, and the number of the important parameters is greater than 1.

[0014] In some embodiments, the land parcels are divided for prediction, and for important parameters, time series data of the important parameters are obtained according to a first sensor in the area to which the land parcels belong;

[0015] If the first sensor does not exist in the area to which the block to be predicted belongs, the time series data of important parameters in the block to be predicted are fitted based on the data collected by the second sensor and the data changes of the first sensor belonging to the surrounding blocks.

[0016] In some embodiments, the contribution weight of each first sensor to the time series data in the block to be predicted is determined according to the terrain information.

[0017] In some embodiments, the contribution weight of each first sensor to the time series data in the block to be predicted is determined according to the terrain information, specifically:

[0018] According to the terrain information, a terrain influence factor between the plot to be predicted and the first sensor is determined, a terrain influence index is determined according to the terrain influence factor, and contribution weight information is determined by multiplying the terrain influence index by the inverse of the distance between the center of the plot to be predicted and the first sensor.

[0019] In some embodiments, the fitting of the time series data of the important parameters in the plot to be predicted based on the data change of the first sensor belonging to the surrounding plots is specifically:

[0020] The data changes of the first sensors of several surrounding plots are selected, and the contribution ratio of the predicted plot in the fitting is determined according to the contribution weight information of each first sensor. The change trend is fitted according to the contribution ratio of each first sensor, and the time series data of the important parameters are fitted through the change trend combined with the data points of the important parameters collected by the second sensor.

[0021] In some embodiments, when training the prediction model, the temporal feature extraction subnetwork connected to the output layer is first pre-trained, and the pre-trained temporal feature extraction subnetwork is removed from the output layer and then added to the network for training together.

[0022] In some embodiments, phenological period prediction is performed based on the first model or the second model.

[0023] On the other hand, the embodiment of the present application provides an accumulated temperature and phenological period information processing system based on the Internet of Things, including:

[0024] Memory, used to store programs;

[0025] A processor is used to load the program to execute the accumulated temperature and phenological period information processing method in the Internet of Things.

[0026] On the other hand, the embodiment of the present application provides an accumulated temperature and phenological period information processing system based on the Internet of Things, including:

[0027] A training system, the training system is used to execute the accumulated temperature and phenological period information processing method based on the Internet of Things;

[0028] The prediction system performs prediction based on the model trained by the training system.

[0029] Through the network structure set by the model of the embodiment of the present application, at least one time series feature extraction subnetwork for extracting time series features of key parameters and a global feature extraction subnetwork for extracting global features are set in parallel, wherein the time series feature extraction subnetwork is used to extract time series features of key parameters, such as accumulated temperature and other important influencing parameters on crop growth; so that the model can learn more information, and then fuse the time series features and global features of important parameters, so that the model can dig out the hidden features of important parameter changes and achieve more accurate prediction; in addition, in addition to the improvement of the model structure, the present application also proposes a training method for crop models of the same type. Through two relatively similar crops, according to the differences in their characteristics, parameters are adjusted on the basis of the existing model before training, which can speed up the training of similar models and reduce the resources required for training. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments.

[0031] Figure 1 is a flow chart of the method provided in the embodiment of the present application;

[0032] Figure 2 It is a schematic diagram of the model structure provided in the embodiment of the present application;

[0033] Figure 3 It is a module block diagram of the system provided in the embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described through the implementation method with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0035] DNN (Deep Neural Network) is an artificial neural network architecture whose core feature is that the network contains multiple hidden layers, which makes the model highly nonlinear and capable of learning complex data patterns. DNN is a powerful and widely used tool in the field of machine learning, especially deep learning, and is particularly suitable for processing complex tasks such as image classification, speech recognition, natural language processing, and recommendation systems.

[0036] LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN) architecture that solves the gradient vanishing or gradient exploding problems that are common in traditional RNNs when processing long sequence data. Therefore, it is very suitable for processing and predicting important events with very long intervals and delays in time series. LSTM controls the flow of information by introducing a "gate" mechanism, so that it can learn long-term dependencies. Therefore, time series feature information can be extracted through the LSTM network.

[0037] Reference Figure 1 and Figure 2 The embodiment of the present application provides a method for processing accumulated temperature and phenological period information based on the Internet of Things. It can be understood that the scheme can be implemented by a computer system. The scheme can construct and train a phenological period prediction model to complete the prediction of related tasks, including:

[0038] S1. Construct a prediction model for crop phenological period, wherein the prediction model includes at least one temporal feature extraction subnetwork for extracting temporal features of key parameters and a global feature extraction subnetwork for extracting global features, and a classification network for classifying fused features of temporal features and global features, wherein the key parameters include at least effective accumulated temperature information.

[0039] It can be understood that only key parameters, such as accumulated temperature, precipitation, nutritional changes and other parameters that will obviously affect crop growth, need to extract time series features, while other parameters mainly appear in the current state or the accumulated state, and the extraction task is completed through the global feature extraction sub-network. The two aspects of the features are fused through the fusion layer to form a fusion feature, in which the fusion layer splices the time series features with the global features according to a certain weight ratio. In this embodiment, a time series feature sub-network of three important parameters is set. The features are then classified through a classification network composed of a deep neural network, and after output through the output layer, the phenological period prediction results are obtained. Through this solution, the model can learn the time series information of key parameters.

[0040] The parameters of the prediction model include: at least multiple information of planting time, effective accumulated temperature, land parameters, fertilization parameters, precipitation, sunshine time, remote sensing data, pest parameters, and humidity information. Among them, remote sensing data includes NDVI (normalized vegetation index), EVI (enhanced vegetation index), NDWI (normalized water index), NIR (near infrared light), etc.

[0041] S2. Training the initialized prediction model based on a first training data set to obtain a first model, wherein the first training data set is training data for a first crop.

[0042] It is understandable that the constructed prediction model can be trained using the first training data set, and the training data set can use species with more representative characteristics among the same type of species as basic training data, or use species with a relatively large amount of data as basic training data. First, a first model whose prediction ability can meet the design requirements is trained.

[0043] In some embodiments, when training the prediction model, the temporal feature extraction subnetwork connected to the output layer is first pre-trained, and the pre-trained temporal feature extraction subnetwork is removed from the output layer and added to the network for training. It is understandable that the data in some training samples can be separated to extract the influence of the temporal features of important parameters on the results. By splitting the local training, the generalization of the model can be improved and the training efficiency can be improved.

[0044] S3. Determine a characteristic difference between the first crop and the second crop, and determine a parameter type corresponding to the characteristic difference.

[0045] In this step, the differences in characteristics between two different varieties are determined based on experience and qualitative analysis. For example, some varieties are less sensitive to precipitation, so it can be determined that the influence of precipitation on the variety is actually relatively low.

[0046] S4. Determine an adjustment direction of the parameter type according to the characteristic difference and the parameter type.

[0047] Continuing with the above example, for the new model, the parameter type corresponding to the characteristic difference can be determined in the model input parameter type. If the difference characteristic indicates that the new variety is more sensitive to the parameter, it is adjusted in the direction of increasing the weight. If it is not sensitive to the parameter, it is adjusted in the direction of reducing the weight.

[0048] S5. Map the parameter types that need to be adjusted in the first model to the weights of at least some parameters in the network, and adjust them according to the adjustment direction to obtain a second model.

[0049] Therefore, this parameter can be lowered to correspond to the input weight or branch weight in the network. Through such manual adjustment, the parameters that need to be adjusted can be adjusted quickly, so that the model can converge relatively quickly in the training data set of the variety, thereby achieving the effect of reducing training resources. This solution can determine the characteristic differences between varieties through qualitative tests, and relying on traditional test methods, it can train the model of the variety when the data of the new variety is insufficient. The proportion of manual adjustment can be adjusted between 10% and 30% according to the size of the characteristic difference.

[0050] For example, the second product is significantly less sensitive to moisture than the first product. Therefore, the weight of precipitation and precipitation time series related to moisture can be reduced to reduce the impact of this correlation on the prediction results. In this way, the impact of related parameters on the model itself can be changed quickly, the changes in product difference characteristics can be reflected in the model, and the speed of model convergence during secondary training can be accelerated, thereby achieving the purpose of reducing the resources required for training.

[0051] S6. Use a second training data set to train the second model until a training condition is met, wherein the second training data set is training data for a second crop.

[0052] It is understandable that, through the adjustment in the previous step, the difference feature is adjusted to a certain extent in the right direction, which can reduce some of the training resources required to correct the difference parameter. However, it is understandable that the above adjustment may not be accurate, so it is necessary to use the data set of the second crop to perform parameter tuning training on the model.

[0053] S7. Predict phenological period based on the first model or the second model.

[0054] Reference Figure 2, in the prediction process, the time series features of the key parameters are processed according to the current time and the window of the preset width for sliding processing, and the time series features of the key parameters of the input model are obtained; the current cumulative features and other fixed features are composed of input vectors, and the input vectors are mapped to the embedding layer (feature embedding), and then sent to the DNN sub-network to extract the global parameter features; after the time series features are fused with the global parameter features, they are sent to the constructed classification network, and the results are output through the output network. In this embodiment, the DNN sub-network is a deep neural network with 128 dimensions as input and 3 hidden layers. The classification network can be composed of a residual neural network of ResNet32. Of course, in theory, other deep neural networks can also be used to construct it. This application does not limit the classification network itself, and the network that can achieve equivalent results is also the scope of protection claimed by this application. The LSTM network can be constructed using ordinary LSTM or Bi-LSTM, and the final output layer outputs to the day interval of a specific phenological period.

[0055] In some embodiments, the dimension of prediction is at the land parcel level. In these embodiments, the land parcels are divided for prediction. For important parameters, the time series data of the important parameters are obtained according to the first sensor in the area to which the land parcel belongs. It needs to be explained that the first sensor refers to a sensor fixedly installed at the collection position. Since it is fixed, the collected data is relatively comprehensive. Data such as accumulated temperature can be recorded by the first sensor at each recording node in the whole process. However, in fact, due to the limited number of first sensors, the distribution interval may be relatively large. Therefore, this solution also uses some second sensors for collection, such as remote sensing, such as drones, etc. to collect data. However, the degree of data collection cannot be as frequent as the first sensor. Of course, the second sensor also has a broader meaning. For example, the second sensor can also refer to a sensor with some data loss. In data collection, data loss is more common. Therefore, for data integrity, the data needs to be filled. Therefore, for some important parameters, the complete time series data can be fitted based on the combination of the first sensor and the second sensor. For some parameters that need to form a time series, data fitting can also be performed in a traditional way according to the characteristics of the parameters. The solution provided in this embodiment is only applicable to some parameters, such as accumulated temperature information.

[0056] If the first sensor does not exist in the area to which the block to be predicted belongs, the time series data of important parameters in the block to be predicted are fitted based on the data collected by the second sensor and the data changes of the first sensor belonging to the surrounding blocks.

[0057] The contribution weight of each first sensor to the time series data in the block to be predicted is determined according to the terrain information. Specifically:

[0058] According to the terrain information, a terrain influence factor between the plot to be predicted and the first sensor is determined, a terrain influence index is determined according to the terrain influence factor, and contribution weight information is determined by multiplying the terrain influence index by the inverse of the distance between the center of the plot to be predicted and the first sensor.

[0059] In some embodiments, the fitting of the time series data of the important parameters in the plot to be predicted based on the data change of the first sensor belonging to the surrounding plots is specifically:

[0060] The data changes of the first sensors of several surrounding plots are selected, and the contribution ratio of the predicted plot in the fitting is determined according to the contribution weight information of each first sensor. The change trend is fitted according to the contribution ratio of each first sensor, and the time series data of the important parameters are fitted through the change trend combined with the data points of the important parameters collected by the second sensor.

[0061] It is understandable that when it is desired to determine the sensor data collected at a certain location, the closer the first sensor is, the more reference value it has. The terrain between the target location and the first sensor will affect the reference value of the first sensor. For the initial combination of certain factors, the reference value will be reduced. Therefore, in addition to considering the distance factor, the terrain factor also needs to be considered. In this way, the reference value of the change trend of several first sensors that are relatively close to the target location to the change trend of the target location is determined. Without considering the terrain difference, the simplest solution is to average the data of several surrounding first sensors, but after referring to the terrain, the reference value of different locations is different. Therefore, in this solution, the contribution weight is described as Wherein, D represents the distance, and xi represents the influencing factors that exist. It is understandable that the idea of ​​the present application is to find several first sensors that can provide references within a certain range. Then the degree to which the first sensor can be used as a reference is judged based on the terrain information and the distance. Assuming that there is no difference between the terrains, the changes in important parameters such as accumulated temperature are close to the average of several surrounding first sensors. For some terrain factors, the influence of the first sensor on the target position will be affected. It is understandable that the influencing factors include the following: there is a significant height difference between the first sensor and the target position, there is water in the middle, there are obvious hills or basins and other negative factors.

[0062] Fitting the change trend, two data endpoints are known, and the data change situation between the data endpoints needs to be fitted, referring to the change trend of several nearby data collection points. For example, three reference points are selected, and the change trends of the three reference points are basically the same, but there are differences in amplitude. Assume that compared with the first data point, the change rate of the first reference point is 3%, the change rate of the second reference point is 3.5%, and the change rate of the third reference point is 4%. Among them, the contribution weights of the three are a, b, and c respectively. When fitting the change of the target point, the calculation method is (a*3%+b*3.5%+c*4%) / (a+b+c). It should be noted that the contribution ratio is determined by dividing the respective contribution weights by the sum of the contribution weights.

[0063] Through this solution, when the first sensor is widely distributed, the data collected by the second sensor can be used, and then the missing data in the middle can be fitted with reference to the change trend of the surrounding first sensors to fill the missing data.

[0064] The embodiment of the present application provides an accumulated temperature and phenological period information processing system based on the Internet of Things, including:

[0065] Memory, used to store programs;

[0066] A processor is used to load the program to execute the accumulated temperature and phenological period information processing method in the Internet of Things.

[0067] The method includes:

[0068] Constructing a prediction model for crop phenological period, the prediction model comprising at least one temporal feature extraction subnetwork for extracting temporal features of key parameters and a global feature extraction subnetwork for extracting global features, and a classification network for classifying fusion features of temporal features and global features, wherein the key parameters at least include effective accumulated temperature information;

[0069] Training the initialized prediction model based on a first training data set to obtain a first model, wherein the first training data set is training data for a first crop;

[0070] Determine a characteristic difference between the first crop and the second crop, and determine a parameter type corresponding to the characteristic difference;

[0071] Determining an adjustment direction of the parameter type according to the characteristic difference and the parameter type;

[0072] Mapping the parameter types that need to be adjusted in the first model to weights of at least some parameters in the network, and adjusting them according to the adjustment direction to obtain a second model;

[0073] The second model is trained using a second training data set until a training condition is met, wherein the second training data set is training data for a second crop.

[0074] Reference Figure 3 The embodiment of the present application provides an accumulated temperature and phenological period information processing system based on the Internet of Things, including:

[0075] A training system, wherein the training system is used to execute the accumulated temperature and phenological period information processing method based on the Internet of Things.

[0076] The prediction system performs prediction based on the model trained by the training system.

[0077] It can be understood that, in the above-mentioned product solutions, corresponding technical effects are achieved by executing the above-mentioned method embodiments.

[0078] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0079] Note that the above are only preferred embodiments of the present application and the technical principles used. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present application. Therefore, although the present application is described in more detail through the above embodiments, the present application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the appended claims.

Claims

1. A method for processing accumulated temperature and phenological period information based on the Internet of Things, characterized in that: include: Constructing a prediction model for crop phenological period, the prediction model comprising at least one temporal feature extraction subnetwork for extracting temporal features of key parameters and a global feature extraction subnetwork for extracting global features, and a classification network for classifying fusion features of temporal features and global features, wherein the key parameters at least include effective accumulated temperature information; Training the initialized prediction model based on a first training data set to obtain a first model, wherein the first training data set is training data for a first crop; Determine a characteristic difference between the first crop and the second crop, and determine a parameter type corresponding to the characteristic difference; Determining an adjustment direction of the parameter type according to the characteristic difference and the parameter type; Mapping the parameter types that need to be adjusted in the first model to weights of at least some parameters in the network, and adjusting them according to the adjustment direction to obtain a second model; The second model is trained using a second training data set until a training condition is met, wherein the second training data set is training data for a second crop.

2. The method for processing accumulated temperature and phenological period information based on the Internet of Things according to claim 1 is characterized in that: The parameters of the prediction model include: at least multiple information of planting time, effective accumulated temperature, land parameters, fertilization parameters, precipitation, sunshine time, remote sensing data, and pest parameters; Among them, the key parameters include at least effective accumulated temperature, and the number of the important parameters is greater than 1.

3. The method for processing accumulated temperature and phenological period information based on the Internet of Things according to claim 1 is characterized in that: Divide the land into plots for prediction, and for important parameters, obtain time series data of the important parameters according to a first sensor in an area of ​​the plot; If the first sensor does not exist in the area to which the block to be predicted belongs, the time series data of important parameters in the block to be predicted are fitted based on the data collected by the second sensor and the data changes of the first sensor belonging to the surrounding blocks.

4. The method for processing accumulated temperature and phenological period information based on the Internet of Things according to claim 3 is characterized in that: The contribution weight of each first sensor to the time series data in the land block to be predicted is determined according to the terrain information.

5. The method for processing accumulated temperature and phenological period information based on the Internet of Things according to claim 4 is characterized in that: Determine the contribution weight of each first sensor to the time series data in the block to be predicted according to the terrain information, specifically: According to the terrain information, a terrain influence factor between the plot to be predicted and the first sensor is determined, a terrain influence index is determined according to the terrain influence factor, and contribution weight information is determined by multiplying the terrain influence index by the inverse of the distance between the center of the plot to be predicted and the first sensor.

6. The method for processing accumulated temperature and phenological period information based on the Internet of Things according to claim 5 is characterized in that: The fitting of the time series data of the important parameters in the plot to be predicted based on the data change of the first sensor belonging to the surrounding plots is specifically: The data changes of the first sensors of several surrounding plots are selected, and the contribution ratio of the predicted plot in the fitting is determined according to the contribution weight information of each first sensor. The change trend is fitted according to the contribution ratio of each first sensor, and the time series data of the important parameters are fitted through the change trend combined with the data points of the important parameters collected by the second sensor.

7. The method for processing accumulated temperature and phenological period information based on the Internet of Things according to claim 1 is characterized in that: When training the prediction model, the time series feature extraction sub-network connected to the output layer is first pre-trained, and the pre-trained time series feature extraction sub-network is removed from the output layer and added to the network for training together.

8. The method for processing accumulated temperature and phenological period information based on the Internet of Things according to claim 1 is characterized in that: The phenological period is predicted based on the first model or the second model.

9. An accumulated temperature and phenological period information processing system based on the Internet of Things, characterized in that: include: Memory, used to store programs; A processor is used to load the program to execute the method for processing accumulated temperature and phenological period information in the Internet of Things as described in any one of claims 1 to 7.

10. An accumulated temperature and phenological period information processing system based on the Internet of Things, characterized in that: comprising a training system for executing the method according to any one of claims 1 to 7; The prediction system performs prediction based on the model trained by the training system.

Images