An early warning monitoring system for smart agriculture

Abstract

The present invention provides an early warning and monitoring system for smart agriculture, which belongs to the field of smart agriculture information processing technology. It includes a data acquisition module for collecting high-frequency visual data and low-frequency environmental data, and performing normalization processing to generate normalized data; a first processing module for determining the volatility of each data source based on the normalized data; a second processing module for calculating the dynamic weight coefficient of each data source according to the volatility and a preset basic influence coefficient; a third processing module for performing weighted fusion based on the dynamic weight coefficient and the normalized data through a preset feature mapping function to generate a farmland system state vector. The present invention constructs an effective mapping of heterogeneous data to a unified state space; for high-frequency visual data and low-frequency environmental data, two data types with very different physical modes and information density, the system uses a special feature mapping function for processing.

Description

Technical Field

[0001] The present invention relates to the technical field of smart agriculture information processing, and in particular to an early warning monitoring system for smart agriculture. Background Art

[0002] Existing smart agricultural monitoring systems face challenges in fusing and processing multi-source heterogeneous data. Typically, the system needs to integrate high-temporal-resolution visual data with low-temporal-resolution environmental sensor data. Current technical solutions often use fixed weights for data fusion. This approach cannot effectively address the dynamic changes in the timeliness and importance of different data sources, resulting in lags or misjudgments in the comprehensive assessment results generated by the system, making it difficult to accurately reflect the true state and evolution of farmland ecosystems. Therefore, existing system functions are mainly limited to discrete event-based threshold alarms, a passive response mechanism that lacks continuous and predictive understanding of the overall equilibrium state of the farmland system.

[0003] The purpose of this invention is to provide a technical solution, the core of which lies in an adaptive multimodal data fusion algorithm. This algorithm dynamically calculates the weight coefficients of each data source in a comprehensive assessment model based on the inherent volatility of the data source and preset agricultural influencing factors. This algorithm effectively fuses high-frequency visual data with low-frequency environmental data into a unified and predictive farmland system state vector. This core mechanism enables the system's functionality to be upgraded from discrete event-based threshold alarms to continuous regulation and guidance for systemic equilibrium in the agricultural environment.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0005] The purpose of the present invention is to provide an early warning monitoring system for smart agriculture to solve the problems raised in the above background technology.

[0006] The technical solution of the present invention comprises: a data acquisition module for collecting high-frequency visual data and low-frequency environmental data, and performing normalization processing to generate normalized data;

[0007] A first processing module is configured to determine the volatility of each data source based on the normalized data;

[0008] The second processing module is used to calculate the dynamic weight coefficient of each data source based on volatility and a preset basic influence coefficient;

[0009] The third processing module is used to generate a farmland system state vector by performing weighted fusion based on the dynamic weight coefficient and the normalized data through a preset feature mapping function;

[0010] The fourth processing module is used to predict the future state vector based on the time series of the farmland system state vector, and to compare the future state vector with a preset ideal equilibrium state to determine whether to generate a control instruction.

[0011] Preferably, the first processing module is used to determine volatility, including:

[0012] The volatility is determined by the finite difference method based on the numerical changes and time intervals of the normalized data at adjacent sampling moments.

[0013] Preferably, the second processing module is used to calculate the dynamic weight coefficient, including:

[0014] The instantaneous activation of each data source is determined by combining volatility, basic influence coefficient and preset volatility sensitivity coefficient. The Softmax function is used to normalize the instantaneous activation of all data sources to generate a dynamic weight coefficient.

[0015] Preferably, the feature mapping function in the third processing module includes:

[0016] Convolutional neural networks for processing high-frequency visual data;

[0017] Multilayer perceptron for processing low-frequency environmental data.

[0018] Preferably, the fourth processing module is used to predict the future state vector, including:

[0019] A recurrent neural network model is used to determine the future state vector based on the current and historical farmland system state vectors and control input vectors.

[0020] Preferably, the recurrent neural network model is obtained by performing deep learning training on historical state vector sequences and corresponding control input sequences.

[0021] Preferably, the fourth processing module is used to determine whether to generate a control instruction, including:

[0022] Calculate the vector norm deviation between the future state vector and the ideal equilibrium state;

[0023] When the vector norm deviation exceeds a preset threshold, a control instruction is generated to minimize the future state deviation;

[0024] When the vector norm deviation does not exceed the preset threshold, the current control strategy is maintained.

[0025] Preferably, the ideal equilibrium state is predefined by an agricultural expert knowledge base according to the crop growth stage, or is determined by analyzing the mean value of the state vector associated with the historical highest yield.

[0026] The present invention provides an early warning monitoring system for smart agriculture through improvement. Compared with the existing technology, it has the following improvements and advantages:

[0027] 1. This invention implements dynamic, adaptive assessment of data influence. Rather than employing a fixed weighting scheme, the system uses the finite difference method to determine the real-time volatility of each data source based on the change in its value between adjacent sampling moments. Based on this volatility, a basic influence coefficient derived from prior agricultural knowledge, and an optimizable fluctuation sensitivity coefficient, a dynamic weight coefficient is calculated for each data source. This enables the system to instantly and quantitatively amplify the influence of currently fluctuating data sources, thereby more accurately capturing key dynamic changes in farmland ecosystems and improving the accuracy and sensitivity of state perception.

[0028] 2. This invention establishes an effective mapping of heterogeneous data into a unified state space. The system utilizes specialized feature mapping functions to process high-frequency visual data and low-frequency environmental data, two data types with vastly different physical modalities and information densities. It also utilizes convolutional neural networks to extract deep spatial semantic features from visual data, while simultaneously employing multi-layer perceptrons to fit nonlinear relationships between environmental data. This approach transforms raw data from diverse sources into feature vectors of the same dimension and semantically comparable, providing the technical prerequisite for subsequent meaningful weighted fusion to generate a farmland system state vector, thus resolving the challenge of heterogeneous data fusion.

[0029] 3. This invention introduces a data-driven adaptive system evolution prediction model. The prediction of the future evolution of the farmland system state vector is not based on fixed physical or biochemical models, but rather on a data-driven nonlinear dynamic model such as a recurrent neural network. The model's key parameters, namely the state transition matrix and the control input matrix, are obtained through systematic identification and learning of historical state vectors and control input sequences. This approach enables the prediction model to autonomously learn the inherent dynamic characteristics of a specific farmland from historical data, thereby achieving more accurate predictions of future state vectors and possessing a high degree of environmental specificity and adaptability.

[0030] 4. The present invention establishes a forward-looking and optimized control decision-making mechanism; the system's control decision is not a passive response to the current state, but is based on the prediction of the future state vector; by calculating the vector norm deviation between the predicted state and the ideal equilibrium state and comparing it with a preset threshold, it is decided whether to generate a control instruction; the ideal equilibrium state itself is dynamic and can be defined by an expert knowledge base or learned from historical high-yield data; this mechanism realizes the transition from passive response to active prediction, and aims to minimize future state deviations, making the control behavior more forward-looking, accurate and cost-effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention will be further explained below in conjunction with the accompanying drawings and examples:

[0032] Figure 1 This is a flow chart of an early warning monitoring system for smart agriculture according to the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0034] Example 1:

[0035] See also Figure 1 ,The present invention provides an early warning monitoring system for smart agriculture, comprising: a data acquisition module for collecting high-frequency visual data and low-frequency environmental data, and performing normalization processing to generate normalized data;

[0036] A first processing module is configured to determine the volatility of each data source based on the normalized data;

[0037] The second processing module is used to calculate the dynamic weight coefficient of each data source based on volatility and a preset basic influence coefficient;

[0038] The third processing module is used to generate a farmland system state vector by performing weighted fusion based on the dynamic weight coefficient and the normalized data through a preset feature mapping function;

[0039] a fourth processing module, configured to predict a future state vector based on the time series of the farmland system state vector, and to compare the future state vector with a preset ideal equilibrium state to determine whether to generate a control instruction;

[0040] The present embodiment discloses an early warning and monitoring system for smart agriculture. The workflow begins with a data acquisition module collecting high-frequency visual data and low-frequency environmental data, and processing them using a maximum-minimum normalization method to generate dimensionless normalized data. Based on this normalized data, a first processing module quantitatively assesses the volatility of each data source. A second processing module calculates a dynamic weight coefficient for each data source based on this volatility and a preset basic influence coefficient derived from prior knowledge in the agricultural field. A third processing module uses the dynamic weight coefficient to perform weighted fusion on the normalized data processed by a specific feature mapping function. This process generates a farmland system state vector that can comprehensively represent the real-time state of the farmland ecosystem. A fourth processing module analyzes the time series of this state vector to predict the future state vector, compares the predicted result with a preset ideal equilibrium state, and determines whether to generate a control instruction based on the comparison result. The system architecture achieves continuous assessment and forward-looking management of the systemic state of the agricultural environment through a complete technical chain from data acquisition, dynamic weighting, state representation, to predictive control.

[0041] Example 2

[0042] The first processing module is used to determine volatility, including:

[0043] The volatility is determined by the finite difference method based on the numerical changes and time intervals of the normalized data at adjacent sampling moments.

[0044] The second processing module is used to calculate the dynamic weight coefficient, including:

[0045] Combining volatility, basic influence coefficient, and preset volatility sensitivity coefficient, the instantaneous activation of each data source is determined. The Softmax function is used to normalize the instantaneous activation of all data sources to generate a dynamic weight coefficient.

[0046] In this embodiment, the function of the first processing module is to quantify the temporal dynamic characteristics of the data source and provide key input for the subsequent adaptive weight allocation. To achieve this goal, the volatility calculation based on the finite difference method is introduced.

[0047] Mathematical realization is defined as:

[0048] ;

[0049] in, : Data Source At time volatility; : Originated from Normalized data at the moment; : Normalized data from the previous moment; : Sampling time interval; The internal logic of this formula is to use the first-order difference to approximate the instantaneous rate of change of discrete data points; Due to the normalized data dimensionless, and The dimension is time , so volatility The physical dimension is , that is, frequency, which is consistent with the physical meaning of the intensity of the change; : data source; :time;

[0050] The output of the first processing module The data is passed to the second processing module to calculate the dynamic weight coefficient of each data source. The calculation process first determines the instantaneous activation through a linear activation model.

[0051] Mathematical realization is defined as:

[0052] ;

[0053] in, : Originated from The instantaneous activation of the moment; : preset dimensionless basic influence coefficient; : Fluctuation sensitivity coefficient; : Volatility of input; to ensure dimensional consistency, due to and are dimensionless, and The dimension of , so the fluctuation sensitivity coefficient The physical dimension must be time ;parameter The value range is , the values ​​are quantified based on the prior knowledge in the agricultural experts’ knowledge base about the importance of a specific data source on the crop.

[0054] A preferred implementation is to use the analytic hierarchy process to determine the basic influence coefficient of each data source. The process includes: at least three agricultural experts compare all data sources, such as soil moisture, light intensity, and leaf images, for a specific crop, such as rice, to build a judgment matrix. For example, the experts need to answer which is more important for the healthy growth of rice during the tillering period, light intensity or soil moisture, and give the relative importance according to the 1-9 scale. The relative weight of each data source is obtained by calculating the maximum eigenvalue of the judgment matrix and its corresponding eigenvector. The weight vector is normalized so that its value falls within Within the interval, the basic influence coefficient of each data source can be obtained ;

[0055] parameter As an adjustable characteristic time constant, its value needs to be determined by optimizing on the historical data set, and the optimization goal is to maximize the accuracy of the final system prediction;

[0056] To determine the fluctuation sensitivity coefficient , an optimization method combining grid search and cross validation can be used; a training set and a validation set can be divided from the historical data set; Set a reasonable physical value range, for example, taking into account the sampling time interval The reciprocal of For 1 hour, The search range can be set to hours, and generate a series of candidate values ​​within this range with a fixed step size, such as 0.1; for each candidate The entire process of the present invention is run on the training set, including the prediction of the fourth processing module, and the prediction accuracy is evaluated on the validation set, for example, using the root mean square error (RMSE) as the evaluation indicator; the one that makes the prediction accuracy on the validation set the highest, that is, the one with the smallest RMSE, is selected. value, as the final fluctuation sensitivity coefficient

[0057] To generate relative weights that can be used for weighted fusion, the instantaneous activations of all data sources need to be normalized;

[0058] Mathematical realization is defined as:

[0059] ;

[0060] in, : Originated from The final dynamic weight coefficient at the moment; : activation level; is the exponential sum of the activations of all data sources j; :All data sources; This normalization step is mathematically equivalent to the Softmax function, ensuring that at any time , the weights of all data sources are dimensionless positive numbers, and their sum is always 1; this set of dynamic weight coefficients This is the final output of the second processing module; It is the key part of realizing the Softmax function. It can convert a set of arbitrary activation values ​​into a probability distribution that sums to 1 and can highlight important items, that is, a dynamic weight coefficient.

[0061] The present invention realizes dynamic adaptive assessment of data influence. Instead of adopting a fixed weight distribution scheme, the system determines the real-time volatility of each data source through the finite difference method based on the numerical change of each data source at adjacent sampling moments. The dynamic weight coefficient of each data source is calculated based on this volatility, the basic influence coefficient derived from agricultural prior knowledge, and the optimizable fluctuation sensitivity coefficient. This enables the system to instantaneously and quantitatively amplify the influence of the current data source that is changing dramatically, thereby more accurately capturing the key dynamic changes of the farmland ecosystem and improving the accuracy and sensitivity of state perception.

[0062] Example 3

[0063] The feature mapping function in the third processing module includes:

[0064] Convolutional neural networks for processing high-frequency visual data;

[0065] Multilayer perceptron for processing low-frequency environmental data;

[0066] In this embodiment, the feature mapping function in the third processing module Used to normalize data of different modalities Mapping to a unified high-dimensional feature space; for high-frequency visual data, the feature mapping function used is a pre-trained convolutional neural network;

[0067] In this embodiment, the convolutional neural network for processing high-frequency visual data can be a custom network containing 5 convolutional layers and 2 fully connected layers; the first three convolutional layers use Each convolution layer is followed by a convolution kernel and ReLU activation function of different sizes. The maximum pooling layer of ; the last two convolutional layers use Convolution kernels are used but no pooling is performed; two fully connected layers are used, the first of which contains 128 neurons and uses the ReLU activation function, and the second fully connected layer outputs a feature vector with the same dimension as the multilayer perceptron output, mapping the extracted features to a unified feature space. Alternatively, as another preferred solution, a lightweight network model pre-trained on the ImageNet dataset, such as ResNet-18, can be used, with its final classification layer removed and replaced with a new fully connected layer to output a feature vector of a specified dimension. Before being applied to this system, this model can be fine-tuned using a specific crop image dataset to improve the targetedness of feature extraction.

[0068] The technology relies on using convolutional and pooling layers to gradually extract spatial hierarchical features of an image, from low-level edge and texture information to high-level semantic features related to crop growth status, such as leaf spot patterns or fruit maturity. For low-frequency environmental data, the feature mapping function used is a multi-layer perceptron.

[0069] In this embodiment, the multilayer perceptron used to process low-frequency environmental data includes an input layer, three hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the low-frequency environmental data. For example, if four types of data, temperature, humidity, light intensity, and soil pH, are collected, the input layer has four neurons. The three hidden layers contain 32, 64, and 32 neurons, respectively, and all use ReLU as the activation function. The output layer outputs a dimensionless feature vector with the same dimension as the feature vector output by the convolutional neural network to ensure that subsequent fusion can be performed.

[0070] The technology relies on learning and fitting the complex nonlinear relationships between environmental parameters, such as temperature, humidity, and light intensity, through multiple hidden layers and nonlinear activation functions. By applying CNN and MLP, raw data from different sources is converted into dimensionless feature vectors of the same dimension and semantically comparable, providing the technical prerequisite for subsequent meaningful weighted fusion.

[0071] The present invention constructs an effective mapping of heterogeneous data to a unified state space. The system uses a special feature mapping function to process high-frequency visual data and low-frequency environmental data, two data types with very different physical modalities and information density. A convolutional neural network is used to extract deep spatial semantic features from visual data, while a multi-layer perceptron is used to fit the nonlinear relationship between environmental data. This converts raw data from different sources into feature vectors of the same dimension and comparable at the semantic level, providing the technical prerequisite for the subsequent execution of meaningful weighted fusion to generate the farmland system state vector, thus solving the problem of heterogeneous data fusion.

[0072] Example 4

[0073] The fourth processing module is used to predict the future state vector, including:

[0074] A recurrent neural network model is used to determine the future state vector based on the current and historical farmland system state vectors and control input vectors.

[0075] The recurrent neural network model is obtained by deep learning training on the historical state vector sequence and the corresponding control input sequence;

[0076] In this embodiment, the technical principle of the model lies in that the recurrent neural network memorizes historical information through its internal hidden states, which can effectively capture the complex nonlinear dynamics and long-term dependencies in time series data. This makes it very suitable for simulating complex time-varying systems such as agricultural ecosystems that change with the growth stage of crops.

[0077] Its mathematical implementation can be conceptually defined as:

[0078] ;

[0079] in, is the predicted state vector for the next moment; is the current state vector; is the current control input vector; is the hidden state of the network at the previous moment, which encodes the historical information of the system; Represents the nonlinear function of the recurrent neural network; Represents all parameters learned through training in the network model, such as weights and biases.

[0080] Parameters in a recurrent neural network model , is obtained by deep learning training on a large amount of historical data; the training process includes: transforming the historical state vector sequence The corresponding control input sequence As input, the network parameters are adjusted through the back-propagation algorithm , to minimize the error between the future state vector predicted by the model and the future state vector actually observed;

[0081] This data-driven deep learning approach enables the prediction model to autonomously learn and internalize the unique, time-varying dynamic characteristics of specific farmland environments, eliminating the need for a fixed linear model. This significantly improves the accuracy and adaptability of future state vector predictions.

[0082] The present invention introduces a data-driven adaptive system evolution prediction model; the prediction of the future evolution of the state vector of the farmland system is not based on a solidified physical or biochemical model, but rather adopts a data-driven nonlinear dynamic model such as a recurrent neural network; the key parameters of the model, namely the state transfer matrix and the control input matrix, are obtained through systematic identification and learning of historical state vectors and control input sequences; this method enables the prediction model to autonomously learn the intrinsic dynamic characteristics of specific farmland from historical data, thereby achieving more accurate predictions of future state vectors, and possessing a high degree of environmental specificity and adaptability.

[0083] Example 5

[0084] The fourth processing module is used to determine whether to generate a control instruction, including:

[0085] Calculate the vector norm deviation between the future state vector and the ideal equilibrium state;

[0086] When the vector norm deviation exceeds a preset threshold, a control instruction is generated to minimize the future state deviation;

[0087] When the vector norm deviation does not exceed the preset threshold, the current control strategy is maintained;

[0088] The ideal equilibrium state is pre-defined by the agricultural expert knowledge base based on the crop growth stage, or determined by analyzing the mean learning of the state vector associated with the historical maximum yield;

[0089] When expert knowledge base is used to define The method is: agricultural experts define an ideal value range for each dimension of the state vector, such as temperature, humidity, etc. For example, the ideal temperature for tomato flowering period is ; then take the median of the range ( ) as the ideal value of the dimension; after the ideal values ​​of all dimensions are normalized in the same way as the data acquisition module, they are combined into the final ideal equilibrium state vector ;

[0090] When determined through machine learning The method is as follows: collect records containing historical yield data; sort all historical periods according to the final yield, select the historical periods corresponding to the 5% with the highest yield; extract all farmland system state vectors in these high-yield periods , and calculate the arithmetic mean of these vectors, that is, the centroid, and use the mean vector as the ideal equilibrium state driven by the data ;Right now:

[0091] ;

[0092] in, is the total number of state vectors during the high-yield period; : sum index, representing the state vector during the high-yield period; : state vector of the farmland system during the high-yield period;

[0093] In this embodiment, the fourth processing module performs decision-making and regulation based on the state prediction result; the core of the decision logic is to convert the predicted future state vector With a preset ideal equilibrium state Compare the two by computing the vector norm deviation between them:

[0094] ;

[0095] This is accomplished by the above formula, where is a scalar that quantifies the extent to which the predicted state deviates from the ideal target; : predicted future state vector; : The preset ideal equilibrium state;

[0096] The deviation value Will be triggered by a preset control threshold Comparison; Threshold The determination of is based on statistical analysis of historical data, aiming to balance the system's sensitivity to state deviations and the stability of control actions, avoiding excessive regulation due to small, harmless fluctuations;

[0097] The threshold The determination of can be achieved through the following statistical method: using historical data to calculate a series of deviation values ​​without regulatory intervention or maintaining a stable control strategy:

[0098] ;

[0099] in It refers to the historical moment The state vector of ; calculate these historical deviation values To avoid frequent adjustments while ensuring sensitivity to significant deviations, the 95th percentile of the distribution or the mean calculated according to the three sigma criterion plus three standard deviations can be used. The value is set as the control trigger threshold ; At a historical moment Deviation value of

[0100] When the calculated deviation Exceeding the threshold When the system determines that the future state will exceed the acceptable range, it will start an optimal control algorithm to generate control instructions aimed at minimizing the deviation of the future state. ;If the deviation Not exceeding the threshold , the system maintains the current control strategy and does not intervene;

[0101] Ideal equilibrium state as a basis for decision-making , itself is a parameter determined dynamically according to the specific situation; The source of can be one of two ways: first, it can be pre-defined by the agricultural expert knowledge base according to the different growth stages of specific crops, for example, different ideal state vectors are provided for different developmental stages; second, it can be determined by machine learning of historical data, that is, analyzing the state vectors corresponding to the period associated with the historical highest yield or best crop quality, and calculating their mean or centroid as the data-driven ideal equilibrium state; this flexible The definition method ensures that the system's control objectives are always scientific and targeted;

[0102] The present invention establishes a forward-looking and optimized control decision-making mechanism; the system's control decision is not a passive response to the current state, but is based on the prediction of the future state vector; by calculating the vector norm deviation between the predicted state and the ideal equilibrium state and comparing it with a preset threshold, it is decided whether to generate a control instruction; the ideal equilibrium state itself is dynamic and can be defined by an expert knowledge base or learned from historical high-yield data; this mechanism realizes the transition from passive response to active prediction, and aims to minimize future state deviations, making the control behavior more forward-looking, accurate and cost-effective.

[0103] 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 the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An early warning monitoring system for smart agriculture, characterized in that: include: The data acquisition module is used to collect high-frequency visual data and low-frequency environmental data, and perform normalization processing to generate normalized data; A first processing module is configured to determine the volatility of each data source based on the normalized data; The second processing module is used to calculate the dynamic weight coefficient of each data source based on volatility and a preset basic influence coefficient; The third processing module is used to generate a farmland system state vector by performing weighted fusion based on the dynamic weight coefficient and the normalized data through a preset feature mapping function; a fourth processing module, configured to predict a future state vector based on the time series of the farmland system state vector, and to compare the future state vector with a preset ideal equilibrium state to determine whether to generate a control instruction; The second processing module is used to calculate the dynamic weight coefficient, including: Combining volatility, basic influence coefficient, and preset volatility sensitivity coefficient, the instantaneous activation of each data source is determined. The Softmax function is used to normalize the instantaneous activation of all data sources to generate a dynamic weight coefficient. Volatility calculations; Mathematical realization is defined as: ； in, : Data Source At time volatility; : Originated from Normalized data at the moment; : Normalized data from the previous moment; : sampling time interval; : data source; :time; The output of the first processing module The data is passed to the second processing module to calculate the dynamic weight coefficient of each data source. The calculation process first determines the instantaneous activation through a linear activation model. Mathematical realization is defined as: ； in, : Originated from The instantaneous activation of the moment; : preset dimensionless basic influence coefficient; : Fluctuation sensitivity coefficient; : Volatility of input; To generate relative weights that can be used for weighted fusion, the instantaneous activations of all data sources need to be normalized; the mathematical implementation is defined as: ； in, : Originated from The final dynamic weight coefficient at the moment; : activation level; is the exponential sum of the activations of all data sources j; : All data sources.

2. The early warning monitoring system for smart agriculture according to claim 1, characterized in that: The first processing module is used to determine volatility, including: The volatility is determined by the finite difference method based on the numerical changes and time intervals of the normalized data at adjacent sampling moments.

3. The early warning monitoring system for smart agriculture according to claim 1, characterized in that: The feature mapping function in the third processing module includes: Convolutional neural networks for processing high-frequency visual data; Multilayer perceptron for processing low-frequency environmental data.

4. The early warning monitoring system for smart agriculture according to claim 1, characterized in that: The fourth processing module is used to predict the future state vector, including: A recurrent neural network model is used to determine the future state vector based on the current and historical farmland system state vectors and control input vectors.

5. The early warning monitoring system for smart agriculture according to claim 4, characterized in that: The recurrent neural network model is obtained by performing deep learning training on historical state vector sequences and corresponding control input sequences.

6. The early warning monitoring system for smart agriculture according to claim 1, characterized in that: The fourth processing module is used to determine whether to generate a control instruction, including: Calculate the vector norm deviation between the future state vector and the ideal equilibrium state; When the vector norm deviation exceeds a preset threshold, a control instruction is generated to minimize the future state deviation; When the vector norm deviation does not exceed the preset threshold, the current control strategy is maintained.

7. An early warning monitoring system for smart agriculture according to claim 1 or 6, characterized in that: The ideal equilibrium state is pre-defined by an agricultural expert knowledge base according to the crop growth stage, or is determined by analyzing the mean learning of the state vector associated with the historical highest yield.

Images