CN-121981536-A - Agricultural drought risk assessment method based on machine learning

CN121981536ACN 121981536 ACN121981536 ACN 121981536ACN-121981536-A

Abstract

The invention provides an agricultural drought risk assessment method based on machine learning, and relates to the technical field of agricultural big data. The method comprises the steps of carrying out time sequence segmentation on multi-source monitoring data based on crop weather period standards to construct a space-time coupling data set, determining a moisture sensitivity coefficient by utilizing development key nodes and physiological water demand threshold values, generating a dynamic attention weight matrix, and extracting a comprehensive drought characteristic vector set through a long-period and short-period memory network. And then, constructing a nonlinear vulnerability fitting curve of the yield response mapping space by adopting a support vector regression algorithm, and constructing a reinforcement learning simulation environment according to the nonlinear vulnerability fitting curve. And constructing a state space vector, performing iterative optimization training of water resource allocation actions in a simulation environment by using a depth Q network algorithm to obtain an optimal risk management and control strategy network, and outputting an agricultural drought risk assessment result and a water resource allocation instruction. The method realizes the accurate quantification of drought risk and the global dynamic optimal configuration of water resources.

Inventors

ZHAO RUXIN
SUN HONGQUAN
Xing Lisong
Tan Zhuoyan
YU HUIQIAN
LI MING

Assignees

应急管理部国家自然灾害防治研究院

Dates

Publication Date: 20260505
Application Date: 20260113

Claims (10)

1. The agricultural drought risk assessment method based on machine learning is characterized by comprising the following steps of: Based on a preset crop weather period standard, carrying out time sequence segmentation on multi-source agricultural monitoring data of a target area to obtain a space-time coupling data set with a plurality of growth stage partitions; Extracting development key node data of all growth stage partitions in the space-time coupling data set and corresponding physiological water requirement threshold values; Determining the moisture sensitivity coefficient of each growth stage subarea according to the development key node data and the corresponding physiological water requirement threshold value, acquiring a meteorological influence factor from the multi-source agricultural monitoring data, and determining a dynamic attention weight matrix according to the meteorological influence factor and the moisture sensitivity coefficient; In the effective weathered time window of each partition, based on the dynamic attention weight matrix, carrying out weighted feature extraction on the space-time coupled data set by using a long-period memory network to obtain a comprehensive drought feature vector set, wherein the effective weathered time window is a growth stage partition remained after invalid data are removed; Constructing a yield response mapping space on the comprehensive drought characteristic vector set, and generating a nonlinear vulnerability fitting curve in the mapping space by adopting a support vector regression algorithm, wherein the mapping relation of the mapping space is a numerical mapping from drought characteristics to yield loss rate; Constructing a reinforcement learning simulation environment according to the nonlinear vulnerability fitting curve; Sequentially constructing state space vectors based on the comprehensive drought characteristic vector set and the time sequence of crop growth and development; Inputting the state space vector into a reinforcement learning simulation environment, and performing iterative optimization training of water resource allocation actions in the reinforcement learning simulation environment based on a depth Q network algorithm to obtain an optimal risk management and control strategy network; And carrying out decision deduction on the real-time monitoring data of the target area according to the optimal risk management and control strategy network to obtain an agricultural drought risk assessment result and a water resource allocation instruction.
2. The machine learning-based agricultural drought risk assessment method of claim 1, wherein the performing time-series segmentation on the multi-source agricultural monitoring data of the target area based on the preset crop weathers criteria to obtain a space-time coupled dataset having a plurality of growth stage partitions comprises: Collecting historical meteorological monitoring values and remote sensing image data of the target area to construct multi-source agricultural monitoring data, and carrying out space-time coordinate alignment and missing value filling on the multi-source agricultural monitoring data to obtain a standardized monitoring sequence; Calculating an effective temperature value and a sunlight accumulation amount in the standardized monitoring sequence, carrying out numerical matching on the effective temperature value and the preset crop weathers standard, and determining a plurality of moments when the development state of the crop is mutated to obtain a weathered partition node; Performing time dimension cutting-off processing on the standardized monitoring sequence by utilizing the weather segmentation nodes to obtain a plurality of independent time segments, and defining each time segment as the growth stage partition; And establishing time sequence indexes of all the growth stage partitions, and packaging the growth stage partitions with the time sequence indexes to obtain the space-time coupling data set.
3. The machine learning based agricultural drought risk assessment method of claim 1, wherein the extracting of development key node data and corresponding physiological water thresholds for all growth stage partitions in a spatiotemporal coupled dataset comprises: acquiring vegetation index time sequence data in each growth stage subarea based on the space-time coupling data set, and carrying out smooth filtering and function fitting on the vegetation index time sequence data to obtain a growth trend curve; calculating a second derivative of the growth trend curve, and determining a moment point with zero second derivative and maximum corresponding curvature as the development key node data; and calling a preset crop water demand knowledge base, carrying out index retrieval on the crop water demand knowledge base by utilizing the development key node data to obtain a corresponding standard water demand standard value, and determining the standard water demand standard value as the physiological water demand threshold value.
4. The machine learning based agricultural drought risk assessment method of claim 1, wherein the determining the moisture sensitivity coefficients of each growth stage partition according to the development key node data and the corresponding physiological water demand threshold, and obtaining weather influencing factors from the multi-source agricultural monitoring data, determining a dynamic attention weight matrix according to the weather influencing factors and the moisture sensitivity coefficients, comprises: calculating theoretical water demand of the current growth stage based on the development key node data; Comparing the theoretical water demand with the physiological water demand threshold value to obtain a water demand gap rate, and mapping the water demand gap rate into the moisture sensitivity coefficient by using a preset sensitivity conversion function; Analyzing precipitation time sequence data, evapotranspiration data and soil moisture content data in the multi-source agricultural monitoring data, and splicing characteristic dimensions of the precipitation time sequence data, the evapotranspiration data and the soil moisture content data to generate the weather influencing factors; Taking the moisture sensitivity coefficient as a weight adjustment parameter, performing element-by-element weighted dot product operation on the weather influence factors to obtain weighted influence vectors, and performing serialization stacking on all the weighted influence vectors according to time step length to generate the dynamic attention weight matrix; The sensitivity transfer function has the expression: ; Wherein, the Is the moisture sensitivity coefficient; The notch rate is the water demand; Is a sensitivity response rate constant; Is a critical threshold constant for water shortage.
5. The machine learning based agricultural drought risk assessment method of claim 4, wherein said calculating theoretical water demand for a current growth stage based on said development key node data comprises: performing physical index matching on a preset crop coefficient database by utilizing the development key node data, and extracting crop coefficients corresponding to the current growth stage partition; Analyzing a temperature characteristic value, a humidity characteristic value and a radiation characteristic value from the weather influencing factors, and calculating the reference crop evapotranspiration of the target area according to the temperature characteristic value, the humidity characteristic value and the radiation characteristic value; constructing a product operation model of the crop coefficient and the reference crop evapotranspiration, and determining an output result of the product operation model as the theoretical water demand; The expression of the product operation model is as follows: ; Wherein, the Is the theoretical water demand of the current growth stage; crop coefficients extracted from the database; net radiation dose for the crop surface; Is soil heat flux; is the dry-wet surface constant; The temperature characteristic value is analyzed; Wind speed at a height of 2 meters; is saturated water vapor pressure; is the actual water vapor pressure.
6. The machine learning-based agricultural drought risk assessment method of claim 1, wherein the performing weighted feature extraction on the spatio-temporal coupled data set by using a long-short term memory network based on the dynamic attention weight matrix in the effective weatherable time window of each partition to obtain a comprehensive drought feature vector set comprises: in the effective physical time window, carrying out time-step characteristic weighting operation on the dynamic attention weight matrix and the space-time coupling data set to obtain a weighted time sequence input sequence; inputting the weighted time sequence input sequence to an input layer of the long-short-period memory network, and carrying out recursion calculation on the weighted time sequence input sequence through a forgetting gate, an input gate and an output gate to update the hidden layer state of the current time step in real time; Extracting the hidden layer state corresponding to the last time step in the effective weather time window, mapping the hidden layer state to a high-dimensional feature space to obtain a comprehensive drought feature vector corresponding to the current growth stage partition, and forming the comprehensive drought feature vector set based on the comprehensive drought feature vectors of all the growth stage partitions.
7. The machine learning-based agricultural drought risk assessment method of claim 1, wherein constructing a yield response mapping space on the comprehensive drought signature vector set and generating a nonlinear vulnerability fitting curve in the mapping space using a support vector regression algorithm comprises: acquiring historical crop yield statistical data of the target area, and calculating a historical yield loss rate based on a preset high yield year benchmark; Performing association matching on the comprehensive drought characteristic vector set and the historical yield loss rate according to years to obtain a regression training sample set; mapping the regression training sample set to a high-dimensional feature space by using a radial basis function, and defining the high-dimensional feature space containing the regression training sample set as the yield response mapping space; and in the yield response mapping space, taking the minimized structural risk as an objective function, carrying out iterative optimization on regression model parameters by using a support vector regression algorithm, solving to obtain an optimal regression hyperplane, and determining the optimal regression hyperplane as the nonlinear vulnerability fitting curve.
8. The machine learning-based agricultural drought risk assessment method of claim 1, wherein the sequentially constructing state space vectors based on the comprehensive drought feature vector set and the time sequence of crop growth and development comprises: performing time sequence index sequencing on the comprehensive drought feature vector set according to the time sequence of the crop growth and development, and positioning a single drought feature vector corresponding to the current simulation time step; performing numerical coding on the growth stage of the current simulation time step to obtain a time embedded feature capable of representing the development process of crops; and reading the current water resource remaining reserve value from the reinforcement learning simulation environment, and performing feature splicing on the water resource remaining reserve value, the single drought feature vector and the time embedded feature to obtain the state space vector.
9. The machine learning-based agricultural drought risk assessment method of claim 1, wherein the inputting the state space vector into a reinforcement learning simulation environment and performing iterative optimization training of water resource allocation actions in the reinforcement learning simulation environment based on a deep Q network algorithm to obtain an optimal risk management and control strategy network comprises: Constructing an evaluation network and a target network with the same structure, inputting the state space vector into the evaluation network, and outputting action value Q values corresponding to all candidate water resource allocation actions in the current state; selecting a target water resource allocation action according to the action value Q value by adopting a preset greedy exploration strategy, inputting the target water resource allocation action into the reinforcement learning simulation environment for execution, and obtaining a feedback single-step rewarding value and a state space vector of the next time step; Packaging the state space vector, the target water resource allocation action, the single step rewarding value and the state space vector of the next time step into an experience quadruple, and storing the experience quadruple into an experience playback buffer area; Randomly sampling batch experience data from the experience playback buffer zone, and analyzing to obtain a current state batch, an action batch, a reward batch and a next state batch; Inputting the next state batch into the target network to obtain the maximum estimated value, and carrying out weighted summation on the maximum estimated value and the rewarding batch to generate a target Q value label; inputting the current state batch into the evaluation network to obtain the predicted values of all actions, and extracting corresponding execution action values from the predicted values of all actions by utilizing the action batch; calculating the mean square error between the target Q value label and the execution action value to obtain a time sequence differential loss; Performing counter-propagation update on the weight parameters of the evaluation network by using a gradient descent algorithm until the time sequence differential loss is converged, and determining the converged evaluation network as the optimal risk management and control strategy network; the computational expression of the evaluation network is: ; the calculation expression of the target network is as follows: ; Wherein, the To evaluate a time-series differential loss value of the network; The size of the sample batch; A target Q value tag generated for a target network; To evaluate network parameters Lower current state Executing an action Is a predictive value of (2); Is the first in the current state batch A state vector; Is the first in the action batch Executing actions; To evaluate current weight parameters of the network; to award the first in the batch A single step prize value; Is a preset discount factor; is the first in the next state batch A state vector; Is a candidate action that may be taken in the next state.
10. The machine learning-based agricultural drought risk assessment method according to claim 1, wherein the performing decision deduction on the real-time monitoring data of the target area according to the optimal risk management and control strategy network to obtain an agricultural drought risk assessment result and a water resource allocation instruction comprises: Collecting real-time monitoring data of the target area at the current moment, and converting the real-time monitoring data into a real-time state space vector and a real-time comprehensive drought characteristic vector; Inputting the real-time comprehensive drought feature vector into the nonlinear vulnerability fitting curve to calculate the theoretical yield loss rate under the current meteorological conditions, and comparing the theoretical yield loss rate with a preset risk classification threshold value to generate the agricultural drought risk assessment result; inputting the real-time state space vector into the optimal risk management and control strategy network, outputting the predicted value distribution of all candidate actions in the current state, and selecting a target action index with the maximum predicted value; and analyzing the target action index into a specific irrigation area position and a corresponding water quantity adjusting parameter by using a preset action decoding dictionary to obtain the water resource allocation instruction.

Description

Agricultural drought risk assessment method based on machine learning Technical Field The invention relates to the technical field of agricultural big data, in particular to an agricultural drought risk assessment method based on machine learning. Background The agricultural drought risk assessment is a key technology for guaranteeing grain safety and efficient utilization of water resources. The current evaluation method mainly relies on calculation of drought indexes such as standardized rainfall indexes and the like, and the frequency and the intensity of drought are quantified by combining a statistical method. For example, in the prior art, a Copula function is often used to construct a joint probability distribution model of drought duration and drought intensity, and disaster conditions are estimated by dividing drought grade thresholds and combining land utilization data such as cultivated land area. Some advanced methods further introduce markov chains to deduce state transitions of drought classes in an attempt to provide basis for drought decision making. With the development of big data and artificial intelligence technology, drought risk assessment is evolving from single weather hydrologic analysis to multi-source data fusion. Research trends tend to change from static post-disaster evaluation to dynamic process identification, strive to capture characteristic changes of drought in real time in the process of occurrence and development of drought, and try to process nonlinear meteorological data by using a machine learning algorithm so as to improve the identification accuracy of extreme drought events. However, the existing agricultural drought risk assessment method still has significant defects in practical application. Firstly, the sensitivity difference of crops to water stress at different growth stages is often ignored in the prior art, so that an evaluation result is disjointed with actual agricultural loss, secondly, a decision model based on a Markov chain usually depends on a fixed transition probability matrix obtained by statistics of historical data, the method cannot adapt to complex and changeable real-time environments, so that large deviation exists in deduction of a future risk state, finally, the traditional method is mostly based on drought grade direct matching with a preset static rule, and lacks multi-objective dynamic balance between limited water resources and irrigation cost and expected benefits of the crops, so that the method is difficult to provide optimized quantitative decision support at administrative management and commercial prediction levels. Disclosure of Invention In order to overcome the defects of the prior art, the invention aims to provide an agricultural drought risk assessment method based on machine learning, and solves the problems of lack of dynamic coupling of crop weathers, static stiffness of a state transition model and lack of global optimizing capability of management decisions in the prior art. In order to achieve the above object, the present invention provides the following solutions: an agricultural drought risk assessment method based on machine learning, comprising: Based on a preset crop weather period standard, carrying out time sequence segmentation on multi-source agricultural monitoring data of a target area to obtain a space-time coupling data set with a plurality of growth stage partitions; Extracting development key node data of all growth stage partitions in the space-time coupling data set and corresponding physiological water requirement threshold values; Determining the moisture sensitivity coefficient of each growth stage subarea according to the development key node data and the corresponding physiological water requirement threshold value, acquiring a meteorological influence factor from the multi-source agricultural monitoring data, and determining a dynamic attention weight matrix according to the meteorological influence factor and the moisture sensitivity coefficient; In the effective weathered time window of each partition, based on the dynamic attention weight matrix, carrying out weighted feature extraction on the space-time coupled data set by using a long-period memory network to obtain a comprehensive drought feature vector set, wherein the effective weathered time window is a growth stage partition remained after invalid data are removed; Constructing a yield response mapping space on the comprehensive drought characteristic vector set, and generating a nonlinear vulnerability fitting curve in the mapping space by adopting a support vector regression algorithm, wherein the mapping relation of the mapping space is a numerical mapping from drought characteristics to yield loss rate; Constructing a reinforcement learning simulation environment according to the nonlinear vulnerability fitting curve; Sequentially constructing state space vectors based on the comprehensive drought characteristic vector set and the