A drought emergency plan optimization method and system based on multi-source data fusion

Abstract

This invention discloses a method and system for optimizing drought emergency response plans based on multi-source data fusion. The method includes collecting multi-source data and historical emergency data from a preset area, and preprocessing the multi-source data and historical emergency data. The multi-source data includes meteorological data, soil data, crop data, and water conservancy data. The multi-source data is input into a lightweight multimodal fusion network to obtain fused data. A dual-index temporal abrupt change detection algorithm and a weakly supervised convolutional autoencoder are used to capture sudden drought features in the fused data to obtain sudden drought features. A machine learning-based drought emergency response plan model is constructed based on the sudden drought features. Predicted emergency data is obtained based on the drought emergency response plan model. Decision bias fit is obtained based on the historical emergency data and the predicted emergency data. Multi-objective reinforcement learning is used to dynamically optimize decisions based on the decision bias fit to obtain the optimized result.

Description

Technical Field

[0001] This invention relates to the field of drought emergency response plan optimization technology, and in particular to a drought emergency response plan optimization method and system based on multi-source data fusion. Background Technology

[0002] In recent years, global climate anomalies have intensified, and flash droughts, characterized by their rapid onset, high intensity, and difficulty in prediction, have become a major threat to agricultural production and water resource allocation. Traditional drought emergency management models are no longer adequate for the new situation. Existing emergency plans mostly rely on single meteorological or soil data, lacking the integration of multi-source information from meteorology, soil, crops, and water resources. This makes it impossible to accurately characterize the coupling relationships between root water uptake, canopy stress, and accessibility of water conservancy facilities, leading to delayed drought identification and significant biases in emergency decision-making. Furthermore, flash droughts exhibit significant suddenness and spatiotemporal heterogeneity, making it difficult for conventional detection methods to capture their rapid abrupt changes, easily missing the optimal emergency response window.

[0003] At the decision-making level, traditional emergency plans are mostly static and fixed processes, failing to dynamically calibrate deviations by combining historical emergency data with real-time forecast results. Faced with complex scenarios such as canal water supply constraints, differences in crop water requirements, and regional water use conflicts, they cannot achieve coordinated optimization of water allocation, irrigation timing, and resource scheduling. Furthermore, existing technologies often rely on single-model predictions, with insufficient integration of physical mechanisms and data-driven approaches. The accuracy of soil moisture simulation and root parameter calibration is limited, making it difficult to support refined emergency decision-making.

[0004] The current drought emergency response system suffers from shortcomings such as shallow data fusion, weak identification of sudden drought, static decision-making, and single optimization objectives. As a result, it is difficult to guarantee the efficiency of emergency response, water resource utilization, and crop yield protection. There is an urgent need for a method that integrates multi-source data, accurately captures the characteristics of sudden drought, and dynamically optimizes emergency decision-making to improve the intelligence and practicality of drought emergency management. Summary of the Invention

[0005] The purpose of this invention is to provide a method for optimizing drought emergency response plans based on multi-source data fusion.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Collect multi-source data and historical emergency data from a preset area, and preprocess the multi-source data and historical emergency data; the multi-source data includes meteorological data, soil data, crop data and water conservancy data; The multi-source data is input into a lightweight multimodal fusion network to obtain fused data. A dual-index temporal mutation detection algorithm and a weakly supervised convolutional autoencoder are used to capture drought features in the fused data to obtain drought features. A drought emergency response plan model based on machine learning is constructed based on the characteristics of the sudden drought, and predictive emergency data is obtained based on the drought emergency response plan model. The decision bias fit is obtained based on the historical emergency data and the predicted emergency data. Multi-objective reinforcement learning is then used to perform dynamic decision optimization based on the decision bias fit to obtain the optimization result.

[0007] Furthermore, the method for inputting the multi-source data into a lightweight multimodal fusion network to obtain fused data includes: A vertical distribution function of root density was established based on crop type. Soil moisture at different depths was weighted according to root distribution to generate a root-weighted soil moisture map, and the water absorption weight was calculated. : ; ; in Surface root density, For the maximum root depth, For distribution morphology parameters, For depth indexing, For depth Root density, For depth Soil moisture conditions, The total number of depth layers, Let d be the normalized depth. For depth Root density; Spatial interaction is performed between hyperspectral and root-weighted soil moisture data to calculate the spatial attention coefficient matrix. A spatial attention map is generated based on this spatial attention. The original stratified soil features and hyperspectral features are then weighted by spatial attention to obtain the root-water coupled spatial attention-enhanced features, expressed as follows: ; ; in This is the spatial attention coefficient matrix. For the spatial gradient of soil moisture, For the Sigmoid function, For feature splicing, For space broadcast multiplication, Characteristics of the original layered soil. For normalized hyperspectral vegetation index, It is a 3×3 convolutional layer; Based on meteorological time series, the daily-scale crop water deficit index is calculated, and the cumulative impact of historical high temperatures is calculated using exponential decay weighting. The expression is as follows: ; ; in For reference evaporation, Let be the rainfall at time t. The highest temperature of the day, This is the critical high-temperature temperature for crops. For indicator functions, This is the daily-scale crop water deficit index. For memory decay factor, The cumulative thermal stress index, For the length of the backtracking time window, The time backtracking step size; Temporal weights are calculated using a self-attention mechanism, and meteorological time-series features are aggregated based on these attention weights. The expression is as follows: , ; ; ; in For time series weights, This is the transpose of the matrix. To accumulate thermal stress mask, For the time series input matrix, To query the learnable weight matrix, The key is a learnable weight matrix. The value is a learnable weight matrix. For querying the matrix, The key matrix, For value matrices, For the feature dimension of the key, It is a temporal aggregation feature; Depthwise separable convolutions are applied to temporal aggregated features to extract local drought symptom features. These local features are then unfolded into a spatial sequence and input into a lightweight mobile vision module. A decomposed attention mechanism is used to calculate the attention coefficients of the lightweight mobile vision module. :

[0008] in It is a spatial sequence. For the dimensions of lightweight mobile vision modules, The jump connection weight matrix; Cross-modal fusion of temporal aggregated features and spatial features is performed. A gating mechanism controls the injection intensity of high-temperature accumulated information. The sequence is folded back to the spatial dimension, and residual connections are executed. The expression is: ; ; ; in For spatial local features, For time-series aggregation features, For feature splicing, For the gated weight matrix, It is a 1×1 convolutional layer. For weighted time series features, For weighted spatial features, For sequence folding operations, For sequence expansion operations, For cross-modal fusion features; Irrigation accessibility maps are constructed using water conservancy facilities. , where nodes For grid cells, edges For irrigation canal connections, a graph attention network encoding is performed to project and broadcast crop types to the spatial dimension via a fully connected layer. The expression is as follows: ; ; in For graph attention networks, For water conservancy expansion feature map, For raw data of water conservancy facilities, This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, For encoding, This is a crop attribute feature map. For broadcast operation; First, align the attention features and crop broadcasting features of the water conservancy map grid by grid, and then stitch them together according to the channel dimension to output unified dimension fused data.

[0009] Furthermore, a method for capturing drought features from the fused data using a dual-index temporal abrupt change detection algorithm and a weakly supervised convolutional autoencoder includes: Based on the evapotranspiration layer and soil moisture layer in the fused data, a dual-index abrupt change detection mechanism is constructed. The Penman formula is used to calculate the daily potential evapotranspiration from the meteorological data of the fused data, and the ratio of actual evapotranspiration to potential evapotranspiration is calculated. Construct a 7-day sliding window anomaly index, expressed as: ; ; in Potential evapotranspiration on a daily scale Net radiation, For soil heat flux, The average temperature The wind speed is 2 meters. For saturated water vapor pressure difference, This is the constant of the wet and dry meter. The slope of the saturated water vapor pressure-temperature curve. Let t be the 7-day sliding window anomaly index. For actual evaporation, This is the actual water vapor pressure. The saturated vapor pressure, This is the average of the previous 7 days as the base period. This represents the standard deviation of the base period of the previous 7 days. Let be the ratio of actual evapotranspiration to potential evapotranspiration at time t. This is the constant of the wet / dry meter; when If the evapotranspiration rate is below the initial threshold for a sudden drought and lasts for 3 days or more, then evapotranspiration is abnormal; the 5-day difference of the standardized soil moisture index is calculated based on the soil moisture content within the fused data. : ; in To standardize the soil moisture index, Water content by volume Field holding capacity The point of wilting. The historical standard deviation of the standardized soil moisture index, Let be the standardized soil moisture index at time t; when And currently If soil moisture decreases rapidly, then soil moisture will decrease rapidly; if both abnormal evapotranspiration and rapid decrease in soil moisture are met simultaneously, then it is determined to be the onset of a sudden drought. By combining logical operations with evapotranspiration anomalies and soil moisture decline, and extracting the spatiotemporal feature vector of the onset of the sudden drought, the expression is as follows: ; ; in This is due to the accumulation of abnormal evaporation over the previous three days. Let be the probability of the drought starting at time t. It is a spatiotemporal feature vector; Input fused data and use a 5-layer convolutional coding structure to extract multi-scale drought features layer by layer, outputting latent features; the latent features are 64-dimensional compressed representations that encode the coupling patterns of root water uptake, canopy temperature, and meteorological stress. The decoder reconstructs the input data using a symmetric structure, constrains feature learning, and uses the standardized soil moisture anomaly index as a weakly supervised signal to identify regions where soil moisture deviates extremely from normal levels. The expression is: ; in The physical prior weak supervision label is used; the weighted sum of reconstruction loss, physical consistency loss, and sparsity constraint is used as the multi-task loss function. A composite agricultural drought index is calculated based on the encoder, and the weighted convolutional autoencoder features are extracted to obtain implicit learning features, which include the composite drought index, feature activation intensity, and feature uncertainty entropy; the expression is: ; ; in The composite agricultural drought index, Let z be the Euclidean norm of the latent feature z. Let z be the information entropy of the latent feature z. This is a feature of implicit learning. is the weight vector of the convolutional autoencoder, and z is the latent feature vector output by the autoencoder; The explicit physical features of the physical drought feature channel are adaptively fused with the implicit learned feature map to obtain the sudden drought fused feature map. The probability of sudden drought occurrence is calculated based on the implicit learned features, as expressed by: ; ; in This is a feature map of sudden drought fusion. These are obvious physical characteristics. This represents the probability of a sudden drought. This is the weight matrix of the fusion layer. This is the bias term for the fusion layer.

[0010] Furthermore, the method for constructing a drought emergency response plan model based on machine learning according to the aforementioned drought characteristics includes: A three-tiered drought emergency response model is constructed, consisting of prediction, correction, and decision-making layers. The correction layer outputs calibrated soil moisture, which serves as the state input for the decision-making layer, forming a complete link between prediction, correction, and decision-making. The drought emergency response model includes a prediction layer, a correction layer, and a decision-making layer. The prediction layer uses a hybrid convolutional neural network and random forest model to predict the drought development trajectory. The correction layer uses physical soil moisture equations superimposed with machine learning for dynamic correction and isotope tracing to calibrate root parameters. The decision-making layer uses binary dynamic programming to solve for the optimal irrigation time window. The characteristics of sudden drought were input into a hybrid convolutional neural network and random forest model. The convolutional neural network was used to extract spatial heterogeneity features and capture field water transfer patterns. The flattened vector output by the convolutional neural network and historical weather data were used as inputs to train a random forest regressor to predict future (7-day) soil moisture and drought level. The expression is: ; ; in The spatial feature map output by the convolutional neural network score. For convolutional blocks, Characterized by sudden drought This is historical meteorological data from the past 30 days. For random forest regressors, The predicted soil moisture value for the next 7 days; The spatial heterogeneity characteristics and predicted soil moisture are integrated and output as follows: When the fusion weight When the value is 0.6, physical prediction is prioritized, and the irrigation demand urgency index is output. ;in This represents the probability of drought risk. To integrate weights, It is a fully connected layer; A physics-data hybrid model is constructed based on Richards equations, with the physical foundation being... Introducing data-driven residual correction terms The correction coefficient is adaptively adjusted according to the prediction uncertainty, and its expression is: ; in Water content by volume For unsaturated hydraulic conductivity, As the matrix potential, For root water absorption, For correction factor, This represents the spatial gradient term of water flux. For depth coordinates, For physical terms, For machine learning residual correction terms, The historical error variance of the physical model; Using stable isotope stratified sampling data, the root water absorption depth distribution is inverted. The solution is obtained using the least squares method, mapping the water absorption ratio to the root water absorption term and correcting the root water absorption term. The expression is as follows:

[0011]

[0012]

[0013] in The stable isotope values ​​of water in plant stems, The stable isotope value of soil water at depth f is given. The percentage of water absorbed by the roots at depth f. For least squares residuals, For regularization hyperparameters, This is the calibrated root water uptake term. The stress coefficient is... For potential transpiration; Decision variables Using binary, a decision variable of 1 indicates that field i is irrigated at time t, and a decision variable of 0 indicates that the field is not irrigated. Given constraints including canal water supply time window, soil moisture balance, and emergency response to sudden drought, the expression is:

[0014]

[0015]

[0016] in The maximum available water flow rate at time t. To force irrigation trigger threshold, Let i be the irrigation flow rate. Let the actual water flow rate of field i be ; Define value function To maximize the cumulative benefit under the initial state, a recursive equation is given. The optimal strategy sequence is solved through back induction, outputting the optimal irrigation time window for each field. The expression is:

[0017] in In the current state Make an executive decision Net income obtained after the current period Next period status The maximum cumulative benefit; The optimal irrigation time window, water allocation scheme, and early warning response level are output as decision-making emergency plans.

[0018] Furthermore, the method for obtaining the decision bias fit degree based on the historical emergency data and the predicted emergency data includes: Historical emergency decisions extracted from historical emergency data and recommended emergency decisions output by a sudden drought feature prediction model are encoded into standardized decision vectors. The weighted Euclidean bias of decisions is calculated based on historical decision variables from historical emergency data and recommended decision variables from predicted emergency data. The standardized decision vectors include time response, water allocation, spatial priority, resource intensity, and risk level. Let a be the importance weight of the a-th decision dimension. To predict the a-th component of the decision vector, Let a be the a-th component of the historical decision vector; Introducing an asymmetric penalty, we calculate the time window offset sensitivity for irrigation time response, and calculate the structure bias index based on the actual execution decision vector and the recommended decision vector:

[0019]

[0020] in To predict the irrigation time response of the decision vector, The irrigation time response is the historical decision vector. For time window offset sensitivity, To allow for early tolerance, The lag penalty coefficient, This is the structural deviation index. To predict the decision vector, For the recommendation decision vector, Let be the magnitude of the vector; Calculate the historical decision performance score, and construct a decision performance correction factor based on the historical decision performance score. The expression is as follows:

[0021]

[0022] in Score the effectiveness of historical decisions. To avoid production loss, For water productivity, For the total emergency cost, , , These are the benefit weights, As a factor to adjust the decision-making effect, Steepness coefficient, The threshold for decision-making effectiveness; When the historical decision performance score is below the decision performance threshold, a slight deviation can actually indicate potential for improvement. Approaching 0 reduces the penalty for bias; when the historical decision performance score is greater than the decision performance threshold, Tendency towards 1, strict requirement for consistency; Calculate the fit of decision bias :

[0023] in , where is the bias penalty coefficient; the decision bias fit is normalized. When the decision bias fit ∈ [0.8,1], the historical emergency decision and the recommended emergency decision are consistent; when the decision bias fit ∈ [0.5,0.8], the historical emergency decision and the recommended emergency decision are biased; when the decision bias fit is less than 0.5, the historical emergency decision and the recommended emergency decision are significantly divergent.

[0024] Furthermore, a method for dynamic decision optimization based on the aforementioned decision bias fit using multi-objective reinforcement learning includes: Multi-objective reinforcement learning consists of an upper layer and a lower layer; the upper layer is the leader, which outputs water supply quotas; the lower layer is the follower, which provides feedback on water demand; and the decision bias fit is used as a dynamic feedback signal. The upper layer is a multi-objective game model, with players being the watershed dispatch center and a set of regional water users; the state space includes reservoir storage, inflow forecast, decision bias fit, and water shortage rate in each region; the action space includes inter-regional water transfer volume, reservoir group dispatch rules, and priority water supply object weights; a multi-objective reward function is constructed, expressed as:

[0025] in For the degree of ecological damage, For fairness index, For total water shortage losses, Let be the water shortage loss coefficient for the s-th region. Let be the water demand of the s-th region. The actual water supply for the s-th region is... Let be the water shortage in the s-th region. For the number of regions, The water supply percentage for each region is as follows, after sorting. For cumulative weighting, Let j be the target value for ecological flow at the j-th river cross-section. The actual flow rate at the j-th river cross-section; To impose fuzzy constraints on engineering capabilities, trapezoidal fuzzy numbers are introduced, and the membership function is given as follows:

[0026] in To achieve the minimum water conveyance capacity of the project, This represents the lower limit of the project's stable water delivery capacity. This represents the upper limit of the project's stable water delivery capacity. To the maximum water conveyance capacity of the project, For water delivery volume, For water delivery volume membership function, Employing a multi-agent deep deterministic policy gradient, where the watershed center and regional users are both adversaries and the environment, the Nash equilibrium policy satisfies: ,in The optimal Nash equilibrium strategy for the watershed scheduling center. For the aggregation of regional water users, For expectation operator, This is the transpose of the multi-objective weight vector. For a multi-objective reward function, The strategy to be optimized for the scheduling center; When the fit of decision bias is less than 0.6, increase the weight of water shortage loss; when the fit of decision bias is greater than 0.8, increase the weight of fairness. The lower layer is a fuzzy interactive programming model, with decision variables including irrigation regime, crop structure adjustment ratio, and water-saving technology deployment rate. A fuzzy objective function is constructed, and interactive constraints are given based on the fuzzy matching between the upper-level water supply allocation and the lower-level demand. The expression is:

[0027]

[0028]

[0029] in To fuzzy output target value, Here, m represents the target value for water productivity, and m is the field partition index. Let be the fuzzy membership function for the yield of the m-th field. Let m be the soil moisture of the m-th plot. Let m be the planting area of ​​the m-th plot. Let m be the potential yield of the m-th plot. To increase the deployment rate of water-saving technologies, To adjust the proportion of crop structure, For water supply distribution to the upper floors, For the needs of the lower levels, This represents the actual evapotranspiration of the m-th field. The planned irrigation water volume for the m-th field is... Let m be the actual yield of the m-th field. For the variance of water supply fluctuations, This is the matching threshold. To match supply and demand by degree of membership; The states in proximal policy optimization reinforcement learning are soil moisture field, crop phenological stage, weather forecast, and decision bias fitness feedback value; the actions are field-level irrigation decisions and crop replacement suggestions; the reward function for proximal policy optimization reinforcement learning is:

[0030] in For the reward function, As a weighted average of production output, As a weight for water productivity, To increase production, For water productivity, Weighting for non-drought conditions, This is the water resource cost penalty coefficient. For water resource costs; When the fit of decision deviation is greater than the decision deviation threshold, the reward is biased towards output and water productivity; when the fit of decision deviation is less than the decision deviation threshold, the water resource cost penalty is increased, and conservative strategies are forced to be explored. Upper-level planning: The basin center issues an initial water supply strategy based on the current DDFI; Lower-level response: Each irrigation district assesses the fit between the quota and local decision-making deviation, solves the fuzzy interactive programming problem, and provides feedback on the actual water demand; if Supply and demand mismatch, adjustments made at the top Weights, re-solve for the Nash equilibrium until convergence. ; When |allocated water volume - actual water demand| is less than ε, the convergence rolling optimization triggering conditions include hard triggering, soft triggering and periodic triggering; Hard triggering: the decision bias fitness is less than 0.5 and retraining is performed immediately, soft triggering is level jump fine-tuning, and periodic triggering is updated every 7 days; Soft triggering: the drought level increases by 1 level or more and activates policy fine-tuning; Periodic triggering: the value network of the upper-layer multi-agent deep deterministic policy gradient is updated at a fixed period; A model predictive control framework is adopted, with optimization in the time domain for 7 days and control in the time domain for 3 days. Optimization is performed every 3 days, and the strategy for the next 4 days is retained as an initial guess.

[0031] Secondly, a drought emergency response plan optimization system based on multi-source data fusion includes: Data acquisition and preprocessing module: used to acquire multi-source data and historical emergency data of a preset area, and to preprocess the multi-source data and historical emergency data; the multi-source data includes meteorological data, soil data, crop data and water conservancy data; The data fusion and feature extraction module is used to input the multi-source data into a lightweight multimodal fusion network to obtain fused data, and to use a dual-index temporal mutation detection algorithm and a weakly supervised convolutional autoencoder to capture drought features in the fused data to obtain drought features. Model building and prediction module: used to build a drought emergency response plan model based on machine learning based on the characteristics of the sudden drought, and to obtain predicted emergency data based on the drought emergency response plan model; Bias adaptation and dynamic optimization module: used to obtain the decision bias adaptation degree based on the historical emergency data and the predicted emergency data, and to perform dynamic decision optimization based on the decision bias adaptation degree using multi-objective reinforcement learning to obtain the optimization result.

[0032] The beneficial effects of this invention are: This invention is a method and system for optimizing drought emergency response plans based on multi-source data fusion. Compared with existing technologies, this invention has the following technical advantages: This invention achieves deep coupling of multi-source data from meteorology, soil, crops, and water resources through a lightweight multimodal fusion network, accurately characterizing root-water coupling and water accessibility features, and improving data utilization efficiency. It employs a fast note-start point detection + weakly supervised convolutional autoencoder to efficiently capture sudden drought abrupt changes, solving the problem of lagging drought identification in traditional methods. A three-level model of prediction-correction-decision is constructed, combining physical mechanisms and machine learning correction to improve the accuracy of drought simulation and irrigation decision-making. The introduction of decision bias fit and multi-objective reinforcement learning enables dynamic optimization of water supply scheduling and irrigation schemes, balancing yield preservation, water conservation, and fairness. This improves the efficiency of emergency response to sudden droughts and water resource utilization. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the steps of an optimization method for drought emergency response plans based on multi-source data fusion, as described in this invention. Detailed Implementation

[0034] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0035] The present invention provides a method and system for optimizing drought emergency response plans based on multi-source data fusion, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Collect multi-source data and historical emergency data from a preset area, and preprocess the multi-source data and historical emergency data; the multi-source data includes meteorological data, soil data, crop data and water conservancy data; In this practical assessment, this embodiment takes Area A1 of a certain province as the research object. Area A1 is a winter wheat-corn planting area, which relies on the Yellow River irrigation system for irrigation and is prone to sudden droughts. The core irrigation area of ​​Area A1 is 300km long. 2 It includes 3 townships, 12 main irrigation canals, and 236 farmland grids (100m×100m); the study aims to optimize the drought emergency response plan during the period from March to June 2025 (the period of jointing to grain filling of winter wheat, the period with the highest risk of sudden drought); Meteorological data includes daily maximum temperature, precipitation, wind speed, net radiation, and relative humidity; soil data includes volumetric water content at 0–20 cm, 20–40 cm, and 40–60 cm depths; crop data includes winter wheat root distribution, phenological stage, and canopy temperature; and water conservancy data includes irrigation canal flow, reservoir storage, and irrigation accessibility. The preprocessing method for meteorological data is linear interpolation of missing values, and uniformity is to daily time series; the preprocessing method for soil data is standardization to SSI index and removal of outliers; the preprocessing method for crop data is root density fitting to generate root distribution map; the preprocessing method for water conservancy data is to construct irrigation canal topology map and encode water supply capacity. Historical emergency data includes four drought emergency cases from 2021 to 2024, including: irrigation start time, water allocation, crop yield reduction rate, emergency costs, and records of canal water supply constraints; preprocessing: coded into standardized decision vectors, which include time response, water allocation, spatial priority, resource intensity, and risk level; The multi-source data is input into a lightweight multimodal fusion network to obtain fused data. A dual-index temporal mutation detection algorithm and a weakly supervised convolutional autoencoder are used to capture drought features in the fused data to obtain drought features. A drought emergency response plan model based on machine learning is constructed based on the characteristics of the sudden drought, and predictive emergency data is obtained based on the drought emergency response plan model. The decision bias fit is obtained based on the historical emergency data and the predicted emergency data. Multi-objective reinforcement learning is then used to perform dynamic decision optimization based on the decision bias fit to obtain the optimization result.

[0036] In this embodiment, the method for inputting the multi-source data into a lightweight multimodal fusion network to obtain fused data includes: A vertical distribution function of root density was established based on crop type. Soil moisture at different depths was weighted according to root distribution to generate a root-weighted soil moisture map, and the water absorption weight was calculated. : ; ; in Surface root density, For the maximum root depth, For distribution morphology parameters, For depth indexing, For depth Root density, For depth Soil moisture conditions, The total number of depth layers, Let d be the normalized depth. For depth Root density; Spatial interaction is performed between hyperspectral and root-weighted soil moisture data to calculate the spatial attention coefficient matrix. A spatial attention map is generated based on this spatial attention. The original stratified soil features and hyperspectral features are then weighted by spatial attention to obtain the root-water coupled spatial attention-enhanced features, expressed as follows: ; ; in This is the spatial attention coefficient matrix. For the spatial gradient of soil moisture, For the Sigmoid function, For feature splicing, For space broadcast multiplication, Characteristics of the original layered soil. For normalized hyperspectral vegetation index, It is a 3×3 convolutional layer; Based on meteorological time series, the daily-scale crop water deficit index is calculated, and the cumulative impact of historical high temperatures is calculated using exponential decay weighting. The expression is as follows: ; ; in For reference evaporation, Let be the rainfall at time t. The highest temperature of the day, This is the critical high-temperature temperature for crops. For indicator functions, This is the daily-scale crop water deficit index. For memory decay factor, The cumulative thermal stress index, For the length of the backtracking time window, The time backtracking step size; Temporal weights are calculated using a self-attention mechanism, and meteorological time-series features are aggregated based on these attention weights. The expression is as follows: , ; ; ; in For time series weights, This is the transpose of the matrix. To accumulate thermal stress mask, For the time series input matrix, To query the learnable weight matrix, The key is a learnable weight matrix. The value is a learnable weight matrix. For querying the matrix, The key matrix, For value matrices, For the feature dimension of the key, It is a temporal aggregation feature; Depthwise separable convolutions are applied to temporal aggregated features to extract local drought symptom features. These local features are then unfolded into a spatial sequence and input into a lightweight mobile vision module. A decomposed attention mechanism is used to calculate the attention coefficients of the lightweight mobile vision module. :

[0037] in It is a spatial sequence. For the dimensions of lightweight mobile vision modules, The jump connection weight matrix; Cross-modal fusion of temporal aggregated features and spatial features is performed. A gating mechanism controls the injection intensity of high-temperature accumulated information. The sequence is folded back to the spatial dimension, and residual connections are executed. The expression is: ; ; ; in For spatial local features, For time-series aggregation features, For feature splicing, For the gated weight matrix, It is a 1×1 convolutional layer. For weighted time series features, For weighted spatial features, For sequence folding operations, For sequence expansion operations, For cross-modal fusion features; Irrigation accessibility maps are constructed using water conservancy facilities. , where nodes For grid cells, edges For irrigation canal connections, a graph attention network encoding is performed to project and broadcast crop types to the spatial dimension via a fully connected layer. The expression is as follows: ; ; in For graph attention networks, For water conservancy expansion feature map, For raw data of water conservancy facilities, This is the weight matrix of the fully connected layer. For the bias term of the fully connected layer, For encoding, This is a crop attribute feature map. For broadcast operation; First, align the attention features and crop broadcasting features of the water conservancy map grid by grid, and then stitch them together according to the channel dimension to output unified dimension fused data; In the actual assessment, the maximum root depth was 80cm, the surface root density was 1.2, and the distribution morphology parameter was 1.8. The weighted value of soil moisture at different depths was calculated according to the formula to generate a root weighted moisture map, highlighting the water absorption weight of the 0–40cm taproot zone. The spatial interaction between hyperspectral data and soil moisture data generates a root-water coupled spatial attention map, which accurately locates drought patches in the field; the daily-scale crop water deficit index (CWD) is calculated, and the cumulative heat stress over 7 days is traced back, and meteorological time-series features are aggregated through self-attention. Construct an irrigation accessibility map with 12 irrigation canals as edges and farmland grids as nodes; output unified dimension fusion data, coupled with four types of features: meteorology, soil, crops, and water conservancy.

[0038] In this embodiment, a method for capturing drought features from the fused data using a dual-index temporal abrupt change detection algorithm and a weakly supervised convolutional autoencoder includes: Based on the evapotranspiration layer and soil moisture layer in the fused data, a dual-index abrupt change detection mechanism is constructed. The Penman formula is used to calculate the daily potential evapotranspiration from the meteorological data of the fused data, and the ratio of actual evapotranspiration to potential evapotranspiration is calculated. Construct a 7-day sliding window anomaly index, expressed as: ; ; in Potential evapotranspiration on a daily scale Net radiation, For soil heat flux, The average temperature The wind speed is 2 meters. For saturated water vapor pressure difference, This is the constant of the wet and dry meter. The slope of the saturated water vapor pressure-temperature curve. Let t be the 7-day sliding window anomaly index. For actual evaporation, This is the actual water vapor pressure. The saturated vapor pressure, This is the average of the previous 7 days as the base period. This represents the standard deviation of the base period of the previous 7 days. Let be the ratio of actual evapotranspiration to potential evapotranspiration at time t. This is the constant of the wet / dry meter; when If the evapotranspiration rate is below the initial threshold for a sudden drought and lasts for 3 days or more, then evapotranspiration is abnormal. The 5-day difference of the standardized soil moisture index is calculated based on the soil moisture content within the fused data. : ; in To standardize the soil moisture index, Water content by volume Field holding capacity The point of wilting. The historical standard deviation of the standardized soil moisture index, Let be the standardized soil moisture index at time t; when And currently If soil moisture decreases rapidly, then the condition is considered to be the onset of a sudden drought. If both abnormal evapotranspiration and rapid decrease in soil moisture are met, then the drought is considered to be initiating. By combining logical operations with evapotranspiration anomalies and soil moisture decline, and extracting the spatiotemporal feature vector of the onset of the sudden drought, the expression is as follows: ; ; in This is due to the accumulation of abnormal evaporation over the previous three days. Let be the probability of the drought starting at time t. It is a spatiotemporal feature vector; Input fused data and use a 5-layer convolutional coding structure to extract multi-scale drought features layer by layer, outputting latent features; the latent features are 64-dimensional compressed representations that encode the coupling patterns of root water uptake, canopy temperature, and meteorological stress. The decoder reconstructs the input data using a symmetric structure, constrains feature learning, and uses the standardized soil moisture anomaly index as a weakly supervised signal to identify regions where soil moisture deviates extremely from normal levels. The expression is: ; in The physical prior weak supervision label is used; the weighted sum of reconstruction loss, physical consistency loss, and sparsity constraint is used as the multi-task loss function. A composite agricultural drought index is calculated based on the encoder, and the weighted convolutional autoencoder features are extracted to obtain implicit learning features, which include the composite drought index, feature activation intensity, and feature uncertainty entropy; the expression is: ; ; in The composite agricultural drought index, Let z be the Euclidean norm of the latent feature z. Let z be the information entropy of the latent feature z. This is a feature of implicit learning. is the weight vector of the convolutional autoencoder, and z is the latent feature vector output by the autoencoder; The explicit physical features of the physical drought feature channel are adaptively fused with the implicit learned feature map to obtain the sudden drought fused feature map. The probability of sudden drought occurrence is calculated based on the implicit learned features, as expressed by: ; ; in This is a feature map of sudden drought fusion. These are obvious physical characteristics. This represents the probability of a sudden drought. This is the weight matrix of the fusion layer. For the fusion layer bias term; In the actual assessment, the evapotranspiration layer was estimated by meteorological elements, the soil moisture layer came from the stage-one fusion results, and the initial threshold for sudden drought was -1.5; the feature dimensions of the implicit learning features were: evapotranspiration anomaly, soil moisture decline rate, initial probability, soil moisture change rate, compound drought index, feature intensity, and feature uncertainty; the classification of sudden drought levels is shown in Table 1. Table 1 Classification of Sudden Drought Levels

[0039] The Penman formula calculates the daily potential evapotranspiration ET0 as 5.2 mm / d, with an actual / potential evapotranspiration ratio of 0.32. The 7-day sliding anomaly index is -1.6, which is less than the drought threshold of -1.5. The index has been maintained for 3 days, indicating an evapotranspiration anomaly. The standardized soil moisture index is 0.35, 0.35 < 0.4, and the 5-day difference is -1.2, -1.2 is less than -1; since both indicators are met, it is determined to be the onset of a sudden drought, and the spatiotemporal feature vector is extracted. Five convolutional encoding layers output 64-dimensional latent features, with weakly supervised labels. The value is 1, indicating extremely abnormal soil moisture. By combining explicit physical characteristics and implicit learning characteristics, the probability of sudden drought is calculated to be 0.83, indicating that the peak of sudden drought has begun.

[0040] In this embodiment, the method for constructing a drought emergency response plan model based on machine learning according to the characteristics of sudden drought includes: A three-tiered drought emergency response model is constructed, consisting of prediction, correction, and decision-making layers. The correction layer outputs calibrated soil moisture, which serves as the state input for the decision-making layer, forming a complete link between prediction, correction, and decision-making. The drought emergency response model includes a prediction layer, a correction layer, and a decision-making layer. The prediction layer uses a hybrid convolutional neural network and random forest model to predict the drought development trajectory. The correction layer uses physical soil moisture equations superimposed with machine learning for dynamic correction and isotope tracing to calibrate root parameters. The decision-making layer uses binary dynamic programming to solve for the optimal irrigation time window. The characteristics of sudden drought were input into a hybrid convolutional neural network and random forest model. The convolutional neural network was used to extract spatial heterogeneity features and capture field water transfer patterns. The flattened vector output by the convolutional neural network and historical weather data were used as inputs to train a random forest regressor to predict future (7-day) soil moisture and drought level. The expression is: ; ; in The spatial feature map output by the convolutional neural network score. For convolutional blocks, Characterized by sudden drought This is historical meteorological data from the past 30 days. For random forest regressors, The predicted soil moisture value for the next 7 days; The spatial heterogeneity characteristics and predicted soil moisture are integrated and output as follows: When the fusion weight When the value is 0.6, physical prediction is prioritized, and the irrigation demand urgency index is output. ;in This represents the probability of drought risk. To integrate weights, It is a fully connected layer; A physics-data hybrid model is constructed based on Richards equations, with the physical foundation being... Introducing data-driven residual correction terms The correction coefficient is adaptively adjusted according to the prediction uncertainty, and its expression is: ; in Water content by volume For unsaturated hydraulic conductivity, As the matrix potential, For root water absorption, For correction factor, This represents the spatial gradient term of water flux. For depth coordinates, For physical terms, For machine learning residual correction terms, The historical error variance of the physical model; Using stable isotope stratified sampling data, the root water absorption depth distribution is inverted. The solution is obtained using the least squares method, mapping the water absorption ratio to the root water absorption term and correcting the root water absorption term. The expression is as follows:

[0041]

[0042]

[0043] in The stable isotope values ​​of water in plant stems, The stable isotope value of soil water at depth f is given. The percentage of water absorbed by the roots at depth f. For least squares residuals, For regularization hyperparameters, This is the calibrated root water uptake term. The stress coefficient is... For potential transpiration; Decision variables Using binary, a decision variable of 1 indicates that field i is irrigated at time t, and a decision variable of 0 indicates that the field is not irrigated. Given constraints including canal water supply time window, soil moisture balance, and emergency response to sudden drought, the expression is:

[0044]

[0045]

[0046] in The maximum available water flow rate at time t. To force irrigation trigger threshold, Let i be the irrigation flow rate. Let the actual water flow rate of field i be ; Define value function To maximize the cumulative benefit under the initial state, a recursive equation is given. The optimal strategy sequence is solved through back induction, outputting the optimal irrigation time window for each field. The expression is:

[0047] in In the current state Make an executive decision Net income obtained after the current period Next period status The maximum cumulative benefit; The optimal irrigation time window, water allocation scheme, and early warning response level are output as a decision-making emergency plan. In actual assessments, the isotopes are It was determined that samples should be taken every 10 days during the crop's growth period, and the calibrated root water absorption rate was calculated. Directly replace the original Richards equation Prediction layer: CNN + Random Forest predicts soil moisture for the next 7 days, with a drought risk probability of 0.78 and an irrigation urgency of 0.86; Correction layer: Richards equation + machine learning residual correction, isotope... Calibrate the root water uptake term; soil moisture simulation error <3%; Decision layer: Solve for the optimal irrigation time window using binary dynamic programming; Actual irrigation decision output: Optimal irrigation time: April 16-17 (canal water supply window); Water allocation: 45m core wheat field 3 / mu, edge plot 35m 3 / mu; Warning level: Orange drought warning, Level II emergency response activated.

[0048] In this embodiment, the method for obtaining the decision bias fit degree based on the historical emergency data and the predicted emergency data includes: Historical emergency decisions extracted from historical emergency data and recommended emergency decisions output by a sudden drought feature prediction model are encoded into standardized decision vectors. The weighted Euclidean bias of decisions is calculated based on historical decision variables from historical emergency data and recommended decision variables from predicted emergency data. The standardized decision vectors include time response, water allocation, spatial priority, resource intensity, and risk level. Let a be the importance weight of the a-th decision dimension. To predict the a-th component of the decision vector, Let a be the a-th component of the historical decision vector; Introducing an asymmetric penalty, we calculate the time window offset sensitivity for irrigation time response, and calculate the structure bias index based on the actual execution decision vector and the recommended decision vector:

[0049]

[0050] in To predict the irrigation time response of the decision vector, The irrigation time response is the historical decision vector. For time window offset sensitivity, To allow for early tolerance, The lag penalty coefficient, This is the structural deviation index. To predict the decision vector, For the recommendation decision vector, Let be the magnitude of the vector; Calculate the historical decision performance score, and construct a decision performance correction factor based on the historical decision performance score. The expression is as follows:

[0051]

[0052] in Score the effectiveness of historical decisions. To avoid production loss, For water productivity, For the total emergency cost, , , These are the benefit weights, As a factor to adjust the decision-making effect, Steepness coefficient, The threshold for decision-making effectiveness; When the historical decision performance score is below the decision performance threshold, a slight deviation can actually indicate potential for improvement. Approaching 0 reduces the penalty for bias; when the historical decision performance score is greater than the decision performance threshold, Tendency towards 1, strict requirement for consistency; Calculate the fit of decision bias :

[0053] in The deviation penalty coefficient is used to normalize the decision deviation fit. When the decision deviation fit is ∈ [0.8,1], the historical emergency decision and the recommended emergency decision are consistent; when the decision deviation fit is ∈ [0.5,0.8], the historical emergency decision and the recommended emergency decision are biased; when the decision deviation fit is less than 0.5, the historical emergency decision and the recommended emergency decision are significantly divergent. In actual evaluation, the decision effectiveness threshold is 70; recommended decision vector: [16-day start, 45 / 35 ratio, core area priority, level 2 response, high risk]; historical decision vector in 2023: [20-day start, 40 / 40 ratio, average allocation, level 3 response, medium risk]; The weighted Euclidean bias was 0.42, the time offset sensitivity was 0.38, the historical decision performance score was 62 (below 70), the correction factor γeff=0.21, and the decision bias fit was 0.47 (<0.5), indicating a significant deviation between the historical judgment and the recommended decision.

[0054] In this embodiment, a method for dynamic decision optimization based on the decision bias fit using multi-objective reinforcement learning includes: Multi-objective reinforcement learning consists of an upper layer and a lower layer; the upper layer is the leader, which outputs water supply quotas; the lower layer is the follower, which provides feedback on water demand; and the decision bias fit is used as a dynamic feedback signal. The upper layer is a multi-objective game model, with players being the watershed dispatch center and a set of regional water users; the state space includes reservoir storage, inflow forecast, decision bias fit, and water shortage rate in each region; the action space includes inter-regional water transfer volume, reservoir group dispatch rules, and priority water supply object weights; a multi-objective reward function is constructed, expressed as:

[0055] in For the degree of ecological damage, For fairness index, For total water shortage losses, Let be the water shortage loss coefficient for the s-th region. Let be the water demand of the s-th region. The actual water supply for the s-th region is... Let be the water shortage in the s-th region. For the number of regions, The water supply percentage for each region is as follows, after sorting. For cumulative weighting, Let j be the target value for ecological flow at the j-th river cross-section. The actual flow rate at the j-th river cross-section; To impose fuzzy constraints on engineering capabilities, trapezoidal fuzzy numbers are introduced, and the membership function is given as follows:

[0056] in To achieve the minimum water conveyance capacity of the project, This represents the lower limit of the project's stable water delivery capacity. This represents the upper limit of the project's stable water delivery capacity. To the maximum water conveyance capacity of the project, For water delivery volume, For water delivery volume membership function, Employing a multi-agent deep deterministic policy gradient, where the watershed center and regional users are both adversaries and the environment, the Nash equilibrium policy satisfies: ,in The optimal Nash equilibrium strategy for the watershed scheduling center. For the aggregation of regional water users, For expectation operator, This is the transpose of the multi-objective weight vector. For a multi-objective reward function, The strategy to be optimized for the scheduling center; When the fit of decision bias is less than 0.6, increase the weight of water shortage loss; when the fit of decision bias is greater than 0.8, increase the weight of fairness. The lower layer is a fuzzy interactive programming model, with decision variables including irrigation regime, crop structure adjustment ratio, and water-saving technology deployment rate. A fuzzy objective function is constructed, and interactive constraints are given based on the fuzzy matching between the upper-level water supply allocation and the lower-level demand. The expression is:

[0057]

[0058]

[0059] in To fuzzy output target value, Here, m represents the target value for water productivity, and m is the field partition index. Let be the fuzzy membership function for the yield of the m-th field. Let m be the soil moisture of the m-th plot. Let m be the planting area of ​​the m-th plot. Let m be the potential yield of the m-th plot. To increase the deployment rate of water-saving technologies, To adjust the proportion of crop structure, For water supply distribution to the upper floors, For the needs of the lower levels, This represents the actual evapotranspiration of the m-th field. The planned irrigation water volume for the m-th field is... Let m be the actual yield of the m-th field. For the variance of water supply fluctuations, This is the matching threshold. To match supply and demand by degree of membership; The states in proximal policy optimization reinforcement learning are soil moisture field, crop phenological stage, weather forecast, and decision bias fitness feedback value; the actions are field-level irrigation decisions and crop replacement suggestions; the reward function for proximal policy optimization reinforcement learning is:

[0060] in For the reward function, As a weighted average of production output, As a weight for water productivity, To increase production, For water productivity, Weighting for non-drought conditions, This is the water resource cost penalty coefficient. For water resource costs; When the fit of decision deviation is greater than the decision deviation threshold, the reward is biased towards output and water productivity; when the fit of decision deviation is less than the decision deviation threshold, the water resource cost penalty is increased, and conservative strategies are forced to be explored. Upper-level planning: The basin center issues an initial water supply strategy based on the current DDFI; Lower-level response: Each irrigation district assesses the fit between the quota and local decision-making deviation, solves the fuzzy interactive programming problem, and provides feedback on the actual water demand; if Supply and demand mismatch, adjustments made at the top Weights, re-solve for the Nash equilibrium until convergence. ; When |allocated water volume - actual water demand| is less than ε, the convergence rolling optimization triggering conditions include hard triggering, soft triggering and periodic triggering; Hard triggering: the decision bias fitness is less than 0.5 and retraining is performed immediately, soft triggering is level jump fine-tuning, and periodic triggering is updated every 7 days; Soft triggering: the drought level increases by 1 level or more and activates policy fine-tuning; Periodic triggering: the value network of the upper-layer multi-agent deep deterministic policy gradient is updated at a fixed period; A model-based predictive control framework is adopted, with optimization over a 7-day time domain and control over a 3-day time domain. Optimization is performed every 3 days, and the strategy for the next 4 days is retained as an initial guess. The optimized irrigation time window, water allocation, and early warning level are output. In the actual assessment, the matching degree threshold was 0.7. At the upper level (basin dispatch center), the decision bias fit was 0.47 (<0.6), so the weight of water shortage loss was increased, and the inter-regional water transfer was adjusted by 120,000 m³. 3 Prioritize core wheat fields; lower level (irrigated area): fuzzy interactive planning optimizes irrigation system, increases water-saving technology deployment rate by 15%, and fine-tunes crop structure; Water supply and demand deviation < 0.5 million m³ 3 The strategy converges; it updates every 7 days, the drought level does not jump, and it does not trigger soft optimization. The optimized emergency plan is as follows: Irrigation period: 10:00 AM on April 16th to 6:00 PM on April 17th; Total water supply: 860,000 m³ 3 Compared to historical plans, it saves 12% of water; the crop yield preservation rate is 94%, an increase of 11% compared to the past; the water supply fairness index of the canal system is 0.87, meeting the ecological flow constraints.

[0061] Secondly, a drought emergency response plan optimization system based on multi-source data fusion includes: Data acquisition and preprocessing module: used to acquire multi-source data and historical emergency data of a preset area, and to preprocess the multi-source data and historical emergency data; the multi-source data includes meteorological data, soil data, crop data and water conservancy data; The data fusion and feature extraction module is used to input the multi-source data into a lightweight multimodal fusion network to obtain fused data, and to use a dual-index temporal mutation detection algorithm and a weakly supervised convolutional autoencoder to capture drought features in the fused data to obtain drought features. Model building and prediction module: used to build a drought emergency response plan model based on machine learning based on the characteristics of the sudden drought, and to obtain predicted emergency data based on the drought emergency response plan model; Bias adaptation and dynamic optimization module: used to obtain the decision bias adaptation degree based on the historical emergency data and the predicted emergency data, and to perform dynamic decision optimization based on the decision bias adaptation degree using multi-objective reinforcement learning to obtain the optimization result.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing drought emergency response plans based on multi-source data fusion, characterized in that, Includes the following steps: Collect multi-source data and historical emergency data from a preset area, and preprocess the multi-source data and historical emergency data; the multi-source data includes meteorological data, soil data, crop data and water conservancy data; The multi-source data is input into a lightweight multimodal fusion network to obtain fused data. A dual-index temporal mutation detection algorithm and a weakly supervised convolutional autoencoder are used to capture drought features in the fused data to obtain drought features. A drought emergency response plan model based on machine learning is constructed based on the characteristics of the sudden drought, and predictive emergency data is obtained based on the drought emergency response plan model. The decision bias fit is obtained based on the historical emergency data and the predicted emergency data. Multi-objective reinforcement learning is then used to perform dynamic decision optimization based on the decision bias fit to obtain the optimization result.

2. The drought emergency response plan optimization method based on multi-source data fusion according to claim 1, characterized in that, A method for obtaining fused data by inputting the multi-source data into a lightweight multimodal fusion network includes: A root density vertical distribution function is established based on crop type. Soil moisture at different depths is weighted according to root distribution to generate a root-weighted soil moisture map and calculate water absorption weights. By spatially interacting hyperspectral data with root-weighted soil moisture data, a spatial attention coefficient matrix is ​​calculated. A spatial attention map is generated based on the spatial attention data. The original stratified soil features and hyperspectral features are weighted by spatial attention to obtain the root-water coupled spatial attention enhanced features. The daily-scale crop water deficit index is calculated based on meteorological time series, the cumulative impact of historical high temperatures is calculated using exponential decay weighting, the time series weights are calculated through a self-attention mechanism, and meteorological time series features are aggregated based on the attention weights. Depthwise separable convolutions are applied to temporal aggregated features to extract local drought symptom features. These local features are then unfolded into a spatial sequence and input into a lightweight mobile vision module. A decomposed attention mechanism is used to calculate the attention coefficients of the lightweight mobile vision module. ; The temporal aggregated features and spatial features are fused across modes. The injection intensity of high temperature accumulation information is controlled by a gating mechanism. The sequence is folded back to the spatial dimension and residual connections are performed. Irrigation accessibility maps are constructed using water conservancy facilities. , where nodes For grid cells, edges To establish irrigation canal connections, a graph attention network encoding is performed, projecting and broadcasting crop types to the spatial dimension via a fully connected layer; First, align the attention features and crop broadcasting features of the water conservancy map grid by grid, and then stitch them together according to the channel dimension to output unified dimension fused data.

3. The drought emergency response plan optimization method based on multi-source data fusion according to claim 1, characterized in that, A method for obtaining drought features by capturing drought features from the fused data using a dual-index temporal abrupt change detection algorithm and a weakly supervised convolutional autoencoder includes: Based on the evapotranspiration layer and soil moisture layer in the fused data, a dual-index abrupt change detection mechanism is constructed. The Penman formula is used to calculate the daily potential evapotranspiration from the meteorological data of the fused data, and the ratio of actual evapotranspiration to potential evapotranspiration is calculated. Construct a 7-day sliding window anomaly index; when If the evapotranspiration is below the initial threshold for sudden drought and lasts for 3 days or more, then the evapotranspiration is abnormal; the 5-day difference of the standardized soil moisture index is calculated based on the soil moisture in the fused data. when And currently If the soil moisture decreases rapidly, the soil moisture will decrease rapidly. If both abnormal evapotranspiration and rapid decrease in soil moisture are met, it is determined that a sudden drought has begun. The spatiotemporal feature vector of the start of the sudden drought is extracted by combining the abnormal evapotranspiration and the decrease in soil moisture through logical AND operation. Input fused data and use a 5-layer convolutional coding structure to extract multi-scale drought features layer by layer, outputting latent features; the latent features are 64-dimensional compressed representations that encode the coupling patterns of root water uptake, canopy temperature, and meteorological stress. The decoder reconstructs the input data using a symmetric structure, constrains feature learning, and uses the standardized soil moisture anomaly index as a weakly supervised signal to identify regions where soil moisture deviates extremely from normal levels. The expression is: ； in Weakly supervised labels based on physical priors; the weighted sum of reconstruction loss, physical consistency loss, and sparsity constraints is used as a multi-task loss function. The composite agricultural drought index is calculated based on the encoder, and the weighted convolutional autoencoder features are extracted to obtain implicit learning features, which include the composite drought index, feature activation intensity, and feature uncertainty entropy. The explicit physical features of the physical drought feature channel are adaptively fused with the implicit learning feature map to obtain the sudden drought fusion feature map, and the probability of sudden drought occurrence is calculated based on the implicit learning features.

4. The drought emergency response plan optimization method based on multi-source data fusion according to claim 1, characterized in that, A method for constructing a drought emergency response plan model based on machine learning according to the aforementioned characteristics of sudden drought includes: A three-tiered drought emergency response model is constructed, consisting of prediction, correction, and decision-making layers. The correction layer outputs calibrated soil moisture, which serves as the state input for the decision-making layer, forming a complete link between prediction, correction, and decision-making. The drought emergency response model includes a prediction layer, a correction layer, and a decision-making layer. The prediction layer uses a hybrid convolutional neural network and random forest model to predict the drought development trajectory. The correction layer uses physical soil moisture equations superimposed with machine learning for dynamic correction and isotope tracing to calibrate root parameters. The decision-making layer uses binary dynamic programming to solve for the optimal irrigation time window. The characteristics of sudden drought are input into a hybrid convolutional neural network and random forest model. The convolutional neural network is used to extract spatial heterogeneity features and capture field water transfer patterns. The flattened vector output by the convolutional neural network and historical weather data are used as inputs to train a random forest regressor to predict future soil moisture and drought level. ; The spatial heterogeneity characteristics and predicted soil moisture are integrated and output as follows: When the fusion weight When the value is 0.6, physical prediction is prioritized, and the irrigation demand urgency index is output. ;in This represents the probability of drought risk. To integrate weights, It is a fully connected layer; A physics-data hybrid model is constructed based on Richards equations, with the physical foundation being... Introducing data-driven residual correction terms The correction coefficient is adaptively adjusted according to the prediction uncertainty; Using stable isotope stratified sampling data, the root water absorption depth distribution is inverted, and the water absorption ratio is mapped to the root water absorption term by solving the least squares method, thus correcting the root water absorption term. Decision variables Using binary, when the decision variable is 1, it means that field i is irrigated at time t, and when the decision variable is 0, it means that the field is not irrigated. Given constraints, the constraints include canal water supply time window, soil moisture balance, and emergency response to sudden drought. Define value function To maximize the cumulative benefit under the initial state, a recursive equation is given. The optimal strategy sequence is solved by back induction, and the optimal irrigation time window for each field is output. The optimal irrigation time window, water allocation scheme, and early warning response level are output as decision-making emergency plans.

5. The drought emergency response plan optimization method based on multi-source data fusion according to claim 1, characterized in that, A method for obtaining decision bias fit based on the historical emergency data and the predicted emergency data includes: Historical emergency decisions extracted from historical emergency data and recommended emergency decisions output by a sudden drought feature prediction model are encoded into standardized decision vectors. The weighted Euclidean bias of decisions is calculated based on historical decision variables from historical emergency data and recommended decision variables from predicted emergency data. The standardized decision vectors include time response, water allocation, spatial priority, resource intensity, and risk level. Let a be the importance weight of the a-th decision dimension. To predict the a-th component of the decision vector, Let a be the a-th component of the historical decision vector; Asymmetric penalty is introduced to calculate the time window offset sensitivity for irrigation time response. The structural deviation index is calculated based on the actual decision vector and the recommended decision vector. The historical decision effect score is calculated, and the decision effect correction factor is constructed based on the historical decision effect score. When the historical decision performance score is below the decision performance threshold, a slight deviation can actually indicate potential for improvement. Approaching 0 reduces the penalty for bias; when the historical decision performance score is greater than the decision performance threshold, Tendency towards 1, strict requirement for consistency; Calculate the fit of decision bias The decision bias fit is normalized. When the decision bias fit is ∈ [0.8,1], the historical emergency decision and the recommended emergency decision are consistent. When the decision bias fit is ∈ [0.5,0.8], the historical emergency decision and the recommended emergency decision are biased. When the decision bias fit is less than 0.5, the historical emergency decision and the recommended emergency decision are significantly divergent.

6. The drought emergency response plan optimization method based on multi-source data fusion according to claim 1, characterized in that, A method for dynamic decision optimization based on the decision bias fit using multi-objective reinforcement learning includes: Multi-objective reinforcement learning consists of an upper layer and a lower layer; the upper layer is the leader, which outputs water supply quotas; the lower layer is the follower, which provides feedback on water demand; and the decision bias fit is used as a dynamic feedback signal. The upper layer is a multi-objective game model, with the players being the watershed dispatch center and the set of regional water users; the state space includes reservoir storage, water inflow forecast, decision bias fit degree, and water shortage rate in each region; the action space includes inter-regional water transfer volume, reservoir group dispatch rules, and priority water supply object weights; a multi-objective reward function is constructed. A trapezoidal fuzzy number is introduced to impose fuzzy constraints on engineering capability constraints, and a membership function is given. Employing a multi-agent deep deterministic policy gradient, where the watershed center and regional users are both adversaries and the environment, the Nash equilibrium policy satisfies: ,in The optimal Nash equilibrium strategy for the watershed scheduling center. For the aggregation of regional water users, For expectation operator, This is the transpose of the multi-objective weight vector. For a multi-objective reward function, The strategy to be optimized for the scheduling center; When the fit of decision bias is less than 0.6, increase the weight of water shortage loss; when the fit of decision bias is greater than 0.8, increase the weight of fairness. The lower layer is a fuzzy interactive planning model, with decision variables including irrigation system, crop structure adjustment ratio, and water-saving technology deployment rate. A fuzzy objective function is constructed, and interactive constraints are given based on the fuzzy matching between the upper layer water supply allocation and the lower layer demand. The states in proximal policy optimization reinforcement learning are soil moisture field, crop phenological stage, weather forecast, and decision bias fitness feedback value; the actions in proximal policy optimization reinforcement learning are field-level irrigation decisions and crop replacement suggestions; the reward function for proximal policy optimization reinforcement learning is given as follows: When the fit of decision deviation is greater than the decision deviation threshold, the reward is biased towards output and water productivity; when the fit of decision deviation is less than the decision deviation threshold, the water resource cost penalty is increased, and conservative strategies are forced to be explored. Upper-level planning: The basin center issues an initial water supply strategy based on the current DDFI; Lower-level response: Each irrigation district assesses the fit between the quota and local decision-making deviation, solves the fuzzy interactive programming problem, and provides feedback on the actual water demand; if Supply and demand mismatch, adjustments made at the top Weights, re-solve for the Nash equilibrium until convergence. ; When |allocated water volume - actual water demand| is less than ε, the convergence rolling optimization triggering conditions include hard triggering, soft triggering and periodic triggering; Hard triggering: the decision bias fitness is less than 0.5 and retraining is performed immediately, soft triggering is level jump fine-tuning, and periodic triggering is updated every 7 days; Soft triggering: the drought level increases by 1 level or more and activates policy fine-tuning; Periodic triggering: the value network of the upper-layer multi-agent deep deterministic policy gradient is updated at a fixed period; A model predictive control framework is adopted, with optimization in the time domain for 7 days and control in the time domain for 3 days. Optimization is performed every 3 days, and the strategy for the next 4 days is retained as an initial guess.

7. A drought emergency response plan optimization system based on multi-source data fusion, used to execute the method described in any one of claims 1-6, characterized in that, include: Data acquisition and preprocessing module: used to acquire multi-source data and historical emergency data of a preset area, and to preprocess the multi-source data and historical emergency data; the multi-source data includes meteorological data, soil data, crop data and water conservancy data; The data fusion and feature extraction module is used to input the multi-source data into a lightweight multimodal fusion network to obtain fused data, and to use a dual-index temporal mutation detection algorithm and a weakly supervised convolutional autoencoder to capture drought features in the fused data to obtain drought features. Model building and prediction module: used to build a drought emergency response plan model based on machine learning based on the characteristics of the sudden drought, and to obtain predicted emergency data based on the drought emergency response plan model; Bias adaptation and dynamic optimization module: used to obtain the decision bias adaptation degree based on the historical emergency data and the predicted emergency data, and to perform dynamic decision optimization based on the decision bias adaptation degree using multi-objective reinforcement learning to obtain the optimization result.

Images