A crop storm disaster quantitative risk assessment method based on a CNN-LSTM hybrid neural network

Abstract

The application discloses a crop storm disaster quantitative risk assessment method based on a CNN-LSTM hybrid neural network, which comprises the following steps: obtaining multi-source data, performing standardization processing, performing time-space superposition to construct a three-dimensional feature matrix, extracting a spatial feature vector and a time sequence feature vector of disaster data based on a CNN-LSTM hybrid neural network model, completing fusion to obtain a fused feature vector, outputting a risk probability and a loss intensity, continuously considering crop loss caused by the risk probability and the loss intensity, completing construction of a storm disaster-crop coupling loss function, and finally generating a crop storm disaster risk probability, a main crop loss intensity and a storm disaster-crop coupling total loss based on intelligent grid precipitation and wind speed prediction data. The application comprehensively estimates crop quantitative loss based on a storm disaster-crop coupling dynamic response, and realizes effective transformation from static disaster assessment to dynamic multi-source fusion response assessment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of meteorological science, and particularly relates to a crop rainstorm disaster quantitative risk assessment method based on a CNN-LSTM hybrid neural network. BACKGROUND

[0002] The frequency and destructiveness of rainstorm disasters have significantly increased, and the threat to agricultural production is increasingly severe. Traditional risk assessment methods mostly rely on historical disaster statistics or static meteorological threshold models, and cannot effectively integrate multi-source heterogeneous data (such as high-resolution remote sensing images, soil moisture time series data, etc.), nor can they capture the complex nonlinear relationship of “meteorology-environment-crop response” in the rainstorm disaster chain. For example, the traditional model based on index weight method often ignores the dynamic changes of surface water bodies and the sensitivity differences of crop growth periods, resulting in insufficient spatial resolution and relatively poor timeliness of risk assessment results.

[0003] Deep learning technology provides a breakthrough tool for rainstorm disaster risk assessment. Studies have shown that machine learning algorithms such as random forest (RF) and support vector machine (SVM) are significantly better than traditional statistical methods in rainstorm disaster risk assessment, especially in handling high-dimensional meteorological data and nonlinear relationships. However, existing researches mostly focus on single disaster factors (such as rainfall intensity) or static risk assessment, and have not fully integrated spatio-temporal coupling characteristics and crop physiological response mechanisms. For example, the sensitivity of rice to waterlogging during the heading stage is 3-5 times higher than that during the seedling stage, and traditional models often ignore such key agronomic parameters. SUMMARY

[0004] The present application provides a precipitation forecast spatio-temporal verification method based on target object diagnosis to better meet the needs of fine meteorological forecast services.

[0005] Technical scheme: In order to achieve the above application purpose, the application adopts the following technical scheme: a crop rainstorm disaster quantitative risk assessment method based on a CNN-LSTM hybrid neural network, comprising the following steps:

[0006] S1, data preparation and preprocessing, obtaining multi-source data, and standardizing the multi-source data to obtain multi-source grid data, the multi-source data including meteorological data, soil moisture content data, crop growth monitoring data, geographic environment data and historical disaster data; the meteorological data including real-time and forecast precipitation, real-time and forecast wind speed;

[0007] S2, spatiotemporal superposition is performed on the multi-source grid data to construct a three-dimensional feature matrix X, X=[T, Lat, Lon, C], wherein T is the number of time steps, Lat is the latitude, Lon is the longitude, and C is the number of channels; the channels include real-time hourly precipitation, real-time daily precipitation, real-time process cumulative precipitation, forecast hourly precipitation, forecast daily precipitation, forecast process cumulative precipitation, real-time hourly maximum wind speed, forecast hourly maximum wind speed, soil moisture content 10 cm below the ground, soil moisture content 20 cm below the ground, soil moisture content 40 cm below the ground, NDVI index, crop growth period code, slope, TDI index, crop disaster area ratio, crop disaster area ratio, crop disaster area ratio, crop yield reduction rate, and agricultural economic loss;

[0008] S3, a CNN-LSTM hybrid neural network model is constructed, spatial feature vectors of meteorological data, geographical environment data, crop growth data and disaster data are extracted based on a CNN neural network model ; time sequence feature vectors of precipitation and wind speed data sequences, soil moisture content, crop growth data and disaster data are extracted based on an LSTM model containing an attention mechanism ; multi-source features of spatial features and time sequence features extracted by the CNN neural network model and the LSTM model are fused to complete model construction to obtain a fusion feature vector , output risk probability and loss intensity data;

[0009] S4, considering the crop loss caused by the risk probability and the loss intensity together, a loss function of a storm disaster-crop coupling is constructed ;

[0010] S5, according to the model of S3 and the loss function of S4, based on intelligent grid precipitation and wind speed prediction data, crop storm disaster risk probability, main crop loss intensity and storm disaster-crop coupling total loss in a future storm process are generated.

[0011] Further, the data preparation and preprocessing of step S1 first performs the following processing on the multi-source data:

[0012] (1) normalization of meteorological data: first, the meteorological data is down-scaled to obtain 1km×1km meteorological grid data, and the meteorological grid data is normalized respectively; the meteorological data includes real-time hourly precipitation, real-time daily precipitation, real-time process cumulative precipitation, real-time hourly maximum wind speed, forecast hourly precipitation, forecast daily precipitation, forecast process cumulative precipitation, and forecast hourly maximum wind speed;

[0013] (2) The three-layer soil moisture data is processed by downscaling to obtain the three-layer soil relative humidity with a spatial resolution of 1 km x 1 km, wherein the soil relative humidity is the ratio of the current moisture content to the field water capacity; the three-layer soil relative humidity includes the soil relative humidity of 10 cm below the surface, 20 cm below the surface and 40 cm below the surface;

[0014] (3) The NDVI index obtained by satellite is processed into daily NDVI index, and the crop growth period information is converted into categorical variable data according to the growth period of the crop to form growth period stage coding data;

[0015] (4) The slope is obtained by the following formula and the microtopographic drainage capacity index TDI,

[0016] ,

[0017] In the formula, represents the elevation standard deviation, represents the horizontal distance, slope represents the slope degree obtained according to the slope analysis in Gis, represents the total runoff within the catchment area, represents the catchment flow per unit area;

[0018] Then, the three-layer soil relative humidity, daily NDVI index, slope tan(slope) and microtopographic drainage capacity index TDI of each grid are respectively standardized;

[0019] Finally, the disaster data at the county scale is processed into standardized grid data with a spatial resolution of 1 km*1 km according to the crop planting area and planting distribution.

[0020] Further, in the three-dimensional feature matrix X in step S2, Lat=3, Lon=3; the time step T=13, including 5 days before the occurrence of the rainstorm disaster to 7 days after the occurrence of the rainstorm disaster; the number of channels C=20.

[0021] Further, step S3 of constructing the CNN-LSTM hybrid neural network model specifically includes the following steps:

[0022] S3.1 The CNN neural network model respectively obtains the spatial feature vectors of the meteorological data, geographical environment data, crop growth data and disaster data, and the spatial feature vector sequence of the live hourly precipitation, live daily precipitation, live process cumulative precipitation, live hourly maximum wind speed, three-layer soil relative humidity and daily NDVI index at each time step, to obtain a 128-dimensional spatial feature vector ;

[0023] The LSTM model comprising the attention mechanism described in S3.2 comprises 13 time steps, and the attention mechanism is applied respectively, and the weight of each attention mechanism is calculated by the following formula, and the LSTM model comprising the attention mechanism is constructed,

[0024] ,

[0025] In the formula, is the weight coefficient of the output value of the tth time step, represents the weight function at the tth time step, represents the weight function at the ith time step, T represents the number of time steps, and takes the value of 13; is the hyperbolic tangent activation function, is the weight matrix, is the bias term, is the transpose of the input vector at this time step; represents the output vector at the tth time step;

[0026] Then, the time sequence feature vector is extracted by the LSTM model comprising the attention mechanism,

[0027] ,

[0028] In the formula, and are the weight coefficient and the output vector at the tth time step, respectively, is a 256-dimensional time sequence feature vector;

[0029] S3.3 fuses the 128-dimensional spatial feature vector extracted in steps S3.1 and S3.2 respectively and the 256-dimensional time sequence feature vector , and then obtains the final multi-source fused feature vector by the attention mechanism weighting,

[0030] S3.4 According to the multi-source fused feature vector obtained in steps S3.1-S3.3, the model finally outputs the risk probability and the loss intensity two variables, wherein: the risk probability is obtained by a binary classification method, and the value range is 0-1; the loss intensity is obtained by the following formula,

[0031] ,

[0032] In the formula, is the basic yield reduction rate, and the value range is 0-1; a growth period sensitive coefficient of a current growth stage of the crop, is the final multi-source fusion feature vector.

[0033] Further, the step S4 comprises the following steps:

[0034] Firstly, the loss L of the risk probability is calculated by using the binary cross entropy (BCE) method based on the risk probability prob，

[0035] ,

[0036] In the formula, L prob represents the loss of the risk probability, is the probability of the model predicting the occurrence of disasters for the sample, is the risk probability;

[0037] Then, the loss L of the loss intensity is calculated by using the mean square error (MSE) of the growth period sensitive coefficient of the crop , which is calculated by the following formula,

[0038] ,

[0039] In the formula, and are the real loss intensity of the sample and the predicted loss intensity of the sample respectively, and N represents the number of samples;

[0040] Finally, the total loss of the storm disaster-crop coupling is established ,

[0041] ,

[0042] In the formula, is a hyperparameter for balancing the two factor terms, and the value is 0.7.

[0043] Further, in the step S5, the crop storm disaster risk probability, the main crop loss intensity and the total loss of the storm disaster-crop coupling are generated by using the intelligent grid precipitation and wind speed prediction data, and the distribution maps of the risk probability, the loss intensity and the total loss of the storm disaster-crop coupling with a spatial resolution of 1km x 1km are drawn.

[0044] Advantages: Compared with the prior art, the present application has the following advantages:

[0045] (1) By using the method of the present application, multi-source data such as meteorological data, soil data, crop data and geographic information are used to construct the model, the dynamic response of the storm disaster-crop coupling to the quantitative loss estimation of the crop is comprehensively considered, and the dynamic capture of the precipitation-soil moisture lag effect is realized, so that the effective transformation from static disaster assessment to dynamic multi-source fusion response assessment is realized.

[0046] (2) The method of the present application constructs a CNN-LSTM hybrid neural network model, uses a deep learning method to consider the effects of various factors in the storm disaster-crop yield reduction disaster chain as much as possible, and finally forms a risk assessment product that greatly enriches existing agricultural meteorological service products. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 It is a general technical flowchart of the present application.

[0048] Figure 2 It is a schematic diagram of the CNN-LSTM hybrid network architecture described in the present application. DETAILED DESCRIPTION

[0049] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that these embodiments are only used to illustrate the present application and not to limit the scope of the present application. After reading the present application, those skilled in the art can make various equivalent modifications to the present application, which fall within the scope defined by the claims attached hereto.

[0050] In view of the above problems and challenges existing in the prior art, the present application proposes a main crop storm disaster quantitative risk assessment method based on a CNN-LSTM hybrid neural network. By integrating observation data of precipitation, strong wind, soil moisture observation data and crop growth monitoring data (NDVI index), a "air-land-farm" three-in-one input feature matrix is constructed, a multi-source data fusion architecture is constructed, and the spatial resolution can reach 1 kilometer, breaking through the limitations of traditional grid scale evaluation. The CNN branch is used to extract the spatial features of farmland micro-topography (such as slope and drainage capacity), the LSTM branch is used to capture the time sequence law of rainfall process and soil water migration, and the attention mechanism is used to dynamically weight the key nodes of the disaster chain, which significantly enhances the response ability of the model to short-time heavy rain. The crop growth period sensitivity coefficient and the disaster loss elasticity model are introduced, the traditional regression problem is converted into a "risk probability-loss intensity" double output prediction task, and the disaster-crop coupling loss function is innovatively used to embed the vulnerability theory in the field of agricultural disaster into the deep learning framework, solving the defect that the existing model is insufficient in representing the physiological characteristics of crops.

[0051] This study reveals the feedback mechanism between the rainstorm disaster chain and the crop system, promotes the cross-innovation of agricultural disaster science and artificial intelligence, and develops a "disaster chain dynamics model" that can simulate the multi-level progressive effects of "rainstorm → waterlogging → soil erosion → crop yield reduction", filling the theoretical gap in the existing risk assessment system. At the same time, it provides an interpretable decision support tool for precise pricing of agricultural insurance and optimal pre-disaster resource scheduling (such as drainage equipment deployment). By integrating deep learning frontier technology with agricultural disaster science theory, not only does it promote the paradigm shift of risk assessment from "static threshold" to "dynamic process", but also provides an innovative solution for global food security and climate-resilient agriculture.

[0052] As shown in Figure 1 and 2 , the main crop rainstorm disaster quantitative risk assessment method based on the CNN-LSTM hybrid neural network of the application comprises the following steps:

[0053] S1: Preprocessing of multi-source data such as meteorological, soil, crop, and geographic information to provide a data foundation for subsequent model construction, specifically including:

[0054] Meteorological data is downscaled to 1km x 1km grid, including 8 variables such as real-time hourly precipitation, real-time daily precipitation, real-time cumulative precipitation during rainstorm process, real-time daily maximum wind speed of each observation site, 3-hourly forecast precipitation, daily forecast precipitation, forecast cumulative precipitation during rainstorm process, and forecast daily maximum wind speed of each grid point, which are normalized respectively.

[0055] Soil moisture data is processed into 1km x 1km spatial resolution grid data, and the daily soil relative humidity (current moisture content / field water holding capacity) of each grid is calculated.

[0056] The NDVI index obtained through MODIS satellite is time-interpolated to obtain daily continuous NDVI index, and the crop growth period information is converted into categorical variable data (such as jointing stage, heading stage, flowering stage, etc.), obtaining growth stage coding data.

[0057] The slope (Slope, degree) and micro-topographic drainage capacity index (TDI) of each grid are calculated using digital elevation model (DEM).

[0058] Slope calculation formula: ,

[0059] Micro-topographic drainage capacity index TDI calculation method: ,

[0060] Wherein, The total runoff in the catchment area can be obtained by processing DEM data, using the ArcGIS hydrological analysis module to determine the surface water flow path, and then calculating the runoff accumulation of each grid according to the water direction data, and converting the runoff accumulation into unit area runoff (SCA) by using the grid calculation tool:

[0061]

[0062] The slope and micro-topographic drainage capacity index of each grid are processed into 1km×1km grid data.

[0063] The daily soil relative humidity, daily NDVI index, slope and micro-topographic drainage capacity index (TDI) of each grid point are standardized.

[0064] According to the crop planting area and distribution, the historical disaster data of county scale are processed into 1km×1km grid data.

[0065] S2: Construct a "air-ground-farm" three-dimensional feature matrix for each 1km×1km grid point. For each 1km×1km grid, a three-dimensional feature matrix X is constructed within a time window T (for example, 5 days before the occurrence of a rainstorm disaster to 7 days after the occurrence, a total of 13 days), X=[T, Lat, Lon, C], wherein T is the time step, Lat is the latitude, Lon is the longitude, and C is the number of channels; the channels include real-time hourly precipitation, real-time daily precipitation, real-time process cumulative precipitation, forecast hourly precipitation, forecast daily precipitation, forecast process cumulative precipitation, real-time hourly maximum wind speed, forecast hourly maximum wind speed, soil moisture content below the ground surface 10cm, soil moisture content below the ground surface 20cm, soil moisture content below the ground surface 40cm, NDVI index, crop growth period code, slope, TDI index, crop disaster area ratio, crop disaster area ratio, crop absolute loss area ratio, crop yield reduction rate and agricultural economic loss. Since the data used is 1km×1km grid data, it is a point in space, but in order to preserve the spatial characteristics, we take the 3×3 neighborhood around the grid, that is, each grid represents the local spatial characteristics with a 3×3 neighborhood, Lat=3, Lon=3; time step T=13, feature channel C=20, including: precipitation, wind speed, soil moisture content, NDVI index, growth period code, slope, TDI index, etc.

[0066] S3: CNN-LSTM hybrid neural network model construction.

[0067] ​The CNN-LSTM hybrid neural network model structure is composed of an input layer, a CNN layer, an LSTM layer and an output layer. By combining CNN and LSTM with attention mechanism, the internal correlation between meteorological data, crop growth monitoring data, geographical environment data, soil moisture data and disaster loss time series data can be fully mined and utilized, so as to obtain quantitative loss assessment results.

[0068] S3.1: Spatial features are extracted by using a convolutional neural network (CNN), that is, a 2D CNN model is used to obtain spatial feature vectors of static data such as geographical environment data (slope and microtopographic drainage capacity index) and disaster data (affected area, disaster area, absolute loss area, yield reduction data and agricultural economic loss), and spatial feature vector sequences of data such as real-time hourly rainfall, real-time daily rainfall, real-time process cumulative rainfall, forecast hourly rainfall, forecast daily rainfall, forecast process cumulative rainfall, real-time hourly maximum wind speed, forecast hourly maximum wind speed, three-layer soil relative humidity, daily NDVI index and the like at each time step. The 2D CNN model is composed of an input layer, a convolution layer, a pooling layer, an activation function, a full connection layer and an output layer. In the convolution layer, a convolution kernel with a kernel number of 64 and a size of 3x3 is used to extract local features, a ReLU (Rectified Linear Unit) activation function is used to improve the problem of small gradient, a 2x2 maximum pooling method is used in the pooling layer to reduce the data dimension, and then a full connection layer is used to obtain a 128-dimensional spatial feature vector of the slope and TDI index and a 128-dimensional spatial feature vector sequence of meteorological data, soil moisture data and crop growth data containing each time step, i.e. a 13x128 vector matrix of the entire time sequence (13 days).

[0069] S3.2: Time sequence features are extracted by using a long short-term memory neural network (LSTM) with attention mechanism. The LSTM model includes an input gate, a forgetting gate and an output gate, adopts a structure containing two LSTM layers, sets 50 LSTM hidden units as the first layer and 32 hidden units as the second layer, and increases a Dropout layer to reduce the risk of overfitting between LSTM layers; the attention mechanism is applied to 13 time steps to calculate the weight of each time step, wherein the attention mechanism weight calculation method is as follows:

[0070] ,

[0071] ,

[0072] wherein, is the transpose of the input vector at the time step, is the output vector at the tth time step, is the hyperbolic tangent activation function, is the weight matrix, is the bias term, is the weight function at the t-th time step, is the weight coefficient of the output value at the t-th time step. The final output of the LSTM model with attention mechanism is the weighted time series feature vector :

[0073] ,

[0074] where, and are the weight coefficient and output at the t-th time step, respectively, is the 256-dimensional time series feature vector.

[0075] S3.3: The spatial features extracted by CNN and the time series features extracted by LSTM are fused to obtain the final multi-source fused feature distribution. The feature vector of the last time step extracted by CNN (128-dimensional) is concatenated with the feature vector of LSTM (256-dimensional) to obtain (384-dimensional), and then the attention mechanism is used to obtain the fused feature vector .

[0076] After multi-source feature fusion, the model outputs two variables, risk probability and loss intensity. The risk probability (range 0-1) is output by the binary classification (disaster occurrence and non-occurrence) method, and the loss intensity (range 0-1) is obtained by combining the crop growth period sensitivity coefficient using a regression model. In the loss intensity calculation process, the fused feature vector passes through a fully connected layer (without activation) to obtain the basic yield reduction rate (range 0-1, representing 0% to 100%), and then the growth period coding of the current grid is combined with the fused feature vector to obtain the sensitivity coefficient (range 0-1, representing 0% to 100%), and then the growth period coding of the current grid is combined with the fused feature vector to obtain the sensitivity coefficient ; the basic yield reduction rate is multiplied again by to obtain the loss intensity , i.e.,

[0077] ,

[0078] ,

[0079] wherein is the fusion feature vector, is the basic yield reduction rate, is determined by the growth period of the crop.

[0080] S4: Establish a storm disaster-crop coupling loss function, the total loss function is composed of two parts, including the loss of risk probability and the loss of loss intensity, the process is as follows:

[0081] S4.1: Calculate the loss of risk probability by using the binary cross entropy (BCE) method :

[0082] ,

[0083] In the formula, is the probability of the model predicting the occurrence of disaster for the sample, is the risk probability.

[0084] Consider the root mean square error (MSE) of the crop growth period sensitivity coefficient to calculate the loss of loss intensity , the present application only calculates the loss intensity loss of the sample actually occurring disaster (i.e. The sample), the loss intensity of the sample not occurring disaster is 0 in theory, so this part of the loss is not calculated (or use 0 loss).

[0085] ,

[0086] In the formula, and are the real loss intensity of the sample and the predicted loss intensity of the sample respectively.

[0087] S4.2: Storm disaster-crop coupling total loss calculation,

[0088] ,

[0089] In the formula, is a hyperparameter, used to balance the two factor terms, generally the experience value is 0.7.

[0090] S5: According to the optimized model, generate crop storm disaster risk probability, main crop loss intensity and storm disaster-crop coupling total loss based on intelligent grid precipitation forecast data, and draw to form a risk probability, loss intensity and storm disaster-crop coupling total loss distribution map with a spatial resolution of 1km×1km.

[0091] The application can also be applied to scenarios including forming the total loss amount at the county scale, and an insurance loss assessment index can also be generated according to the county predicted loss as a reference. The specific method is as follows:

[0092] By combining the crop planting area distribution data, the total loss amount at the county scale can be calculated :

[0093]

[0094] wherein, is the crop area of the unit grid, is the average yield in the past five years, is the predicted yield reduction rate. In addition, the insurance loss assessment index can also be generated according to the county predicted loss as a reference.

[0095]

[0096] wherein, is the average yield of a crop in the county.

[0097] The present application can effectively improve the solution ability of the pain points such as low precision, poor timeliness and weak mechanism in the agricultural disaster assessment through the three innovations of multi-source data fusion, mechanism model embedding and double output architecture.

[0098] Based on the above ideal embodiments according to the present application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the present application. The technical scope of the present application is not limited to the contents of the specification, and the technical scope must be determined according to the scope of claims.

Claims

1. A quantitative risk assessment method for crop rainstorm disasters based on CNN-LSTM hybrid neural networks, characterized in that... Includes the following steps: S1, Data preparation and preprocessing: Acquire multi-source data and standardize the multi-source data to obtain multi-source grid data. The multi-source data includes meteorological data, soil moisture data, crop growth monitoring data, geographical environment data, and historical disaster data. The meteorological data includes actual and forecasted precipitation and actual and forecasted wind speed. S2, the multi-source grid data is spatiotemporally overlaid to construct a three-dimensional feature matrix X, X=[T, Lat, Lon,C], where T is the time step, Lat is the latitude, Lon is the longitude, and C is the number of channels; the channels include actual hourly precipitation, actual daily precipitation, actual cumulative precipitation, forecasted hourly precipitation, forecasted daily precipitation, forecasted cumulative precipitation, actual hourly maximum wind speed, forecasted hourly maximum wind speed, soil moisture content 10cm below the surface, soil moisture content 20cm below the surface, soil moisture content 40cm below the surface, NDVI index, crop growth period code, slope, TDI index, crop disaster area ratio, crop disaster area ratio, crop crop failure area ratio, crop yield reduction rate, and agricultural economic loss. S3. Construct a CNN-LSTM hybrid neural network model, and extract spatial feature vectors from meteorological data, geographical environment data, crop growth data, and disaster data based on the CNN neural network model. Temporal feature vectors of precipitation and wind speed data sequences, soil moisture content, crop growth data, and disaster data are extracted based on an LSTM model incorporating an attention mechanism. The model is constructed by fusing the spatial and temporal features extracted by the CNN neural network model and the LSTM model respectively to obtain the fused feature vector. Output risk probability and loss intensity data; S4. Taking into account both the risk probability and the intensity of loss, a loss function coupling rainstorm disaster and crop loss is constructed. ; S5, based on the model of S3 and the loss function of S4, generates the probability of crop rainstorm disaster risk, the intensity of loss of major crops, and the total coupled loss of rainstorm disaster and crops during future rainstorms based on intelligent grid precipitation and wind speed forecast data; Step S3, constructing the CNN-LSTM hybrid neural network model, specifically includes the following steps: Risk probability is obtained using a binary classification method, with a value ranging from 0 to 1; loss intensity It is obtained from the following formula, ， In the formula, The base production reduction rate, ranging from 0 to 1; This represents the sensitivity coefficient of the crop's current growth stage. It is the final feature vector of multi-source fusion; Step S4 includes the following steps: First, the loss L based on the risk probability is calculated using the binary cross-entropy (BCE) method. prob， ， In the formula, L prob Loss represents the probability of risk. The model is used to predict the probability of a disaster occurring in a sample. Risk probability; Then, the loss intensity is calculated from the root mean square error (MSE) of the crop growth period sensitivity coefficient. Calculated using the following formula, ， In the formula, and These represent the actual loss strength of the sample and the predicted loss strength of the sample, respectively, where N represents the number of samples; Finally, establish the total loss coupled with rainstorm disaster and crop losses. , ， In the formula, It is a hyperparameter used to balance the two factor terms, with a value of 0.

7.

2. The method for quantitative risk assessment of crop rainstorm disasters based on CNN-LSTM hybrid neural networks according to claim 1, characterized in that: The data preparation and preprocessing described in step S1 first involves the following processing of the multi-source data: (1) Normalization of meteorological data: First, the meteorological data is downscaled to obtain 1km×1km meteorological grid data, and then the meteorological grid data is normalized respectively; the meteorological data includes actual hourly precipitation, actual daily precipitation, actual cumulative precipitation, actual hourly maximum wind speed, forecasted hourly precipitation, forecasted daily precipitation, forecasted cumulative precipitation, and forecasted hourly maximum wind speed; (2) The soil moisture content data of the three layers are downscaled to obtain the relative humidity of the three layers of soil with a spatial resolution of 1km×1km. The relative humidity of the soil is the ratio of the current moisture content to the field capacity. The relative humidity of the three layers of soil includes the relative humidity of the soil 10cm, 20cm and 40cm below the surface. (3) The NDVI index obtained by satellite is processed into the daily NDVI index, and the crop growth period information is converted into categorical variable data according to the growth period of the crop to form growth period stage coding data; (4) The slope is obtained by the following formula. and the micro-topography drainage capacity index TDI, ， In the formula, Indicates the standard deviation of elevation. The horizontal distance is represented by 'slope', and the slope is represented by the slope degree obtained from the slope analysis in GIS. This represents the total runoff within the catchment area. This represents the flow rate per unit area. Next, the relative humidity, daily NDVI index, slope (tan), and micro-topographic drainage capacity index (TDI) of the three soil layers in each grid were standardized. Finally, based on the crop planting area and distribution, the disaster data at the county level was processed into standardized grid data with a spatial resolution of 1km*1km.

3. The method for quantitative risk assessment of crop rainstorm disasters based on CNN-LSTM hybrid neural networks according to claim 1, characterized in that: In step S2, the three-dimensional feature matrix X has Lat=3 and Lon=3; the time step T=13, including 5 days before the rainstorm disaster and 7 days after the rainstorm disaster; and the number of channels C=20.

4. The method for quantitative risk assessment of crop rainstorm disasters based on CNN-LSTM hybrid neural networks according to claim 1, characterized in that: The CNN neural network model described in S3.1 acquires spatial feature vectors from meteorological data, geographical environment data, crop growth data, and disaster data, as well as spatial feature vector sequences for each time step, including actual hourly precipitation, actual daily precipitation, actual cumulative precipitation, actual hourly maximum wind speed, relative humidity of three soil layers, and daily NDVI index, resulting in a 128-dimensional spatial feature vector. ; The LSTM model with attention mechanism described in S3.2 contains 13 time steps, each applying an attention mechanism. The weights of each attention mechanism are calculated using the following formula, thus completing the construction of the LSTM model with attention mechanism. ， In the formula, The weighting coefficients for the output value at time step t. This represents the weight function at time step t. Let T represent the weight function at the i-th time step, and let T represent the number of time steps, with a value of 13. The hyperbolic tangent activation function is used. It is a weight matrix. It is a bias term. It is the input vector at this time step. transpose; This represents the output vector at the t-th time step; Then, the temporal feature vector is extracted using an LSTM model that incorporates an attention mechanism. , ， In the formula, and These are the weight coefficients and output vector at the t-th time step, respectively. It is a 256-dimensional temporal feature vector; S3.3 The 128-dimensional spatial feature vectors extracted in steps S3.1 and S3.2 are then processed. and 256-dimensional temporal feature vectors The data is then fused and weighted using an attention mechanism to obtain the final multi-source fused feature vector. , S3.4 Based on the feature vectors obtained from multi-source fusion in steps S3.1-S3.3 The model ultimately outputs the risk probability. and loss intensity Two variables.

5. The method for quantitative risk assessment of crop rainstorm disasters based on CNN-LSTM hybrid neural networks according to claim 1, characterized in that: In step S5, the probability of crop rainstorm disaster risk, the intensity of loss of major crops and the total loss coupled with rainstorm disaster are generated by intelligent grid precipitation and wind speed forecast data, and a 1km×1km spatial resolution distribution map of the probability of risk, the intensity of loss and the total loss coupled with rainstorm disaster is drawn.

Images