Soil drought risk early warning system and early warning method based on machine learning

Abstract

The invention discloses a soil drought risk early warning system and early warning method based on machine learning, and belongs to the field of soil drought early warning, and the system comprises a data collection module, a data restoration and fusion module, a space-time drought coupling prediction model construction module, and a real-time early warning and model updating module. Wherein the data acquisition module is used for acquiring wind cloud meteorological data and MODIS optical image data; the data restoration and fusion module is used for generating an all-weather soil humidity chart; the space-time drought coupling prediction model construction module is used for constructing a space-time drought coupling prediction model; and the real-time early warning and model updating module is used for obtaining the dynamic drought index and dynamically updating the dynamic drought index. By adopting the soil drought risk early warning system and early warning method based on machine learning, high-precision drought risk prediction and decision support are realized through multi-source data fusion, a space-time coupling prediction model and a dynamic grading early warning mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of soil drought early warning, and particularly to a soil drought risk early warning system and method based on machine learning. Background Art

[0002] Traditional drought early warning methods rely on single meteorological indicators or static models, and it is difficult to capture the spatio-temporal heterogeneity and suddenness of soil drought. These methods often only focus on simple meteorological parameters such as precipitation and temperature, ignoring multi-dimensional factors such as soil moisture and vegetation status, resulting in insufficient understanding of the complexity of drought development. The existing technologies have the following limitations:

[0003] 1. Insufficient data integration: Traditional models fail to fully integrate remote sensing soil moisture, meteorological reanalysis data, and vegetation dynamic indicators, lacking effective integration of these multi-source data, resulting in difficulty for prediction models to comprehensively capture the complex mechanism of drought evolution, thus limiting the prediction accuracy.

[0004] 2. Weak model generalization ability: Drought is a complex dynamic process, and its development is not only affected by historical climate conditions but also closely related to neighboring regions in the geographical space. However, although static machine learning models (such as random forests) perform well in processing structured data, they are less adaptable when dealing with the temporal dependence and spatial correlation of drought evolution.

[0005] 3. Poor early warning timeliness: The occurrence and development of drought is a gradual but rapidly changing process, and timely monitoring and early warning are required to take effective measures. However, many early warning systems rely on fixed-cycle data updates and cannot quickly respond to environmental changes, resulting in late warnings and missing the best intervention opportunity. In addition, the lack of support for real-time data streams also limits the iterative optimization of the model, further weakening the prediction performance of the system. That is, the existing systems generally lack a dynamic update mechanism and it is difficult to achieve real-time early warning at the sub-seasonal (1 - 12 weeks) scale. Summary of the Invention

[0006] The purpose of the present invention is to provide a soil drought risk early warning system and method based on machine learning to solve the above technical problems.

[0007] To achieve the above purpose, the present invention provides a soil drought risk early warning system based on machine learning, including a data acquisition module, a data repair and fusion module, a spatio-temporal drought coupling prediction model construction module, and a real-time early warning and model update module;

[0008] Among them, the data acquisition module is used to collect Fengyun meteorological data and MODIS optical image data;

[0009] The data repair and fusion module is used to repair the data in the cloudy area by using the generative adversarial network, and fuse the Fengyun meteorological data and MODIS optical image data to generate an all-weather soil moisture map;

[0010] The spatio-temporal drought coupling prediction model construction module is used to construct a spatio-temporal drought coupling prediction model based on the all-weather soil moisture map;

[0011] The real-time warning and model update module is used to input the real-time collected Fengyun meteorological data and MODIS optical image data into the spatio-temporal drought coupling prediction model to obtain a dynamic drought index, verify the accuracy through ground station data, and realize the dynamic update of the spatio-temporal drought coupling prediction model in combination with the federated learning framework.

[0012] The warning method of the soil drought risk warning system based on machine learning includes the following steps:

[0013] S1. Collect Fengyun meteorological data and MODIS optical image impact data, and perform feature extraction to form a data set;

[0014] S2. Use the generative adversarial network to repair the data in the cloudy area, fuse the Fengyun meteorological data and MODIS optical image impact data collected in step S1, eliminate the cloud pollution interference, and generate an all-weather soil moisture map;

[0015] S3. Based on the all-weather soil moisture map described in step S2, construct a spatio-temporal drought coupling prediction model;

[0016] S4. Input the real-time collected Fengyun meteorological data and MODIS optical image impact data into the spatio-temporal drought coupling prediction model constructed in step S3 to obtain a dynamic drought index;

[0017] S5. Verify the accuracy of the dynamic drought index obtained in step S4 through ground station data, and realize the dynamic update of the spatio-temporal drought coupling prediction model in combination with the federated learning framework.

[0018] Preferably, in step S1, the Fengyun meteorological satellite data includes microwave soil moisture data M sm ;

[0019] The MODIS optical image data includes the vegetation index NDVI and the land surface temperature LST;

[0020] Eliminate noise through wavelet decomposition and extract the drought cycle characteristics of different time scales:

[0021]

[0022] In the formula, W j,k represents the decomposed drought mutation signal and drought cycle trend component; T represents the total time; x(t) represents the original time series; ψj,k (t) represents the discrete wavelet basis function, j represents the decomposition level, and k represents the translation parameter.

[0023] Preferably, step S2 specifically includes the following steps:

[0024] S21. Data preprocessing: Unify the FY meteorological data and MODIS optical image data into the WGS84 coordinate system, and use the MODIS cloud mask to label the cloud-covered areas in the MODIS optical images;

[0025] S22. Construct a generative adversarial network and input historical data for training;

[0026] S23. Repair of cloudy areas: In the cloud-covered areas, replace the original MODIS optical image data with the cloud-free optical images output by the generator of the generative adversarial network, and use Poisson fusion at the edge transition area of the MODIS optical image data to eliminate the stitching traces;

[0027] S24. Fuse the repaired MODIS optical image data and the FY meteorological satellite data, and generate a 1km resolution all-weather soil moisture map SM through a multiple regression model:

[0028] SM = α×DVI + β×M sm + γ×LST + ∈(2);

[0029] In the formula, α, β, and γ all represent regression coefficients; DVI represents the meteorological-vegetation coupling index; ∈ represents the error term;

[0030] Among them,

[0031]

[0032] In the formula, NDVI max and NDVI min represent the maximum and minimum values of the vegetation index respectively; ET represents the actual evapotranspiration, that is, the sum of plant transpiration and surface water evaporation; ET max represents the potential maximum evapotranspiration;

[0033] S25. Verification and iterative optimization: Calculate the root mean square error between the measured soil moisture data at the meteorological station and the soil moisture predicted in step S24, and compare whether the root mean square error is within the set threshold range. If it is, it proves that the all-weather soil moisture map constructed in step S24 passes the verification; otherwise, return to step S22 for update and iteration.

[0034] Preferably, the generator of the generative adversarial network adopts a U-Net structure, and the input of the generator is the cloud-covered MODIS optical image x input and the FY meteorological data M sm, output cloudless optical image x out ;

[0035] The discriminator of the generative adversarial network uses a convolutional neural network to distinguish between the real cloudless optical image x real and the output cloudless optical image x out differences;

[0036] And the total loss function L of the generative adversarial network total The expression is as follows:

[0037] L total = L GAN + λL phy (4);

[0038] In the formula, L GAN represents the adversarial loss; L phy represents the physical constraint loss; λ represents the physical constraint weight;

[0039] Among them,

[0040] L GAN = E[logD(x real )] + E[log(1 - D(G(M sm , x input )))] (5);

[0041] L phy = ||M sm - f inversion (x out )|| (6);

[0042] In the formula, D(x real ) represents the real data output by the discriminator; E represents the expected value operator; D(G(M sm , x input )) represents the probability that false data is judged to be true; f inversion represents the soil moisture inversion function based on the vegetation index NDVI and the land surface temperature LST.

[0043] Preferably, the spatio-temporal coupled drought prediction model described in step S3 is constructed based on a temporal convolutional network and a graph attention network, where the temporal convolutional network is used to extract the long-range dependence of the historical soil moisture X t-L:t :

[0044]

[0045] In the formula, represents the feature vector obtained after the temporal convolutional network extracts the features of the input sequence at time t; TCN represents the extraction operation of the temporal convolutional network; L represents the historical time window;

[0046] The graph attention network is used to model the drought propagation relationship between regions:

[0047]

[0048] Wherein, represents the drought prediction result of region v at time t; N(v) represents the set of neighbor regions of region v; α uv represents the influence intensity of region v on its neighbor region u; represents the historical soil moisture data of region u at time t, which is obtained by temporal convolutional network extraction;

[0049] And the main task objective of the spatio-temporal coupled drought prediction model is set as drought level classification, and the auxiliary task objective is soil moisture numerical regression.

[0050] Preferably, the expression of soil moisture numerical regression is as follows:

[0051]

[0052] Wherein, L 回归 represents the quantity of the regression degree after soil moisture numerical regression; N represents the total number of regions; i represents the index variable; represents the predicted soil moisture value; represents the actually measured soil moisture value.

[0053] Preferably, the expression of the dynamic drought index EDI described in step S4 is as follows:

[0054]

[0055] Wherein, LAI 实际 represents the actual leaf area index; LAI 潜在 represents the potential leaf area index; ρ d represents the logistic regression probability;

[0056] Among them,

[0057] LAI 潜在 = LAI max ×(1 - e -0.6·积温 ) (11);

[0058]

[0059] Wherein, LAI max represents the maximum leaf area index; β0, β1, β2 and β3 all represent regression coefficients.

[0060] Preferably, the expression for realizing model dynamic update in the federated learning framework described in step S5 is as follows:

[0061]

[0062] In the formula, W global represents the global warning value; k represents; K represents the total number of iterations; n k represents the number of data points used in the k-th iteration; represents the warning value calculated in the k-th iteration; ΔW meta represents the weight update amount.

[0063] Therefore, the soil drought risk warning system and warning method based on machine learning adopted by the present invention have the following beneficial effects:

[0064] 1. Sufficient data integration: It integrates multi-source information such as Fengyun meteorological data and MODIS optical image data, effectively improving the capture ability of the prediction model for the complex mechanism of drought evolution;

[0065] 2. Strong model generalization ability: By constructing a spatio-temporal coupled drought prediction model and combining a temporal convolutional network and a graph attention network, it can simultaneously process the temporal dependence and spatial correlation of drought, enhancing the generalization performance of the model;

[0066] 3. Good warning timeliness: The federal learning framework is adopted to realize the dynamic update of the model, support real-time data stream processing, help to quickly respond to environmental changes, and improve the timeliness of warning.

[0067] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0068] Figure 1 is a flowchart of a soil drought risk warning system based on machine learning according to the present invention;

[0069] Figure 2 is an all-weather soil humidity map generated by the simulation experiment of the present invention. Detailed Embodiments

[0070] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the following further describes the embodiments of the present invention in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention, and are not used to limit the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of this application. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout.

[0071] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0072] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0073] A soil drought risk early warning system based on machine learning includes a data acquisition module, a data repair and fusion module, a spatio-temporal drought coupling prediction model construction module, and a real-time early warning and model update module; wherein, the data acquisition module is used to acquire FY meteorological data and MODIS optical image data; the data repair and fusion module is used to repair the data in the cloudy area by using a generative adversarial network and fuse the FY meteorological data and MODIS optical image data to generate an all-weather soil moisture map; the spatio-temporal drought coupling prediction model construction module is used to construct a spatio-temporal drought coupling prediction model based on the all-weather soil moisture map; the real-time early warning and model update module is used to input the real-time acquired FY meteorological data and MODIS optical image data into the spatio-temporal drought coupling prediction model to obtain a dynamic drought index, verify the accuracy through ground station data, and realize the dynamic update of the spatio-temporal drought coupling prediction model in combination with the federated learning framework.

[0074] As Figure 1 shown, the early warning method of the soil drought risk early warning system based on machine learning includes the following steps:

[0075] S1. Acquire FY meteorological data and MODIS optical image data, and perform feature extraction to form a data set;

[0076] In step S1, the FY meteorological satellite data includes microwave soil moisture data M sm ;

[0077] The MODIS optical image data includes the vegetation index NDVI and the land surface temperature LST;

[0078] Eliminate noise through wavelet decomposition and extract drought cycle characteristics at different time scales:

[0079]

[0080] In the formula, W j,k represents the decomposed drought mutation signal and drought cycle trend component; T represents the total time; x(t) represents the original time series; ψ j,k (t) represents the discrete wavelet basis function, j represents the decomposition level, and k represents the translation parameter.

[0081] S2. Use a generative adversarial network to repair data in cloudy areas, fuse the FY meteorological data and MODIS optical image data collected in step S1, eliminate cloud pollution interference, and generate an all-weather soil moisture map;

[0082] Step S2 specifically includes the following steps:

[0083] S21. Data preprocessing: Unify the FY meteorological data and MODIS optical image data into the WGS84 coordinate system, and use the MODIS cloud mask to label the cloud-covered areas in the MODIS optical images;

[0084] S22. Construct a generative adversarial network and input historical data for training;

[0085] S23. Repair of cloudy areas: In the cloud-covered areas, replace the original MODIS optical image data with the cloud-free optical images output by the generator of the generative adversarial network, and use Poisson fusion at the edge transition area of the MODIS optical image data to eliminate the stitching marks;

[0086] S24. Fuse the repaired MODIS optical image data and FY meteorological satellite data, and generate a 1-km resolution all-weather soil moisture map SM through a multiple regression model:

[0087] SM = α × DVI + β × M sm + γ × LST + ∈ (2);

[0088] In the formula, α, β, and γ all represent regression coefficients; DVI represents the meteorological-vegetation coupling index; ∈ represents the error term;

[0089] Among them,

[0090]

[0091] In the formula, NDVI max and NDVI min respectively represent the maximum and minimum values of the vegetation index; ET represents the actual evapotranspiration, that is, the sum of plant transpiration and surface water evaporation; ET max represents the potential maximum evapotranspiration;

[0092] S25. Verification and iterative optimization: Calculate the root mean square error between the measured soil moisture data at the meteorological station and the soil moisture predicted in step S24, and compare whether the root mean square error is within the set threshold range. If it is, it proves that the all-weather soil moisture map constructed in step S24 passes the verification, otherwise return to step S22 for update and iteration.

[0093] Preferably, the generator of the generative adversarial network adopts a U-Net structure, and the input of the generator is the cloud-covered MODIS optical image xinput and the Fengyun meteorological data M sm , output the cloudless optical image x out ;

[0094] The discriminator of the generative adversarial network uses a convolutional neural network to distinguish the real cloudless optical image x real from the output cloudless optical image x out ;

[0095] And the total loss function L of the generative adversarial network total is expressed as follows:

[0096] L total = L GAN + λL phy (4);

[0097] In the formula, L GAN represents the adversarial loss; L phy represents the physical constraint loss; λ represents the physical constraint weight, and λ = 0.5 is taken in this embodiment;

[0098] Among them,

[0099] L GAN = E[logD(x real )] + E[log(1 - D(G(M sm , x input )))] (5);

[0100] L phy = ||M sm - f inversion (x out )|| (6);

[0101] In the formula, D(x real ) represents the real data output by the discriminator; E represents the expected value operator; D(G(M sm , x input )) represents the probability that the fake data is judged to be true; f inversion represents the soil moisture inversion function based on the vegetation index NDVI and the land surface temperature LST.

[0102] S3. Based on the all-weather soil moisture map described in step S2, construct a spatio-temporal drought coupling prediction model;

[0103] The spatio-temporal coupling drought prediction model described in step S3 is constructed based on a temporal convolutional network and a graph attention network, where the temporal convolutional network is used to extract the long-range dependence of the historical soil moisture X t-L:t :

[0104]

[0105] In the formula, represents the feature vector obtained by the temporal convolutional network for feature extraction of the input sequence at time t; TCN represents the extraction operation of the temporal convolutional network; L represents the historical time window;

[0106] The graph attention network is used to model the drought propagation relationship between regions:

[0107]

[0108] In the formula, represents the drought prediction result of region v at time t; N(v) represents the set of neighbor regions of region v; α uv represents the influence intensity of region v on its neighbor region u; represents the historical soil moisture data of region u at time t, which is obtained by extraction of the temporal convolutional network;

[0109] And the main task objective of the spatio-temporal coupled drought prediction model is set as drought level classification, and the auxiliary task objective is soil moisture numerical regression.

[0110] Preferably, the expression of the soil moisture numerical regression is as follows:

[0111]

[0112] In the formula, L 回归 represents the quantity of the regression degree after the soil moisture numerical regression; N represents the total number of regions; i represents the index variable; represents the predicted soil moisture value; represents the actually measured soil moisture value.

[0113] S4. Input the real-time collected Fengyun meteorological data and MODIS optical image data into the spatio-temporal drought coupling prediction model constructed in step S3 to obtain the dynamic drought index;

[0114] The expression of the dynamic drought index EDI described in step S4 is as follows:

[0115]

[0116] In the formula, LAI 实际 represents the actual leaf area index; LAI 潜在 represents the potential leaf area index; ρ d represents the logistic regression probability;

[0117] Among them,

[0118] LAI 潜在 = LAI max ×(1 - e -0.6·积温 ) (11);

[0119]

[0120] Wherein, LAI max represents the maximum leaf area index; β0, β1, β2 and β3 all represent regression coefficients.

[0121] In this embodiment, it is set that when EDI < 0.3, it is set as the red warning level, and the corresponding countermeasures are: starting emergency irrigation and allocating water resources; when 0.3 ≤ EDI < 0.6, it is set as the orange warning level, and the corresponding countermeasures are: restricting non-agricultural water use; when EDI ≥ 0.6, it is set as the yellow warning level, and the corresponding countermeasures are: issuing drought reminders.

[0122] S5. Verify the accuracy of the dynamic drought index obtained in step S4 through ground station data, and realize the dynamic update of the spatio-temporal drought coupling prediction model in combination with the federated learning framework.

[0123] The expression for realizing the dynamic update of the model by the federated learning framework described in step S5 is as follows:

[0124]

[0125] Wherein, W global represents the global warning value; k represents; K represents the total number of iterations; n k represents the number of data points used in the k-th iteration; represents the warning value calculated in the k-th iteration; ΔW meta represents the weight update amount.

[0126] Simulation experiment

[0127] Data preparation: Collect Fengyun meteorological data (including microwave soil moisture data, MODIS optical image data (vegetation index, surface temperature)) in different regions and different time periods (the time span and spatial range of the data should be representative, for example, select data in multiple climate zones and different topographical regions, covering both dry and wet periods in time), and at the same time obtain the measured soil moisture data, actual leaf area index, accumulated temperature, etc. of the corresponding regional ground stations for model training, verification and result comparison.

[0128] Environment setting: Build the software environment (Python programming environment) required for the experiment, including programming languages, deep learning frameworks, and libraries related to data processing and analysis. In terms of hardware, equip a computer with sufficient computing power.

[0129] The all-weather soil moisture map generated by using the warning method described in the present invention is as Figure 2 shown.

[0130] Table 1 Warning accuracy result table

[0131]

[0132] Combined with Table 1 Figure 2 It can be seen that the minimum coincidence degree between the early warning method described in the present invention and the actual data is 85%, and the accuracy meets the requirements, which proves the effectiveness of the present invention.

[0133] And during this process, as the number of training rounds increases, the classification accuracy of drought levels gradually rises. It rises relatively fast in the initial stage and tends to be stable in the later stage, and stabilizes above 85% at the end of training, indicating that the model's ability to classify drought levels is continuously enhanced and tends to be stable.

[0134] At the same time, during the training process, the MSE gradually decreases. It decreases significantly in the early stage, and then the decreasing speed slows down, and finally stabilizes below 0.05, meaning that the accuracy of the model's prediction of soil moisture content values is continuously improving.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A soil drought risk early warning system based on machine learning, characterized by: It includes data acquisition module, data repair and fusion module, spatiotemporal drought coupling prediction model construction module and real-time warning and model update module; Among them, the data acquisition module is used to collect wind and cloud meteorological data and MODIS optical image data; The data restoration and fusion module is used to restore cloudy area data using a generative adversarial network and fuse wind and cloud meteorological data with MODIS optical image data to generate an all-weather soil moisture map. The spatiotemporal drought coupling prediction model construction module is used to construct a spatiotemporal drought coupling prediction model based on the all-weather soil moisture map; The real-time warning and model updating module is used to input the real-time collected wind and cloud meteorological data and MODIS optical image data into the spatiotemporal drought coupling prediction model to obtain a dynamic drought index, verify the accuracy through ground station data, and combine the federated learning framework to realize the dynamic update of the spatiotemporal drought coupling prediction model.

2. The early warning method of the soil drought risk early warning system based on machine learning according to claim 1, characterized in that: The following steps are involved: S1, collect wind and cloud meteorological data and MODIS optical image impact data, and perform feature extraction to form a data set; S2, using the generative adversarial network to repair the cloudy area data, integrating the wind and cloud meteorological data collected in step S1 and the MODIS optical image impact data, eliminating cloud pollution interference, and generating an all-weather soil moisture map; S3, constructing a spatiotemporal drought coupling prediction model based on the all-weather soil moisture map described in step S2; S4, inputting the real-time collected wind and cloud meteorological data and MODIS optical image impact data into the spatiotemporal drought coupling prediction model constructed in step S3 to obtain a dynamic drought index; S5. Verify the accuracy of the dynamic drought index obtained in step S4 through ground station data, and combine the federated learning framework to realize the dynamic update of the spatiotemporal drought coupling prediction model.

3. The early warning method of the soil drought risk early warning system based on machine learning according to claim 2 is characterized in that: In step S1, the Fengyun meteorological satellite data includes microwave soil moisture data M sm ; MODIS optical image data includes vegetation index NDVI and land surface temperature LST; Wavelet decomposition is used to eliminate noise and extract drought cycle characteristics at different time scales: Where W j,k represents the decomposed drought mutation signal and drought cycle trend component; T represents the total time; x(t) represents the original time series; ψ j,k (t) represents the discrete wavelet basis function, j represents the number of decomposition levels, and k represents the translation parameter.

4. The early warning method of the soil drought risk early warning system based on machine learning according to claim 3 is characterized by: Step S2 specifically includes the following steps: S21. Data preprocessing: Unify the wind and cloud meteorological data and MODIS optical image impact data into the WGS84 coordinate system, and use the MODIS cloud mask to annotate the cloud coverage area in the MODIS optical image; S22, build a generative adversarial network and input historical data for training; S23, cloudy area restoration: In the cloud-covered area, the cloud-free optical image output by the generator of the generative adversarial network is used to replace the original MODIS optical image data. At the same time, Poisson fusion is used to eliminate the splicing traces in the edge transition area of ​​the MODIS optical image data. S24, integrate the repaired MODIS optical image data and Fengyun meteorological satellite data, and generate a 1km resolution all-weather soil moisture map SM through a multivariate regression model: SM=α×DVI+β×M sm +γ×LST+∈ (2); In the formula, α, β, and γ all represent regression coefficients; DVI represents the meteorological-vegetation coupling index; ∈ represents the error term; in, In the formula, NDVI max and NDVI min They represent the maximum and minimum values ​​of the vegetation index respectively; EET represents actual evapotranspiration, which is the sum of plant transpiration and surface water evaporation; ET max Represents the potential maximum evapotranspiration; S25, verification and iterative optimization: calculate the root mean square error between the soil moisture data measured by the meteorological station and the soil moisture predicted in step S24, and compare whether the root mean square error is within the set threshold range. If so, it proves that the all-weather soil moisture map constructed in step S24 has passed the verification, otherwise return to step S22 for update and iteration.

5. The early warning method of the soil drought risk early warning system based on machine learning according to claim 4 is characterized in that: The generator of the generative adversarial network adopts the U-Net structure, and the input of the generator is the cloud-covered MODIS optical image x input Hefengyun Weather Data M sm , output cloud-free optical image x out ; The discriminator of the generative adversarial network uses a convolutional neural network to distinguish between real cloud-free optical images x real Output cloud-free optical image x out differences; And the total loss function L of the generated adversarial network total The expression is as follows: THE total =L GAN +λL phy (4); Where, L GAN Indicates resistance loss; L phy represents the physical constraint loss; λ represents the physical constraint weight; in, L GAN =E[logD(x real )]+E[log(1-D(G(M sm ,x input )))] (5); L phy =||M sm -f inversion (x out )|| (6); In the formula, D(x real ) represents the real data output by the discriminator; E represents the expected value operator; D(G(M sm ,x input )) represents the probability of false data being judged as true; f inversion Represents the soil moisture inversion function based on the vegetation index NDVI and the land surface temperature LST.

6. The early warning method of the soil drought risk early warning system based on machine learning according to claim 5 is characterized in that: The spatiotemporal coupled drought prediction model described in step S3 is constructed based on a temporal convolutional network and a graph attention network, wherein the temporal convolutional network is used to extract historical soil moisture X t-L:t Long-range dependencies: In the formula, It represents the feature vector obtained after the temporal convolutional network extracts the features of the input sequence at time L; TCN represents the temporal convolutional network extraction operation; L represents the historical time window; Graph attention network is used to model drought propagation relationships among regions: In the formula, represents the drought prediction result of region v at time t; N(v) represents the set of neighboring regions of region v; α uv Indicates the influence strength of region v on its neighbor region u; represents the historical soil moisture data of region u at time t, which is extracted by the time series convolutional network; The main task objective of the spatiotemporal coupled drought prediction model is set as drought grade classification, and the auxiliary task objective is soil moisture numerical regression.

7. The early warning method of the soil drought risk early warning system based on machine learning according to claim 6 is characterized by: The numerical regression expression of soil moisture is as follows: Where, L 回归 It represents the regression degree after soil moisture value regression; N represents the total number of regions; i represents the index variable; represents the predicted soil moisture value; Indicates the actual measured soil moisture value.

8. The early warning method of the soil drought risk early warning system based on machine learning according to claim 7 is characterized in that: The dynamic drought index EDI expression described in step S4 is as follows: Where LAI 实际 Stands for actual leaf area index; LAI 潜在 represents the potential leaf area index; ρ d represents logistic regression probability; in, LAI 潜在 =LAI max ×(1-e -0.6·积温 ) (11); Where LAI max represents the maximum leaf area index; β0, β1, β2 and β3 represent regression coefficients.

9. The early warning method of the soil drought risk early warning system based on machine learning according to claim 8 is characterized by: The federated learning framework described in step S5 implements the dynamic update expression of the model as follows: Where W global represents the global warning value; k represents; K represents the total number of iterations; n k represents the number of data points used in the kth iteration; represents the warning value calculated in the kth iteration; ΔW meta Represents the weight update amount.

Images