A method, system and storage medium for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms

Abstract

This application discloses a drought-flood transition prediction method, system, and storage medium that integrates artificial intelligence and physical mechanisms. The method includes: collecting multi-source data within a selected drought-flood transition study area; optimizing key factors affecting terrestrial water storage at each grid point within the study area; constructing and calibrating a temporal convolutional network deep learning model; using a long series of key factor sets to drive the deep learning model and reconstruct a long series of terrestrial water storage inversion datasets; extracting atmospheric circulation characteristics during drought-flood transition events and drought-flood transition events to construct a series of atmospheric circulation-drought-flood transition intensity pairings; deriving a nonlinear scaling relationship between drought-flood transition intensity and atmospheric circulation characteristics; and predicting the intensity and risk of drought-flood transitions at each grid point in the future. This invention provides a new perspective for drought-flood transition prediction, has important scientific significance for disaster prevention and mitigation, and provides a theoretical and technical foundation for assessing the environmental and hazard effects of Earth system evolution.

Description

Technical Field

[0001] The present application relates to the technical field of climate prediction, and in particular to a method, system, and storage medium for predicting sudden transitions from drought to flood that integrate artificial intelligence and physical mechanisms. Background Art

[0002] Global climate change is altering the energy budget and water cycle of the land-atmosphere system. Extreme climate disasters such as heat waves, droughts, and floods are becoming more frequent, posing significant challenges to the sustainable development of socioeconomic systems and the ecological environment. my country is one of the regions most severely affected by floods and droughts. With warming rates far exceeding the global average, temperatures could rise by 4°C by the end of this century, posing a serious threat to my country's flood control, water supply, food security, energy security, and ecological and environmental security. A deeper understanding of the socioeconomic impacts of floods and droughts under climate change scenarios is crucial for future drought and flood risk prediction, disaster prevention and mitigation, and adaptation management.

[0003] Continuous high temperatures and droughts cause soil dehydration, leading to loose soil or hardening and caking. This leads to a significant accumulation of atmospheric energy, making even the slightest disturbance highly susceptible to severe convective weather and concentrated, short-term, heavy rainfall. This can easily trigger flash floods and extreme disasters caused by the superposition of multiple hazard factors. In recent years, droughts, accompanied by a period of no or low precipitation, have been followed by a short period of intense rainfall and flooding. This phenomenon, known as drought-flood abrupt transition, is known as a drought-flood abrupt transition. Due to the rapid nature of the transition, the short duration, and the high rainfall intensity, it can easily trigger flash floods, sudden river surges, and river intrusions. This abrupt transition poses significant challenges to flood and water disaster prevention, necessitating early prevention. However, how to monitor and predict future drought-flood abrupt transitions remains unclear.

[0004] In March 2002, the Gravity Recovery and Climate Experiment (GRACE) satellite was successfully launched, providing a continuous, high-precision, direct observational tool for global, large-scale mass transport on the Earth's surface. The gravity field model derived from GRACE satellite signals can extract information on variations in Earth's lunar gravity field on a spatial scale of 300 km × 300 km. By eliminating the influence of factors such as crustal material movement, atmospheric motion, ocean currents, and tides, it effectively reflects gravity variations caused by ice and snow, surface water, soil water, groundwater, and human factors, and comprehensively monitors terrestrial water storage anomalies (TWSA) signals. In May 2018, following the one-year hiatus of the GRACE gravity satellites, the GRACE-FO (GRACE Follow-On) gravity satellite was successfully launched, continuing the GRACE mission. The GRACE / GRACE-FO gravity satellites effectively address the challenges of shallow ground observation ranges, uneven spatial distribution, and difficulty in acquiring data, demonstrating great potential for tracking regional and global flood and drought events.

[0005] Although GRACE / GRACE-FO satellites are beginning to be used for drought and flood monitoring and assessment, their short observation timeframes make retrospective drought and flood risk assessment over longer timescales difficult. Currently, research on predicting future droughts based on terrestrial water storage anomalies is still relatively new internationally, and reports in China are even rarer. Furthermore, these studies typically focus on changes in a single disaster type, such as drought or waterlogging, with limited consideration of the evolution of drought-flood transitions, and even less on the full potential of artificial intelligence technology for drought-flood transition forecasting. Summary of the Invention

[0006] In view of this, the present application provides a method, system and storage medium for predicting sudden transitions from drought to flood that integrate artificial intelligence and physical mechanisms, which can effectively and accurately combine artificial intelligence and atmospheric thermal dynamic mechanisms to predict sudden transitions from drought to flood disasters under climate change.

[0007] In a first aspect, the present invention provides a method for predicting drought-flood transitions by integrating artificial intelligence and physical mechanisms, comprising the following steps:

[0008] Step 1: Select a drought-flood transition study area and collect multi-source data within the study area, including meteorological and hydrological data, terrestrial water storage data inverted by the GRACE gravity satellite, atmospheric circulation data, large-scale climate indices, and CMIP6 global climate model output data;

[0009] Step 2: Calculate saturated vapor pressure deficit and specific humidity; use a multi-scale spatial feature extraction method, XGBoost model, and spatial attention mechanism to select the key factors affecting terrestrial water storage at each grid point in the study area;

[0010] Step 3: Based on the selected key factors and the terrestrial water storage data inverted by the GRACE gravity satellite, a temporal convolutional network deep learning model is constructed and calibrated for simulating terrestrial water storage; a long series of key factor sets are used to drive the deep learning model to reconstruct a long series of terrestrial water storage inversion datasets;

[0011] Step 4: Based on the long series of terrestrial water storage inversion datasets reconstructed in Step 3, we combined the run theory to extract drought-flood transition events. We then applied the empirical orthogonal function analysis method to process atmospheric circulation data and large-scale climate indices, extracted atmospheric circulation characteristics during drought-flood transition events, and constructed a paired series of atmospheric circulation and drought-flood transition intensity.

[0012] Step 5: Based on the bin scaling function, the atmospheric circulation characteristics and drought-flood transition intensity pairs of all grid points in the study area are divided into M bins with equal sample sizes, and the nonlinear scaling relationship between drought-flood transition intensity and atmospheric circulation characteristics is derived;

[0013] Step 6: Based on the CMIP6 global climate model output data and the nonlinear scaling relationship obtained in Step 5, predict the intensity and risk of drought-flood transitions at each grid point in the future. Emergence constraint technology is used to reduce the uncertainty of the prediction, and a quantitative regression method is used to provide a probabilistic prediction of the risk of drought-flood transitions.

[0014] Furthermore, the meteorological and hydrological data in step 1 include: 2m air temperature, dew point temperature, near-ground pressure, precipitation, soil moisture, water vapor flux divergence, cloud cover, shortwave radiation and leaf area index, 500hPa geopotential height field, 200hPa horizontal wind field in the ERA5 reanalysis dataset; impervious area at each grid point on a monthly scale.

[0015] Furthermore, the step 2 includes:

[0016] The Clausius-Clapeyron thermodynamic equation was used to calculate the saturation vapor pressure deficit based on the 2-meter air temperature and dew point temperature in the ERA5 reanalysis dataset, and the specific humidity was calculated using the near-surface pressure and dew point temperature.

[0017] 2m air temperature, specific humidity, saturated vapor pressure deficit, precipitation, soil moisture, water vapor flux divergence, cloud cover, shortwave radiation, leaf area index, and impervious area were selected as driving factors;

[0018] For each driving factor, the average value within multiple sliding windows is calculated. The driving factors of multiple months are selected as the initial input variables of the XGBoost model. The spatial attention mechanism is introduced to allow the XGBoost model to automatically learn the importance weights of different positions and rank the importance of each input variable to the simulation variable.

[0019] Set a threshold and select variables within the ranking threshold range as key factors.

[0020] Furthermore, the step 3 includes:

[0021] Build and calibrate a temporal convolutional network deep learning model to simulate terrestrial water storage;

[0022] Based on the key factors selected in step 2, the long series of data sets are used to drive the temporal convolutional network deep learning model established in step 3 to reconstruct a long series of terrestrial water storage inversion data sets.

[0023] Furthermore, the step 4 includes:

[0024] Based on the terrestrial water storage data in the long series of terrestrial water storage inversion dataset reconstructed in step 3, calculate the drought intensity based on terrestrial water storage anomalies;

[0025] Set up a standardized index to measure the intensity of drought-flood transition events;

[0026] Combined with the travel theory, drought thresholds and flood thresholds were set respectively. Based on the drought intensity of the abnormal land water storage, whether it is a dry month or a flood month was determined. The drought-flood transition event judgment criteria were set to extract drought-flood transition events and calculate the drought-flood transition event intensity measurement index.

[0027] The atmospheric circulation field data corresponding to the drought-flood transition event were extracted to construct data matrices, and the characteristic roots of the covariance matrix were calculated and sorted. The atmospheric circulation field data with variance contribution rate and cumulative variance contribution rate greater than the set threshold were selected as the main atmospheric circulation characteristics of the drought-flood transition event.

[0028] The drought-flood transition intensity measurement index of each grid point in the forecast area is paired with the drought-flood transition intensity measurement index of all previous drought-flood transition events to form an atmospheric circulation-drought-flood transition intensity pairing series.

[0029] Furthermore, the step 5 includes:

[0030] The atmospheric circulation characteristics and drought-flood abrupt transition intensity paired series of all grid points in the study area are used as inputs to the box element scaling method;

[0031] Respectively characterize the atmospheric circulation characteristics of the box element and the average intensity of drought-flood transition events; obtain the average value of atmospheric circulation characteristics of group M - the average intensity of drought-flood transition;

[0032] The nonlinear scaling relationship between the intensity of drought-flood transition and atmospheric circulation characteristics is deduced based on the average value of atmospheric circulation characteristics and the average intensity of drought-flood transition.

[0033] Furthermore, the step 6 includes:

[0034] At each grid point in the study area, for each global climate model under each global shared economic pathway, calculate the intensity of drought-flood abrupt transitions in the future period and extract atmospheric circulation characteristic data for the future period;

[0035] Using the nonlinear scaling relationship obtained in step 5, a preliminary forecast of the intensity of drought-flood abrupt transitions in the future period is made;

[0036] For each grid point in the study area, the trend of the principal component comprehensive score in the historical period of each global climate model is calculated;

[0037] By integrating data from all grid points in the study area under multiple global climate models, we can obtain a pairing of the historical period's principal component comprehensive score change trend and the intensity of drought-flood abrupt transitions in the future period.

[0038] An emergence constraint model was constructed. Based on the reanalysis data in the ERA5 reanalysis dataset and the large-scale climate index provided by NOAA, the trend of the PCS principal component comprehensive score at each grid point was derived and the trend term was substituted into the emergence constraint model.

[0039] The predicted results of the intensity of drought-flood transition in the future at all grid points in the study area under multiple global climate models are collected and probability distribution fitting is performed to calculate the probability of occurrence of drought-flood transition events of different intensity levels.

[0040] In a second aspect, the present invention provides a drought-flood transition prediction system that integrates artificial intelligence and physical mechanisms, comprising:

[0041] Data acquisition module: It is used to select the drought-flood transition research area and collect multi-source data in the research area, including: meteorological and hydrological data, terrestrial water storage data inverted by GRACE gravity satellite, atmospheric circulation data, large-scale climate indices and CMIP6 global climate model output data;

[0042] Key Factor Optimization Module: This module is used to calculate saturated vapor pressure deficit and specific humidity. It uses a multi-scale spatial feature extraction method, XGBoost model, and spatial attention mechanism to optimize the key factors affecting terrestrial water storage at each grid point in the study area.

[0043] A dataset reconstruction module is used to construct and calibrate a temporal convolutional network deep learning model for simulating terrestrial water storage based on the preferred key factors and terrestrial water storage data inverted by the GRACE gravity satellite; and to drive the deep learning model with a long series of key factor sets to reconstruct a long series of terrestrial water storage inversion datasets.

[0044] Pairing module: This module is used to extract drought-flood transition events based on a reconstructed long-series terrestrial water storage inversion dataset combined with run theory. It also applies empirical orthogonal function analysis methods to process atmospheric circulation data and large-scale climate indices, extracting atmospheric circulation characteristics in drought-flood transition events and constructing a pairing series of atmospheric circulation and drought-flood transition intensity.

[0045] Scaling relationship derivation module: It is used to divide the atmospheric circulation characteristics and drought-flood abrupt change intensity pairs of all grid points in the study area into M bins with equal sample capacity based on the bin scaling function, and derive the nonlinear scaling relationship between drought-flood abrupt change intensity and atmospheric circulation characteristics;

[0046] The risk prediction module is used to predict the intensity and risk of drought-flood transitions at each grid point in the future based on the CMIP6 global climate model output data and the nonlinear scaling relationship obtained in step 5. It uses emergent constraint technology to reduce the uncertainty of the prediction and provides a probabilistic drought-flood transition risk prediction through quantitative regression methods.

[0047] In a third aspect, the present invention provides a computer-readable medium, wherein the computer-readable medium is a server workstation;

[0048] The server workstation stores a computer program executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of the drought-flood rapid transition prediction method integrating artificial intelligence and physical mechanisms according to any one of claims 1 to 7.

[0049] Compared with existing technologies, the present invention has the following beneficial effects: This application provides a method, system, and storage medium for predicting sudden drought-flood transitions that integrates artificial intelligence and physical mechanisms. This method collects multi-source data, optimizes key factors using the XGBoost model and spatial attention mechanism, and constructs a temporal convolutional network model in conjunction with GRACE satellite data to reconstruct long series of terrestrial water storage data. Run-length theory and EOF analysis are further applied to extract drought-flood transition events and atmospheric circulation characteristics. Nonlinear scaling relationships are derived using a box element scaling function, and finally, the risk of future sudden drought-flood transitions is predicted based on CMIP6 data and an ensemble learning model. This method innovatively combines multi-source data with advanced algorithms, providing a new perspective for predicting sudden drought-flood transitions. This method has important scientific significance for disaster prevention and mitigation, and provides a theoretical and technical foundation for assessing the environmental and hazard effects of Earth system evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1 Provides a specific flow chart of the estimation method for the embodiment of the present application;

[0052] Figure 2 Schematic diagram of the spatial sliding window method. DETAILED DESCRIPTION

[0053] To make the above-mentioned objects, features, and advantages of the present application more clearly understood, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the scope of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0054] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules clearly listed, but may include other steps or modules that are not clearly listed or that are inherent to these processes, methods, products or devices. The naming or numbering of steps in this application does not mean that the steps in the method flow must be executed in the time / logical sequence indicated by the naming or numbering. The process steps that have been named or numbered can be changed in the execution order according to the technical purpose to be achieved, as long as the same or similar technical effects can be achieved. The division of units described in this application is a logical division. In actual implementation, other divisions may be used. For example, multiple units may be combined or integrated into another system, or some features may be ignored or not implemented. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through interfaces, and the indirect coupling or communication connection between units may be electrical or other similar forms, all of which are not limited in this application. Furthermore, the units or subunits described as separate components may or may not be physically separated, may or may not be physical units, or may be distributed across multiple circuit units. Some or all of these units may be selected based on actual needs to achieve the objectives of this application.

[0055] Example 1

[0056] This embodiment provides a method for predicting sudden transitions from drought to flood that integrates artificial intelligence and physical mechanisms. The method collects multi-source data, selects key factors through the XGBoost model and spatial attention mechanism, and constructs a temporal convolutional network model in combination with GRACE satellite data to achieve reconstruction of long series of terrestrial water storage data; further applies run theory and EOF analysis to extract sudden transitions from drought to flood events and atmospheric circulation characteristics; derives nonlinear scaling relationships through box element scaling functions, and finally predicts future sudden transitions from drought to flood based on CMIP6 data and the XGBoost model. This method innovatively combines multi-source data and advanced algorithms to provide a new perspective for the prediction of sudden transitions from drought to flood, has important scientific significance for disaster prevention and mitigation, and provides a theoretical and technical basis for evaluating the environmental and disaster effects of the evolution of the Earth system. The specific process is detailed in [1]. Figure 1 .

[0057] The technical solution of the present application is further described below through embodiments and in conjunction with the accompanying drawings:

[0058] Based on the above embodiment, step 1 includes:

[0059] First, a set of meteorological, hydrological, vegetation and land use data since 1950 was collected, including monthly meteorological, hydrological and atmospheric circulation data from the fifth generation atmospheric reanalysis dataset (ERA5) of the European Centre for Medium-Range Weather Forecasts, specifically including 2m air temperature, dew point temperature, air pressure, precipitation, soil moisture, water vapor flux divergence, cloud cover, shortwave radiation and leaf area index (LAI), 500hPa geopotential height field, and 200hPa horizontal wind field; the land use dataset uses land use datasets published by international and domestic scholars. The present invention extracts the monthly scale impervious surface area (ISA) of each grid point using Google Earth Engine technology.

[0060] Satellite gravity data is used to invert terrestrial water storage. Internationally, major institutions such as the Potsdam Geoscience Center in Germany, the Jet Propulsion Laboratory (JPL) at the California Institute of Technology, the Center for Space Research at the University of Texas, Austin (CSR), and NASA's Goddard Space Flight Center (GSFC) are responsible for interpreting the data and publishing monthly-scale global gravity field output data. This paper simultaneously uses the latest sixth-generation (RL06) products released by JPL, CSR, and GSFC. These three data sources have different spatial resolutions and all provide mascon monthly-scale equivalent water heights (excluding the 2004-2009 average field) based on the lumped-mass gravity field solution, ultimately outputting a long-sequence TWSA dataset. To account for the uncertainties that may arise from different data products, the three GRACE / GRACE-FO gravity satellite datasets were interpolated to a 0.25° × 0.25° spatial grid. The average of each product was taken within each time step, ultimately resulting in a monthly dataset of terrestrial water storage from 2002 to the present.

[0061] We further acquired a large-scale climate index data set from 1950 to the present, including monthly time series of the Pacific Decadal Oscillation (PDO), Arctic Oscillation (AO), North Atlantic Oscillation (NAO), and three ENSO (El Nino Southern Oscillation) indices provided by the Physical Sciences Laboratory (PSL) and Climate Prediction Center (CPC) of the National Oceanic and Atmospheric Administration (NOAA), including the NINO 1+2 region sea surface temperature anomaly index, the NINO 3 region sea surface temperature anomaly index, and the NINO 3.4 region sea surface temperature anomaly index.

[0062] To project future climate scenarios, five recently released global climate models from the Coupled Model Inter-comparison Project Phase 6 (CMIP6) were used: GFDL-ESM4, IPSL-CM6A-LR, MPI-ESM1-2-HR, MRI-ESM2-0, and UKESM1-0-LL. Compared to its predecessor (CMIP5), CMIP6 utilizes a matrix framework based on Shared Socioeconomic Pathways (SSPs) and Representative Concentration Pathways (RCPs). This implementation uses model output data from the highest greenhouse gas emissions scenario, SSP5-8.5. The output variables of the five global climate models include daily temperature data, with the historical period from 1985 to 2014 (a total of 30 years). To project meteorological characteristics under future climate change scenarios, the final 30-year window of this century (2071-2100) was selected as the future period.

[0063] Based on the above embodiment, step 2 includes:

[0064] The Clausius-Clapeyron thermodynamic equation can quantitatively describe the saturated water vapor pressure e sat Nonlinear relationship with temperature T:

[0065]

[0066] Among them, T0 and e s0 L is the first integral constant and the second integral constant, which are 273.16K and 611Pa respectively; vis the latent heat of vaporization constant, which is 2.5×10 6 J kg -1 ; R v is the water vapor gas constant, which is 461 J kg -1 K -1 .

[0067] Dew point temperature represents the temperature when air is cooled to water vapor saturation under the conditions of constant water vapor content and air pressure. Substituting it into the Clausius-Clapeyron equation can measure the actual water vapor pressure. 2m ) and dew point temperature (T dew ) are substituted into formula (1) to deduce the near-saturation vapor pressure deficit

[0068] VPD=e sat (T 2m )-e sat (T 2m )

[0069] Specific humidity q is the ratio of water vapor mass to the total mass of the air mass. It is derived using ERA5 ground pressure p and dew point temperature. The formula is as follows:

[0070]

[0071] Further, if Figure 2 As shown in the figure, a multi-scale spatial feature extraction method is used. For each grid point, 3×3, 5×5, and 7×7 spatial sliding window thresholds are set simultaneously, and the window is slid sequentially along these intervals. This multi-scale method can capture information at different spatial scales, improving the robustness of machine learning models and their sensitivity to local and regional scale changes.

[0072] The XGBoost algorithm is used to construct a relationship model between the driving factors in the above multi-scale spatial range and the GRACE gravity satellite terrestrial water storage. The driving factors include air temperature, specific humidity, saturated water vapor pressure deficit, precipitation, soil moisture, water vapor flux divergence, cloud cover, shortwave radiation, leaf area index, and impervious area (a total of 10 driving factors). For each driving factor, the average value in 3×3, 5×5, and 7×7 sliding windows is calculated to form 30 spatial features (10 factors × 3 sliding window averages). The input data of the XGBoost model are all GRACE inversion data period (April 2002 to present), and the time lag effect of each driving factor on terrestrial water storage is taken into account. For each month, the driving factors of that month and the previous two months are selected as model input. For this embodiment, a total of 30×3=90 variables are used as initial model input. At the same time, the spatial attention mechanism is introduced to allow the model to automatically learn the importance weights of different positions. The spatial attention mechanism can be expressed as:

[0073] A=softmax(f(X))Y=A☉X (3)

[0074] Where X is the input spatial feature, a four-dimensional tensor [B, T, HW, C], where B is the batch size, T is the time step (the current month plus the previous two months, a total of 3 months), HW is the number of spatial grid points in the study area, and C is 30 features (10 driving factors × 3 sliding window sizes), containing a total of 90 variables (10 × 3 × 3). f is the attention scoring function, which is used to calculate the importance score of each spatial location and is implemented using a single-layer neural network, that is, f(X) = W*X+b, where W is the weight matrix and b is the bias vector. A is the attention weight matrix, which is obtained by normalizing the output of the attention scoring function using the softmax function. Y is the spatial feature output after the spatial attention mechanism has the same shape as X. Each feature in Y is assigned a different weight, highlighting the factors and spatial locations that are more important for land water storage prediction. ⊙ represents matrix multiplication.

[0075] The XGBoost model ranks the importance of each input variable to the simulation variables and, combined with the weights of the spatial attention mechanism, selects the most important factors affecting terrestrial water storage. In this example, a threshold of 20% is set, meaning that after ranking 90 variables, the top 18 variables are selected as key factors.

[0076] Based on the above embodiment, step 3 includes:

[0077] Based on these optimized key factors, a temporal convolutional network (TCN) deep learning model was constructed using key factors and GRACE gravity satellite data since April 2002. The TCN model's structure includes multiple dilated convolutional layers, residual connections, and skip connections, effectively capturing long-term temporal dependencies.

[0078] The constructed temporal convolutional network deep learning model is used to simulate the terrestrial water storage since April 2002, which is expressed as:

[0079] TWS(t)=F[QM(t),QM(t-1),QM(t-2),QM(t-3)] (4)

[0080] Where TWS(t) represents the terrestrial water storage simulated by the TCN model at time t, QM(t) represents the input variable at time t, QM(t-1) represents the input variable at time t-1, QM(t-2) represents the input variable at time t-2, and QM(t-3) represents the input variable at time t-3; F represents the TCN model.

[0081] Based on the key factors selected in step 2, the long series of data sets since 1940 are used to drive the temporal convolutional network deep learning model established in step 3 to reconstruct the long series of terrestrial water storage data sets since 1940.

[0082] Based on the above embodiment, step 4 includes:

[0083] 4.1 Based on the long series of terrestrial water storage data since 1940 reconstructed in step 3, calculate the long series of drought intensity as follows:

[0084]

[0085] Where DWI(j) represents the drought intensity based on terrestrial water storage anomaly in month j, TWSA j represents the reconstructed terrestrial water storage data for month j, and σ m are the mean and standard deviation of TWSA in month m, where m = 1, 2…, 12, determined according to the calendar month in which month j falls.

[0086] Furthermore, a standardized drought-wetness arupt alternation intensity index (SDWAI) is proposed, as follows:

[0087] SDWAI=[(|DWI i |+|DWI i+1 |)*W*D*S*I] α (6)

[0088] Where: DWI i and DWI i+1 are the drought intensities based on the abnormal land water storage for two consecutive months; W is the weight term, which aims to appropriately increase the asymmetry of drought-flood transition, and W is defined as: W = [max(|DWI i |,|DWI i+1 |) / min(|DWI i |,|DWI i+1 |)] β ,β is the weight adjustment parameter, which is set to 0.5; D is the persistence factor, defined as: D = 1 + ln (1 + N), where N is the number of consecutive drought or flood months; S is the seasonal factor, defined as: S = 1 + 0.5 * sin (2π * m / 12), where m is the month of the event (1-12); I is the intensity change rate, defined as: I = |DWI i+1 -DWI i | / max(|DWI i|,|DWI i+1 |); α is the nonlinear adjustment parameter, which is set to 1.5.

[0089] Combined with the run theory and based on the long series of DWI, the drought threshold and flood threshold are -0.8 and 0.8, respectively, to determine whether it is a dry month or a flood month. The drought-flood transition event is defined as a transition between a dry month and a flood month within two consecutive months. Specifically, it is classified as follows:

[0090] Drought to flood (D to W) event, DWI i <-0.8,DWI i+1 >0.8

[0091] Flood-to-drought (W to D) event, DWI i >0.8,DWI i+1 <-0.8

[0092] Using this method, we extracted drought-flood transition events since 1940 for every grid point in the study area. Furthermore, for each identified drought-flood transition event, we calculated its SDWAI value. A larger SDWAI value indicates a more severe drought-flood transition event. The SDWAI-based classification criteria for drought-flood transition events are shown in Table 1.

[0093] Table 1

[0094]

[0095]

[0096] 4.2 Apply the EOF (Empirical Orthogonal Function) analysis method to process atmospheric circulation data and large-scale climate indices, extract the atmospheric circulation characteristics during drought-flood transition events, and construct a pairing series of atmospheric circulation and drought-flood transition intensity. The specific steps are as follows:

[0097] For each month of drought-flood abrupt transition events since 1940, we extracted the corresponding atmospheric circulation data, including the 500hPa geopotential height field, the 200hPa horizontal wind field, the Pacific Decadal Oscillation (PDO) index, the Arctic Oscillation (AO) index, the North Atlantic Oscillation (NAO) index, and the three El Niño Southern Oscillation (ENSO) indices from 1940 to the present. For each atmospheric circulation data and climate index, we constructed a data matrix X(n×m), where n is the number of spatial grid points (for climate indices, n=1) and m is the number of time samples. We then calculated the covariance matrix to obtain the square matrix C. n×n , as follows:

[0098]

[0099] Then calculate the square matrix C n×nThe characteristic roots (λ1,…,λ n ) and the eigenvector V n×n , both satisfy

[0100] C n×n ×V n×n =V n×n ×∧ n×n (8)

[0101] Where ∧ is an n×n dimensional diagonal matrix, that is

[0102]

[0103] Arrange the characteristic roots λ from large to small, that is, λ1>λ2>…>λ m ≥0, the eigenvector corresponding to each non-zero eigenvalue is an EOF mode. Projecting EOF onto the data matrix X, we can obtain the time coefficients (principal components) corresponding to all spatial eigenvectors, that is,

[0104]

[0105] Furthermore, the variance contribution rate and cumulative variance contribution rate of each mode are calculated. The top p main EOF modes are selected so that their cumulative variance contribution rate reaches 80%. The selected p EOF modes are regarded as the main atmospheric circulation characteristics of the drought-flood abrupt transition event. The calculation formulas for the variance contribution rate and cumulative variance contribution rate are as follows:

[0106]

[0107]

[0108] The SDWAI index of each grid point in the forecast area is paired with the p principal component scores of the corresponding time to form an atmospheric circulation-drought-flood transition intensity pairing series. The pairing format is as follows:

[0109] {time, event type, SDWAI, EOF_PC1, EOF_PC2, ..., EOF_PC p}

[0110] Where: EOF_PC1, EOF_PC2, ..., EOF_PC p They represent the scores of the first p EOF principal components of the corresponding time, EOF_PC1>EOF_PC2>…>EOF_PC p .

[0111] Based on the above embodiment, step 5 includes:

[0112] Binning scaling is a method that effectively captures nonlinear relationships between variables and has been widely used in recent years to analyze the climate response to extreme hydrological events. This step uses the atmospheric circulation characteristics and drought-flood transition intensity pairs for all grid points within the study area as input to the binning scaling method.

[0113] Specifically, the principal component scores (EOF_PC1, EOF_PC2, ..., EOF_PCp) of the main EOF modes obtained in step 4 are used, and the principal component synthetic score (PCS) is used as the ranking criterion:

[0114] PCS=w1*EOF_PC1+w2*EOF_PC2+...+Wp*EOF_PCp(13)

[0115] Where: wi is the weight of the i-th principal component, and its corresponding variance contribution rate is used.

[0116] For each grid point in the study area, we pair the calculated PCS value with the previously calculated SDWAI value to form a series of PCS-SDWAI pairs. Each pair contains the following information:

[0117] {time, event type, PCS, SDWAI}

[0118] Combine the PCS-SDWAI pairing series of all grid points and sort them by PCS value from small to large. Divide these pairs into 12 bins of equal sample size; calculate the median PCS of each PCS in each bin median and the median SDWAI median , respectively characterizing the atmospheric circulation characteristics of the box element and the average intensity of drought-flood transition events, and obtaining 12 sets of atmospheric circulation characteristics-drought-flood transition intensity series.

[0119] Further, using these 12 groups (PCS median ,SDWAI median ) data points, the following formula is used to measure and predict the response function of the drought-flood transition intensity to the changes in atmospheric circulation characteristics in the study area:

[0120] SDWAI2=SDWAI1*(1+α SDWAI *δPCS) β (14)

[0121] Where: SDWAI1 and SDWAI2 are the SDWAI of adjacent bins respectively. median ; α SDWAIis the scaling factor (Scalingrate), which indicates the response strength; δPCS is the PCS between adjacent bins median β is a power exponent used to capture nonlinear relationships. Specifically, β>1 indicates an upward convex (convex function) relationship, and O<β<1 indicates a downward concave (concave function) relationship.

[0122] The local weighted sliding regression smoothing method is used to fit equation (14). Previous studies have found that the climate response types of extreme hydrological events can be divided into monotonically rising type, monotonically falling type and Hook structure (the inflection point is defined as the critical point, PCS critical ). The present invention first identifies the climate response type and PCS critical , and then derive the scaling coefficient α by the least squares method SDWAI and power index β. In this process, only PCS is used <PCS critical The data were fitted in bins of the data to ensure the stability and reliability of the obtained nonlinear scaling relationship.

[0123] Based on the above embodiment, step 6 includes:

[0124] 6.1 For each global climate model under each global shared economy pathway at each grid point in the study area, the drought-flood transition intensity (SDWAII) for the future period (2070–2100) was calculated using Equation (6). Atmospheric circulation characteristic data for the future period were extracted, the corresponding EOF mode was calculated, and the principal component score (PCS) for the future period was calculated using Equation (13). Using the nonlinear scaling relationship obtained in Step 5, a preliminary prediction of the drought-flood transition intensity for the future period is as follows:

[0125] SDWAI fut =SDWAI his *(1+α*δPCS fut ) β (15)

[0126] Where SDWAI fut SDWAI represents the average intensity of drought-flood transition in the future period under a global climate model at this grid point. his The average value of drought-flood transition intensity in the historical period obtained based on the reconstructed DWI index at this grid point, δPCS fut is the PCS change value in the future period relative to the historical period, and α and β are the scaling coefficient and power exponent obtained in step 5.

[0127] For each grid point in the study area, the trend of PCS variation in the historical period under each global climate model (denoted as PCS trendFurthermore, the data of all grid points in the study area under the five global climate models were integrated to obtain the trend of PCS in the historical period and the trend of SDWAI in the future period. fut The emergence constraint model constructed by the pairing combination is as follows:

[0128] SDWA Ifut =a*PCS trend +b (16)

[0129] Where: a and b represent the parameters of the emergence constraint model;

[0130] The least squares method is used to solve the parameters a and b of the emergence constraint model.

[0131] Furthermore, based on the reanalysis data of the ERA5 dataset and the large-scale climate index provided by NOAA, the changing trend of the PCS of each grid point during 1985-2014 was derived, and this trend term was substituted into the constructed emergence constraint model, as follows:

[0132] SDWAI fut-corr =a*OBS trend +b (17)

[0133] Where: SDWAI fut-corr The OBS represents the forecast results of the intensity of drought-flood transition in the future after correction. trend It represents the observed PCS trend at this grid point calculated based on ERA5 reanalysis data and large-scale climate indices provided by NOAA.

[0134] Finally, the prediction results of drought-flood abrupt transition intensity of all grid points in the study area in the future period under the five global climate models SDWAI fut-corr , set, perform probability distribution fitting, and calculate the probability of occurrence of drought-flood sudden transition events of different intensity levels, as follows:

[0135] P(SDWAI fut-corr >threshold)=1-CDF(threshold)

[0136] Where: CDF is the cumulative distribution function, threshold is the predetermined intensity threshold (multiple thresholds can be set to correspond to different levels, see Table 1 for details).

[0137] Example 2

[0138] This embodiment provides a drought-flood transition prediction system that integrates artificial intelligence and physical mechanisms, including:

[0139] Data acquisition module: It is used to select the drought-flood transition research area and collect multi-source data in the research area, including: meteorological and hydrological data, terrestrial water storage data inverted by GRACE gravity satellite, atmospheric circulation data, large-scale climate indices and CMIP6 global climate model output data;

[0140] Key Factor Optimization Module: This module is used to calculate saturated vapor pressure deficit and specific humidity. It uses a multi-scale spatial feature extraction method, XGBoost model, and spatial attention mechanism to optimize the key factors affecting terrestrial water storage at each grid point in the study area.

[0141] A dataset reconstruction module is used to construct and calibrate a temporal convolutional network deep learning model for simulating terrestrial water storage based on the preferred key factors and terrestrial water storage data inverted by the GRACE gravity satellite; and to drive the deep learning model with a long series of key factor sets to reconstruct a long series of terrestrial water storage inversion datasets.

[0142] Pairing module: This module is used to extract drought-flood transition events based on a reconstructed long-series terrestrial water storage inversion dataset combined with run theory. It also applies empirical orthogonal function analysis methods to process atmospheric circulation data and large-scale climate indices, extracting atmospheric circulation characteristics in drought-flood transition events and constructing a pairing series of atmospheric circulation and drought-flood transition intensity.

[0143] Scaling relationship derivation module: It is used to divide the atmospheric circulation characteristics and drought-flood abrupt change intensity pairs of all grid points in the study area into M bins with equal sample capacity based on the bin scaling function, and derive the nonlinear scaling relationship between drought-flood abrupt change intensity and atmospheric circulation characteristics;

[0144] The risk prediction module is used to predict the intensity and risk of drought-flood transitions at each grid point in the future based on the CMIP6 global climate model output data and the nonlinear scaling relationship obtained in step 5. It uses emergent constraint technology to reduce the uncertainty of the prediction and provides a probabilistic drought-flood transition risk prediction through quantitative regression methods.

[0145] Example 3

[0146] This embodiment provides a computer-readable medium.

[0147] The computer readable medium is a server workstation;

[0148] The server workstation stores a computer program executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of the drought-flood rapid transition prediction method based on artificial intelligence and satellite remote sensing in an embodiment of the present invention.

[0149] The above is only a preferred specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed in this application should be covered by the scope of protection of the present application.

[0150] It should be understood that parts not elaborated in detail in this specification belong to the prior art.

[0151] It should be understood that the above description of the preferred embodiment is relatively detailed and cannot be regarded as limiting the scope of protection of the patent of the present invention. Under the guidance of the present invention, ordinary technicians in this field can also make substitutions or modifications without departing from the scope of protection of the claims of the present invention, which all fall within the scope of protection of the present invention. The scope of protection requested by the present invention shall be based on the attached claims.

Claims

1. A method for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms, characterized in that: The following steps are involved: Step 1: Select a drought-flood transition study area and collect multi-source data within the study area, including meteorological and hydrological data, terrestrial water storage data inverted by the GRACE gravity satellite, atmospheric circulation data, large-scale climate indices, and CMIP6 global climate model output data; Step 2: Calculate saturated vapor pressure deficit and specific humidity; use a multi-scale spatial feature extraction method, XGBoost model, and spatial attention mechanism to select the key factors affecting terrestrial water storage at each grid point in the study area; Step 3: Based on the selected key factors and the terrestrial water storage data inverted by the GRACE gravity satellite, a temporal convolutional network deep learning model is constructed and calibrated for simulating terrestrial water storage; a long series of key factor sets are used to drive the deep learning model to reconstruct a long series of terrestrial water storage inversion datasets; Step 4: Based on the long series of terrestrial water storage inversion datasets reconstructed in Step 3, we combined the run theory to extract drought-flood transition events. We then applied the empirical orthogonal function analysis method to process atmospheric circulation data and large-scale climate indices, extracted atmospheric circulation characteristics during drought-flood transition events, and constructed a paired series of atmospheric circulation and drought-flood transition intensity. Step 5: Based on the bin scaling function, the atmospheric circulation characteristics and drought-flood transition intensity pairs of all grid points in the study area are divided into M bins with equal sample sizes, and the nonlinear scaling relationship between drought-flood transition intensity and atmospheric circulation characteristics is derived; Step 6: Based on the CMIP6 global climate model output data and the nonlinear scaling relationship obtained in Step 5, predict the intensity and risk of drought-flood transitions at each grid point in the future. Emergence constraint technology is used to reduce the uncertainty of the prediction, and a quantitative regression method is used to provide a probabilistic prediction of the risk of drought-flood transitions.

2. The method for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms according to claim 1, characterized in that: The meteorological and hydrological data in step 1 include: 2m air temperature, dew point temperature, near-ground pressure, precipitation, soil moisture, water vapor flux divergence, cloud cover, shortwave radiation and leaf area index, 500hPa geopotential height field, 200hPa horizontal wind field in the ERA5 reanalysis dataset; and impervious area at each grid point on a monthly scale.

3. The method for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms according to claim 2, characterized in that: The step 2 includes: The Clausius-Clapeyron thermodynamic equation was used to calculate the saturation vapor pressure deficit based on the 2-meter air temperature and dew point temperature in the ERA5 reanalysis dataset, and the specific humidity was calculated using the near-surface pressure and dew point temperature. 2m air temperature, specific humidity, saturated vapor pressure deficit, precipitation, soil moisture, water vapor flux divergence, cloud cover, shortwave radiation, leaf area index, and impervious area were selected as driving factors; For each driving factor, the average value within multiple sliding windows is calculated. The driving factors of multiple months are selected as the initial input variables of the XGBoost model. The spatial attention mechanism is introduced to allow the XGBoost model to automatically learn the importance weights of different positions and rank the importance of each input variable to the simulation variable. Set a threshold and select variables within the ranking threshold range as key factors.

4. The method for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms according to claim 1, characterized in that: The step 3 includes: Build and calibrate a temporal convolutional network deep learning model to simulate terrestrial water storage; Based on the key factors selected in step 2, the long series of data sets are used to drive the temporal convolutional network deep learning model established in step 3 to reconstruct a long series of terrestrial water storage inversion data sets.

5. The method for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms according to claim 1, characterized in that: The step 4 comprises: Based on the terrestrial water storage data in the long series of terrestrial water storage inversion dataset reconstructed in step 3, calculate the drought intensity based on terrestrial water storage anomalies; Set up a standardized index to measure the intensity of drought-flood transition events; Combined with the travel theory, drought thresholds and flood thresholds were set respectively. Based on the drought intensity of the abnormal land water storage, whether it is a dry month or a flood month was determined. The drought-flood transition event judgment criteria were set to extract drought-flood transition events and calculate the drought-flood transition event intensity measurement index. The atmospheric circulation field data corresponding to the drought-flood transition event were extracted to construct data matrices, and the characteristic roots of the covariance matrix were calculated and sorted. The atmospheric circulation field data with variance contribution rate and cumulative variance contribution rate greater than the set threshold were selected as the main atmospheric circulation characteristics of the drought-flood transition event. The drought-flood transition intensity measurement index of each grid point in the forecast area is paired with the drought-flood transition intensity measurement index of all previous drought-flood transition events to form an atmospheric circulation-drought-flood transition intensity pairing series.

6. The method for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms according to claim 5, characterized in that: The step 5 comprises: The atmospheric circulation characteristics and drought-flood abrupt transition intensity paired series of all grid points in the study area are used as inputs to the box element scaling method; Respectively characterize the atmospheric circulation characteristics of the box element and the average intensity of drought-flood transition events; obtain the average value of atmospheric circulation characteristics of group M - the average intensity of drought-flood transition; The nonlinear scaling relationship between the intensity of drought-flood transition and atmospheric circulation characteristics is deduced based on the average value of atmospheric circulation characteristics and the average intensity of drought-flood transition.

7. The method for predicting sudden drought-flood transitions by integrating artificial intelligence and physical mechanisms according to claim 5, characterized in that: The step 6 comprises: At each grid point in the study area, for each global climate model under each global shared economic pathway, calculate the intensity of drought-flood abrupt transitions in the future period and extract atmospheric circulation characteristic data for the future period; Using the nonlinear scaling relationship obtained in step 5, a preliminary forecast of the intensity of drought-flood abrupt transitions in the future period is made; For each grid point in the study area, the trend of the principal component comprehensive score in the historical period of each global climate model is calculated; By integrating data from all grid points in the study area under multiple global climate models, we can obtain a pairing of the historical period's principal component comprehensive score change trend and the intensity of drought-flood abrupt transitions in the future period. An emergence constraint model was constructed. Based on the reanalysis data in the ERA5 reanalysis dataset and the large-scale climate index provided by NOAA, the trend of the PCS principal component comprehensive score at each grid point was derived and the trend term was substituted into the emergence constraint model. The predicted results of the intensity of drought-flood transition in the future at all grid points in the study area under multiple global climate models are collected and probability distribution fitting is performed to calculate the probability of occurrence of drought-flood transition events of different intensity levels.

8. A drought-flood rapid transition prediction system integrating artificial intelligence and physical mechanisms, characterized by: include: Data acquisition module: It is used to select drought-flood rapid transition research areas and collect multi-source data in the research areas, including: meteorological and hydrological data, terrestrial water storage data inverted by GRACE gravity satellite, atmospheric circulation data, large-scale climate indices and CMIP6 global climate model output data; Key Factor Optimization Module: This module is used to calculate saturated vapor pressure deficit and specific humidity. It uses a multi-scale spatial feature extraction method, XGBoost model, and spatial attention mechanism to optimize the key factors affecting terrestrial water storage at each grid point in the study area. A dataset reconstruction module is used to construct and calibrate a temporal convolutional network deep learning model for simulating terrestrial water storage based on the preferred key factors and terrestrial water storage data inverted by the GRACE gravity satellite; and to drive the deep learning model with a long series of key factor sets to reconstruct a long series of terrestrial water storage inversion datasets. Pairing module: This module is used to extract drought-flood transition events based on a reconstructed long-series terrestrial water storage inversion dataset combined with run theory. It also applies empirical orthogonal function analysis methods to process atmospheric circulation data and large-scale climate indices, extracting atmospheric circulation characteristics in drought-flood transition events and constructing a pairing series of atmospheric circulation and drought-flood transition intensity. Scaling relationship derivation module: It is used to divide the atmospheric circulation characteristics and drought-flood abrupt change intensity pairs of all grid points in the study area into M bins with equal sample capacity based on the bin scaling function, and derive the nonlinear scaling relationship between drought-flood abrupt change intensity and atmospheric circulation characteristics; The risk prediction module is used to predict the intensity and risk of drought-flood transitions at each grid point in the future based on the CMIP6 global climate model output data and the nonlinear scaling relationship obtained in step 5. It uses emergent constraint technology to reduce prediction uncertainty and provides a probabilistic drought-flood transition risk prediction through quantitative regression methods. The drought-flood transition prediction system integrating artificial intelligence and physical mechanisms is used to execute the steps in the drought-flood transition prediction method integrating artificial intelligence and physical mechanisms described in any one of claims 1-7.

9. A computer-readable medium, characterized in that The computer readable medium is a server workstation; The server workstation stores a computer program executed by an electronic device. When the computer program runs on the electronic device, the electronic device executes the steps of the drought-flood rapid transition prediction method integrating artificial intelligence and physical mechanisms according to any one of claims 1 to 7.

Images