Method and system for identifying intensity of compound high-temperature drought event and interpreting driving factors

Abstract

The application discloses a composite high-temperature drought event intensity identification and driving factor interpretation method and system. The method comprises the following steps: collecting multi-source variable sample data of a target area; taking the standardized precipitation evapotranspiration index as a drought index, combining the standardized temperature index and the multi-source variable sample data, constructing a composite high-temperature drought event index including a standardized composite event index and a composite drought high-temperature evapotranspiration index by calculating a joint probability distribution, and defining a composite high-temperature drought event intensity classification rule for dividing intensity categories; constructing and training a machine learning model as a composite high-temperature drought event intensity identification model to obtain intensity identification results of multi-source variable target data of the target area; and obtaining driving factor interpretation results of the intensity identification results by using SHAP. The problems in the prior art, such as inaccurate composite event representation, large identification deviation under unbalanced category conditions, and lack of interpretability of model results, are solved, and the method is suitable for intelligent monitoring and risk identification of complex terrains.

Description

Technical Field

[0001] This application relates to the field of data synchronization technology, and in particular to a method and system for identifying the intensity of complex high-temperature drought events and interpreting driving factors. Background Technology

[0002] Against the backdrop of global warming, the risk of high temperatures, droughts, and their combined events continues to rise. Compared with single high temperatures or droughts, compound high-temperature and drought events (CHTDEs) are often accompanied by the combined effects of processes such as soil moisture deficit, abnormal surface energy distribution, and persistent atmospheric circulation anomalies, exhibiting stronger persistence, wider impact, and amplification of disasters.

[0003] The changes in CHTDEs levels in complex terrain areas are influenced by multiple factors. These factors exhibit nonlinearity, lag, and interaction, making it difficult for traditional linear statistical methods to fully characterize the complex relationship between them and CHTDEs levels.

[0004] Machine learning methods have advantages in high-dimensional feature fusion, nonlinear relationship learning, and complex classification tasks, and have been gradually applied to the monitoring, prediction, and driver identification of extreme weather and climate events. However, model accuracy is not equivalent to scientific explanation. Due to a lack of interpretability analysis, machine learning results tend to remain at the level of "whether a distinction can be made," failing to answer questions such as "which factors contribute to the level distinction" and "whether different intensity levels depend on different driving paths."

[0005] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main purpose of this application is to provide a method and system for identifying the intensity of complex high-temperature and drought events and interpreting their driving factors, so as to solve the problem that the above-mentioned machine learning methods lack interpretability analysis in predicting the intensity of complex high-temperature and drought events in complex terrain areas.

[0007] To achieve the above objectives, this application proposes a method for identifying the intensity of complex high-temperature drought events and interpreting their driving factors, comprising the following steps:

[0008] Collect multi-source variable sample data of the target area;

[0009] Using the standardized precipitation evapotranspiration index SPEI as the drought index DI, and combining the standardized temperature index STI and multi-source variable sample data, a composite high temperature drought event index CHTDI, including the standardized composite SPEI event index SCSEI and the composite drought high temperature evapotranspiration index BDHEI, is constructed by calculating the joint probability distribution, so as to define the intensity classification rules of composite high temperature drought events for classifying intensity categories.

[0010] Based on the Composite High Temperature and Drought Event Index (CHTDI), the intensity classification rules of composite high temperature and drought events, and multi-source variable sample data, a machine learning model was constructed and trained as a composite high temperature and drought event intensity identification model.

[0011] Collect multi-source variable target data of the target area and input them into the composite high temperature and drought event intensity identification model to obtain the intensity identification results corresponding to the multi-source variable target data;

[0012] The SHAP method was used to decompose the output of the composite high temperature and drought event intensity identification model to obtain the driving factor interpretation results.

[0013] Furthermore, the multi-source variable sample data includes:

[0014] Geographic static factors, monthly-scale variables of ERA5 single levels, monthly-scale circulation variables of ERA5 pressure levels, and large-scale ocean-atmosphere and teleconnection indices;

[0015] The geographic static factors include longitude, latitude, and elevation;

[0016] The monthly scale variables of the ERA5 single levels include 2m air temperature t2m, total precipitation tp, first soil volumetric water content swvl1, and evapotranspiration stress index ESI calculated from actual evapotranspiration and potential evapotranspiration.

[0017] The monthly circulation variables of the ERA5 pressure levels include 500hPa and 850hPa geopotential heights, wind field components and vertical velocities, as well as the corresponding high-pressure and subtropical high-pressure related indices.

[0018] The large-scale air-sea and teleconnection indices include AO, NAO, MEI.v2, AMO, IOD, PDO, and the Barents Sea ice concentration index SICBS.

[0019] Furthermore, a method for constructing the composite high-temperature drought event index CHTDI, which includes the standardized composite SPEI event index SCSEI and the composite drought-temperature evapotranspiration index BDHEI, by calculating the joint probability distribution includes:

[0020] Based on multi-source variable sample data, the marginal distributions of the drought index (DI) and the standardized temperature index (STI) are estimated respectively, and the marginal cumulative distribution functions of the drought index and the temperature index are calculated based on the marginal distributions.

[0021] By combining the marginal cumulative distribution function of the drought index and the marginal cumulative distribution function of the temperature index, a joint distribution characteristic is generated;

[0022] Based on the joint distribution characteristics, the joint probabilities corresponding to the standardized composite SPEI event index SCSEI and the composite drought-high temperature evapotranspiration index BDHEI are constructed respectively.

[0023] By using normal quantile transformation to map the joint probability to a standard normal distribution, a composite high-temperature drought event index CHTDI is generated, which includes the standardized composite SPEI event index SCSEI and the composite drought-temperature evapotranspiration index BDHEI.

[0024] Furthermore, based on multi-source variable sample data, the marginal distributions of the drought index (DI) and the standardized temperature index (STI) are estimated separately. Methods for calculating the marginal cumulative distribution functions of the drought index and the temperature index based on these marginal distributions include:

[0025] Based on multi-source variable sample data, the probability density function of the drought index DI(x) was calculated using kernel density estimation. The probability density function of the standardized temperature index STI(y) With probability density function and probability density function The marginal distribution is expressed as follows:

[0026]

[0027]

[0028] Where n is the sample size, and x and y are the specific values ​​to be calculated, representing a possible value between the drought index DI and the standardized temperature index STI. and These are the i-th drought index (DI) and standardized temperature index (STI) values ​​in the sample data, respectively; h is the bandwidth, calculated using Scott's Rule; and K is the kernel function.

[0029] Integrating the marginal distributions yields the marginal cumulative distribution functions u (drought index) and v (temperature index), expressed as:

[0030]

[0031] .

[0033] Furthermore, by combining the Copula function with the marginal cumulative distribution function of the drought index and the marginal cumulative distribution function of the temperature index, a joint distribution characteristic is generated, expressed as:

[0034] ,

[0035] in, is the Copula function, u is the marginal cumulative distribution function of the drought index, and v is the marginal cumulative distribution function of the temperature index;

[0036] The joint probability of the standardized composite SPEI event index SCSEI is:

[0037] ;

[0039] The joint probability corresponding to the combined drought high temperature evapotranspiration index (BDHEI) is:

[0040]

[0041] The expression for the Standardized Composite SPEI Event Index (SCSEI) is:

[0042] The expression for the high-temperature evapotranspiration index in complex drought is:

[0043]

[0044] ,in, F is the inverse function of the standard Gaussian distribution, and F is the empirical Gringorten position partitioning formula.

[0045] Furthermore, the intensity classification rules for the complex high-temperature drought event include:

[0046] Based on the magnitude of the combined drought high temperature evapotranspiration index (BDHEI), the intensity of combined high temperature drought events (CHTDEs) is divided into four intensity categories from Class 0 to Class 3.

[0047] When the combined high-temperature evapotranspiration index (BDHEI) is not greater than -1.5, the combined high-temperature drought event intensity (CHTDEs) is Class 0, indicating severe or worse events.

[0048] When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -1.5 and not greater than -1.0, the combined high-temperature drought event intensity (CHTDEs) is Class 1, indicating a moderate event.

[0049] When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -1.0 and not greater than -0.5, the combined high-temperature drought event intensity (CHTDEs) is Class 2, indicating a mild event.

[0050] When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -0.5, the combined high-temperature drought event intensity (CHTDEs) is Class 3, indicating that no combined event or a weak combined event has occurred.

[0051] The three composite drought high temperature evapotranspiration index (BDHEI) level thresholds of -0.5, -1.0, and -1.5 are further flanked by a ±0.05 buffer zone to remove samples from multi-source variable data that fall into the buffer zone.

[0052] Furthermore, the machine learning model is a tree model, an ensemble model, or a deep learning model, and the method for training the machine learning model includes:

[0053] Multi-source variable sample data are stratified and randomly divided according to intensity category to construct training set and independent test set;

[0054] Parameter optimization is performed only on the training set and uses 5-fold hierarchical cross-validation.

[0055] Furthermore, the machine learning models include RF, XGBoost, LightGBM, CatBoost, CNN, LSTM, and TabNet. The method for constructing and training machine learning models as a composite high-temperature drought event intensity identification model also includes:

[0056] According to the preset model evaluation criteria, the strength recognition test results output by each machine learning model on the independent test set are evaluated. The model evaluation criteria include accuracy, and supplementary evaluation indicators including Kappa coefficient, Matthews correlation coefficient, Jaccard coefficient, Macro-F1 and Weighted-F1.

[0057] Based on the evaluation results, the optimal machine learning model was selected as the intensity identification model for the complex high temperature and drought event in the target area.

[0058] Furthermore, the method for decomposing the output of the composite high-temperature drought event intensity identification model using the SHAP method to obtain the driving factor interpretation results includes:

[0059] Obtain the baseline probability Pbase(k) for intensity category k;

[0060] The baseline probability Pbase(k) is converted into a first probability, expressed as:

[0061]

[0062] And, the SHAP value f is converted from logarithmic odds to second odds, expressed as:

[0063] ;

[0065] Based on the first and second probabilities, the predicted probability of observing intensity category k after considering the influence of target feature values ​​is calculated as follows:

[0066] ;

[0068] The predicted probability is converted into a probability to characterize the predicted probability of observing intensity category k after considering the influence of the target feature value. The expression is:

[0069] ;

[0071] Based on the predicted probability, the SHAP value, expressed in probability units, is calculated as the interpretation result of the driving factor to characterize the contribution of the target feature value to the probability change of the observed intensity category k. The expression is as follows:

[0072] .

[0074] Furthermore, to achieve the above objectives, this application also proposes a system for identifying the intensity of complex high-temperature drought events and interpreting their driving factors, comprising:

[0075] The data acquisition module is used to collect multi-source variable sample data of the target area;

[0076] The intensity classification rule definition module is used to construct the composite high temperature and drought event index CHTDI, which includes the standardized composite SPEI event index SCSEI and the composite drought and high temperature evapotranspiration index BDHEI, by using the standardized precipitation evapotranspiration index SPEI as the drought index DI, combined with the standardized temperature index STI and multi-source variable sample data, and by calculating the joint probability distribution, so as to define the intensity classification rules of composite high temperature and drought events.

[0077] The identification model modeling module is used to construct and train a machine learning model as an identification model for the intensity of composite high temperature and drought events based on the Composite High Temperature and Drought Event Index (CHTDI), the rules for classifying the intensity of composite high temperature and drought events, and multi-source variable sample data.

[0078] The intensity identification module is used to collect multi-source variable target data of the target area and input it into the composite high temperature and drought event intensity identification model to obtain the intensity identification results corresponding to the multi-source variable target data.

[0079] The driving factor interpretation module is used to perform contribution decomposition on the output of the composite high temperature and drought event intensity identification model using the SHAP method to obtain the driving factor interpretation results.

[0080] The method and system for identifying the intensity of complex high-temperature and drought events and interpreting driving factors provided in this application use a Copula function to characterize the joint dependence structure of drought and high temperature. Compared with a single drought index or temperature index, it can more accurately characterize the joint occurrence probability and comprehensive intensity of complex high-temperature and drought events. It solves the technical problem in the prior art that "drought and high temperature are assessed independently and cannot reflect the complex effect". It breaks through the limitations of traditional statistical methods that can only perform event identification or only perform attribution analysis. It realizes the technical integration of "automatic intensity level discrimination" and "quantitative interpretation of key driving factors", making the identification results interpretable.

[0081] By transforming precipitation and temperature variables of different dimensions into a standardized probability space, regional and seasonal differences are eliminated, making the index comparable in time and space. This makes it suitable for unified monitoring of large areas and has constructed a complete technical chain from "monitoring index construction → intensity level identification → driving factor interpretation". This has formed an integrated solution for intelligent monitoring and attribution analysis of complex high temperature and drought events, and has improved the scientific support capability for responding to complex extreme events in complex terrain areas. Attached Figure Description

[0083] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0084] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0085] Figure 1 This is a flowchart illustrating the method for identifying the intensity of complex high-temperature drought events and interpreting driving factors in Embodiment 1 of this application.

[0086] Figure 2 This is a schematic diagram of the TopK set overlap results provided by the method for identifying the intensity of complex high-temperature and drought events and interpreting driving factors in Embodiment 1 of this application;

[0087] Figure 3 This is a schematic diagram illustrating the consistency of importance ranking provided by the method for identifying the intensity of complex high-temperature and drought events and interpreting driving factors in Embodiment 1 of this application;

[0088] Figure 4 This is a schematic diagram of the XGBoost-SHAP global importance ranking provided by the method for identifying the intensity of complex high-temperature and drought events and interpreting driving factors in Embodiment 1 of this application;

[0089] Figure 5 A schematic diagram of the SHAP beeswarm results of the XGBoost model provided by the method for identifying the intensity of compound high temperature and drought events and interpreting driving factors in Embodiment 1 of this application at four compound high temperature and drought levels (Class 0–3);

[0090] Figure 6 This is a schematic diagram of the structure provided for the method of identifying the intensity of complex high-temperature drought events and interpreting driving factors in Embodiment 2 of this application.

[0091] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0093] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0094] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0095] In existing technologies, the changes in CHTDEs levels in complex topographic regions are influenced by multiple factors. For example, anterior precipitation and soil moisture can alter evapotranspiration processes and sensible heat fluxes, thereby affecting the maintenance of high temperatures. The 500 hPa and 850 hPa circulation fields, the South Asian High and the Western Pacific Subtropical High can modulate subsidence, water vapor transport, and regional thermal background. Teleconnection signals such as ENSO, AMO, PDO, and IOD may also affect regional hydrothermal anomalies through cross-seasonal or cross-annual scales. These factors exhibit nonlinearity, hysteresis, and interaction, making it difficult for traditional linear statistical methods to fully characterize the complex relationship between them and CHTDEs levels.

[0096] Example 1

[0097] This application provides a method for identifying the intensity of complex high-temperature drought events and interpreting their driving factors, referring to... Figure 1 This includes the following steps:

[0098] Step S1: Collect multi-source variable sample data of the target area;

[0099] Step S2: Using the standardized precipitation evapotranspiration index SPEI as the drought index DI, and combining the standardized temperature index STI and multi-source variable sample data, construct the composite high temperature drought event index CHTDI, which includes the standardized composite SPEI event index SCSEI and the composite drought high temperature evapotranspiration index BDHEI, by calculating the joint probability distribution, in order to define the intensity classification rules of composite high temperature drought events.

[0100] Step S3: Based on the Composite High Temperature and Drought Event Index (CHTDI), the composite high temperature and drought event intensity classification rules, and multi-source variable sample data, construct and train a machine learning model as a composite high temperature and drought event intensity identification model.

[0101] Step S4: Collect multi-source variable target data of the target area and input it into the composite high temperature and drought event intensity identification model to obtain the intensity identification results corresponding to the multi-source variable target data;

[0102] Step S5: Use the SHAP method to decompose the output of the composite high temperature and drought event intensity identification model to obtain the driving factor interpretation results.

[0103] Specifically, in this embodiment, the target area is the southwest region, with a geographical range of approximately 20°54′–34°19′N and 97°31′–112°03′E. The collected multi-source variable sample data includes the following four categories:

[0104] Geographic static factors: longitude, latitude, and elevation, sourced from the Chinese Academy of Sciences Geospatial Data Cloud Platform;

[0105] ERA5 single-level monthly scale variables: 2m air temperature (t2m), total precipitation (tp), first soil volumetric water content (swvl1), and evapotranspiration stress index (ESI), where ESI is calculated from actual evapotranspiration and potential evapotranspiration;

[0106] ERA5 pressure levels monthly circulation variables: 500hPa geopotential height, 850hPa geopotential height, U / V wind field components, vertical velocity, and the South Asian High and Western Pacific Subtropical High indices calculated from these.

[0107] Large-scale ocean-atmosphere and teleconnection indices: Arctic Oscillation (AO), North Atlantic Oscillation (NAO), Multivariate ENSO Index (MEI.v2), Atlantic Multidecadal Oscillation (AMO), Indian Ocean Dipole (IOD), Pacific Decadal Oscillation (PDO), and Barents Sea Ice Concentration Index (SICBS).

[0108] In addition, to characterize the impact of prior conditions on the complex high-temperature and drought event, lag terms were constructed for each time-varying variable. Based on the prediction time, features such as lag of 1 month (_lag1), lag of 2 months (_lag2), and lag of 3 months (_lag3) were extracted. For example, tp_lag1 represents the total precipitation in the month before the prediction time, AMO_lag1 represents the AMO index in the month before the prediction time, and MEI.v2_lag12 represents the MEI.v2 index in the 12 months before the prediction time. The final feature set includes longitude, latitude, elevation, and multiple lag terms of the above multi-source variables, totaling N features.

[0109] By adopting this implementation method, a multi-scale, multi-factor feature system is constructed by integrating geographic static factors, local meteorological variables, atmospheric circulation background and large-scale ocean-atmosphere teleconnection signals. This system can more comprehensively characterize the formation conditions of complex high-temperature and drought events than a single meteorological variable.

[0110] By introducing lag term features (such as tp_lag1, AMO_lag1, etc.), the cumulative impact and lag effect of previous hydrothermal conditions and low-frequency climate signals on the event are effectively captured, improving the model's predictability of the event's development process.

[0111] Geographic static factors (longitude, latitude, and elevation) serve as spatial proxy variables, compensating for the heterogeneity of the underlying surface in complex topographic regions and enabling the model to identify event differences among different geomorphic units in Southwest China.

[0112] Furthermore, in this embodiment, the standardized precipitation evapotranspiration index SPEI on a 3-month scale is used as the drought index DI. The calculation of SPEI is based on the water surplus and deficit sequence of monthly precipitation and monthly potential evapotranspiration, fitted with a log-logistic distribution, and the standardized index value is obtained through standard normal quantile transformation.

[0113] The Standardized Temperature Index (STI) is calculated as follows: First, the 30-day moving average of monthly near-surface air temperature is calculated. Then, the moving average temperature series is standardized. Since the temperature data approximately follows a normal distribution, the STI is calculated as follows:

[0114]

[0115] , where X is the 30-day moving average near-surface temperature, μ is the climatological mean of the temperature series, and σ is the climatological standard deviation.

[0116] In this embodiment, the method for constructing the composite high-temperature drought event index CHTDI, which includes the standardized composite SPEI event index SCSEI and the composite drought-temperature evapotranspiration index BDHEI, by calculating the joint probability distribution includes:

[0117] Based on multi-source variable sample data, the marginal distributions of the drought index (DI) and the standardized temperature index (STI) are estimated respectively, and the marginal cumulative distribution functions of the drought index and the temperature index are calculated based on the marginal distributions.

[0118] By combining the marginal cumulative distribution function of the drought index and the marginal cumulative distribution function of the temperature index, a joint distribution characteristic is generated;

[0119] Based on the joint distribution characteristics, the joint probabilities corresponding to the standardized composite SPEI event index SCSEI and the composite drought-high temperature evapotranspiration index BDHEI are constructed respectively.

[0120] By using normal quantile transformation to map the joint probability to a standard normal distribution, a composite high-temperature drought event index CHTDI is generated, which includes the standardized composite SPEI event index SCSEI and the composite drought-temperature evapotranspiration index BDHEI.

[0121] Furthermore, based on multi-source variable sample data, the marginal distributions of the drought index (DI) and the standardized temperature index (STI) are estimated separately. Methods for calculating the marginal cumulative distribution functions of the drought index and the temperature index based on these marginal distributions include:

[0122] Based on multi-source variable sample data, the probability density function of the drought index DI(x) was calculated using kernel density estimation. The probability density function of the standardized temperature index STI(y) With probability density function and probability density function The marginal distribution is expressed as follows:

[0123]

[0124]

[0125] Where n is the sample size, and x and y are the specific values ​​to be calculated, representing a possible value between the drought index DI and the standardized temperature index STI. and These are the i-th drought index (DI) and standardized temperature index (STI) values ​​in the sample data, respectively; h is the bandwidth, calculated using Scott's Rule; and K is the kernel function.

[0126] Integrating the marginal distributions yields the marginal cumulative distribution functions u (drought index) and v (temperature index), expressed as:

[0127]

[0128] .

[0130] Specifically, bandwidth Where σ is the sample standard deviation; K is the Gaussian kernel function:

[0131]

[0132] Where t is the standardized distance, or .

[0133] In this implementation method, the standardized temperature index (STI) is selected, which can effectively convert the absolute value of temperature into a standardized form to compare temperature anomalies at different times and in different regions.

[0134] Using kernel density estimation (KDE) to calculate the marginal cumulative distribution function of these two exponential variables can transform the original variables into standardized probability values.

[0135] Of course, depending on the actual situation of the target area, the Standardized Precipitation Index (SPI) or the Meteorological Drought Composite Index (MCI) can be selected as the drought index DI to replace the Standardized Precipitation Evapotranspiration Index (SPEI).

[0136] Furthermore, to describe the dependency between DI and STI, a Copula function is used in conjunction with the marginal cumulative distribution function of the drought index and the marginal cumulative distribution function of the temperature index to generate joint distribution characteristics. The optimal Copula function type is automatically selected based on the data distribution characteristics, such as fitting five Copula functions: Gaussian, t, Clayton, Gumbel, and Frank. The Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) are used for model selection, and the Copula function with the smallest AIC and BIC values ​​is selected as the optimal function. Through automatic selection, this embodiment ultimately adopts the Gumbel Copula as the optimal Copula function, with the expression as follows:

[0137]

[0138] ,in, is the Copula function, u is the marginal cumulative distribution function of the drought index, and v is the marginal cumulative distribution function of the temperature index;

[0139] The joint probability of the standardized composite SPEI event indices SCSEI is:

[0140] ;

[0142] The joint probability corresponding to the combined drought high temperature evapotranspiration index (BDHEI) is:

[0143] ;

[0145] The expression for the Standardized Composite SPEI Event Index (SCSEI) is:

[0146]

[0147] The expression for the high-temperature evapotranspiration index (BDHEI) of compound drought is:

[0148]

[0149] ,in, Let F be the inverse function of the standard Gaussian distribution, and let F be the empirical Gringorten partitioning formula. p is the cumulative empirical probability after sorting, and n is the sample size.

[0150] By adopting the above implementation method, kernel density estimation (KDE) is used instead of parametric distribution assumptions (such as Gamma, log-logistic), which avoids fitting bias caused by improper distribution assumptions. It is especially suitable for non-normal and multi-modal distribution characteristics of precipitation and temperature in complex terrain areas.

[0151] The optimal bandwidth is automatically calculated using Scott's Rule, which balances the smoothness and fidelity of density estimation. No manual parameter tuning is required, which improves the automation and regional adaptability of the method.

[0152] The smoothing properties of the Gaussian kernel function enable the marginal distribution estimation to have good numerical stability, providing a reliable marginal input for the construction of the Copula joint distribution;

[0153] Through a three-step standardization process of "marginal distribution estimation → Copula joint distribution → normal quantile transformation", the original meteorological variables are transformed into a composite index that follows a standard normal distribution, giving the index a clear probabilistic meaning and a unified dimension.

[0154] The Copula function is used to flexibly characterize the nonlinear dependency between drought and high temperature, breaking through the limitations of the traditional linear correlation coefficient assumption. It can accurately describe the tail correlation between variables (i.e. the joint probability of extreme drought and extreme high temperature occurring simultaneously). At the same time, the Copula function separates the marginal distribution from the dependency structure, so that the calculation of joint probability is not limited by the type of marginal distribution, thus enhancing the versatility and scalability of the method.

[0155] The probability ranking is performed using Gringorten's empirical position formula, which is more adaptable to small samples and tail extreme values ​​than simple average ranking, thus improving the robustness of joint probability estimation.

[0156] Furthermore, in this embodiment, the intensity classification rule for the complex high-temperature drought event includes:

[0157] Based on the magnitude of the combined drought high temperature evapotranspiration index (BDHEI), the intensity of combined high temperature drought events (CHTDEs) is divided into four intensity categories from Class 0 to Class 3.

[0158] When the combined high-temperature evapotranspiration index (BDHEI) is not greater than -1.5, the combined high-temperature drought event intensity (CHTDEs) is Class 0, indicating severe or worse events.

[0159] When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -1.5 and not greater than -1.0, the combined high-temperature drought event intensity (CHTDEs) is Class 1, indicating a moderate event.

[0160] When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -1.0 and not greater than -0.5, the combined high-temperature drought event intensity (CHTDEs) is Class 2, indicating a mild event.

[0161] When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -0.5, the combined high-temperature drought event intensity (CHTDEs) is Class 3, indicating that no combined event or a weak combined event has occurred.

[0162] The three composite drought high temperature evapotranspiration index (BDHEI) level thresholds of -0.5, -1.0, and -1.5 are further flanked by a ±0.05 buffer zone to remove samples from multi-source variable data that fall into the buffer zone.

[0163] Specifically, this implementation method transforms continuous indices into operationally meaningful level labels, making the model task an ordered multi-classification problem, which facilitates integration with meteorological disaster warning levels (blue / yellow / orange / red).

[0164] Since samples near the BDHEI threshold may be affected by random fluctuations, observation errors, or exponential uncertainties, forcibly assigning a fixed level would introduce label noise. To reduce the interference of boundary samples on model training, a ±0.05 buffer is set on both sides of the -0.5, -1.0, and -1.5 level thresholds. Samples falling into the buffer do not participate in training and testing. After removal, the final dataset contains 1,638,832 samples, covering the records of 12,442 spatial grid points in Southwest China from 1980 to 2024 summers. The number of samples in Class 0, Class 1, Class 2, and Class 3 are 11,121, 23,938, 70,789, and 1,532,984, respectively, accounting for 0.68%, 1.46%, 4.32%, and 93.54%, showing significant class imbalance. This implementation method avoids the training bias caused by forcibly assigning a fixed level to boundary samples and improves the model's stability in recognizing real intensity jumps.

[0165] Furthermore, in this embodiment, the machine learning model is a tree model, an ensemble model, or a deep learning model, and the method for training the machine learning model includes:

[0166] Multi-source variable sample data are stratified and randomly divided according to intensity category to construct training set and independent test set;

[0167] Parameter optimization is performed only on the training set and uses 5-fold hierarchical cross-validation.

[0168] By adopting this implementation method, the stratified random division according to intensity category ensures that the proportion of the four intensity categories Class 0-3 in the training set and the test set is consistent with the total, avoiding the problem that random sampling will cause the minority classes (such as Class 0) to be missing or have an excessively low proportion in the training set, thus ensuring the model's ability to learn minority classes.

[0169] Hierarchical cross-validation makes full use of data under limited sample conditions, and evaluates the model’s generalization performance through multiple splits, which can more reliably estimate the model’s true recognition ability than a single split.

[0170] Parameter optimization is performed only on the training set, while the test set is completely independent, effectively preventing information leakage and overfitting, and ensuring the unbiasedness of the evaluation results.

[0171] Furthermore, the machine learning models include RF, XGBoost, LightGBM, CatBoost, CNN, LSTM, and TabNet. The method for constructing and training machine learning models as a composite high-temperature and drought event intensity identification model also includes:

[0172] According to the preset model evaluation criteria, the strength recognition test results output by each machine learning model on the independent test set are evaluated. The model evaluation criteria include accuracy, and supplementary evaluation indicators including Kappa coefficient, Matthews correlation coefficient, Jaccard coefficient, Macro-F1 and Weighted-F1.

[0173] Based on the evaluation results, the optimal machine learning model was selected as the intensity identification model for the complex high temperature and drought event in the target area.

[0174] Specifically, RF is based on the Bagging idea to build multiple decision trees and integrate the output, which usually has strong stability; XGBoost, LightGBM and CatBoost belong to the Boosting framework, which can characterize complex nonlinear relationships by iteratively correcting errors; CNN, LSTM and TabNet provide comparisons from the perspectives of local structure, temporal dependence and attention-based tabular feature selection, respectively.

[0175] Multi-source variable sample data are stratified and randomly divided according to category, with 70% used as the training set and 30% as the independent test set. Parameter optimization is only performed on the training set and 5-fold stratified cross-validation is used. Considering that the proportion of severe and above event samples is relatively low, a class weight mechanism is introduced during the training phase to increase the contribution of minority class samples in the loss function and reduce the risk of the model being overly biased towards the majority class. For some parameter-sensitive models, Optuna is used for automated hyperparameter search, and comprehensive indicators such as weighted Kappa are used as the basis for parameter tuning.

[0176] To avoid overestimating model performance under class imbalance conditions by relying solely on accuracy, this embodiment introduces other supplementary evaluation metrics. Accuracy reflects the overall classification accuracy, Macro-F1 emphasizes the average recognition ability of each class and is suitable for minority class evaluation, Weighted-F1 considers class sample weights, the Kappa coefficient considers the ordered distance relationship between levels and can measure the consistency between predicted and true levels, MCC integrates confusion matrix information and has high robustness in class imbalance scenarios, and the Jaccard coefficient is used to measure the similarity between two sets.

[0177] The seven models exhibited significant differences under a unified feature system and sample partitioning. The training set results are shown in Table 1. RF, LightGBM, and XGBoost all demonstrated strong fitting capabilities, with RF achieving accuracy, macro-F1, weighted-F1, Kappa coefficient κw, and MCC of 0.997, 0.988, 0.997, 0.989, and 0.976, respectively. The test set results better reflect generalization ability, with RF remaining the best, achieving accuracy, macro-F1, weighted-F1, κw, and MCC of 0.985, 0.907, 0.986, 0.950, and 0.884, respectively. LightGBM followed closely, with corresponding metrics of 0.982, 0.901, 0.982, 0.940, and 0.863. XGBoost's test set metrics were 0.978, 0.895, 0.980, 0.929, and 0.845, slightly lower than RF and LightGBM, but still significantly better than the other models.

[0178] In comparison, CNN, LSTM, CatBoost, and TabNet all showed varying degrees of shortcomings in their overall performance. CNN achieved an accuracy of 0.980 on the test set, but its Macro-F1 score was only 0.792, indicating that its high accuracy was mainly affected by majority class recognition, and its ability to balance discrimination of minority event class samples was insufficient. LSTM's Macro-F1 and MCC scores were 0.735 and 0.695, respectively, indicating limited discrimination ability at medium to high intensity levels. CatBoost's test set accuracy, Macro-F1, κw, and MCC scores were 0.941, 0.821, 0.829, and 0.685, respectively, which were generally lower than the first three tree models. TabNet's test set Macro-F1 and MCC scores were 0.790 and 0.644, respectively, failing to demonstrate any advantage over traditional tree models.

[0179] Overall, RF, LightGBM, and XGBoost form the first tier. All three have a κw of over 0.929, indicating that they can not only maintain a high overall accuracy, but also better characterize the ordered relationship between different intensity levels. RF has the best overall stability, while LightGBM and XGBoost perform similarly.

[0180]

[0181] Table 1 Comparison of classification performance of different models on the training and test sets.

[0182] To avoid over-reliance on a single model for interpretability conclusions, we further compared the feature importance of three types of tree models: RF, XGBoost, and LightGBM, such as... Figure 2As shown, the TopK set overlap results indicate that the Jaccard coefficients of the overall Top 20 models are approximately 0.30–0.46, and those of the Top 50 models are approximately 0.28–0.37, indicating that there is some consensus among the models, but the degree of complete overlap at the variable name level is not high.

[0183] Pairwise comparisons show that the consistency between XGBoost and LightGBM is significantly higher than that between the combination of RF and Boosting models. In the Top 20 cases, the Jaccard coefficients of LightGBM-XGBoost are 0.67, 0.74, 0.90 and 0.74 in Class 0, Class 1, Class 2 and Class 3, respectively. In the Top 50 cases, it still maintains a high degree of overlap of about 0.54-0.64. In contrast, the overlap between RF-XGBoost and RF-LightGBM is generally low, indicating that RF provides a relatively independent feature selection perspective.

[0184] like Figure 3 As shown, the consistency of importance ranking further indicates that the Spearman correlation coefficients between LightGBM and XGBoost across the four levels are approximately 0.79–0.88, while the correlations between RF and the two Boosting models are mostly between 0.51 and 0.67. The Bootstrap stability test shows that RF has a generally smaller confidence interval width across the four levels, indicating strong stability in importance ranking; XGBoost has moderate and relatively balanced stability, while LightGBM has relatively higher uncertainty in Class 0.

[0185] In addition, since XGBoost has mature SHAP support in subsequent interpretation analysis, considering prediction performance, cross-model consistency and interpretation stability, the optimal machine learning model in this embodiment is XGBoost. Optuna is used for automated hyperparameter search, parameter optimization is only performed on the training set, 5-fold hierarchical cross-validation is used, and the weighted Kappa coefficient (Quadratic Weighted Kappa, κw) is used as the evaluation index for hyperparameter optimization.

[0186] During the model training phase, an intensity category weighting mechanism is introduced, assigning different weights to each intensity category. The weights are inversely proportional to the number of samples in that category.

[0187]

[0188] Where wk is the weight of intensity class k, N is the total number of samples, Nk is the number of samples of intensity class k, and K is the total number of intensity classes (K=4 in this embodiment); during model training, the intensity class weights are introduced into the loss function to increase the contribution of minority class samples (especially Class0) in the loss calculation.

[0189] Furthermore, after determining that the optimal machine learning model for the target region in this embodiment is XGBoost, the method of using the SHAP method to decompose the output of the composite high-temperature drought event intensity identification model to obtain the driving factor interpretation results includes:

[0190] Obtain the baseline probability Pbase(k) for intensity category k;

[0191] The baseline probability Pbase(k) is converted into a first probability, expressed as:

[0192]

[0193] And, the SHAP value f is converted from logarithmic odds to second odds, expressed as:

[0194] ;

[0196] Based on the first and second probabilities, the predicted probability of observing intensity category k after considering the influence of target feature values ​​is calculated as follows:

[0197] ;

[0199] The predicted probability is converted into a probability to characterize the predicted probability of observing intensity category k after considering the influence of the target feature value. The expression is:

[0200] ;

[0202] Based on the predicted probability, the SHAP value, expressed in probability units, is calculated as the interpretation result of the driving factor to characterize the contribution of the target feature value to the probability change of the observed intensity category k. The expression is as follows:

[0203] .

[0205] Specifically, such as Figure 4As shown, the XGBoost-SHAP global importance ranking indicates that the discrimination of different levels of CHTDEs is not controlled by a single factor, but is driven by spatial location, anterior hydrothermal conditions, and teleconnection / circulation background. In Class 0, longitude, latitude, tp_lag1, and tp_lag2 are at the top, indicating that the identification of strong events primarily depends on spatial differences within the study area and precipitation conditions in the preceding 1-2 months. Meanwhile, variables such as AMO_lag1, MEI.v2_lag12, t2m_lag3, and ESI_lag1 enter the important feature set, indicating that large-scale background and land surface evapotranspiration stress play a supplementary role in the discrimination of strong events. The ranking structure of Class 1 is similar to that of Class 0, with tp_lag1 and tp_lag2 still ranking highly, latitude and longitude maintaining high importance, and ESI_lag1, AMO_lag7, the lag term of the western extension point of the Western Pacific Subtropical High, and the lag term of the 850 hPa geopotential height also contributing to the classification. Compared to Class 0, the importance distribution of Class 1 is more dispersed, indicating that the distinction between moderate events is influenced by multiple factors, but the advantage of a single dominant factor is weaker than that of strong events.

[0206] The importance structure of Class 2 shows a significant change, with AMO_lag2 ranking first, and the importance of latitude and longitude, tp_lag3, IOD_lag4, and the lag term of the western extension point of the Western Pacific Subtropical High increasing relatively. This indicates that the discrimination of mild events may rely more on low-frequency background signals and longer-term climate information, rather than being controlled solely by near-monthly precipitation. In Class 3, tp_lag1, AMO_lag1, tp_lag2, latitude, and longitude are at the top, with previous precipitation and spatial location remaining the basic discriminant factors. However, the importance of the AMO signal has significantly increased, suggesting that the level of non-event or weak events is also influenced by large-scale background modulation.

[0207] Figure 5 The results of the XGBoost model on SHAPbeeswarm at four composite high-temperature and drought levels (Class 0–3) are presented. From the perspective of the direction of action, the explanation of the antecedent precipitation factor is the clearest. In Class 0, low-value samples of tp_lag1 and tp_lag2 mostly correspond to positive SHAP contributions, while high-value samples mostly correspond to negative SHAP contributions. This indicates that less precipitation in the preceding 1-2 months strengthens the tendency of samples to be classified as severe or worse CHTDEs, while more precipitation suppresses the classification of strong events. A similar trend exists in Class 1, but the contribution magnitude is weaker than in Class 0. In Class 3, the direction of action of tp_lag1 and tp_lag2 is approximately mirrored in Class 0, with high-value samples corresponding to positive contributions and low-value samples corresponding to negative contributions, indicating that more precipitation in the preceding period is beneficial for the classification of weak events or non-events.

[0208] Spatial factors remain highly important across multiple levels. Latitude and longitude do not directly represent a single physical mechanism, but they can serve as spatial proxies for differences in complex topography, climate background, underlying surface, and circulation influences. Their high SHAP contribution indicates that even with the inclusion of multiple hydrothermal, circulation, and teleconnection factors, the model still needs to utilize spatial location information to identify regional heterogeneity at the CHTDEs level in Southwest China.

[0209] By adopting this implementation method, the performance differences between tree models (RF, XGBoost, etc.) and deep learning models (CNN, LSTM, etc.) on ordered multi-classification tasks are compared systematically, and the most suitable model type for identifying the intensity of complex high temperature and drought events is selected, thus avoiding the blind selection of models.

[0210] The comprehensive evaluation using multiple metrics such as Macro-F1, Weighted-F1, κw, and MCC is more effective than a single Accuracy metric in reflecting the true recognition ability under class imbalance conditions, with particular attention to the recognition performance of a few intensity classes (Class 0).

[0211] Choosing the optimal model as the final recognition model ensures recognition accuracy and provides a single, stable model foundation for SHAP interpretability analysis, avoiding the interpretability complexity brought about by multi-model integration.

[0212] By converting the SHAP value from log odds units to probability units (range 0-1), the feature contribution has an intuitive probabilistic meaning, making it easier for business personnel to understand "a certain factor increases the probability of an event by X%";

[0213] By calculating the difference between the baseline probability and the predicted probability, the marginal contribution of a single feature to the discrimination result of a specific intensity category is clearly quantified, realizing the direct mapping from "factor to probability change".

[0214] The probability unit SHAP value supports comparisons across features and samples, making it easy to identify global factors and local anomalies that play a key role in the discrimination of specific intensity categories (such as Class 0).

[0215] In summary, by using this embodiment to construct CHTDI (including SCSEI and BDHEI) through joint probability distribution, it is possible to simultaneously consider the joint occurrence probability of drought and high temperature, and more accurately characterize the joint intensity features of complex high temperature and drought events than a single drought index or temperature index.

[0216] By setting a grade boundary buffer and removing boundary samples, label noise was effectively reduced, and the recognition accuracy of the machine learning model under the condition of intensity class imbalance was improved.

[0217] By introducing an intensity category weighting mechanism and a hierarchical random partitioning strategy, the adverse effects of insufficient minority class samples, such as Class 0 (severe and above event categories), on model training are mitigated.

[0218] Through SHAP interpretability analysis, the model output is decomposed into the probability contribution of each input feature to each intensity category, realizing the transformation from "black box prediction" to "white box attribution". This clarifies the direction of action, nonlinear response characteristics and threshold effect of key driving factors such as anterior precipitation and large-scale ocean-atmosphere background.

[0219] Example 2

[0220] like Figure 6 As shown, this application also proposes an embodiment of a system for identifying the intensity of complex high-temperature drought events and interpreting driving factors, including:

[0221] The data acquisition module is used to collect multi-source variable sample data of the target area;

[0222] The intensity classification rule definition module is used to construct the composite high temperature and drought event index CHTDI, which includes the standardized composite SPEI event index SCSEI and the composite drought and high temperature evapotranspiration index BDHEI, by using the standardized precipitation evapotranspiration index SPEI as the drought index DI, combined with the standardized temperature index STI and multi-source variable sample data, and by calculating the joint probability distribution, so as to define the intensity classification rules of composite high temperature and drought events.

[0223] The identification model modeling module is used to construct and train a machine learning model as an identification model for the intensity of composite high temperature and drought events based on the Composite High Temperature and Drought Event Index (CHTDI), the rules for classifying the intensity of composite high temperature and drought events, and multi-source variable sample data.

[0224] The intensity identification module is used to collect multi-source variable target data of the target area and input it into the composite high temperature and drought event intensity identification model to obtain the intensity identification results corresponding to the multi-source variable target data.

[0225] The driving factor interpretation module is used to perform contribution decomposition on the output of the composite high temperature and drought event intensity identification model using the SHAP method to obtain the driving factor interpretation results.

[0226] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for identifying the intensity of complex high-temperature drought events and interpreting their driving factors, characterized in that, Includes the following steps: Collect multi-source variable sample data of the target area; Using the standardized precipitation evapotranspiration index SPEI as the drought index DI, and combining the standardized temperature index STI and multi-source variable sample data, a composite high temperature drought event index CHTDI, including the standardized composite SPEI event index SCSEI and the composite drought high temperature evapotranspiration index BDHEI, is constructed by calculating the joint probability distribution, so as to define the intensity classification rules of composite high temperature drought events for classifying intensity categories. Based on the Composite High Temperature and Drought Event Index (CHTDI), the intensity classification rules of composite high temperature and drought events, and multi-source variable sample data, a machine learning model was constructed and trained as a composite high temperature and drought event intensity identification model. Collect multi-source variable target data of the target area and input them into the composite high temperature and drought event intensity identification model to obtain the intensity identification results corresponding to the multi-source variable target data; The SHAP method was used to decompose the output of the composite high temperature and drought event intensity identification model to obtain the driving factor interpretation results.

2. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 1, characterized in that, The multi-source variable sample data includes: Geographic static factors, monthly-scale variables of ERA5 single levels, monthly-scale circulation variables of ERA5 pressure levels, and large-scale ocean-atmosphere and teleconnection indices; The geographic static factors include longitude, latitude, and elevation; The monthly scale variables of the ERA5 single levels include 2m air temperature t2m, total precipitation tp, first soil volumetric water content swvl1, and evapotranspiration stress index ESI calculated from actual evapotranspiration and potential evapotranspiration. The monthly circulation variables of the ERA5 pressure levels include 500hPa and 850hPa geopotential heights, wind field components and vertical velocities, as well as the corresponding high-pressure and subtropical high-pressure related indices. The large-scale air-sea and teleconnection indices include AO, NAO, MEI.v2, AMO, IOD, PDO, and the Barents Sea ice concentration index SICBS.

3. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 1, characterized in that, A method for constructing the composite high-temperature drought event index (CHTDI) by calculating the joint probability distribution, which includes the standardized composite SPEI event index (SCSEI) and the composite drought-temperature evapotranspiration index (BDHEI), includes: Based on multi-source variable sample data, the marginal distributions of the drought index (DI) and the standardized temperature index (STI) are estimated respectively, and the marginal cumulative distribution functions of the drought index and the temperature index are calculated based on the marginal distributions. By combining the marginal cumulative distribution function of the drought index and the marginal cumulative distribution function of the temperature index, a joint distribution characteristic is generated; Based on the joint distribution characteristics, the joint probabilities corresponding to the standardized composite SPEI event index SCSEI and the composite drought-high temperature evapotranspiration index BDHEI are constructed respectively. By using normal quantile transformation to map the joint probability to a standard normal distribution, a composite high-temperature drought event index CHTDI is generated, which includes the standardized composite SPEI event index SCSEI and the composite drought-temperature evapotranspiration index BDHEI.

4. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 3, characterized in that, Based on multi-source variable sample data, the marginal distributions of the drought index (DI) and the standardized temperature index (STI) are estimated separately. Methods for calculating the marginal cumulative distribution functions of the drought index and the temperature index based on these marginal distributions include: Based on multi-source variable sample data, the probability density function of the drought index DI(x) was calculated using kernel density estimation. The probability density function of the standardized temperature index STI(y) With probability density function and probability density function The marginal distribution is expressed as follows: ， Where n is the sample size, and x and y are the specific values ​​to be calculated, representing a possible value between the drought index (DI) and the standardized temperature index (STI). and These are the i-th drought index (DI) and standardized temperature index (STI) values ​​in the sample data, respectively; h is the bandwidth, calculated using Scott's rule; and K is the kernel function. Integrating the marginal distributions yields the marginal cumulative distribution functions u (drought index) and v (temperature index), expressed as: 。 5. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 1, characterized in that: By combining the Copula function with the marginal cumulative distribution functions of the drought index and the temperature index, a joint distribution characteristic is generated, expressed as follows: ， in, is the Copula function, u is the marginal cumulative distribution function of the drought index, and v is the marginal cumulative distribution function of the temperature index; The joint probability of the standardized composite SPEI event index SCSEI is: ； The joint probability corresponding to the combined drought high temperature evapotranspiration index (BDHEI) is: ； The expression for the Standardized Composite SPEI Event Index (SCSEI) is: ， The expression for the high-temperature evapotranspiration index (BDHEI) in compound drought is: ， in, F is the inverse function of the standard Gaussian distribution, and F is the empirical Gringorten position partitioning formula.

6. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 1, characterized in that, The rules for classifying the intensity of complex high-temperature and drought events include: Based on the magnitude of the combined drought high temperature evapotranspiration index (BDHEI), the intensity of combined high temperature drought events (CHTDEs) is divided into four intensity categories from Class 0 to Class 3. When the combined high-temperature evapotranspiration index (BDHEI) is not greater than -1.5, the combined high-temperature drought event intensity (CHTDEs) is Class 0, indicating severe or worse events. When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -1.5 and not greater than -1.0, the combined high-temperature drought event intensity (CHTDEs) is Class 1, indicating a moderate event. When the combined high-temperature evapotranspiration index (BDHEI) of drought is greater than -1.0 and not greater than -0.5, the combined high-temperature drought event intensity (CHTDEs) is Class 2, indicating a mild event. When the combined high temperature evapotranspiration index (BDHEI) of drought is greater than -0.5, the combined high temperature drought event intensity (CHTDEs) is Class 3, indicating that no combined event or a weak combined event has occurred. The three composite drought high temperature evapotranspiration index (BDHEI) level thresholds of -0.5, -1.0, and -1.5 are further flanked by a ±0.05 buffer zone to remove samples from multi-source variable data that fall into the buffer zone.

7. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 1, characterized in that, The machine learning model is a tree model, an ensemble model, or a deep learning model, and the methods for training the machine learning model include: Multi-source variable sample data are stratified and randomly divided according to intensity category to construct training set and independent test set; Parameter optimization is performed only on the training set and uses 5-fold hierarchical cross-validation.

8. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 7, characterized in that, The machine learning models include RF, XGBoost, LightGBM, CatBoost, CNN, LSTM, and TabNet. Methods for constructing and training machine learning models as composite high-temperature and drought event intensity identification models also include: According to the preset model evaluation criteria, the strength recognition test results output by each machine learning model on the independent test set are evaluated. The model evaluation criteria include accuracy, and supplementary evaluation indicators including Kappa coefficient, Matthews correlation coefficient, Jaccard coefficient, Macro-F1 and Weighted-F1. Based on the evaluation results, the optimal machine learning model was selected as the intensity identification model for the complex high temperature and drought event in the target area.

9. The method for identifying the intensity of complex high-temperature drought events and interpreting driving factors as described in claim 8, characterized in that, A method for decomposing the output of the composite high-temperature drought event intensity identification model using the SHAP method to obtain the driving factor interpretation results includes: Obtain the baseline probability Pbase(k) for intensity category k; The baseline probability Pbase(k) is converted into a first probability, expressed as: ， And, the SHAP value f is converted from logarithmic odds to second odds, expressed as: ； Based on the first and second probabilities, the predicted probability of observing intensity category k after considering the influence of target feature values ​​is calculated as follows: ； The predicted probability is converted into a probability to characterize the predicted probability of observing intensity category k after considering the influence of the target feature value. The expression is: ； Based on the predicted probability, the SHAP value, expressed in probability units, is calculated as the interpretation result of the driving factor to characterize the contribution of the target feature value to the probability change of the observed intensity category k. The expression is as follows: 。 10. A system for identifying the intensity of complex high-temperature drought events and interpreting driving factors, characterized in that, include: The data acquisition module is used to collect multi-source variable sample data of the target area; The intensity classification rule definition module is used to construct the composite high temperature and drought event index CHTDI, which includes the standardized composite SPEI event index SCSEI and the composite drought and high temperature evapotranspiration index BDHEI, by using the standardized precipitation evapotranspiration index SPEI as the drought index DI, combined with the standardized temperature index STI and multi-source variable sample data, and by calculating the joint probability distribution, so as to define the intensity classification rules of composite high temperature and drought events. The identification model modeling module is used to construct and train a machine learning model as an identification model for the intensity of composite high temperature and drought events based on the Composite High Temperature and Drought Event Index (CHTDI), the rules for classifying the intensity of composite high temperature and drought events, and multi-source variable sample data. The intensity identification module is used to collect multi-source variable target data of the target area and input it into the composite high temperature and drought event intensity identification model to obtain the intensity identification results corresponding to the multi-source variable target data. The driving factor interpretation module is used to perform contribution decomposition on the output of the composite high temperature and drought event intensity identification model using the SHAP method to obtain the driving factor interpretation results.

Images