Global drought spatio-temporal variation AI prediction method, system and equipment

By combining wavelet transform and LSTM model, the SPEI index is used to analyze the spatiotemporal changes of global droughts, the shortcomings of existing drought prediction models in nonlinear feature capture and long-term prediction are solved, and a more accurate and comprehensive drought prediction effect is achieved.

CN118916695APending Publication Date: 2024-11-08INST OF AGRI RESOURCES & REGIONAL PLANNING CHINESE ACADEMY OF AGRI SCI
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202410957160.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing drought prediction models are inadequate in capturing nonlinear features and long-term predictions, and the current meteorological drought changes are complex, requiring a more accurate and comprehensive prediction method.

Method used

A mixed model (WT-LSTM) combining wavelet transform and long and short-term memory neural network (LSTM) is proposed, and the spatial and temporal changes of global drought are analyzed using standardized precipitation evapotranspiration index (SPEI) to comprehensively consider precipitation, surface temperature and potential evapotranspiration Deep learning is performed in multi-factor data to predict trend changes in drought.

Benefits of technology

It improves the accuracy and comprehensiveness of drought prediction, can more accurately capture drought trends and patterns on a global scale, and enhances the stability and convergence of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916695A_ABST
    Figure CN118916695A_ABST
Patent Text Reader

Abstract

The invention provides a global drought spatio-temporal variation AI prediction method, system and equipment, and the method comprises the steps: obtaining prediction factors of a to-be-predicted region, and the prediction factors comprise precipitation, potential evapotranspiration and surface temperature; inputting the prediction factor into a preset drought spatio-temporal variation prediction model to obtain a drought prediction result; wherein the drought spatio-temporal variation prediction model is constructed based on a WT-LSTM model and is obtained by training based on a constructed data set, and the data set comprises a historical prediction factor, a historical SPEI index and a corresponding historical drought event. According to the method, a WT-LSTM model is utilized to comprehensively consider precipitation, surface temperature, potential evapotranspiration and other multi-factor data to carry out deep learning so as to predict the trend change of drought; therefore, more comprehensive drought change prediction is carried out.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of meteorological forecasting, and in particular relates to an AI forecasting method, system and device for global drought spatiotemporal changes. Background Art

[0002] Meteorological drought, as one of the most common natural disasters in the world, poses a serious threat to agricultural production, water resources management, and ecosystem stability. As climate change continues to intensify, the frequency and intensity of drought events are on the rise, posing unprecedented challenges to global food security and sustainable management of water resources.

[0003] In the process of monitoring and analyzing drought, it is crucial to select appropriate indicators to determine the spatiotemporal distribution and changing trends of drought. At present, more than 160 drought indices have been developed to analyze the spatiotemporal changes of drought, and each index shows its unique advantages and limitations. In particular, the Standardized Precipitation Evapotranspiration Index (SPEI), which takes into account precipitation and potential evapotranspiration, can more comprehensively reflect the hydroclimatic conditions of a region, rather than relying on a single climatic factor. Studies have shown that compared with the simple Standardized Precipitation Index (SPI), SPEI is more accurate in identifying the severity and duration of drought. For example, in a study of the Kaffi River Basin in northern Zambia, SPEI more effectively captured the temporal changes of drought between 1960 and 2015 than SPI (Tirivarombo et al., 2018). In addition, other indices such as SCPDSI, although providing estimates of drought severity, often show deviations from actual conditions and exhibit a certain lag in the evolution of drought. Therefore, SPEI was selected as the main indicator for analyzing global droughts in order to provide a more accurate and comprehensive drought assessment.

[0004] Given the complexity of time series data types, developing accurate prediction models often faces significant challenges. Over the years, many prediction models have been established, including regression models (Sun et al., 2012), conditional probability models (Hao et al., 2016), data-driven models (Xu et al., 2020), and machine learning algorithms (Xu et al., 2022). Each of these methods has its own advantages and disadvantages. For example, among data-driven models, the autoregressive integrated moving average (ARIMA) model performs well in long-term drought prediction (Belayneh et al., 2014), but performs poorly in capturing nonlinear characteristics (Kim et al., 2022; Wang, 2022). In recent years, machine learning has become a core component of data-driven research in the field of earth sciences and has been widely used in the study of the atmosphere, land surface, and ocean (Shen et al., 2018, Ham et al., 2019, Reichstein et al., 2019, Chen et al., 2019, Retsch et al., 2022; Liu et al., 2022; Zhang et al., 2022b; Zheng et al., 2022; Asgarimehr et al., 2022). In particular, in the field of meteorology, methods such as random forest (RF) models have been widely used to solve earth science and remote sensing problems (Gislason et al., 2006).

[0005] In the field of deep learning, artificial neural networks (ANNs) are widely used due to their excellent ability in time series forecasting, especially in capturing nonlinear information of data. However, with the increase in the number of parameters and the extension of training time during training, people have begun to turn to other nonlinear models, such as recurrent neural networks (RNNs), which are suitable for tasks based on short-term memory but are less sensitive to old data (Le et al., 2017). To achieve long-term prediction, the RNN-based long short-term memory (LSTM) model came into being, which is characterized by a more complex neuronal structure (Dikshit et al., 2021). In addition, wavelet transform, as a data preprocessing tool, has been proven to successfully decompose raw data and effectively predict nonlinear and non-stationary time series (Madhu et al., 2004), greatly improving the performance of machine learning models. The WT-LSTM model, which combines wavelet transform with LSTM model, performs well in hydrological forecasting, although its application in short-term meteorological drought forecasting is relatively rare.

[0006] In the current context of complex changes in global meteorological drought, the rapid development of artificial intelligence (AI) technology has brought new breakthroughs in improving the accuracy of drought prediction. In particular, technologies such as deep learning have shown significant advantages in processing large-scale and high-dimensional meteorological data analysis. By learning historical meteorological data, these technologies can mine and identify complex patterns and potential information in drought development, thereby significantly improving the accuracy of drought prediction. Therefore, it is urgent to propose an AI prediction method, system and equipment for global drought spatiotemporal changes. Summary of the invention

[0007] The purpose of the present invention is to propose an AI prediction method, system and device for the spatiotemporal changes of global drought, and to analyze the spatiotemporal changes of global drought using the Standardized Precipitation Evapotranspiration Index (SPEI). In order to improve the accuracy of global drought prediction, a hybrid model (WT-LSTM) combining wavelet transform and long short-term memory neural network (LSTM) is further constructed. The WT-LSTM model is used to conduct deep learning on multi-factor data such as precipitation, surface temperature and potential evapotranspiration to predict the trend changes of drought; thereby making a more comprehensive prediction of drought changes.

[0008] To achieve the above objectives, the present invention provides an AI prediction method, system and device for global drought spatiotemporal changes, including:

[0009] Obtain prediction factors for the area to be predicted, wherein the prediction factors include: precipitation, potential evapotranspiration and surface temperature;

[0010] The prediction factors are input into a preset drought spatiotemporal change prediction model to obtain drought prediction results; wherein the drought spatiotemporal change prediction model is constructed based on the WT-LSTM model and is trained based on a constructed data set, and the data set includes: historical prediction factors, historical SPEI index and corresponding historical drought events.

[0011] Optionally, constructing the data set includes:

[0012] Obtaining a historical SPEI index of a relevant region, and identifying historical drought events based on the historical SPEI index;

[0013] The historical prediction factors are screened, and the data set is constructed based on the historical prediction factors, the historical SPEI index and the corresponding historical drought events; wherein the historical prediction factors include: historical precipitation, historical potential evapotranspiration and historical surface temperature.

[0014] Optionally, obtaining the historical SPEI index of the relevant region includes:

[0015] Obtain daily precipitation data from ERA5 and daily potential evapotranspiration data calculated by the modified Penman-Monteith formula;

[0016] The daily precipitation data and the daily potential evapotranspiration data are added together and standardized to obtain the monthly data of the SPEI index.

[0017] Optionally, the SPEI index is:

[0018]

[0019] Among them, W is calculated by the standardized distribution function value (probability), c0=2.515517, c1=0.802853, c2=0.010328, d1=1.432788, d2=0.189269, d3=0.001308.

[0020] Optionally, historical drought event identification includes:

[0021] Drought duration = DI0-DI(t)

[0022]

[0023] Among them, DI0 and DI(t) represent the end and start time of the drought event respectively, D represents the end time, and t represents time.

[0024] Optionally, constructing the WT-LSTM model includes: combining the LSTM model with discrete wavelet transform to form the WT-LSTM model.

[0025] Optionally, training the WT-LSTM model based on the constructed data set includes:

[0026] Based on the discrete wavelet transform, the data in the data set is decomposed into a high-pass filter and a low-pass filter to extract the high-frequency component and the low-frequency component of the data sequence respectively;

[0027] The low-frequency component is estimated using the LSTM model, and a high-pass filter is trained separately using a neural network.

[0028] Optionally, after training the WT-LSTM model, the method further includes:

[0029] The trained WT-LSTM model is evaluated based on preset evaluation indicators; wherein the preset evaluation indicators include: mean absolute error, mean square error, root mean square error, ratio of RMSE to standard deviation of observed values, standard deviation ratio, Nash-Sutcliffe efficiency coefficient, and correlation coefficient.

[0030] Optionally, screening the historical prediction factors includes:

[0031] Analyze and visualize the spatiotemporal trends of predicted variables and drought, and perform significance tests to analyze the potential development trend of drought;

[0032] By visualizing and quantifying the duration, scale, and severity of droughts over a preset time period, drought events are identified and the historical predictors are screened.

[0033] To achieve the above-mentioned purpose, the present invention also proposes a global drought spatiotemporal change AI prediction system, which is used to implement the global drought spatiotemporal change AI prediction method, and the system comprises: a prediction factor acquisition module and a drought spatiotemporal change prediction module;

[0034] The prediction factor acquisition module is used to acquire prediction factors of the area to be predicted, wherein the prediction factors include: precipitation, potential evapotranspiration and surface temperature;

[0035] The drought spatiotemporal change prediction module is used to input the prediction factors into a preset drought spatiotemporal change prediction model to obtain a drought prediction result; wherein the drought spatiotemporal change prediction model is constructed based on a WT-LSTM model and is trained based on a constructed data set, and the data set includes: historical prediction factors, historical SPEI index and corresponding historical drought events.

[0036] To achieve the above-mentioned purpose, the present invention also proposes an electronic device, which includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the global drought spatiotemporal change AI prediction method is implemented.

[0037] The present invention has the following beneficial effects:

[0038] The present invention uses the Standardized Precipitation Evapotranspiration Index (SPEI) to analyze the spatiotemporal changes of global drought. In order to improve the accuracy of global drought prediction, a hybrid model (WT-LSTM) combining wavelet transform and long short-term memory neural network (LSTM) is further constructed. The WT-LSTM model is used to conduct deep learning on multi-factor data such as precipitation, surface temperature and potential evapotranspiration to predict the trend change of drought; thus, a more comprehensive prediction of drought changes can be made. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0040] Figure 1 A schematic diagram of the study area of ​​an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the analysis and prediction of the temporal and spatial characteristics of meteorological drought according to an embodiment of the present invention;

[0042] Figure 3 A schematic diagram of the run theory of an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of an LSTM model of an embodiment of the present invention;

[0044] Figure 5 Schematic diagram of spatial distribution of correlation coefficient and root mean square error between 1-month SPEI using SPEIBase and ERA5 datasets in an embodiment of the present invention;

[0045] Figure 6 is the correlation coefficient between the potential variables of the embodiment of the present invention and the SPEI index, wherein (a) is the potential variable including precipitation (Pre), (b) is the potential evaporation (PET), (c) is the land surface temperature (LST), (d) is the 10m U wind component (u10), (e) is the 10m V wind component (v10), (f) is the evaporation (e), and (g) is the volume soil water layer 1 (swvl1);

[0046] Figure 7 Schematic diagram of the research of multi-source data from 1979 to 2022 in an embodiment of the present invention; wherein (a) is annual precipitation (Pre), (b) is annual potential evapotranspiration (PET), (c) is annual land surface temperature (LST), and (d) is Sen's slope trend analysis of SPEI;

[0047] Figure 8 This is a schematic diagram of the annual trend analysis of the four seasons in the world from 1979 to 2022 according to an embodiment of the present invention; wherein (a, e, i, m) represents spring, (b, f, j, n) represents summer, (c, g, k, o) represents autumn, and (d, h, l, p) represents winter;

[0048] Fig. 9 Schematic diagram of the SPEI index data research from 1979 to 2022 according to an embodiment of the present invention; wherein, (a, c, e and g) the annual average drought duration, severity, intensity and frequency of the SPEI index from 1979 to 2022; (b, d and f) the spatial trend of the annual average drought duration, severity and intensity of the SPEI index from 1979 to 2022;

[0049] Fig.10Schematic diagram of the duration, severity, intensity and frequency of drought in four seasons around the world from 1979 to 2022 according to an embodiment of the present invention; wherein (a, e, i, m) represents spring, (b, f, j, n) represents summer, (c, g, k, o) represents autumn, and (d, h, l, p) represents winter;

[0050] Fig.11 Schematic diagram of statistical comparison of two prediction models of the embodiment of the present invention; wherein (a) is the first scheme of Q1, and (b) is the second scheme of Q2;

[0051] Fig.12 This is a statistical comparison diagram of the prediction models (LSTM and WT-LSTM models) in the first solution Q1 of the embodiment of the present invention;

[0052] Fig.13 This is a statistical comparison diagram of the prediction models (LSTM and WT-LSTM models) in the second solution Q2 of the embodiment of the present invention;

[0053] Fig.14 A schematic diagram of the mean square error (MSE) loss of the Q1 and Q2 prediction model WT-LSTM for global land in an embodiment of the present invention;

[0054] Fig.15 This is a scatter plot of typical regional site verification of the Q1 solution of an embodiment of the present invention;

[0055] Fig.16 This is a scatter plot of typical regional site verification of the Q2 solution of an embodiment of the present invention;

[0056] Fig.17 It is a schematic diagram comparing the prediction results of the two scheme models of the embodiment of the present invention with the results of existing products; among which, (ac spring; df summer; gi autumn; jl winter). DETAILED DESCRIPTION

[0057] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0058] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0059] This embodiment proposes an AI prediction method for global drought spatiotemporal changes, including:

[0060] Obtain prediction factors for the area to be predicted, where the prediction factors include: precipitation, potential evapotranspiration and surface temperature;

[0061] The prediction factors are input into the preset drought spatiotemporal change prediction model to obtain the drought prediction results; the drought spatiotemporal change prediction model is constructed based on the WT-LSTM model and trained based on the constructed data set, which includes: historical prediction factors, historical SPEI index and corresponding historical drought events.

[0062] Furthermore, constructing the dataset includes:

[0063] Obtain the historical SPEI index of the relevant area and identify historical drought events based on the historical SPEI index;

[0064] Historical prediction factors were screened, and a data set was constructed based on historical prediction factors, historical SPEI index, and corresponding historical drought events; among them, historical prediction factors included historical precipitation, historical potential evapotranspiration, and historical surface temperature.

[0065] Furthermore, obtaining the historical SPEI index of relevant regions includes:

[0066] Obtain daily precipitation data from ERA5 and daily potential evapotranspiration data calculated by the modified Penman-Monteith formula;

[0067] The daily precipitation data and daily potential evapotranspiration data were added together and standardized to obtain the monthly data of the SPEI index.

[0068] Furthermore, constructing the WT-LSTM model includes: combining the LSTM model with discrete wavelet transform to form the WT-LSTM model.

[0069] Furthermore, training the WT-LSTM model based on the constructed dataset includes:

[0070] Based on discrete wavelet transform, the data in the data set is decomposed into high-pass filter and low-pass filter to extract the high-frequency component and low-frequency component of the data sequence respectively;

[0071] The LSTM model is used to estimate the low-frequency component, and the high-pass filter is trained separately using a neural network.

[0072] Furthermore, the WT-LSTM model is trained to include:

[0073] The trained WT-LSTM model is evaluated based on the preset evaluation indicators, including mean absolute error, mean square error, root mean square error, ratio of RMSE to standard deviation of observed values, standard deviation ratio, Nash-Sutcliffe efficiency coefficient, and correlation coefficient.

[0074] Further, screening historical predictors include:

[0075] Analyze and visualize the spatiotemporal trends of predicted variables and drought, and perform significance tests to analyze the potential development trend of drought;

[0076] Identify drought events and screen historical predictors by visualizing and quantifying drought duration, scale, and severity over a preset time period.

[0077] The research area of ​​this embodiment covers the main land grain producing areas in the world, spanning Asia, Europe, South America, North America, Oceania and Africa, a total of six continents, excluding Antarctica. Figure 1 As shown. In order to improve the computational efficiency and ensure the accuracy of the research results, an effective data processing method was adopted: the original data was bilinearly interpolated to a resolution of 1×1. The choice of this method not only takes into account the huge amount of data, but also takes into account the geographical differences on a global scale. By interpolating the data to a lower spatial resolution, the amount of data can be greatly reduced while retaining sufficient geographic information, thereby reducing the computational complexity and improving the computational efficiency. Bilinear interpolation is a commonly used and effective interpolation method. It converts the data into a grid of target resolution by linearly interpolating the original data in the horizontal and vertical directions. This processing method can not only reduce the spatial complexity of the data, but also retain the spatial correlation of the data to a certain extent, making the interpolated data closer to the actual situation. The global spatial analysis and prediction of droughts reveals the key characteristics of drought frequency, intensity and duration in different regions for this embodiment. This helps to better understand the spatiotemporal distribution of drought events, thereby providing a scientific basis for disaster risk assessment and the formulation of response measures in relevant areas.

[0078] To verify the reliability of the calculated SPEI in this example, the SPEIbase (https: / / digital.csic.es / data / 10261 / 332007) for the period 1979-2022 was used. The SPEI database is based on monthly precipitation and potential evapotranspiration from the Climate Research Group at the University of East Anglia, based on FAO-56 Penman-Monteith estimates of potential evapotranspiration. The SPEIbase provides SPEI data at multiple time scales from 1 to 48 months, with a spatial resolution of 0.5 and a monthly temporal resolution. To date, it covers the period from January 1901 to December 2020. The SPEIbase has been evaluated and applied by many studies. In the comparative analysis, the spatial resolution of the newly developed SPEIBase was resampled to 1 based on the bilinear method in order to match the spatial resolution of the SPEI.

[0079] This study uses weather station data provided by the National Centers for Environmental Information (NCEI) under the National Oceanic and Atmospheric Administration (NOAA) to study the impact of climate change on surface water resources. The precipitation data covers continuous observation records from 1929 to the present, with a frequency of daily or monthly, depending on the observation strategy of each weather station. The monthly precipitation data of some typical regions from 1979 to 2022, as well as the monthly average temperature (Tavg), maximum temperature (Tmax), minimum temperature (Tmin), precipitation (Pre), relative humidity (RHU), wind speed (WIN) and sunshine duration (SSD) data are extracted and calculated using the Penman-Monteith formula modified by FAO. The SPEI index is calculated by using the monthly precipitation data of the station and potential evapotranspiration to verify the accuracy of the predicted data.

[0080] Previous studies have mainly focused on time series analysis of regional average variables, and relatively few studies have been conducted on the joint evolution of drought in time and space (Mishra et al., 2011). Traditionally, these studies have focused on average time series analysis based on data. In addition, although existing studies have mainly focused on drought prediction on long time scales, analysis and prediction on short time scales have not received sufficient attention. Therefore, the focus of this example analysis is on the Figure 1As shown in the figure, firstly, multi-source data sets, including daily precipitation data from ERA5 and daily potential evapotranspiration data calculated by the modified Penman-Monteith formula of the Food and Agriculture Organization of the United Nations (FAO), are used to obtain monthly data for calculating drought index by adding these daily data. And standardized calculations are performed to obtain monthly data of SPEI index. Compared with the existing SPEI index product SPEIBase, the differences between the two in terms of data source, time scale, calculation method and spatial range are compared to evaluate the accuracy and reliability of the self-calculated SPEI index. Secondly, predictor variables are screened, and the spatiotemporal trend changes of predictor variables and drought are analyzed and visualized, and significance tests are performed to analyze the potential development trend of drought; drought events are effectively identified by visualizing and quantifying the duration, scale and severity of drought from 1979 to 2022. Finally, data conversion is performed to extract features from multi-source data, and data processing is normalized as needed, sliding window operation is performed, and DMeyer wavelet decomposition is used to smooth the data to increase the richness and diversity of the data to ensure the stability and convergence of the model. Two schemes were used to combine it with the LSTM model for prediction, and the accuracy was compared with that of a single LSTM model. Site validation and cross validation were also performed to evaluate the prediction effect.

[0081] As an improved drought index, SPEI takes into account atmospheric precipitation and potential evapotranspiration factors and is suitable for studying the impact of global warming on drought severity. It is based on the probability distribution of accumulated precipitation and potential evapotranspiration differences at different time scales. The multi-scale nature of SPEI can identify different drought types and drought impacts (Vicente-Serrano, 2012). The FAO-56 Penman-Monteith model estimates PET using relevant data to verify the PET data of ERA5 at the grid scale. SPEI enables users to assess the occurrence of short-term (duration of 1 month), medium-term (3 to 12 months) and long-term droughts (12 months and longer) (Guttman, 1999; Labedzki, 2007; Trnka et al., 2009).

[0082] The calculation of SPEI depends on the accumulated deficit or surplus of water balance on different time scales (D i ). i It can be determined from the precipitation (Pre) and potential evapotranspiration (PET) on a given day:

[0083] D i =Pre-PET (1)

[0084] The water balance needs to be normalized to a probability distribution to obtain the SPEI series. The best distribution for SPEI calculation is the generalized extreme value (GEV) distribution (Stagge et al., 2015; Wang et al., 2021), which can overcome the limitations of the original SPEI through the generalized logistic distribution of short-term accumulation (1 to 2 months) (Vicente-Serrano et al., 2010). Therefore, this example uses the GEV distribution to normalize the D series to the SPEI data series (Wang et al., 2021).

[0085] The probability distribution function of the D series is given by 2F(x)

[0086]

[0087] P is the probability of determining the D value, calculated by formula 3:

[0088] P=1-F(x) (3)

[0089] P≤0.5, W is calculated by formula 4,

[0090]

[0091] P>0.5, W is calculated by formula 6

[0092]

[0093] In formula 7, SPEI is the drought index, c0=2.515517, c1=0.802853, c2=0.010328, d1=1.432788, d2=0.189269, and d3=0.001308.

[0094] The severity classification of the SPEI index is shown in Table 1:

[0095] Table 1 Classification of severity levels according to SPEI index

[0096]

[0097] The Theil-Sen Median slope estimation and Mann-Kendall test are performed below;

[0098] Theil-Sen Median method, also known as Sen slope estimation, is widely used in time dynamic analysis and exploration of interannual variation characteristics (Peng et al., 2016). Theil-Sen median trend analysis is a statistical method based on nonparametric statistics (Sen, 1968; Theil, 1950), which is suitable for analyzing long time series trends. Its calculation formula is:

[0099]

[0100] Where: x j and x i is time series data. β greater than 0 indicates that the time series presents an upward trend; β less than 0 indicates that the time series presents a downward trend.

[0101] The nonparametric MK test (Mann, 1945) is often used to assess the significance of monotonic trends in time series in climate, meteorological, and hydrological data (Cailas et al., 1986, Hipel et al., 1988, Zetterqvist, 1991, Chiew, and McMahon, 1993, Yu et al., 1993, Gan, 1998, Lins, 1999, Douglas et al., 2000, Yue et al., 2002, Onoz, 2003, Yue, 2004, Topaloglu et al., 2012). The MK test statistic (S) is calculated in the following equation:

[0102]

[0103] In the formula, x i and and x j S and S are the data values ​​at i and j, and n and S represent the length of the data set. A positive value of S indicates an upward trend, while a negative value of S indicates a downward trend. The following expressions are used as an assumption for a sequence with data length n>10 and data that is approximately normally distributed (variance (σ2=1) and mean (μ=0) values).

[0104]

[0105] In this equation, P is the number of parallel groups, and the aggregation symbol (∑sum) represents the sum of all parallel groups. i The sum is the number of data values ​​in the Pth group. If there is no tied group, this summary process can be ignored. After calculating the variance of the time series data using formula (12), the standard Z value is calculated according to formula (13).

[0106]

[0107] The calculated standard Z and values ​​are compared with the standard normal distribution table with two-tailed confidence level (α=10%, α=5% and α=1%). If |Z|>|Z1-α / 2|, the null hypothesis (H0) is invalid. Therefore, the trend is statistically significant. Otherwise, H0 and the hypothesis is accepted that the trend is not statistically significant and there is no trend in the time series (trendless time series). In this study, the MK method was used with a two-tailed confidence level of 95%.

[0108] Drought event identification

[0109] Run theory is a method to extract drought events by setting relevant thresholds based on the characteristics of drought index in time series (Yevjevich, 1967). The specific process of drought event identification using the threshold method is as follows: the threshold is set to -0.5 (Wu, J., 2019), and values ​​below the threshold indicate the beginning of a drought event, while values ​​returning to or exceeding the threshold indicate the end of a drought event. Drought duration is defined as the length of time that the value remains below the threshold, drought severity refers to the sum of the absolute values ​​of SPEI during the drought event, and drought intensity is the drought severity divided by the drought duration, such as Figure 3 shown.

[0110] Drought duration = DI0-DI(t)(14)

[0111]

[0112]

[0113] Where DI0 and DI(t) represent the end and start of a drought event, respectively, SPEI represents the drought index, D represents the end time, and t represents time.

[0114] Discrete Wavelet

[0115] Wavelet analysis can be divided into continuous wavelet transform (CWT) and discrete wavelet transform (DWT). Continuous wavelet transform is not often used for prediction due to its computational difficulty and processing time requirements (Kisi, 2011). Discrete wavelet transform is often used for prediction due to its small computational complexity and simple application. DWT is given by equation (16):

[0116]

[0117] Where m and n are integers controlling scale and time; m(t) is the mother wavelet. Parameter S o and τ oThe most common choices for are 2 and 1, respectively. According to Mallat's theory, a discrete time series can be decomposed into a series of linear neutral approximate signals and detail signals by inverse discrete wavelet transform. Equation (17) gives the inverse DWT according to Mallat (1989):

[0118]

[0119] in The scale is s = 2 m and τ = 2 m The wavelet coefficients of the discrete wavelet at n and .

[0120] LSTM model based on wavelet transform

[0121] As a special recurrent neural network (RNN) model, the long short-term memory network (LSTM) has excellent performance in time series data modeling. Compared with the traditional RNN, the special feature of LSTM is the complex structure of its hidden layer and the introduction of the cell state gate structure, which can control the flow of memory information and realize the long-term transmission and memory of information (Yang et al., 2019). Due to these characteristics, the LSTM model is widely used in sequence data that needs to consider long-term information and can carry longer time steps (Cui et al., 2022), such as Figure 4 shown.

[0122]

[0123] Where σ represents the sigmoid activation function. Indicates the multiplication of the corresponding elements of the matrix, * indicates convolution, and the information passes through the forget gate f t To decide whether to continue to pass backward. The information that continues to be passed to the input gate and the content that needs to be updated are updated through the tanh layer. t 、h t-1 Perform convolution calculations, respectively with c t-1 Combination to get h t and c t .W xi , W hi , W ci , W xf , W hf , W cf , W xc , W hc , W xo , W ho and W co represents the weight, b i 、b f 、b c and bo . In the field of global drought prediction, data such as the drought index SPEI for one month and multi-source meteorological factors have obvious time scale volatility. In practical applications, the accuracy of using a single LSTM model to predict these non-stationary signals may not be high. In order to improve the accuracy, this embodiment introduces the Wavelet Transform-LSTM (WT-LSTM) model. WT-LSTM combines wavelet analysis and LSTM model, which is suitable for modeling and prediction of non-stationary signals and can better capture the change pattern in time series. The method includes two stages, namely decomposition and reconstruction. First, the sequence is decomposed into high-pass (details) and low-pass (approximation) filters to extract the high-frequency and low-frequency components of the sequence, respectively. The low-frequency signal (approximation) is estimated using the LSTM prediction model, and the detail signal (noise) is trained separately using a neural network. The data are divided into a training set (70%) and a validation set (30%). A new signal is reconstructed from the approximation and details of the predicted values ​​of the sub-series using an inverse wavelet transform. Finally, the joint prediction is obtained by adding the two component predictions. The main advantage of the wavelet transform is that it can neutrally analyze different methods at different time scales. In this study, MATLAB was used to perform Dmeyer mother wavelet decomposition on the time series of drought index, and wavelet decomposition was performed at decomposition level 3. It has been proven through multiple studies that the drought prediction accuracy of the hybrid model is significantly higher than that of the single model. By applying this model to drought monitoring, this embodiment can more accurately capture drought trends and patterns on a global scale.

[0124] Model performance evaluation

[0125] To evaluate the performance of each prediction model, this embodiment uses the correlation coefficient (R), mean absolute error (MAE), and Nash efficiency coefficient (NSE). These indicators provide a quantitative assessment of the prediction accuracy and the agreement between the predicted value and the observed value. These include standard deviation and correlation metrics to match the observed data and the predicted data. To evaluate the accuracy of the model in the test set, this embodiment uses the following evaluation indicators: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), ratio of RMSE to standard deviation of observations (RSR), standard deviation ratio (RSD), Nash-Sutcliffe efficiency coefficient (NSE), correlation coefficient (R): used to evaluate the prediction accuracy of the spatial distribution of the model at each lead time, and measure the degree to which the model explains the variation of the observed data. This embodiment uses these evaluation indicators to comprehensively evaluate the performance of each model on the test set, and visually compares their performance based on the charts.

[0126] The following is a description of the research results and discussions of this embodiment:

[0127] SPEIbase Assessment:

[0128] After calculating the SPEI index, the SPEI index of the new dataset calculated in this embodiment is compared with the existing SPEI product. There is obvious spatial heterogeneity in the correlation. The correlation is higher in the Americas, Europe and China, while the correlation is lower in the Amazon River Basin, Central Africa and Central Asia. Figure 5 As shown. Generally, SPEI-GD and SPEIbase are more consistent on long-term time scales. This is because the long-term time scale integrates more water deficit information while losing more details about water changes, making the two SPEI series smoother and therefore more correlated. The data of SPEIbase mainly rely on meteorological stations, but the number of stations is limited, which leads to the relatively low accuracy of SPEIbase. In contrast, the newly developed SPEI-GD data in this study uses reanalysis data and integrates data from multiple sources such as satellites, observation sites and models. This combination of multi-source data greatly reduces the impact of missing data and improves the accuracy and reliability of the data.

[0129] Predictor screening

[0130] The selection of predictor variables is very important in developing machine learning and deep learning models. The Pearson correlation coefficient is used to identify a set of important predictor variables. The results are as follows Figure 6 As shown. Meteorological drought is a complex process that may be affected by many factors. Seven meteorological variables, namely precipitation (Pre), potential evaporation (PET), land surface temperature (LST), 10m U wind component (u10), 10mV wind component (v10), evaporation (e), and volumetric soil water layer 1 (swvl1), were used as potential influencing factors of the SPEI index. The results showed that precipitation, land surface temperature, and potential evaporation were the main factors of SPEI. The occurrence of drought usually stems from long-term accumulated precipitation deficiency. Precipitation provides input, and the positive correlation between precipitation and SPEI also shows that drought will decrease with increasing precipitation, with an average correlation coefficient of 0.51. The trend of temperature change is opposite. Temperature affects soil moisture loss by affecting potential evapotranspiration and heat flux, thereby affecting drought. The results show that potential evapotranspiration is negatively correlated with the SPEI index, and drought will increase with increasing potential evapotranspiration. In this study, potential evapotranspiration, land surface temperature, and precipitation were selected as predictors for drought prediction of the SPEI index. Figure 6 The correlation coefficient between latent variables and SPEI index. The latent variables include precipitation (Pre) such as Figure 6 As shown in (a), the potential evaporation capacity (PET) is as follows Figure 6 As shown in (b), the land surface temperature (LST) is Figure 6 As shown in (c), the 10m U wind component (u10) is Figure 6As shown in (d), the 10m V wind component (v10) is Figure 6 As shown in (e), the evaporation amount (e) is as follows Figure 6 As shown in (f), the volume soil water layer 1 (swvl1) is Figure 6 As shown in (g).

[0131] Spatial and temporal trend analysis:

[0132] like Figure 7 As shown in the figure, a total of 44 years of multi-source data from 1979 to 2022 were studied to analyze the spatiotemporal trend changes of precipitation (Pre) (a), potential evapotranspiration (PET) (b), land surface temperature (LST) (c), and SPEI index (SPEI) (d) time series. Figure 7 It is obvious that the global land annual precipitation (Pre) and annual potential evapotranspiration (PET) show extremely uneven spatial trend distribution. Specifically, the trend coefficient of annual precipitation (Pre) is less than -0.2 mm (a -1 ) are mainly distributed in southern Africa, central and southern South America including parts of Brazil, Bolivia, Argentina and other countries or regions, and western North America, accounting for 5.23%. The trend coefficient of annual precipitation (Pre) is greater than 0.3mm (a -1 ) are mainly distributed in the southern United States, tropical rain forests in South America, South Asia, southern North Africa, and islands near the equator, accounting for 6.68%. The annual potential evapotranspiration (PET) has an increasing trend coefficient greater than 0.15 mm (a -1 ) are mainly concentrated in the eastern part of South America, Brazil and southern Argentina, southern Australia, Europe, Central Asia, most of China, northern and southern Africa, accounting for 27.82%. The trend coefficient is less than -0.05mm (a -1 ) mainly include a small part of western South America, the United States, South Asia and other regions, accounting for 3.25%. The annual surface temperature (LST) increase trend coefficient is greater than 0.05T (a -1 ) are concentrated in large areas such as northern Asia and Europe, accounting for 11.47%; the trend coefficient is less than -0.005T (a -1 ) include central North America and a small part of South America, accounting for 0.35%. These trend results emphasize the significance of climate change in these regions. Figure 7(d) SPEI index trend study, the trend coefficient is less than -0.01 (a-1) in the northern and southern Africa, western Europe, West Asia, Central Asia, western Australia, western North America, showing a trend of drought, due to the decrease in precipitation and increase in evaporation in these regions. The results show that compared with the past, Africa and Asia still show a significant trend of drought, while precipitation in parts of North America and South America shows a significant increase. After analysis, the overall trend shows a slow upward trend.

[0133] This embodiment also calculates the average annual trend of each season of the multi-source data. Figure 8 It shows that drought varies significantly in different seasons and factors. Figure 8 (a) The precipitation change trend is mainly concentrated in the range of -0.6 to 0.6 (a)-1. A small number of areas in southern Africa, southern China (including southern China and southwestern China), and central and southern South America show a decreasing trend, among which the trends in South America and China are particularly significant. The increasing trend is mainly distributed in northern South America and tropical islands, especially the significant increasing trend. Figure 8 The potential evapotranspiration trend in (e) is mainly concentrated in the range of -0.1 to 0.4(a)-1, accounting for 91.60%. The regions with significant increases are distributed in northern Africa, western Europe, Central Asia and parts of northern China, while there is a significant decrease trend. Figure 8 (i) Areas showing significant temperature increases include most of Asia. Figure 8 The SPEI trend of (m) shows that it is mainly distributed in northern Africa, South Asia and northwest China. Taking northern Africa as an example, the aridification of the region is mainly caused by the increase in potential evapotranspiration and the rise in temperature in some areas. In summer, Figure 8 (b) The areas with a significant decrease in precipitation mainly include a small part of Africa, central China, and parts of Southeast Asia and South Asia. The trend of increasing precipitation in Africa gradually moves from central to southern Africa with the change of seasons, and is concentrated in tropical archipelagos, northern South America, African tropical rainforests, South Asia, eastern China, etc., with an increasing trend coefficient exceeding 0.6 (a) -1 . Figure 8 In (f), the regions with significant increase in potential evapotranspiration are located in Europe and Central Asia, accounting for 8.16%, while countries or regions such as Ethiopia and South Asia show a significant decreasing trend. Figure 8 (n) shows that the areas showing drought in summer are mainly distributed in Brazil, northern Africa and West Asia, mainly due to reduced precipitation and increased potential evapotranspiration. Figure 8 (c) The precipitation trend is consistent with that in spring, and the area is larger than that in spring; Figure 8(g) The potential evapotranspiration increases in Brazil and parts of North and South Africa, and decreases in South Asia. Figure 8 (k) shows that the temperature trend in the Arctic coastal areas in autumn has increased significantly and over a larger area, which is mainly caused by global warming. Figure 8 (o) shows that in winter, drought will occur in countries or regions such as the Middle East, South Africa, and Brazil in northern Africa, which is closely related to precipitation. Figure 8 In (d), the areas with a significant decrease in precipitation are mainly concentrated in Africa, Brazil, Argentina and other countries or regions. Compared with autumn, the areas with increased precipitation remain the same. Figure 8 (h) The areas with increased potential evapotranspiration are mainly distributed in a few areas at low latitudes; Figure 8 (l) The area with increased temperature in autumn is larger than that in autumn, accounting for 5.91%. Figure 8 (p) The areas showing drought in mid-winter are mainly consistent with those in autumn, but their area is only 2.92%. From a global perspective, the areas with greater volatility in global drought trends are mainly concentrated in Africa, Central Asia and South America. The frequency density map shows that the seasonal trend of precipitation remains consistent, and the potential evapotranspiration density trend first shifts to the right and then to the left, indicating that the potential evapotranspiration shows a trend of first increasing and then decreasing, and the temperature also shows the same trend. Figure 8 Annual trend analysis of the four seasons around the world from 1979 to 2022. Among them, Figure 8 (a, e, i, m) represents spring, Figure 8 (b, f, j, n) represents summer, Figure 8 (c, g, k, o) represents autumn, Figure 8 (d, h, l, p) represents winter.

[0134] Drought characteristics

[0135] Fig. 9 (a, c, e and g) Annual average drought duration, severity, intensity and frequency of SPEI index from 1979 to 2022; Fig. 9 (b, d, and f) Spatial trends of annual average drought duration, severity, and intensity of the SPEI index from 1979 to 2022. This study conducted a comprehensive spatiotemporal analysis of the SPEI index over global land and deeply explored the trends and characteristics of drought events during the period 1979-2022. Fig. 9 As shown in (a, b), most of the global droughts last from January to February, accounting for 62.34%. Droughts lasting more than 2 months are mainly distributed in Africa, northern and southern South America, Asia, western North America, Australia and other regions, accounting for 11.68%. Through trend analysis, it is found that North Africa, Brazil in South America, Argentina, Central Asia, the Middle East and other regions have a trend of increasing drought duration. Fig. 9(c) shows that the spatial distribution of drought severity is similar to the spatial distribution of drought duration. The area with a drought severity greater than 3 is relatively large, accounting for 11.20%. It can be seen from 9d that the total area where the severity of drought has decreased is much larger than the area where the degree of drought has increased, especially in Asia and North America, where the severity of drought has shown a clear downward trend. The areas with increased severity are mainly concentrated in a small part of Brazil in eastern South America, northern and southern Africa, Australia, Central Asia, the Middle East and other drought-prone areas, accounting for 34.79%. This phenomenon shows that these areas may have a large cumulative impact of meteorological drought in a short period of time due to the long sunshine hours and high potential evaporation in these areas, thereby aggravating the impact of drought. Fig. 9 As shown in (e), 90.45% of the areas with drought intensity are within the range of 0.6-0.8, and the areas with drought intensity greater than 0.8 are mainly distributed in parts of Asia, Africa, South America, etc., accounting for 5.19%. Fig. 9 (f) shows that the total area where drought intensity increased is larger than the area where intensity decreased. The areas where drought intensity increased significantly are distributed in Africa, North America, Asia, and a small part of South America, accounting for 4.83%.

[0136] This embodiment also calculates the average annual drought duration, severity and intensity for each season. Since some pixels do not experience drought in a season in a given year, the spatiotemporal data is discontinuous, so this embodiment does not calculate the trend. Fig.10 The results show that drought intensity varies significantly in different seasons. Fig.10 In (ad), this example shows that in spring, droughts last longer in a few areas such as northern Africa, Brazil in South America, South Asia and East Asia, accounting for 10.66%. This is consistent with the regions where droughts occur more frequently. Globally, the spatial pattern of drought duration and severity is opposite, mainly distributed in certain areas of North America, Asia, South America and Africa, accounting for 15.58%. The severity is similar to that of higher intensity, but the area of ​​higher intensity is smaller than that of higher severity, accounting for 8.64%. Fig.10 In the (eh) of the 2013 World Climate Change Summit, droughts were more frequent, more severe and lasted longer in parts of northern Africa in the summer because of higher potential evapotranspiration and lower precipitation ( Figure 8 At the same time, the intensity, severity and frequency of drought in China, North Asia, the United States, Brazil and other countries or regions are also high. Fig.10The (il) shows that in autumn, the areas with longer duration and higher frequency of drought are concentrated in North Africa, South Africa, Asia, central and southern South America, western Australia, the United States, Europe and other countries or regions. This is caused by the decrease in precipitation. At the same time, the spatial pattern of severity and intensity in these countries or regions is opposite to that of the continents. Fig.10 The (mp) shows that in winter, the duration, frequency and intensity of droughts in Brazil, North Africa and Asia are greater. The greater evapotranspiration in these regions may be the cause of the droughts ( Figure 8 (h)).

[0137] Overall performance of the prediction model

[0138] according to Fig.11 , by evaluating the two models, the WT-LSTM model has a relatively higher prediction accuracy than the single model LSTM. The WT-LSTM model accuracy of the first solution is 65% higher than the LSTM model, and the second solution is 52% higher. Therefore, WT-LSTM can more effectively capture information in the time and space dimensions; secondly, the accuracy of the WT-LSTM model of the two solutions is compared. The first solution (MAE=0.031, MSE=0.002, RMSE=0.039, RSR=0.321, RSD=0.959, NSE=0.897, R=0.947) and the second solution (MAE=0.028, MSE=0.001, RMSE=0.036, RSR=0.295, RSD=0.904, NSE=0.913, R=0.960) (Table 2-3). The prediction accuracy of the WT-LSTM model of the second solution is higher.

[0139] This example also evaluates the performance in each season. Figure 12-13 It shows that for each season, all models perform about the same Fig.11 The results are similar to those shown in , among which the WT-LSTM model has the highest accuracy. Compared with the two solutions, the second solution has higher accuracy than the first solution. Compared with the single LSTM model, the WT-LSTM model can extract the multi-scale features of the signal including high-frequency and low-frequency signals by wavelet transform, which can better capture the patterns and trends in time series data. At the same time, wavelet can also perform noise reduction on the data, improve the model's ability to recognize the true pattern of the signal, and use the complementary advantages of the two methods to improve the generalization ability and prediction accuracy of the model.

[0140] Table 2 Performance indicators of the first scheme Q1 comparing observed and predicted SPEI in different time periods

[0141]

[0142]

[0143] Table 3 Performance indicators of the second scheme Q2 comparing observed and predicted SPEI in different time periods

[0144]

[0145] Comparison of model solutions

[0146] like Fig.14 As shown, the easy-to-converge MSE is used as the loss function of the data to evaluate the training and testing process of the model. It is found that the loss function values ​​of the training set and test set of all data are constantly decreasing and converging, indicating that the training set and test set are well fitted without overfitting (Q1 represents scheme one, and Q2 represents scheme two).

[0147] Site Verification

[0148] In this study, this embodiment selected multiple sites from around the world to verify the Standardized Precipitation Evapotranspiration Index (SPEI). Through detailed analysis and screening, six representative sites were selected for in-depth research and analysis. These sites are distributed in six typical regions, including China, Kenya, Italy, Australia, the United States, and Argentina. Table 4 lists the performance evaluation indicators of the observed values ​​and predicted values ​​in these regions, including the correlation coefficient (R), Nash-Sutcliffe efficiency coefficient (NSE), and mean absolute error (MAE), where Q1 represents Scheme 1, Q2 represents Scheme 2, Figure 15-16The correlation diagrams of six typical regions of the two schemes are listed. First, for the site verification of six geographical locations including China (A), Kenya (B), Italy (C), Australia (D), the United States (E), and Argentina (F), this embodiment carried out SPEI prediction for the Q1 and Q2 schemes. In terms of the correlation coefficient (R), for the Q1 scheme, China's R reached 0.802, Kenya's was 0.841, Italy's was 0.822, Australia's was 0.798, the United States' was 0.820, and Argentina's was 0.812; the results of the Q2 scheme were 0.918, 0.921, 0.916, 0.912, 0.916, and 0.920, respectively. This shows that in terms of R, the prediction effect of the Q2 scheme is significantly better than that of the Q1 scheme. Secondly, this embodiment evaluates the prediction results using the Nash-Sutcliffe efficiency coefficient (NSE) and the mean absolute error (MAE). For NSE, the results of Q1 scheme in China, Kenya, Italy, Australia, the United States, and Argentina are 0.621, 0.686, 0.670, 0.632, 0.661, and 0.653, respectively, while those of Q2 scheme are 0.833, 0.827, 0.837, 0.821, 0.830, and 0.840, respectively. In terms of MAE, the results of Q1 scheme are 0.482, 0.440, 0.452, 0.485, 0.447, and 0.453, and those of Q2 scheme are 0.312, 0.328, 0.310, 0.330, 0.306, and 0.299. Overall, Q2 scheme shows better prediction performance in both NSE and MAE indicators, and is closer to the actual observations. Q2 scheme combines multiple meteorological elements to more comprehensively describe climate change and hydrological cycle. The single prediction SPEI index series of the Q1 scheme cannot capture the complex relationship between different meteorological elements. The integration of multiple meteorological elements into the prediction model can better understand the dynamic changes of the climate system, capture the impact of climate change, better describe the complexity of the climate system, and help improve the model's ability to explain and predict drought changes.

[0149] Table 4 Performance evaluation of observed and predicted values ​​in representative regions

[0150]

[0151]

[0152] Product Cross Validation

[0153] Cross-validation is crucial in identifying spatial differences between products and forecasts. It needs to consider the spatial resolution of different data sets, the way the data is generated, and the changes over time. Specifically, by selecting the test set time period, this embodiment can group the data in the study area into four seasons and calculate the average value of each season for cross-validation. This can study the differences between forecast results and products from multiple time scales, providing a more comprehensive data basis for comprehensive evaluation, such as Fig.17 (ac spring; df summer; gi autumn; jl winter). The comparison of the prediction results of SPEIBase products and the two schemes shows a relatively consistent spatial distribution, and there are differences in some areas. The correlation coefficients R of the first scheme in the four seasons product verification results are 0.778, 0.818, 0.755, and 0.762, and the MAE errors are 0.090, 0.087, 0.091, and 0.070; the correlation coefficients R of the second scheme verification results are 0.853, 0.874, 0.837, and 0.815, and the MAE errors are 0.069, 0.070, 0.070, and 0.094. In contrast, the prediction method of the second scheme is more accurate. From the perspective of spatial scope, the four seasons are mainly drought-prone areas in northern and southern Africa, South America, and parts of Asia. There are obvious differences in the prediction results, which leads to changes in the drought level. The frequency histograms of the four seasonal differences show that the differences in the forecast results are normally distributed, with good error stability and small differences in changes between data points.

[0154] Fig.17 It shows the comparison between the prediction results of the two models and the existing product results (ac spring; df summer; gi autumn; jl winter); the histogram represents the frequency of the difference (Q1 represents the difference of the first scheme, Q2 represents the difference of the second scheme, and the error accuracy of the two schemes is marked).

[0155] Water resources management strategies and countermeasures in drought areas

[0156] For areas with frequent and severe droughts, it is recommended to actively prepare for drought resistance, such as (1) adopting water-saving irrigation technology: introducing efficient water-saving irrigation technology, such as drip irrigation and sprinkler irrigation systems, to improve agricultural water resource utilization efficiency and reduce water resource waste; (2) water resource reuse: promoting water resource reuse technology, such as rainwater collection and wastewater treatment, to provide additional water supply; (3) improving soil moisture management: improving soil management practices, such as maintaining tillage, reducing soil moisture evaporation loss, and improving soil moisture retention capacity; (4) building reservoirs and water conservancy projects: building reservoirs and water conservancy projects in drought-prone areas to increase water resource reserves and water supply capacity and alleviate drought; (5) water resource management mechanism: establishing a sound water resource management mechanism, including resource allocation and water use management, to ensure the sustainable use of resources; (6) water resource monitoring and early warning system: establishing a strengthened water resource monitoring and drought early warning system to monitor water resource conditions in real time and warn of drought in advance; (7) agricultural production efficiency: improving agricultural production efficiency and reducing dependence on water resources by selecting appropriate farming methods, crop planting structures and irrigation methods. In addition, for those areas that have been affected by drought for a long time, the following options can also be considered: (1) Regional water restriction measures: Implement regional water restriction policies to limit water use, reduce waste, and ensure the sustainable use of water resources; (2) Water resource allocation: Optimize the allocation of water resources and allocate water resources from water-rich areas to water-scarce areas to alleviate drought problems and ensure water supply; (3) Strengthen drought monitoring and early warning: Strengthen drought monitoring and early warning work, timely release drought early warning information, and guide residents and farmers to take effective measures to reduce losses caused by drought.

[0157] Meteorological drought is the antecedent of hydrological drought and agricultural drought. Monitoring and forecasting short-time-scale meteorological drought can effectively avoid the harm and impact caused by drought. The spatiotemporal distribution characteristics of meteorological drought on a one-month time scale are analyzed using precipitation (Pre), potential evapotranspiration (PET), land surface temperature (LST), and SPEI index. The prediction of global land data using a hybrid model (WT-LSTM) of two schemes is compared to verify the applicability of the hybrid model in the field of drought prediction. The differences and similarities of the two prediction methods are also compared. From 1979 to 2022, due to the decrease in precipitation and the increase in potential evapotranspiration, Africa, South America, Asia and other regions still show a clear trend of drought, while precipitation in parts of North America and South America shows a significant increase. After analysis, the overall trend shows a slow upward trend.

[0158] The temporal and spatial distribution of meteorological droughts shows significant differences in different seasons. The duration, frequency and intensity of droughts in summer and autumn are greatest in northern Africa, while in spring and winter, droughts in Brazil, North Africa, South Asia and other regions of South America are long in duration, high in frequency and high in intensity.

[0159] The first scheme has the best test performance among the two schemes, and the R of the second scheme and the two schemes are both above 0.9. The WT-LSTM hybrid model can well solve the problem of unsteady meteorological time series in short time scales and has applicability. The comparison of the two schemes reveals the high consistency of spatial differences. Both have obvious advantages in drought prediction and are good at predicting different types of drought conditions.

[0160] This embodiment also proposes a global drought spatiotemporal change AI prediction system, which is used to implement the global drought spatiotemporal change AI prediction method, and the system includes: a prediction factor acquisition module and a drought spatiotemporal change prediction module;

[0161] A prediction factor acquisition module is used to obtain prediction factors of the area to be predicted, wherein the prediction factors include: precipitation, potential evapotranspiration and surface temperature;

[0162] The drought spatiotemporal change prediction module is used to input the prediction factors into the preset drought spatiotemporal change prediction model to obtain the drought prediction results; wherein, the drought spatiotemporal change prediction model is constructed based on the WT-LSTM model and is trained based on the constructed data set, and the data set includes: historical prediction factors, historical SPEI index and corresponding historical drought events.

[0163] This embodiment also proposes an electronic device, which includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the global drought spatiotemporal change AI prediction method is implemented.

[0164] In the above embodiments, all or part of the functions can be implemented by software, hardware, firmware or any combination thereof. When implemented in the form of a computer program product, the computer program product includes one or more computer instructions. When these instructions are loaded or executed on a computer, the process or function described in accordance with the embodiment of the present invention can be generated in whole or in part. We can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. These computer instructions can be stored in a computer-readable storage medium or transmitted between computer-readable storage media. For example, a computer instruction can be transmitted from a website, a computer, a server or a data center to another website, a computer, a server or a data center by wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, radio, microwave) mode. The computer-readable storage medium can be any available medium accessible by a computer, or a data storage device such as a server, a data center, etc. that integrates one or more available media. Available media include magnetic media (such as floppy disks, hard disks, tapes), optical media (such as DVDs) or semiconductor media (such as solid-state drives (SSDs)).

[0165] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An AI prediction method for global drought spatiotemporal changes, characterized in that: include: Obtain prediction factors for the area to be predicted, wherein the prediction factors include: precipitation, potential evapotranspiration and surface temperature; The prediction factors are input into a preset drought spatiotemporal change prediction model to obtain drought prediction results; wherein the drought spatiotemporal change prediction model is constructed based on the WT-LSTM model and is trained based on a constructed data set, and the data set includes: historical prediction factors, historical SPEI index and corresponding historical drought events.

2. The AI ​​prediction method for global drought spatiotemporal variation according to claim 1 is characterized in that: Constructing the dataset includes: Obtaining a historical SPEI index of a relevant region, and identifying historical drought events based on the historical SPEI index; The historical prediction factors are screened, and the data set is constructed based on the historical prediction factors, the historical SPEI index and the corresponding historical drought events; wherein the historical prediction factors include: historical precipitation, historical potential evapotranspiration and historical surface temperature.

3. The AI ​​prediction method for global drought spatiotemporal variation according to claim 2 is characterized in that: Obtain the historical SPEI index of relevant regions including: Obtain daily precipitation data from ERA5 and daily potential evapotranspiration data calculated by the modified Penman-Monteith formula; The daily precipitation data and the daily potential evapotranspiration data are added together and standardized to obtain the monthly data of the SPEI index.

4. The AI ​​prediction method for global drought spatiotemporal variation according to claim 3 is characterized in that: The SPEI index is: Wherein, W represents rainfall data, where: c0==2.515517, c1==0.802853, c2==0.010328, d1==1.432788, d2=0.189269, d3=0.001308; Identification of historical drought events includes: Drought duration = DI0-DI(t) Among them, DI0 and DI(t) represent the end and start time of the drought event respectively, D represents the end time, and t represents time.

5. The AI ​​prediction method for global drought spatiotemporal variation according to claim 1, characterized in that: Constructing the WT-LSTM model includes: combining the LSTM model with discrete wavelet transform to form the WT-LSTM model.

6. The AI ​​prediction method for global drought spatiotemporal variation according to claim 5 is characterized in that: Training the WT-LSTM model based on the constructed data set includes: Based on the discrete wavelet transform, the data in the data set are filtered using a high-pass filter and a low-pass filter to extract high-frequency components and low-frequency components of the data sequence respectively; The low-frequency component is estimated using an LSTM model, and then the high-pass component is trained and tested separately using the LSTM model, and finally the low-frequency component and the high-frequency component are analyzed separately.

7. The AI ​​prediction method for global drought spatiotemporal variation according to claim 1, characterized in that: After training the WT-LSTM model, the following steps are also included: The trained WT-LSTM model is evaluated based on preset evaluation indicators; wherein the preset evaluation indicators include: mean absolute error, mean square error, root mean square error, ratio of RMSE to standard deviation of observed values, standard deviation ratio, Nash-Sutcliffe efficiency coefficient, and correlation coefficient.

8. The AI ​​prediction method for global drought spatiotemporal variation according to claim 2 is characterized in that: Screening for the historical predictors includes: Analyze and visualize the spatiotemporal trend changes of predicted variables and drought, and perform significance tests to analyze the potential development trend of drought; wherein the predicted variables include: rainfall, surface temperature, and evapotranspiration; Drought events are identified by visualizing and quantifying the duration, scale, and severity of droughts over a preset time period, and the historical predictors are screened out through data association analysis.

9. An AI prediction system for global drought spatiotemporal changes, characterized in that: Used to implement the AI ​​prediction method for global drought spatiotemporal changes as described in any one of claims 1 to 8, the system comprises: a prediction factor acquisition module and a drought spatiotemporal change prediction module; The prediction factor acquisition module is used to acquire prediction factors of the area to be predicted, wherein the prediction factors include: precipitation, potential evapotranspiration and surface temperature; The drought spatiotemporal change prediction module is used to input the prediction factors into a preset drought spatiotemporal change prediction model to obtain a drought prediction result; wherein the drought spatiotemporal change prediction model is constructed based on a WT-LSTM model and is trained based on a constructed data set, and the data set includes: historical prediction factors, historical SPEI index and corresponding historical drought events.

10. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for predicting the global drought spatiotemporal changes according to any one of claims 1 to 8 is implemented.

Citation Information

Cited By

  • Ground surface settlement monitoring method and system fusing Beidou and InSAR data

    CN119437159A

  • Target area composite high-temperature drought prediction method based on dynamics

    CN120355022A

  • Kinetic-based target area compound high temperature drought prediction method

    CN120355022B

  • Time sequence prediction model-based arid region SPEI index prediction method and system

    CN121880792A

  • An interpretable drought prediction method based on time-frequency decomposition and deep learning

    CN122903871A