Basin sudden drought short-impending prediction method and system based on soil water content cumulative deficit

Abstract

The application discloses a kind of based on soil moisture content accumulated deficit watershed sudden drought short-impending prediction method and system, comprising: obtaining target watershed historical meteorological hydrology and soil moisture content data;Soil moisture content space-time prediction model is constructed, with the soil moisture content data of the first preset time length of history and the seasonal core driving factor as input, and the future second preset time length of daily soil moisture content prediction data is output;Based on the training label of multiple constraint sudden drought identification standard construction;Extract the accumulated deficit of accumulated percentage value of soil moisture content in the preset time window before the start of target prediction period relative to critical threshold as core feature, construct probability discriminant model, and output the probability of sudden drought in the future second preset time length in watershed.This application captures the rapid loss characteristics of soil moisture before sudden drought occurs, realizes short-impending scale prediction, provides technical support for watershed water resources management, agricultural disaster prevention and mitigation, water safety guarantee of middle route of South-to-North Water Diversion.

Description

Technical Field

[0001] This invention relates to the field of short-term prediction technology for extreme hydrological events, specifically to a method and system for short-term prediction of sudden droughts in watersheds based on the cumulative deficit of soil moisture content. Background Technology

[0002] Sudden droughts are frequent extreme hydrological events against the backdrop of global warming. They are characterized by rapid onset, short development cycles, and high disaster intensity, posing a serious threat to watershed water resource management, agricultural production safety, and ecosystem stability.

[0003] Current research on the prediction of sudden droughts in watersheds has made some progress, but there are still three key shortcomings: First, the accuracy of existing long-term soil moisture content prediction models decreases rapidly with the extension of the forecast period, and the error of long-term prediction values ​​is large. Directly identifying sudden droughts based on the prediction values ​​will lead to serious error propagation, resulting in frequent misjudgments and missed judgments. Second, existing methods for identifying sudden droughts fall into two categories. One type directly uses meteorological factors as input to construct a black-box classification model, skipping soil moisture content, a core physical indicator of sudden drought occurrence. The model has poor interpretability and cannot adapt to the local characteristics of the watershed. The other type directly applies a fixed threshold for identifying sudden droughts based on the predicted soil moisture content, which cannot avoid the identification bias caused by prediction errors. Third, existing methods generally use a universal fixed threshold for identifying sudden droughts, without considering the seasonal differentiation characteristics of the driving mechanism of sudden droughts. This makes it impossible to balance the missed judgment rate and the false alarm rate, and the prediction adaptability is poor for the high incidence of sudden droughts in spring and summer. Therefore, constructing a short-term prediction method for sudden droughts that can avoid the propagation of prediction errors and adapt to the seasonal differentiation characteristics of watersheds has important engineering application value for watershed drought prevention and control. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned deficiencies in existing technologies by providing a short-term prediction method for sudden droughts in watersheds based on the cumulative deficit of soil moisture content. This method effectively avoids the error propagation problem caused by the decline in the accuracy of long-term soil moisture content prediction, and solves the pain points of traditional fixed thresholds being unable to adapt to the seasonal differentiation characteristics of sudden droughts and the imbalance between false alarm and missed detection rates, thereby achieving accurate short-term prediction of sudden droughts in watersheds.

[0005] To address the aforementioned technical problems, the present invention adopts the following technical solution: A short-term prediction method for sudden droughts in watersheds based on cumulative soil moisture deficit includes the following steps: Historical long-term hydrological and meteorological grid data of the target watershed are collected, the historical long-term hydrological and meteorological grid data are standardized and preprocessed, and the preprocessed data is divided into training dataset and validation dataset. Based on the training dataset, a spatiotemporal prediction model for soil moisture content is constructed, and daily grid-point prediction data of root zone soil moisture content within a second preset time period is output. Based on historical long-series observation data, a binary classification training label is constructed using a multi-constraint sudden drought identification standard. Based on the daily root zone soil moisture content grid prediction data within the second preset time period, the cumulative percentile value of soil moisture content within the second preset time period is calculated with the prediction start time as the benchmark. The cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold is extracted as the input feature. The training label is used to train a probability discrimination model, and the trained probability discrimination model is used to output the probability of sudden drought occurrence in the target watershed in the future period. The optimal threshold for sudden drought in the target watershed is determined by the Youden index and the corresponding seasonal threshold. Based on the output of the occurrence of sudden drought and the optimal threshold of the corresponding season, the binary classification of sudden drought is completed.

[0006] Furthermore, the historical long-term hydro-meteorological gridded data includes root zone soil moisture content, total precipitation, potential evaporation, daily average temperature, 10m wind speed, relative humidity, net surface heat radiation, surface air pressure, snowmelt, etc., with a spatial resolution of 0.1°×0.1°.

[0007] Furthermore, the construction of a spatiotemporal prediction model for soil moisture content includes: employing a feature selection mechanism based on causal inference, using an enhanced regression tree model to quantify the causal contribution of historical long-term hydrological and meteorological grid data to changes in root zone soil moisture content, eliminating spurious correlation factors, and selecting seasonal core driving factors; using daily root zone soil moisture content grid data of the first preset historical duration and the aforementioned seasonal core driving factors as joint inputs, and using daily root zone soil moisture content grid data of the second preset duration as the prediction target, to construct a spatiotemporal prediction model for soil moisture content. The parameters of the spatiotemporal prediction model for soil moisture content are optimized using the training dataset, and the generalization ability of the model is evaluated using data from the validation dataset. A multidimensional accuracy evaluation index system is used to quantify the predictive performance of the model under different forecast periods.

[0008] Furthermore, the spatiotemporal prediction model for soil moisture content is configured as a neural network architecture with spatiotemporal feature extraction capabilities. Internally, it includes a spatial convolution module for capturing the spatial correlation of soil moisture within the watershed, and a temporal recursive module for capturing the memory and evolution patterns of soil moisture. The preset hyperparameters of the spatiotemporal prediction model for soil moisture content are set, including convolution kernel size, hidden layer dimension, optimizer type, learning rate, batch size, and iteration rounds, to complete the initialization of the spatiotemporal prediction model for soil moisture content.

[0009] Furthermore, the multidimensional accuracy evaluation index system includes at least a combination of determination coefficient, error measurement index and signal quality index; when the performance of the soil moisture spatiotemporal prediction model meets the preset accuracy threshold, the model parameters are fixed, and real-time data is input to output daily root zone soil moisture grid prediction data for a second preset duration.

[0010] Furthermore, the aforementioned multi-constraint sudden drought identification criteria include: (1) Initial determination: When the cumulative percentile value of soil moisture content drops from above the first preset high threshold to below the second preset low threshold at a rate greater than the preset rate of decrease, it is determined that a sudden drought has begun. (2) Termination judgment: When the cumulative percentile value of soil moisture content recovers to above the third preset recovery threshold, the drought is judged to have ended; (3) Duration determination: A sudden drought event is confirmed as valid and marked as 1 only if the duration from the start to the end of the drought is greater than or equal to the preset minimum duration. Otherwise, it is marked as 0.

[0011] Furthermore, training the probabilistic discriminative model includes: Based on the daily root zone soil moisture content grid prediction data within the second preset time period, the average soil moisture content of each watershed grid point is calculated according to a 5-day pentad time scale. Based on the empirical cumulative distribution function of historical long-sequence soil moisture content, the cumulative percentile value of soil moisture content for each pentad is calculated. Calculate the cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold; The cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold is used as the core input feature, and the historical anomaly values ​​of the average precipitation and potential evaporation of the corresponding phenological period are used as auxiliary features. Whether a sudden drought occurs within the second preset time period is used as a binary classification label to train the probability discrimination model. The parameters of the probabilistic discrimination model are optimized and trained, and the probability of sudden drought in the target watershed in the future is finally output.

[0012] Furthermore, the formula for calculating the cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold is as follows:

[0013] In the formula, denoted as t, where is the cumulative deficit of the cumulative percentile values ​​of soil moisture content over t year, and m represents the climatological prediction accuracy of soil moisture content. The cumulative percentile value of soil moisture content for the i-th gestation period. The threshold is the cumulative percentile value of soil moisture content at the critical point of sudden drought.

[0014] Furthermore, the optimal threshold for identifying sudden droughts in different seasons of the target watershed based on the Youden index includes: Plot an ROC curve with the false positive rate (FPR) of the validation dataset on the x-axis and the true positive rate (TPR) on the y-axis, and take the Youden index J = TPR. The probability value corresponding to the maximum FPR is used as the optimal discrimination threshold for the corresponding season; The steps use the area under the ROC curve (AUC), prediction accuracy, precision, recall, and F1 score as accuracy evaluation metrics to comprehensively assess the model's predictive performance.

[0015] On the other hand, the present invention provides a method for short-term prediction of sudden drought in watersheds based on cumulative soil moisture deficit, comprising: The data acquisition and preprocessing module is used to collect historical long-term hydrological and meteorological grid data of the target watershed, perform standardized preprocessing on the historical long-term hydrological and meteorological grid data, and divide the preprocessed data into training dataset and validation dataset. Soil moisture content prediction module is used to construct a spatiotemporal prediction model of soil moisture content based on the training dataset and output daily grid prediction data of soil moisture content in the root zone within a second preset time period. The probability discrimination module is used to construct binary classification training labels based on historical long-sequence observation data and multi-constraint sudden drought identification criteria; based on the daily root zone soil moisture content grid prediction data within the second preset time period, it calculates the cumulative percentile value of soil moisture content within the second preset time period with the prediction start time as the reference, extracts the cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold as the input feature, trains the probability discrimination model using the training labels, and outputs the probability of sudden drought occurrence in the target watershed in the future period using the trained probability discrimination model; The sudden drought discrimination module is used to determine the optimal discrimination threshold for sudden drought in the target watershed by season based on the Youden index. Based on the output of the sudden drought occurrence and the optimal threshold of the corresponding season, it completes the binary classification discrimination output of sudden drought.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention uses only data within the time window that the soil moisture content prediction model can accurately output, avoiding the error propagation problem caused by the decay of long-term prediction accuracy with the forecast period. It significantly reduces the false alarm rate and missed detection rate of sudden drought identification, and can provide reliable technical support for watershed water resources management, agricultural disaster prevention and mitigation, and water security guarantee of water source areas for inter-basin water transfer projects. Attached Figure Description

[0017] Figure 1This is a flowchart of a short-term prediction method for sudden drought in watersheds based on the cumulative deficit of soil moisture content, according to the present invention. Figure 2 This is a map showing the geographical location and water system distribution of a certain watershed according to an embodiment of the present invention; Figure 3 This is a graph showing the relationship between the threshold values ​​for sudden drought in spring, summer, autumn, and winter in a certain watershed and the model evaluation index in an embodiment of the present invention. It is used to illustrate the process of determining the optimal threshold values ​​for each season. Figure 4 This is a spatial distribution map of model identification indicators at grid points where a sudden drought occurred in a watershed in the summer of 2022, according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] Example 1 This embodiment provides a short-term prediction method for sudden droughts in watersheds based on the cumulative deficit of soil moisture content, including the following steps: Step 1: Collect historical long-term hydrological and meteorological gridded data of the target watershed, perform standardization preprocessing on the historical long-term hydrological and meteorological gridded data, and divide the preprocessed data into training dataset and validation dataset; Step 2: Construct a spatiotemporal prediction model for soil moisture content based on the training dataset, and output daily grid-point prediction data of soil moisture content in the root zone within the second preset time period; Step 3: Based on historical long-sequence observation data, construct binary classification training labels using multi-constraint sudden drought identification criteria; based on the daily root zone soil moisture content grid prediction data within the second preset time period, calculate the cumulative percentile value of soil moisture content within the second preset time period with the prediction start time as the benchmark, extract the cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold as the input feature, train the probability discrimination model using the training labels, and output the probability of sudden drought occurrence in the target watershed in the future period using the trained probability discrimination model; Step 4: Determine the optimal threshold for sudden drought in the target watershed by season based on the Youden index. Based on the output of the occurrence of sudden drought and the optimal threshold of the corresponding season, complete the binary classification and output of sudden drought.

[0020] The specific process of step 1 above is as follows: Step 1.1: Collect daily gridded hydrological and meteorological data for the target watershed over a long period of time, including root zone soil moisture content, total precipitation, potential evaporation, daily average temperature, 10m wind speed, relative humidity, net surface heat radiation, surface air pressure, snowmelt, etc. The spatial resolution of the data is 0.1°×0.1°. Step 1.2: Preprocess the collected long-term daily gridded hydrological and meteorological data of the target watershed, use linear interpolation to fill in missing values, and use the min-max normalization method to standardize all input factors to eliminate dimensional differences. Step 1.3: In accordance with the principle of no leakage in time series, divide the dataset into training dataset, validation dataset and test dataset to ensure that there is no time overlap between training data and validation data.

[0021] The specific process of step 2 above is as follows: Step 2.1: The specific architecture of the spatiotemporal prediction model for soil moisture content is not strictly limited; any existing model capable of spatiotemporal prediction of watershed soil moisture content can be used. In a preferred embodiment of the present invention, a Causal-ConvLSTM model incorporating LPCMCI causal priors is adopted. Step 2.2: Using a feature screening mechanism based on causal inference, the causal contribution of each candidate meteorological and hydrological factor to the change in soil moisture content in the root zone is quantified, spurious correlation factors are eliminated, and seasonal core driving factors are screened out. In a preferred embodiment of the present invention, the enhanced regression tree (BRT) model is used. Step 2.3: Construct a deep learning-based spatiotemporal prediction model for soil moisture content; using the historical first preset duration of daily root zone soil moisture content grid data and the seasonal core driving factors as joint inputs, and the future second preset duration of daily root zone soil moisture content grid data as the prediction target; the model is configured as a neural network architecture with spatiotemporal feature extraction capabilities, internally including a spatial convolution module to capture the spatial correlation of soil moisture within the watershed, and a temporal recursive module to capture the memory and evolution patterns of soil moisture; set the model's preset hyperparameters, including convolution kernel size, hidden layer dimension, optimizer type, learning rate, batch size, and iteration rounds, to complete the initialization of the model structure; Step 2.4: Perform model training and accuracy verification, and output prediction results; optimize the parameters of the model constructed in Step 2.3 using training data, and evaluate the model's generalization ability using verification data; quantify the model's prediction performance under different forecast periods using a multi-dimensional accuracy evaluation index system, wherein the index system includes at least a combination of multiple combinations of the coefficient of determination, error measurement index, and signal quality index; when the model performance meets the preset accuracy threshold, solidify the model parameters, and input real-time data to output daily root zone soil moisture content grid prediction data for the second preset time period in the future.

[0022] The specific process of step 3 above is as follows: Step 3.1: Based on the daily soil moisture content prediction data obtained in Step 2, calculate the average soil moisture content of the watershed grid points according to the time scale of 5 days and 1 pentad. Based on the empirical cumulative distribution function of historical long-term soil moisture content, calculate the cumulative percentile value (CP) of soil moisture content for each pentad. Step 3.2: Extract the cumulative percentile value of soil moisture content as the core input feature for sudden drought detection. The calculation formula is as follows:

[0023] In the formula, denoted as t, where is the cumulative deficit of the cumulative percentile values ​​of soil moisture content over t year, and m represents the climatological prediction accuracy of soil moisture content. The cumulative percentile value of soil moisture content for the i-th gestation period. The threshold is the cumulative percentile value of soil moisture content at the critical point of sudden drought.

[0024] Step 3.3: Construct a probabilistic discriminant model, using the critical threshold CP cumulative deficit obtained in Step 3.2 as the core input feature, supplemented by the historical anomaly values ​​of the corresponding phenological factors such as average precipitation and potential evaporation in the basin as auxiliary features, and using whether a sudden drought has occurred as the binary classification label to construct the model training set; Step 3.4: Optimize parameters and complete model training, and finally output the probability of sudden drought in the target watershed in the future period.

[0025] The specific process of step 4 above is as follows: Step 4.1: Determine the optimal threshold for seasonal sudden drought in the target watershed based on the Youden index. The specific method is as follows: Plot an ROC curve with the false positive rate (FPR) of the validation set as the horizontal axis and the true positive rate (TPR) as the vertical axis, and take the Youden index J = TPR. The probability value corresponding to the maximum FPR is used as the optimal discrimination threshold for the corresponding season; Step 4.2: The area under the ROC curve (AUC), prediction accuracy, precision, recall, and F1 score are used as accuracy evaluation metrics to comprehensively assess the model's predictive performance. Accuracy: The proportion of samples correctly classified by the model out of the total sample. Precision: The proportion of samples classified as "sudden drought" by the model that actually experience a sudden drought; measures the model's false alarm rate. Recall: The proportion of samples that actually experience a sudden drought that are correctly identified by the model; measures the model's false negative rate. F1 score: The harmonic mean of precision and recall, ranging from [0,1]. The closer to 1, the better the model's overall performance in imbalanced samples. ROC curve and AUC: The ROC curve is plotted with false positive rate on the horizontal axis and true positive rate on the vertical axis. AUC is the area under the ROC curve, ranging from [0,1]. The closer to 1, the stronger the model's binary classification ability; it is a core evaluation metric in imbalanced sample scenarios. In the formula, TP represents a true positive (correctly identifying sudden drought), TN represents a true negative (correctly identifying no sudden drought), FP represents a false positive (misjudging sudden drought), and FN represents a false negative (missing to diagnose sudden drought).

[0026] This river basin is a core water source area for the South-to-North Water Diversion Project and an important grain-producing area in the middle reaches of the Yangtze River. The basin covers an area of ​​159,000 km² and is located in a subtropical monsoon climate zone. Rainfall is unevenly distributed throughout the year, with over 65% of the annual rainfall occurring from June to September. Figure 2 The occurrence and development of floods and droughts in the Hanjiang River Basin are directly related to regional water security and sustainable socio-economic development, making it a key research area for flood and drought disaster prevention and control in my country. Against the backdrop of global warming, sudden drought events are frequent in the Hanjiang River Basin. These events, occurring during the peak spring and summer seasons, not only threaten the water supply security of the South-to-North Water Diversion Project but also cause widespread drought damage to farmland within the basin, resulting in severe agricultural yield reductions. Existing methods for predicting sudden droughts suffer from problems such as declining accuracy over long time series and poor adaptability to fixed thresholds, failing to meet the engineering requirements for short-term early warning of sudden droughts in the basin. Therefore, this invention takes the Hanjiang River Basin as the research object and conducts short-term prediction of sudden droughts based on the characteristics of cumulative CP deficit. The specific process is as follows: The entire watershed of a certain river basin was selected as the study area. Daily gridded hydro-meteorological data from ERA5-Land, collected from January 1, 1981 to December 31, 2022, were collected, including soil moisture content in the three root zones, total precipitation, potential evaporation, total evaporation, 10m wind speed, snowmelt, net surface heat radiation, and surface air pressure. The spatial resolution of the data was 0.1° × 0.1°. Missing measurements were filled using linear interpolation, and the data was standardized using min-max normalization. Following the principle of no time series leakage, the dataset was divided into a training set (1981-2010), a validation set (2011-2014), and a test set (2015-2022). The prediction time window was set to 15 days.

[0027] After collecting and preprocessing the model input data, a Causal-ConvLSTM soil moisture spatiotemporal prediction model integrating LPCMCI causal priors was first constructed. The model takes the historical 20-day daily root zone soil moisture content and seasonal core driving factors as inputs and outputs gridded data of the daily root zone soil moisture content for the next 15 days. The model training and validation were completed. The average coefficient of determination R² of the model validation set from 1 to 15 days reached 0.73, the peak signal-to-noise ratio PSNR was 21.83dB, and the root mean square error RMSE was 0.0177.

[0028] Subsequently, based on the predicted soil moisture content data, the CP value of soil moisture content for each pentad was calculated. The cumulative CP deficit of the first three pentads was extracted as the core input feature, and a LightGBM probabilistic discriminant model was constructed. Five-fold cross-validation was used to optimize the parameters and complete the model training. The model validation set AUC value reached 0.788, and the accuracy reached 0.715. Based on the Youden index, the optimal discrimination threshold for seasonal sudden drought in the Hanjiang River Basin was determined as follows: spring 0.499, summer 0.500, autumn 0.475, and winter 0.494. Figure 3 ).

[0029] The verification was carried out using a typical sudden drought event in a certain river basin in the summer of 2022 as an example. Figure 4 The results show that the method of the present invention can accurately identify the occurrence and development process of the sudden drought event 15 days in advance, effectively solving the problems of long-term prediction error transmission, and imbalance between missed detection rate and false alarm rate, and can provide reliable technical support for the prevention and control of sudden drought in the Hanjiang River Basin.

[0030] Example 2 This embodiment provides a short-term prediction method for sudden droughts in watersheds based on cumulative soil moisture deficit, including: The data acquisition and preprocessing module is used to collect historical long-term hydrological and meteorological grid data of the target watershed, perform standardized preprocessing on the historical long-term hydrological and meteorological grid data, and divide the preprocessed data into training dataset and validation dataset. Soil moisture content prediction module is used to construct a spatiotemporal prediction model of soil moisture content based on the training dataset and output daily grid prediction data of soil moisture content in the root zone within a second preset time period. The probability discrimination module is used to construct binary classification training labels based on historical long-sequence observation data and multi-constraint sudden drought identification criteria; based on the daily root zone soil moisture content grid prediction data within the second preset time period, it calculates the cumulative percentile value of soil moisture content within the second preset time period with the prediction start time as the reference, extracts the cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold as the input feature, trains the probability discrimination model using the training labels, and outputs the probability of sudden drought occurrence in the target watershed in the future period using the trained probability discrimination model; The sudden drought discrimination module is used to determine the optimal discrimination threshold for sudden drought in the target watershed by season based on the Youden index. Based on the output of the sudden drought occurrence and the optimal threshold of the corresponding season, it completes the binary classification discrimination output of sudden drought.

[0031] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0032] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations to the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.

[0033] All other parts not described in detail are existing technologies.

Claims

1. A method for predicting short-term droughts in watersheds based on cumulative soil moisture deficit, characterized in that, Includes the following steps: Historical long-term hydrological and meteorological grid data of the target watershed are collected, the historical long-term hydrological and meteorological grid data are standardized and preprocessed, and the preprocessed data is divided into training dataset and validation dataset. Based on the training dataset, a spatiotemporal prediction model for soil moisture content is constructed, and daily grid-point prediction data of root zone soil moisture content within a second preset time period is output. Based on historical long-series observation data, a binary classification training label is constructed using a multi-constraint sudden drought identification standard. Based on the daily root zone soil moisture content grid prediction data within the second preset time period, the cumulative percentile value of soil moisture content within the second preset time period is calculated with the prediction start time as the benchmark. The cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold is extracted as the input feature. The training label is used to train the probability discrimination model, and the trained probability discrimination model is used to output the probability of sudden drought in the target watershed in the future period. The optimal threshold for sudden drought in the target watershed is determined by the Youden index and the corresponding seasonal threshold. Based on the output of the occurrence of sudden drought and the optimal threshold of the corresponding season, the binary classification of sudden drought is completed.

2. The method for short-term prediction of sudden drought in a watershed based on cumulative soil moisture deficit as described in claim 1, characterized in that, The historical long-term hydro-meteorological gridded data includes root zone soil moisture content, total precipitation, potential evaporation, daily average temperature, 10m wind speed, relative humidity, net surface heat radiation, and surface air pressure, with a spatial resolution of 0.1°×0.1°.

3. The method for short-term prediction of sudden drought in a watershed based on cumulative soil moisture deficit according to claim 2, characterized in that, The construction of a spatiotemporal prediction model for soil moisture content includes: A feature screening mechanism based on causal inference was adopted, and the causal contribution of each historical long-term hydro-meteorological grid data to the change of soil moisture content in the root zone was quantified by using an enhanced regression tree model. False correlation factors were eliminated and seasonal core driving factors were screened out. Using the daily root zone soil moisture content grid data of the first preset duration and the seasonal core driving factor as joint inputs, and the daily root zone soil moisture content grid data of the second preset duration as the prediction target, a spatiotemporal prediction model for soil moisture content is constructed. The parameters of the spatiotemporal prediction model for soil moisture content are optimized using the training dataset, and the generalization ability of the model is evaluated using data from the validation dataset. A multidimensional accuracy evaluation index system is used to quantify the predictive performance of the model under different forecast periods.

4. The method for short-term prediction of sudden drought in a watershed based on cumulative soil moisture deficit according to claim 3, characterized in that, The spatiotemporal prediction model for soil moisture content is configured as a neural network architecture with spatiotemporal feature extraction capabilities. Internally, it includes a spatial convolution module for capturing the spatial correlation of soil moisture within the watershed, and a temporal recursive module for capturing the memory and evolution patterns of soil moisture. The model is initialized by setting preset hyperparameters, including convolution kernel size, hidden layer dimension, optimizer type, learning rate, batch size, and iteration rounds.

5. The method for predicting short-term drought in a watershed based on cumulative soil moisture deficit according to claim 3, characterized in that, The multidimensional accuracy evaluation index system includes at least a combination of determination coefficient, error measurement index and signal quality index; when the performance of the soil moisture spatiotemporal prediction model meets the preset accuracy threshold, the model parameters are fixed, and real-time data is input to output daily root zone soil moisture grid prediction data for a second preset duration.

6. The method for predicting short-term drought in a watershed based on cumulative soil moisture deficit according to claim 1, characterized in that, The aforementioned multi-constraint sudden drought identification criteria include: (1) Initial determination: When the cumulative percentile value of soil moisture content drops from above the first preset high threshold to below the second preset low threshold at a rate greater than the preset rate of decrease, it is determined that a sudden drought has begun. (2) Termination judgment: When the cumulative percentile value of soil moisture content recovers to above the third preset recovery threshold, the drought is judged to have ended; (3) Duration determination: A sudden drought event is confirmed as valid and marked as 1 only if the duration from the start to the end of the drought is greater than or equal to the preset minimum duration. Otherwise, it is marked as 0.

7. The method for short-term prediction of sudden drought in a watershed based on cumulative soil moisture deficit according to claim 1, characterized in that, The training probability discrimination model includes: Based on the daily root zone soil moisture content grid prediction data within the second preset time period, the average soil moisture content of each watershed grid point is calculated according to a 5-day pentad time scale. Based on the empirical cumulative distribution function of historical long-sequence soil moisture content, the cumulative percentile value of soil moisture content for each pentad is calculated. Calculate the cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold; The cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold is used as the core input feature, and the historical anomalies of the corresponding phenological average precipitation and potential evaporation are used as auxiliary features. The probability discrimination model is trained using whether a sudden drought occurs within the second preset time period as a binary classification label. The parameters of the probabilistic discrimination model are optimized and trained, and the probability of sudden drought in the target watershed in the future is finally output.

8. A method for predicting short-term drought in a watershed based on cumulative soil moisture deficit according to claim 7, characterized in that, The formula for calculating the cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold is as follows: In the formula, denoted as t, where is the cumulative deficit of the cumulative percentile values ​​of soil moisture content over t year, and m represents the climatological prediction accuracy of soil moisture content. The cumulative percentile value of soil moisture content for the i-th gestation period. The threshold is the cumulative percentile value of soil moisture content at the critical point of sudden drought.

9. A method for predicting short-term drought in a watershed based on cumulative soil moisture deficit according to claim 1, characterized in that, The optimal threshold for identifying sudden drought in different seasons in a target watershed based on the Youden index includes: Plot an ROC curve with the false positive rate (FPR) of the validation dataset on the x-axis and the true positive rate (TPR) on the y-axis, and take the Youden index J = TPR. The probability value corresponding to the maximum FPR is used as the optimal discrimination threshold for the corresponding season; The steps use the area under the ROC curve (AUC), prediction accuracy, precision, recall, and F1 score as accuracy evaluation metrics to comprehensively assess the model's predictive performance.

10. A method for short-term prediction of sudden drought in watersheds based on cumulative soil moisture deficit, characterized in that, include: The data acquisition and preprocessing module is used to collect historical long-term hydrological and meteorological grid data of the target watershed, perform standardized preprocessing on the historical long-term hydrological and meteorological grid data, and divide the preprocessed data into training dataset and validation dataset. Soil moisture content prediction module is used to construct a spatiotemporal prediction model of soil moisture content based on the training dataset and output daily grid prediction data of soil moisture content in the root zone within a second preset time period. The probability discrimination module is used to construct binary classification training labels based on historical long-sequence observation data and using multi-constraint sudden drought identification criteria; Based on the daily root zone soil moisture content grid prediction data within the second preset time period, the cumulative percentile value of soil moisture content within the second preset time period is calculated with the prediction start time as the benchmark. The cumulative deficit of the cumulative percentile value of soil moisture content relative to the preset critical threshold is extracted as the input feature. The training label is used to train the probability discrimination model, and the trained probability discrimination model is used to output the probability of sudden drought in the target watershed in the future period. The sudden drought discrimination module is used to determine the optimal discrimination threshold for sudden drought in the target watershed by season based on the Youden index. Based on the output of the sudden drought occurrence and the optimal threshold of the corresponding season, it completes the binary classification discrimination output of sudden drought. The watershed drought short-term prediction system based on cumulative soil moisture deficit is used to perform the steps in the watershed drought short-term prediction method based on cumulative soil moisture deficit as described in any one of claims 1-9.

Images