Machine learning based drought monitoring system

Abstract

The application discloses a sudden drought monitoring system based on machine learning, which comprises a meteorological data acquisition module, an environment feature construction module, a sudden drought time series feature coding module, a sudden drought grade semantic matching module, a sudden drought boundary smoothing discrimination module, an abnormal environment data tail suppression module, a parameter training module and a real-time sudden drought monitoring output module. The application belongs to the field of data processing and specifically refers to a sudden drought monitoring system based on machine learning. The scheme improves the anti-data deviation capability through error variance driven multi-source soil moisture fusion. A nonlinear boundary smoothing function is constructed to process matching and realize gradual change of grade boundaries. Then, the sudden drought monitoring deviation is reduced. The influence of abnormal environment features is suppressed through a tail adjustment function. When abnormal extreme matching occurs, the weight is gradually reduced. A boundary smoothing mechanism is added in the training loss to control the grade boundary change and compress the critical region fluctuation. Then, the sudden drought monitoring effect is improved.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically to a drought monitoring system based on machine learning. Background Technology

[0002] A flash drought monitoring system is an information system that uses multi-source environmental data, such as meteorological observations, remote sensing, and reanalysis, combined with drought index calculations and data analysis models, to identify, assess, and dynamically monitor rapidly occurring drought events within a region over a short period. However, typical flash drought monitoring systems suffer from problems such as insufficient ability to express water change states, erroneous responses to abnormal meteorological data leading to large monitoring biases, misclassification of normal areas as flash droughts, over-prediction of extreme levels, and poor ability to distinguish between normal changes and anomalous shifts. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a machine learning-based drought monitoring system. Addressing the problems of insufficient representation of water change states and erroneous responses to abnormal meteorological data in general drought monitoring systems, leading to significant monitoring bias, this solution improves resistance to data bias by using multi-source soil moisture fusion driven by error variance; it constructs a nonlinear boundary smoothing function to process matching, achieving gradual changes in grade boundaries, thereby reducing drought monitoring bias; and it addresses the problems of normal areas being misclassified as drought, extreme grades being over-predicted, and poor ability to distinguish between normal changes and abnormal shifts in general drought monitoring systems. This solution suppresses the influence of abnormal environmental features through a tail adjustment function: when abnormal extreme matching occurs, the weight is gradually reduced; and a boundary smoothing mechanism is added to the training loss to control grade boundary changes and compress critical region fluctuations, thereby improving the effectiveness of drought monitoring.

[0004] The technical solution adopted by the present invention is as follows: The machine learning-based drought monitoring system provided by the present invention includes a meteorological data acquisition module, an environmental feature construction module, a drought time series feature encoding module, a drought level semantic matching module, a drought boundary smoothing function module, an abnormal environmental data tail suppression module, a parameter training module, and a real-time drought monitoring output module.

[0005] The meteorological data acquisition module acquires environmental monitoring data and constructs an environmental state vector after preprocessing.

[0006] The environmental feature construction module, based on environmental monitoring data, performs error variance weighted fusion on soil moisture, calculates soil humidity index, precipitation evapotranspiration index and water vapor pressure difference, and incorporates them into the environmental state vector for standardization processing to construct a meteorological monitoring sample set.

[0007] The drought time series feature encoding module is based on the meteorological monitoring sample set and uses a multi-layer LSTM network for encoding to extract drought time series features.

[0008] The drought level semantic matching module establishes different level labels based on historical drought samples, generates drought level semantic features through category embedding, and calculates the matching score between drought time series features and each drought level using feature interaction mapping.

[0009] The sudden drought boundary smoothing function module constructs a boundary smoothing function based on the grade matching score, and performs nonlinear mapping on continuously changing environmental states.

[0010] The abnormal environment data tail suppression module constructs a tail adjustment function to dynamically suppress abnormal matching results and calculates a comprehensive score for each drought level in combination with the boundary smoothing results.

[0011] The parameter training module constructs a training dataset based on historical drought samples and constructs an objective function for parameter training.

[0012] The real-time drought monitoring output module performs drought monitoring on real-time environmental monitoring data after parameter training is completed.

[0013] Furthermore, the meteorological data acquisition module acquires environmental monitoring data, including CLDAS land surface reanalysis data, ERA5 global meteorological reanalysis data, and SMAP satellite soil moisture data;

[0014] Different data sources are resampled at a uniform time interval and mapped to a unified monitoring grid using a spatial interpolation method;

[0015] For each monitoring area, an environmental state vector is constructed at time t, including precipitation, 2m air temperature, 2m dew point temperature, soil moisture, vegetation index, actual evapotranspiration, and potential evapotranspiration.

[0016] Furthermore, the environmental feature construction module specifically includes:

[0017] Soil moisture fusion calculation: The contribution of soil moisture data is dynamically adjusted using a weighted fusion method driven by error variance; the error variance is calculated and expressed as: ; and These are the c-th and d-th soil moisture data sources, respectively; It is variance; The error variance of the c-th data source; the fusion weights are calculated based on the error variance. ; It is the weight of the c-th data source fusion; The variance of the d-th data source is used to obtain the final soil moisture content. ; It is the merged soil moisture; the merged soil moisture replaces the soil moisture dimension in the environmental state vector; It is a smoothing term;

[0018] Soil moisture index calculation: Based on the historical soil moisture statistics, the current soil moisture is standardized and used as the soil moisture index;

[0019] Calculate the water deficit and standardize it as the precipitation evapotranspiration index;

[0020] Vapor pressure difference calculation: Based on the 2m air temperature and 2m dew point temperature, calculate the saturated vapor pressure and actual vapor pressure, denoted as follows: and Final water vapor pressure difference ;

[0021] Normalization: Z-score normalization is used to normalize each dimension of the vector to construct a meteorological monitoring sample set.

[0022] Furthermore, the drought time-series feature encoding module uses a recurrent neural network to extract the rapid evolution pattern of drought for the meteorological monitoring sample set; it constructs an input sequence by building a time window and combining the features of the meteorological monitoring sample set for L consecutive days. Input a temporal network; the network structure uses a three-layer LSTM, and the output is a sudden temporal feature.

[0023] Furthermore, the drought level semantic matching module establishes a mapping relationship between the environmental evolution state and the drought level; specifically including:

[0024] Based on the comprehensive drought index in historical drought samples and actual monitoring results, drought level labels are established, and drought status is divided into M levels.

[0025] The discrete drought levels are transformed into a level representation in a continuous feature space using a category embedding layer;

[0026] The time-series characteristics of sudden drought are interactively mapped with the characteristics of sudden drought level to calculate the degree of matching between the current environmental evolution process and different sudden drought levels; and then the process is standardized.

[0027] Furthermore, the drought boundary smoothing function module constructs a boundary smoothing function based on the degree of matching between the drought time series characteristics and the drought level status, identifying the gradual process of mild drought developing into severe drought. The boundary smoothing function is as follows: ; It is a boundary smoothing function; It is the boundary tolerance parameter; It is a parameter sensitive to boundary changes; It is a smoothing adjustment parameter; It is the standardized grade matching score.

[0028] Furthermore, the abnormal environmental data tail suppression module designs a tail adjustment function for the abnormal environmental feature shift caused by abnormal observations and extreme weather conditions, reducing the impact of abnormal matching results on the final drought level judgment, and obtaining a drought score.

[0029] Tail adjustment function: ; It is a tail adjustment function; It is the abnormal influence of the initial threshold; It is the decay rate;

[0030] Sudden drought score ; It is the final score corresponding to the i-th sample and the j-th drought level; the drought level corresponding to the highest score is selected as the predicted level.

[0031] Furthermore, the parameter training module uses an objective function that incorporates a boundary smoothing mechanism for parameter training;

[0032] Training objective function: ; It is a loss adjustment parameter; is the actual drought level label; n is the total number of samples.

[0033] Furthermore, after the parameters are trained, the real-time drought monitoring output module monitors the real-time environmental data for drought, calculates scores for each level, selects the level corresponding to the highest score for reporting, and thus realizes drought monitoring.

[0034] The beneficial effects achieved by the present invention using the above solution are as follows:

[0035] (1) In view of the problems that general drought monitoring systems have insufficient ability to express the state of water change and are prone to erroneous responses to abnormal meteorological data, which leads to large monitoring deviations, this scheme improves the ability to resist data bias by using multi-source soil moisture fusion driven by error variance; constructs a nonlinear boundary smoothing function to process the matching and realizes the gradual change of the level boundary; thereby reducing the monitoring deviation of drought.

[0036] (2) In response to the problems of normal areas being misjudged as sudden drought, extreme levels being over-predicted, and poor ability to distinguish between normal changes and abnormal shifts in general drought monitoring systems, this scheme suppresses the influence of abnormal environmental features through tail adjustment function: when abnormal extreme matching occurs, the weight is gradually reduced; a boundary smoothing mechanism is added to the training loss: control the change of level boundary and compress the fluctuation of critical area; thereby improving the drought monitoring effect. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the machine learning-based drought monitoring system provided by the present invention.

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0040] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0041] Example 1, see Figure 1 The machine learning-based drought monitoring system provided by this invention includes a meteorological data acquisition module, an environmental feature construction module, a drought time series feature encoding module, a drought level semantic matching module, a drought boundary smoothing function module, an abnormal environmental data tail suppression module, a parameter training module, and a real-time drought monitoring output module.

[0042] The meteorological data acquisition module acquires environmental monitoring data and constructs an environmental state vector after preprocessing.

[0043] The environmental feature construction module, based on environmental monitoring data, performs error variance weighted fusion on soil moisture, calculates soil humidity index, precipitation evapotranspiration index and water vapor pressure difference, and incorporates them into the environmental state vector for standardization processing to construct a meteorological monitoring sample set.

[0044] The drought time series feature encoding module is based on the meteorological monitoring sample set and uses a multi-layer LSTM network for encoding to extract drought time series features.

[0045] The drought level semantic matching module establishes different level labels based on historical drought samples, generates drought level semantic features through category embedding, and calculates the matching score between drought time series features and each drought level using feature interaction mapping.

[0046] The sudden drought boundary smoothing function module constructs a boundary smoothing function based on the grade matching score, and performs nonlinear mapping on continuously changing environmental states.

[0047] The abnormal environment data tail suppression module constructs a tail adjustment function to dynamically suppress abnormal matching results and calculates a comprehensive score for each drought level in combination with the boundary smoothing results.

[0048] The parameter training module constructs a training dataset based on historical drought samples and constructs an objective function for parameter training.

[0049] The real-time drought monitoring output module performs drought monitoring on real-time environmental monitoring data after parameter training is completed.

[0050] Example 2, see Figure 1 This embodiment is based on the above embodiment. The meteorological data acquisition module acquires environmental monitoring data, including CLDAS land surface reanalysis data, ERA5 global meteorological reanalysis data, and SMAP satellite soil moisture data. Among them, CLDAS data is used to provide regional high spatiotemporal resolution land surface water and heat status; ERA5 data is used to provide large-scale meteorological driving factors; and SMAP data is used to provide satellite inversion of soil moisture status.

[0051] Different data sources are resampled at a uniform time interval so that each time point contains the corresponding environmental variables; and spatial interpolation is used to map data with different spatial resolutions to a unified monitoring grid.

[0052] For each monitoring area, an environmental state vector is constructed at time t, including precipitation (mm), 2m air temperature (°C), 2m dew point temperature (°C), soil moisture (m³ / m³), vegetation index, actual evapotranspiration (mm), and potential evapotranspiration (mm).

[0053] Example 3, see Figure 1 This embodiment is based on the above embodiment. The environmental feature construction module calculates soil moisture anomaly features, precipitation evapotranspiration anomaly features, and atmospheric dryness based on the environmental state vector, incorporates them into the environmental state vector, and standardizes the data dimensions; specifically including:

[0054] Soil moisture fusion calculation: Due to varying degrees of uncertainty in CLDAS, ERA5, and SMAP soil moisture data, an error variance-driven weighting fusion method is used to dynamically adjust contributions. The error variance is calculated to characterize the degree of inconsistency between different data sources, and is expressed as: ; and These are the c-th and d-th soil moisture data sources, corresponding to three data sources: CLDAS, ERA5, and SMAP. It is variance; The error variance of the c-th data source; the fusion weights are calculated based on the error variance. ; It is the weight of the c-th data source fusion; The variance of the d-th data source is used to obtain the final soil moisture content. ; It is the merged soil moisture; the merged soil moisture replaces the soil moisture dimension in the environmental state vector; It is a smooth term, a very small positive number;

[0055] Soil moisture index calculation: Based on historical soil moisture statistics for the same period, the current soil moisture is standardized and expressed as follows: ; It is the soil moisture index; It is to integrate soil moisture at the current moment; It is the average soil moisture value for the same period in history; It represents the standard deviation of soil moisture for the same historical period; indicating the degree to which current soil moisture deviates from historical climatic background.

[0056] The moisture deficit is calculated and standardized, and expressed as follows: ; ; It is the precipitation evapotranspiration index; It is the water surplus or deficit. It is the historical average of water level gains and losses. This is the historical standard deviation of moisture gain / loss, in mm; This is the rainfall amount, in mm; It is the potential evapotranspiration;

[0057] Water vapor pressure difference calculation, characterizing the degree of atmospheric dryness: saturated water vapor pressure and actual water vapor pressure are calculated based on 2m air temperature and 2m dew point temperature, and denoted as follows: and (Unit: kPa); Final water vapor pressure difference ;

[0058] Normalization is performed by using Z-score normalization to normalize each dimension of the vectors, since different features have different dimensions, to construct a meteorological monitoring sample set.

[0059] Example 4, see Figure 1 This embodiment, based on the above embodiment, uses a recurrent neural network to extract the rapid evolution pattern of sudden drought for a meteorological monitoring sample set; specifically, it includes:

[0060] Constructing the input sequence: Since the sudden drought is not a single-point anomaly but a continuous meteorological change process, a time window is constructed, incorporating the characteristics of a continuous L-day meteorological monitoring sample set. Input a time series network; L is the time window length, ranging from 30 to 60;

[0061] The network architecture uses a three-layer LSTM:

[0062] First layer ;

[0063] Second floor ;

[0064] Third layer ;

[0065] The network structure parameters are shown in Table 1:

[0066]

[0067] It is a characteristic of the sudden drought sequence; and These are the short-term environmental change characteristics of the first LSTM output and the intermediate time series characteristics of the second LSTM output, respectively. , and It is a three-layer LSTM.

[0068] Example 5, see Figure 1 This embodiment, based on the above embodiment, establishes a mapping relationship between the environmental evolution state and the drought level using a sudden drought level semantic matching module; specifically, it includes:

[0069] Based on the comprehensive drought index in historical drought samples and actual monitoring results, drought level labels were established, and drought status was divided into M levels, including normal status, mild drought, moderate drought, severe drought and extreme drought.

[0070] The discrete drought levels are transformed into a level representation in a continuous feature space using a category embedding layer. ; It is the level feature vector corresponding to the j-th level of sudden drought; It is a learnable embedding function;

[0071] The temporal characteristics of sudden drought are interactively mapped with the characteristics of sudden drought severity to calculate the degree of matching between the current environmental evolution process and different sudden drought severity levels: ; is the degree of matching between the i-th sample and the j-th drought level; W is the state mapping matrix, with matrix elements ranging from -1 to 1; b is the bias, ranging from -0.5 to 0.5; T is the transpose operation; for Standardization process: ; It is the standardized grade matching score; It is the mean of all matching scores in the historical drought sample. That is the corresponding standard deviation.

[0072] Example 6, see Figure 1 This embodiment is based on the above embodiment. The sudden drought boundary smoothing function module addresses the continuous transitional states during the change of sudden drought levels. Instead of using a fixed drought threshold to discretely divide the environmental state, it constructs a boundary smoothing function based on the degree of matching between the sudden drought temporal characteristics and the level state, identifying the gradual process of development from mild drying to severe sudden drought. The boundary smoothing function is as follows: ; It is a boundary smoothing function; This is the boundary tolerance parameter, which takes a value of 0 to 1; It is a parameter sensitive to boundary changes, with a value ranging from 5 to 20; This is a smoothing adjustment parameter, with a value ranging from 0.05 to 0.5;

[0073] The Sigmoid function term is used to achieve gradual changes in the boundary region of the drought level, and the tanh function term is used to further compress the small fluctuations in the critical region, so that a continuous response relationship is formed between adjacent drought levels.

[0074] By performing the above operations, this solution addresses the problems of insufficient ability to express water change status and the tendency for abnormal meteorological data to generate erroneous responses, leading to large monitoring biases, in general drought monitoring systems. It improves the resistance to data bias by using multi-source soil moisture fusion driven by error variance; it also constructs a nonlinear boundary smoothing function to process the matching, achieving gradual changes in the grade boundaries; thereby reducing drought monitoring bias.

[0075] Example 7, see Figure 1 This embodiment is based on the above embodiment. The abnormal environment data tail suppression module designs a tail adjustment function for the abnormal observation values ​​generated during the fusion of multi-source remote sensing data and the abnormal environmental feature shift caused by extreme weather conditions. It uses a nonlinear decay function to reduce the impact of abnormal matching results on the final drought level judgment and obtains a drought score.

[0076] Tail adjustment function: ; It is a tail adjustment function; This is the threshold for the initial impact of abnormalities, ranging from 1.8 to 2.5. This is the decay rate, with a value ranging from 0.3 to 1.0;

[0077] Sudden drought score ; It is the final score corresponding to the i-th sample and the j-th drought level; the drought level corresponding to the highest score is selected as the predicted level.

[0078] Example 8, see Figure 1 This embodiment is based on the above embodiment, and the parameter training module uses an objective function with a fusion boundary smoothing mechanism for parameter training;

[0079] Training objective function: ; This is the loss adjustment parameter, with a value ranging from 0.01 to 0.5; This is the actual drought level label. If the i-th sample is at the j-th drought level, then... ,otherwise n is the total number of samples;

[0080] Parameter training process:

[0081] Data partitioning: historical data is divided into 70% training set, 20% validation set, and 10% test set;

[0082] Forward propagation ;

[0083] Parameter updates are optimized using Adam.

[0084] Example 9, see Figure 1 This embodiment is based on the above embodiment. After the parameters are trained, the real-time drought monitoring output module monitors the real-time environmental data for drought, calculates the scores of each level, selects the level corresponding to the highest score for reporting, and thus realizes drought monitoring.

[0085] By performing the above operations, this solution addresses the problems of normal areas being misclassified as sudden droughts, extreme levels being over-predicted, and the poor ability to distinguish between normal changes and abnormal shifts in general drought monitoring systems. It suppresses the influence of abnormal environmental features through a tail adjustment function: when abnormal extreme matching occurs, the weights are gradually reduced; a boundary smoothing mechanism is added to the training loss to control level boundary changes and compress critical region fluctuations; thereby improving the effectiveness of drought monitoring.

[0086] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0087] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A machine learning-based drought monitoring system, characterized in that: The system includes a meteorological data acquisition module, an environmental feature construction module, a drought time series feature encoding module, a drought level semantic matching module, a drought boundary smoothing function module, an abnormal environmental data tail suppression module, a parameter training module, and a real-time drought monitoring output module. The meteorological data acquisition module acquires environmental monitoring data and constructs an environmental state vector after preprocessing. The environmental feature construction module, based on environmental monitoring data, performs error variance weighted fusion on soil moisture, calculates soil humidity index, precipitation evapotranspiration index and water vapor pressure difference, and incorporates them into the environmental state vector for standardization processing to construct a meteorological monitoring sample set. The drought time series feature encoding module is based on the meteorological monitoring sample set and uses a multi-layer LSTM network for encoding to extract drought time series features. The drought level semantic matching module establishes different level labels based on historical drought samples, generates drought level semantic features through category embedding, and calculates the matching score between drought time series features and each drought level using feature interaction mapping. The sudden drought boundary smoothing function module constructs a boundary smoothing function based on the grade matching score, and performs nonlinear mapping on the continuously changing environmental state. The abnormal environment data tail suppression module constructs a tail adjustment function to dynamically suppress abnormal matching results and calculates a comprehensive score for each drought level in combination with the boundary smoothing results. The parameter training module constructs a training dataset based on historical drought samples and constructs an objective function for parameter training. The real-time drought monitoring output module performs drought monitoring on real-time environmental monitoring data after parameter training is completed. The environmental feature construction module includes: soil moisture calculation: dynamically adjusting the contribution of soil moisture data using an error variance-driven weight fusion method; calculating the error variance, expressed as: ; and These are the c-th and d-th soil moisture data sources, respectively; It is variance; The error variance of the c-th data source; the fusion weights are calculated based on the error variance. ; It is the weight of the c-th data source fusion; The variance of the d-th data source is used to obtain the final soil moisture content. ; It is the merged soil moisture; the merged soil moisture replaces the soil moisture dimension in the environmental state vector; It is a smoothing term; The sudden drought boundary smoothing function module constructs a boundary smoothing function based on the degree of matching between the sudden drought temporal characteristics and the severity status, identifying the gradual process of mild drought developing into severe sudden drought. The boundary smoothing function is as follows: ; It is a boundary smoothing function; It is the boundary tolerance parameter; It is a parameter sensitive to boundary changes; It is a smoothing adjustment parameter; It is the standardized grade matching score; The tail adjustment function is expressed as follows: ; It is a tail adjustment function; The abnormal impact starts at the threshold. It is the decay rate.

2. The machine learning-based drought monitoring system according to claim 1, characterized in that: The environmental feature construction module also includes: Soil moisture index calculation: Based on the historical soil moisture statistics, the current soil moisture is standardized and used as the soil moisture index; Calculate the water deficit and standardize it as the precipitation evapotranspiration index; Vapor pressure difference calculation: Based on the 2m air temperature and 2m dew point temperature, calculate the saturated vapor pressure and actual vapor pressure, denoted as follows: and Final water vapor pressure difference ; Normalization: Z-score normalization is used to normalize each dimension of the vector to construct a meteorological monitoring sample set.

3. The machine learning-based drought monitoring system according to claim 2, characterized in that: The abnormal environment data tail suppression module designs a tail adjustment function for the abnormal environmental feature shift caused by abnormal observations and extreme weather conditions, reduces the impact of abnormal matching results on the final drought level judgment, and obtains a drought score. Sudden drought score ; It is the final score corresponding to the i-th sample and the j-th drought level; the drought level corresponding to the highest score is selected as the predicted level.

4. The machine learning-based drought monitoring system according to claim 3, characterized in that: The parameter training module uses an objective function with a fusion boundary smoothing mechanism for parameter training. Training objective function: ; It is a loss adjustment parameter; is the actual drought level label; n is the total number of samples; M is the total number of drought levels.

5. The machine learning-based drought monitoring system according to claim 1, characterized in that: The meteorological data acquisition module acquires environmental monitoring data, including CLDAS land surface reanalysis data, ERA5 global meteorological reanalysis data, and SMAP satellite soil moisture data. Different data sources are resampled at a uniform time interval and mapped to a unified monitoring grid using a spatial interpolation method; For each monitoring area, an environmental state vector is constructed at time t, including precipitation, 2m air temperature, 2m dew point temperature, soil moisture, vegetation index, actual evapotranspiration, and potential evapotranspiration.

6. The machine learning-based drought monitoring system according to claim 1, characterized in that: The drought time-series feature encoding module uses a recurrent neural network to extract the rapid evolution pattern of drought for a meteorological monitoring sample set; it constructs an input sequence by building a time window and combining the features of a continuous L-day meteorological monitoring sample set. Input a temporal network; the network structure uses a three-layer LSTM, and the output is a sudden temporal feature.

7. The machine learning-based drought monitoring system according to claim 1, characterized in that: The drought level semantic matching module establishes a mapping relationship between environmental evolution state and drought level; specifically including: Based on the comprehensive drought index in historical drought samples and actual monitoring results, drought level labels are established, and drought status is divided into M levels. The discrete drought levels are transformed into a level representation in a continuous feature space using a category embedding layer; The time-series characteristics of sudden drought are interactively mapped with the characteristics of sudden drought level to calculate the degree of matching between the current environmental evolution process and different sudden drought levels; and then the process is standardized.

8. The machine learning-based drought monitoring system according to claim 1, characterized in that: After the parameters are trained, the real-time drought monitoring output module monitors the drought in real-time environmental data, calculates scores for each level, selects the level corresponding to the highest score for reporting, and thus realizes drought monitoring.

Images