High-temperature drought composite disaster monitoring and early warning method and system

Abstract

The invention provides a high-temperature and drought composite disaster monitoring and early warning method and system, and relates to the technical field of disaster monitoring and early warning, and the method comprises the steps: 2, extracting a data subset from a fusion data set, constructing a feature analysis region according to a preset reference relation, selecting a core analysis unit in the feature analysis region as a focus of an ellipse, and carrying out the calculation of the data subset; selecting an external reference unit as another focus of the ellipse in an adjacent spatial range, and calculating the spatial form and dynamic change of an eccentricity quantization region of the ellipse; according to the incidence relation of the core analysis unit and the external reference unit along with the time change, an analysis path is generated, an environment parameter correction coefficient is calculated to calibrate related elements in the fusion data set, and a calibrated fusion data set is obtained. According to the invention, probability prediction and dynamic early warning of high-temperature drought composite disasters are realized, and the reliability and real-time performance of disaster early warning are improved.

Description

Technical Field

[0001] This invention relates to the field of disaster monitoring and early warning technology, and in particular to methods and systems for monitoring and early warning of combined high-temperature and drought disasters. Background Technology

[0002] In the field of monitoring and early warning of combined high temperature and drought disasters, most existing technical methods rely on single or limited data sources to independently monitor high temperature or drought. For example, based on temperature and precipitation observation data from meteorological stations, high temperature index and drought index are calculated separately, and early warning is issued independently accordingly. However, high temperature and drought often interact and evolve together in the actual process of occurrence, forming a combined disaster. The existing separate monitoring and early warning methods have some limitations in reflecting this interaction and spatial co-evolution characteristics.

[0003] For example, in monitoring practices targeting a specific region, such as a hilly agricultural area, existing technologies may primarily rely on temperature and precipitation data from a limited number of stations within the region to separately determine high-temperature events and meteorological drought events. This method struggles to fully capture the spatial heterogeneity of high-temperature and drought distribution caused by uneven underlying surfaces, such as differences in vegetation cover and soil moisture, as well as the spatial coupling patterns and dynamic changes between the two. For instance, in the early stages of drought, the spread of high temperatures and the development of drought are not uniformly synchronized due to the influence of solar radiation, wind speed, and soil conditions at different locations in the region. Their spatial correlation may exhibit asymmetrical or directional evolution. Existing static index analyses based on fixed stations or regular grids largely lack comprehensive quantitative methods for understanding the dynamics of this spatial pattern and its correlation with temporal evolution, which may affect the identification and timely warning of early signals in the occurrence and development of complex disasters. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for monitoring and early warning of combined high temperature and drought disasters, so as to realize the probability prediction and dynamic early warning of combined high temperature and drought disasters and improve the reliability and real-time performance of disaster early warning.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for monitoring and early warning of combined high-temperature and drought disasters, the method comprising: Step 1: Collect multi-source monitoring data in real time, process the multi-source monitoring data, and generate a fused dataset; Step 2: Extract a subset of data from the fused dataset, construct a feature analysis region based on a preset reference relationship, select a core analysis unit as one focus of an ellipse within the feature analysis region, and select an external reference unit in the adjacent spatial range as the other focus of the ellipse. Calculate the eccentricity of the ellipse to quantify the spatial morphology and dynamic changes of the region. Based on the correlation between the core analysis unit and the external reference unit over time, generate an analysis path, and calculate environmental parameter correction coefficients to calibrate the relevant elements in the fused dataset, obtaining the calibrated fused dataset. Step 3: Based on the calibrated fusion dataset, construct a time-series dynamic analysis structure, extract time-series dynamic feature vectors of key indicators of high temperature and drought; analyze the change trajectory of key indicators, calculate the torsion angle of evolution over time, quantify trend turning points and fluctuation characteristics, and fuse the time-series dynamic feature vectors and torsion angle features to form a comprehensive feature matrix. Step 4: Input the comprehensive feature matrix into the pre-trained machine learning classification model to calculate the probability of occurrence of the combined high temperature and drought disaster and determine the warning level; Step 5: Issue warning information according to the warning level, collect actual disaster information from external feedback as new training data, and periodically retrain the machine learning classification model to achieve adaptive updates.

[0006] Secondly, the monitoring and early warning system for combined high-temperature and drought disasters includes: The fusion module is used to collect multi-source monitoring data in real time, process the multi-source monitoring data, and generate a fused dataset. The calibration module is used to extract a subset of data from the fused dataset, construct a feature analysis region based on a preset reference relationship, select a core analysis unit as one focus of an ellipse within the feature analysis region, and select an external reference unit in the adjacent spatial range as the other focus of the ellipse. It calculates the eccentricity of the ellipse to quantify the spatial morphology and dynamic changes of the region. Based on the correlation between the core analysis unit and the external reference unit over time, it generates an analysis path and calculates environmental parameter correction coefficients to calibrate relevant elements in the fused dataset, thus obtaining the calibrated fused dataset. The feature module is used to construct a time-series dynamic analysis structure based on the calibrated fusion dataset, extract time-series dynamic feature vectors of key indicators of high temperature and drought, analyze the change trajectory of key indicators, calculate the torsional angle of evolution over time, quantify trend turning points and fluctuation characteristics, and fuse the time-series dynamic feature vectors and torsional angle features to form a comprehensive feature matrix. The early warning module is used to input the comprehensive feature matrix into a pre-trained machine learning classification model, calculate the probability of occurrence of the combined high temperature and drought disaster, and determine the early warning level. The update module is used to issue early warning information according to the warning level, collect actual disaster information from external feedback as new training data, and periodically retrain the machine learning classification model to achieve adaptive updates.

[0007] The above-described solution of the present invention has at least the following beneficial effects: By real-time acquisition and fusion processing of multi-source monitoring data, the limitations of a single data source are overcome. Elliptical analysis is constructed by combining core and external reference units to quantify the dynamic changes in spatial morphology using eccentricity and to calibrate the data using environmental parameter correction coefficients, thereby improving the accuracy and reliability of the dataset. Simultaneously, by constructing a time-series dynamic analysis structure, the time-series feature vectors of key indicators of high temperature and drought are integrated with the torsion angle of the change trajectory to capture the trend turning point and fluctuation characteristics of disaster evolution. The resulting comprehensive feature matrix provides comprehensive support for disaster identification. Furthermore, by using a pre-trained machine learning classification model, the probability of disaster occurrence is calculated and graded early warning is achieved, thereby improving the intelligence of early warning and enabling adaptive updates. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the monitoring and early warning method for combined high-temperature and drought disasters provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of a high-temperature and drought combined disaster monitoring and early warning system provided in an embodiment of the present invention. Detailed Implementation

[0010] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0011] like Figure 1 As shown, embodiments of the present invention propose a method for monitoring and early warning of combined high-temperature and drought disasters, the method comprising the following steps: Step 1: Collect multi-source monitoring data in real time, process the multi-source monitoring data, and generate a fused dataset; Step 2: Extract a subset of data from the fused dataset, construct a feature analysis region based on a preset reference relationship, select a core analysis unit as one focus of an ellipse within the feature analysis region, and select an external reference unit in the adjacent spatial range as the other focus of the ellipse. Calculate the eccentricity of the ellipse to quantify the spatial morphology and dynamic changes of the region. Based on the correlation between the core analysis unit and the external reference unit over time, generate an analysis path, and calculate environmental parameter correction coefficients to calibrate the relevant elements in the fused dataset, obtaining the calibrated fused dataset. Step 3: Based on the calibrated fusion dataset, construct a time-series dynamic analysis structure, extract time-series dynamic feature vectors of key indicators of high temperature and drought; analyze the change trajectory of key indicators, calculate the torsion angle of evolution over time, quantify trend turning points and fluctuation characteristics, and fuse the time-series dynamic feature vectors and torsion angle features to form a comprehensive feature matrix. Step 4: Input the comprehensive feature matrix into the pre-trained machine learning classification model to calculate the probability of occurrence of the combined high temperature and drought disaster and determine the warning level; Step 5: Issue warning information according to the warning level, collect actual disaster information from external feedback as new training data, and periodically retrain the machine learning classification model to achieve adaptive updates.

[0012] In this embodiment of the invention, the limitations of a single data source are overcome by real-time acquisition and fusion processing of multi-source monitoring data. Elliptical analysis is constructed by combining core and external reference units to quantify the dynamic changes in spatial morphology using eccentricity and to calibrate the data using environmental parameter correction coefficients, thereby improving the accuracy and reliability of the dataset. Simultaneously, by constructing a time-series dynamic analysis structure, the time-series feature vectors of key indicators of high temperature and drought are integrated with the torsion angle of the change trajectory to capture the trend turning point and fluctuation characteristics of disaster evolution. The resulting comprehensive feature matrix provides comprehensive support for disaster identification. Furthermore, a pre-trained machine learning classification model is used to calculate the probability of disaster occurrence and to implement graded early warning, thereby improving the intelligence of early warning and enabling adaptive updates.

[0013] In a preferred embodiment of the present invention, step 1 above, which involves real-time acquisition of multi-source monitoring data and processing of the multi-source monitoring data to generate a fused dataset, may include: In this embodiment of the invention, real-time acquisition of multi-source monitoring data is first carried out. Combined with the monitoring needs of combined high-temperature and drought disasters, multiple types and dimensions of monitoring data are collected simultaneously to ensure data coverage of core disaster-related elements. Specific data collection includes: ground meteorological station data, which acquires real-time meteorological element data such as temperature, precipitation, relative humidity, and wind speed in the area where the station is located, automatically collecting and uploading data at fixed time intervals using the station's built-in sensors; remote sensing image data, which relies on a satellite remote sensing platform to receive real-time image data covering the monitoring area and extract spatial element information such as surface temperature, vegetation cover, and surface albedo; and soil moisture data, which is collected in real-time using soil moisture sensors deployed at different depths in the monitoring area to collect soil volume content data. Data such as water volume and soil temperature were collected. In addition, runoff data from hydrological stations and vegetation growth data from ecological monitoring points within the area were supplemented to form a multi-dimensional data source matrix. During the collection process, the collection time, spatial coordinates, and data source information of each data point were recorded simultaneously to ensure that each data point has traceable spatiotemporal attributes. The collected multi-source monitoring data were processed in steps. First, data preprocessing was carried out to remove invalid information and standardize data. The first step involved data denoising and outlier removal. Based on the normal fluctuation range of various data types, abnormal data caused by equipment failure or signal interference were screened out. Minor abnormal data were supplemented using the average of adjacent time periods, while severely abnormal data were directly removed. Simultaneously, data that was retained was also removed. The first step is data recording; the second step is data deduplication, comparing data from different data sources with completely consistent spatiotemporal coordinates, retaining entries with higher data accuracy and collection times closer to real-time, and deleting duplicate and redundant data; the third step is data format and unit standardization, unifying the time format of different data sources to a standard timestamp format, and standardizing the units of elements such as temperature, precipitation, and soil moisture to industry-standard units, ensuring that cross-data source data is comparable and integrateable; after preprocessing, spatiotemporal alignment processing is performed to solve the problem of inconsistent spatiotemporal resolution of multi-source data. In the time dimension, the data from other data sources is interpolated or sampled based on the data from the data source with the highest collection frequency, so that all data correspond to the same time node, forming a time sequence with equal time intervals. In terms of data structure, spatially, based on the geographic coordinate system of the monitoring area, areal data such as remote sensing images and point data such as station observations are uniformly projected onto the same spatial coordinate system to fill the spatial gaps in point data, ensuring continuous coverage of data in spatial distribution and guaranteeing that there is corresponding monitoring data for each time period and spatial location. Finally, data integration and fusion are performed, integrating various types of monitoring data after preprocessing and spatiotemporal alignment into a unified dataset according to preset rules. During the integration process, the original attributes and source identifiers of each data are preserved. At the same time, according to the monitoring needs of high temperature and drought, cross-validation of similar element data is performed to ensure data consistency. By associating the spatiotemporal attributes of data in various dimensions, a fused dataset with three-dimensional association of time, space, and elements is formed.

[0014] In a preferred embodiment of the present invention, step 2 above involves extracting a data subset from the fused dataset, constructing a feature analysis region based on a preset reference relationship, selecting a core analysis unit as one focus of an ellipse within the feature analysis region, and selecting an external reference unit as the other focus of the ellipse in the adjacent spatial range. The spatial morphology and dynamic changes of the eccentricity quantification region of the ellipse are then calculated. Based on the correlation between the core analysis unit and the external reference unit over time, an analysis path is generated, and environmental parameter correction coefficients are calculated to calibrate relevant elements in the fused dataset, resulting in a calibrated fused dataset, which may include: In this embodiment of the invention, step 220 involves selecting target areas affected by high temperatures or drought based on the geospatial distribution information in the fused dataset, and defining these target areas as feature analysis areas. Specifically, this includes: firstly, extracting complete geospatial distribution information from the fused dataset, including latitude and longitude coordinates, county-level and above administrative boundaries, topographical zones (mountains, hills, plains), and watershed extent, corresponding to all monitoring data; and simultaneously associating each spatial location with core monitoring elements such as daily average temperature, daily maximum temperature, 24-hour precipitation, cumulative precipitation over the past 7 days, soil moisture in the 0-10cm and 10-20cm layers, and vegetation cover at the corresponding time point. Data ensures that the spatiotemporal information and element data of each spatial location correspond one-to-one and are traceable; combining the formation patterns of high-temperature and drought disasters and the response characteristics of different underlying surfaces, stratified screening criteria are set to adapt to different terrain scenarios. Regarding temperature, in plain areas, if the daily average temperature exceeds the average of the same period in the past 30 years by 3°C or more for three consecutive days, or if the daily maximum temperature exceeds 37°C and lasts for more than 4 hours, it is considered a sign of high temperature; in hilly and mountainous areas, due to altitude differences, the temperature threshold is lowered by 1°C, meaning that if the daily average temperature exceeds the historical average for the same period by 2°C or more for three consecutive days, or if the daily maximum temperature exceeds 36°C and lasts for more than 4 hours, it is considered a sign of high temperature; regarding precipitation... Regardless of topography, any cumulative rainfall in the past 7 days that is 50% or more below the average for the same period in the past 30 years, or any cumulative rainfall in the past 15 days that is 60% or more below the historical average for the same period, is considered a sign of drought. Regarding soil moisture, any soil moisture in the 0-10cm layer that is below 60% for 5 consecutive days or more, or any soil moisture in the 10-20cm layer that is below 55% for 7 consecutive days or more, is considered a sign of drought. Deep soil moisture has a higher weight than surface soil moisture because it better reflects long-term drought conditions. For all spatial locations, each area is compared against the above stratification criteria to select those that meet any one of the criteria. Then, isolated, scattered areas are eliminated. The final determination criteria are... Each area must be less than 1 square kilometer and have a straight-line distance of more than 5 kilometers from surrounding eligible areas. To avoid interference caused by single-point equipment failures or localized anomalies, the remaining eligible areas will be integrated according to administrative boundaries, topographical continuity, and watershed integrity. For example, hilly areas within the same county that are connected in topography will be merged, and arid areas that span multiple townships but belong to the same watershed will be integrated to form target areas with clear boundaries and continuous range. Finally, this target area will be identified as the feature analysis area for feature analysis, ensuring that the analysis scope is accurately focused on the core areas that are potentially affected by disasters or have already been affected, while also adapting to the disaster distribution characteristics of different terrains and watersheds.

[0015] Step 221: Within the feature analysis area, based on the outliers of meteorological and environmental elements in the fused dataset, identify and determine the spatial units with prominent outliers as core analysis units. Specifically, this includes: first, using geographic information tools, dividing the feature analysis area into several uniformly sized, clearly defined, and non-overlapping square spatial units at a fixed spatial resolution of 1km × 1km. Each spatial unit corresponds to a unique spatial code. Simultaneously, it is associated with complete element data for the current day and the past 7 days within that unit, covering daily average temperature, daily maximum temperature, daily minimum temperature, 24-hour precipitation, 0-10cm and 10-20cm soil moisture, vegetation cover, etc., ensuring the completeness and continuity of element data for each unit. Continuing; For each spatial unit, the degree of anomaly of each core element is calculated one by one, with clear calculation logic for each element. The percentage of temperature anomaly is calculated by subtracting the overall average daily temperature of all spatial units within the feature analysis area over the past 7 days from the average daily temperature of that unit over the past 7 days, yielding the temperature deviation. If the deviation is negative, it indicates that the temperature of that unit is below the regional average and is not included in the high-temperature anomaly consideration; only positive deviations are retained. This positive deviation is then divided by the historical normal fluctuation range of the region's temperature over the same period in the past 30 years, i.e., the difference between the highest and lowest average daily temperatures in the same period historically. This is typically 8-10℃ in plains areas and 6-8℃ in hilly and mountainous areas, yielding the percentage of temperature anomaly. The result is rounded to two decimal places. Soil moisture anomaly... The degree of abnormality is calculated by analyzing the soil moisture of all spatial units within the region over the past 7 days. The overall average of the 0-10cm and 10-20cm moisture values ​​is taken, and the average soil moisture of that unit over the past 7 days is subtracted to obtain the moisture deviation. Only positive deviations are retained; negative values ​​indicate sufficient moisture and no drought anomalies. This positive deviation is then divided by the historical normal fluctuation range of soil moisture in the corresponding region: 20-25% for plains and 18-22% for hilly and mountainous areas, to obtain the percentage of abnormal soil moisture. The percentage of abnormal vegetation cover is calculated using the same logic as soil moisture, with a historical normal fluctuation range of 15-20%. Only the deviation of vegetation cover being lower than the regional average is considered, as reduced vegetation cover is usually accompanied by drought. Impact of drought; combining the core influencing factors of high temperature and drought disasters, fixed weights are set: temperature accounts for 40%, soil moisture accounts for 40%, and vegetation cover accounts for 20%. Among the soil moisture weights, the moisture content of the 10-20cm deep layer accounts for 25%, and the moisture content of the 0-10cm surface layer accounts for 15%. When calculating the comprehensive anomaly value of each spatial unit, the anomaly degree of each element is first multiplied by its corresponding weight to obtain the weighted anomaly value of each element. Then, the weighted anomaly values ​​of temperature, soil moisture, and vegetation cover are added together to obtain the comprehensive anomaly value of the unit, and the result is rounded to two decimal places. The comprehensive anomaly values ​​of all spatial units are sorted from high to low, and the unit with the highest comprehensive anomaly value is selected first.If multiple units have the same overall outlier value, compare the degree of anomaly of the core elements and select the unit where both the percentage of temperature anomaly and the percentage of soil moisture anomaly exceed 30%, and the anomaly of deep soil moisture is more significant. If they still cannot be distinguished, select the unit whose spatial location is near the geometric center of the feature analysis area to ensure that the core analysis unit can represent the area with the most prominent element anomalies while also taking into account the overall representativeness of the area. Finally, the selected unit is determined as the core analysis unit.

[0016] Step 222: Based on the spatial correlation characteristics of the elements in the fused dataset, select an external reference unit within the vicinity of the core analysis unit, centered on the core analysis unit. Specifically, this includes: first, based on the fused dataset, systematically analyze the spatial correlation of elements between the core analysis unit and all surrounding spatial units, focusing on three core elements: temperature, soil moisture, and vegetation cover. The strength of the correlation is judged from two dimensions: consistency of change trends and similarity of values. For the consistency of change trends, compare the element change trends of the core analysis unit and surrounding units over the past 7 days, at each 24-hour node. Temperature trends are divided into three categories: rising, falling, and stable. A continuous rise in the daily average temperature for 3 consecutive days indicates an upward trend, while a continuous fall for 3 consecutive days indicates a downward trend. The temperature trend is downward, and fluctuations within 1°C are considered stable. Similarly, soil moisture trends are considered stable with fluctuations within 5%. Vegetation cover trends are judged according to soil moisture standards; for a trend correlation, two or more of the three elements must show completely consistent trends. For numerical similarity, the mean differences of elements between the core analysis unit and surrounding units over the past 7 days are calculated. The difference in mean temperature should not exceed 2°C, the difference in mean soil moisture should not exceed 10%, and the difference in mean vegetation cover should not exceed 8%. Meeting two or more of these three criteria constitutes a numerical correlation. Units that simultaneously meet both trend and numerical correlation criteria are considered to have a strong spatial correlation with the core analysis unit. Subsequently, a neighboring spatial range is defined, with the center of the core analysis unit as the reference point. Using a point as the center, the radius is set based on the overall size of the feature analysis area and its terrain features. For county-level feature analysis areas, the radius for plains is set to 8-10 km, while for hilly and mountainous areas it is reduced to 5-7 km. For city-level feature analysis areas, the radius for plains is 15-20 km, and for hilly and mountainous areas it is 10-15 km. The boundaries of these areas must not exceed the boundaries of the feature analysis area. If the radius exceeds these boundaries, a sector-shaped area is selected based on the boundaries of the feature analysis area. Within the defined adjacent spatial range, spatial units with comprehensive outliers exceeding 1.5 times the overall mean of the feature analysis area are first eliminated, followed by units with values ​​completely identical to the core analysis unit's elements. From the remaining units, priority is given to selecting those with spatial similarities to the core analysis unit. The unit with the strongest correlation is the one where the trends of the three elements are completely consistent and the difference between the mean values ​​of the three elements is the smallest. If there are multiple units with similar correlation, the unit with a comprehensive outlier at a medium level in the vicinity, and whose element status is closer to the normal level of the same period in history, and whose mean values ​​of temperature, soil moisture, and vegetation cover are all within half of the normal fluctuation range of the mean values ​​of the same period in history, is selected as the external reference unit. After selection, the spatial code of the external reference unit, its distance from the core analysis unit, and the details of the element correlation are recorded to ensure that the reference unit can not only form an effective comparison with the core analysis unit, but also truly reflect the basic state of the elements in the surrounding area, providing a reliable reference benchmark for spatial morphology analysis and data calibration.

[0017] Step 223: Using the core analysis unit and a selected external reference unit as the two foci of the ellipse, calculate the eccentricity based on the geometric relationship between the distance between the two foci and the length of the ellipse's major axis. Specifically, this involves: first, using geographic information tools to locate the latitude and longitude coordinates of the center point of the core analysis unit and the center point of the external reference unit; then, calculating the straight-line distance between the two points using these coordinates, which is the total distance between the two foci of the ellipse, and recording this distance value; next, dividing this total distance by 2 to obtain the semi-focal length of the ellipse, i.e., the distance from each focus to the center of the ellipse. This calculation process involves only simple division, and the result retains the same precision as the total distance. Subsequently, determine the length of the ellipse's major axis, which needs to be considered in conjunction with the diffusion range, spatial distribution characteristics, and topographical constraints of high-temperature and drought elements within the feature analysis area. The specific operation is as follows: using the core analysis unit and an external reference unit as the two foci... Using the center point of each unit as the baseline, extend the line outwards to both sides. During the extension, cover at least three spatial units around the core analysis unit that are spatially related to the core unit, while avoiding non-disaster-affected areas, i.e., areas not included in the feature analysis area and terrain-blocked areas, such as mountains and rivers, which may block the diffusion of elements. After extending to a suitable range, measure the total length of the baseline, which is the total length of the major axis of the ellipse. Divide the total length of the major axis by 2 to obtain the semi-major axis of the ellipse, i.e., the distance from the center of the ellipse to the endpoint of the major axis. The calculation logic is the same as that for the semi-focal length. According to the geometric relationship, divide the calculated semi-focal length by the semi-major axis to obtain the eccentricity of the ellipse. The result is rounded to two decimal places. This ratio directly reflects the flatness of the ellipse. The larger the ratio, the flatter the ellipse; the smaller the ratio, the closer the ellipse is to a circle.

[0018] Step 224: Based on eccentricity, quantitatively characterize the spatial structure of high-temperature or drought elements within the region; and dynamically describe the evolution of spatial structure over time by analyzing the changes in eccentricity over a continuous time series; specifically, this includes: firstly, combining the spatial distribution patterns of high-temperature and drought disasters, dividing the region into four intervals according to eccentricity values, and determining the spatial structure and scene adaptation corresponding to each interval. The first interval has an eccentricity of 0.8-1.0 (extremely flat), indicating that the spatial distribution of elements is extremely dispersed and uneven, with elements mainly extending and spreading along the major axis of the ellipse, often appearing in hilly areas or areas distributed along watersheds, influenced by topography (valleys) and watershed orientation. High temperature or drought factors can only diffuse along a specific direction, forming a narrow, elongated distribution band. This indicates that the disaster's impact is spreading along a specific path, and the diffusion is uneven. Secondly, an eccentricity of 0.5-0.8 (moderate flattening) indicates that the spatial distribution of factors is relatively dispersed, with a clear diffusion direction. This often occurs in the boundary area between plains and hills. Factors diffuse in a main direction centered on the core analysis unit, while also spreading slightly in other directions, with the diffusion range and intensity showing a gradient. Thirdly, an eccentricity of 0.2-0.5 (slight flattening) indicates that the spatial distribution of factors is relatively concentrated, with a more even diffusion range. This often occurs in plain areas, with factors centered on the core analysis unit. The disaster spreads evenly in all directions without obvious direction, and the affected area expands in a roughly circular pattern; fourth, the eccentricity is 0-0.2 (close to a circle), indicating that the spatial distribution of the elements is highly concentrated and uniform, mostly in drought core areas or high-temperature concentrated outbreak areas. The elements are densely distributed in the core area, with a small diffusion range and high intensity. At this time, the disaster is in a concentrated outbreak state and needs to be closely monitored; then, continuous time series analysis is carried out. According to a time node of 6 hours, the core analysis unit and external reference unit element data of each node are collected synchronously. The calculation process of step 223 is repeated to calculate the eccentricity corresponding to each time node one by one, forming a continuous eccentricity time series. Each point corresponds to a unique eccentricity value and timestamp. By comparing and analyzing the changes in eccentricity between two adjacent time points, the correlation between the change pattern and the evolution of the disaster is determined. If the eccentricity at a later time point increases compared to the previous time point, and the increase exceeds 0.1, it indicates that the spatial distribution of the element is evolving from concentrated to dispersed, the scope of the disaster's impact is rapidly spreading, and the directionality of the spread is gradually increasing. If the eccentricity decreases, and the decrease exceeds 0.1, it indicates that the spatial distribution of the element is evolving from dispersed to concentrated, the disaster's impact is continuously focusing, and the disaster intensity in the core area may further increase. If the fluctuation range of the eccentricity is within 0.1, it indicates that the spatial morphology of the element is relatively stable, and the disaster is in a stable development or maintenance state. At the same time, the time-series change data of eccentricity is summarized every 24 hours, and the maximum, minimum, and average values ​​of the eccentricity for the day are statistically analyzed. The daily eccentricity change trend is analyzed, and combined with the meteorological conditions of the day, the reasons for the changes in eccentricity are explained, forming a complete record of spatial morphological evolution, clearly capturing the dynamic change process of the spatial structure of high temperature or drought elements.

[0019] Step 225: Based on the element values ​​of the core analysis unit and the external reference unit in multiple time series, calculate the correlation degree between the core analysis unit and the external reference unit at each time point; based on the correlation degree, generate a correlation sequence describing the change of their relationship over time; specifically, this includes: first determining the time series range and core elements, selecting a continuous 72-hour period as the analysis period, setting 12 nodes in total, each 6 hours long, namely 0:00, 6:00, 12:00, 18:00, and cycling sequentially up to 72 hours, with each node corresponding to a timestamp, ensuring that the time nodes are evenly distributed and cover different time periods of day and night; the core elements are still selected as temperature, soil moisture, and vegetation cover, where temperature is taken as the value of the corresponding time node. Real-time temperature and soil moisture were taken as the real-time averages of 0-10cm and 10-20cm. Vegetation cover was taken as the remote sensing inversion value for the corresponding time period. If there was no remote sensing data at night, the remote sensing value of the previous time period and the average value of the current day were interpolated to ensure data continuity. For each time node, the correlation baseline value of the same element in two units was calculated one by one, and the calculation logic was clearly defined for each element. For the temperature correlation baseline value, the absolute value of the difference between the real-time temperature of the core analysis unit and the real-time temperature of the external reference unit was calculated first. Then, the historical normal fluctuation range of temperature for the corresponding time period in the region was queried, such as 6:00 in summer and 12:00 in winter. The difference between the maximum and minimum temperature values ​​for the same period in the past 30 years was usually 8-10℃. The fluctuation range for the nighttime period was narrowed down to 4-6℃; Subtract the absolute value of the difference from the historical normal fluctuation range to obtain the temperature correlation baseline value. If the absolute value of the difference exceeds the historical normal fluctuation range, it indicates that the temperature states of the two are too different and there is no correlation. In this case, the temperature correlation baseline value is set to 0. If the calculation result is negative, it is also set to 0. The correlation baseline value cannot be negative. The soil moisture correlation baseline value is calculated using the same logic as the temperature. The historical normal fluctuation range is set according to the regional topography: 20-25% for plains and 18-22% for hilly and mountainous areas. The absolute value of the difference is calculated using the real-time average soil moisture of the core and reference unit. If the absolute value of the difference exceeds the fluctuation range, the correlation baseline value is set to 0. The vegetation cover correlation baseline value has a historical normal fluctuation range of 15-20%, because vegetation... If vegetation cover changes little in the short term, and the difference in vegetation cover between adjacent time points exceeds 5%, it is considered an anomaly. The average of the two consecutive time points is used as a substitute, and the correlation baseline value is calculated again according to the above logic. After calculating the correlation baseline values ​​of the three elements, the three values ​​are added together and then divided by 3 to obtain the comprehensive correlation degree for that time point. The result is rounded to two decimal places, and the value range is limited to 0-1. If the calculated result is greater than 1, it is taken as 1; if it is less than 0, it is taken as 0. The comprehensive correlation degree value corresponds to the correlation strength. 0.8-1.0 is a strong correlation, indicating that the states of the two unit elements are highly similar and the disaster impact trend is consistent. 0.5-0.8 is a medium correlation, indicating that there are some differences in the state of the elements, but the overall trend is consistent. 0-0.A correlation of 5 indicates a weak correlation, with significant differences in element states and potentially different disaster impact trends. The comprehensive correlation degrees of the 12 time nodes are arranged chronologically, with each node corresponding to three pieces of information: timestamp, comprehensive correlation degree, and correlation strength level, forming a complete sequence of correlation relationships. Simultaneously, key nodes in the sequence are labeled, such as the maximum correlation degree and its corresponding time point, the minimum correlation degree and its corresponding time point, and a brief explanation of the reasons for changes in the correlation degrees of key nodes, fully presenting the dynamic changes in the element correlation relationships between the core analysis unit and external reference units over time.

[0020] Step 226: Based on the correlation sequence, plot points in the spatiotemporal coordinate system with time as the horizontal axis and correlation degree as the vertical axis, and connect the coordinate points corresponding to each time point to generate an analysis path for characterizing the evolution trajectory of the correlation relationship. Specifically, this includes: first, constructing a standardized two-dimensional spatiotemporal coordinate system, determining the coordinate axis setting and labeling rules, with the horizontal axis representing time, its length evenly divided into 72-hour intervals, and a time node labeled every 6 hours, the labeling content being the time point and time period, such as 0:00-early morning, 6:00-morning, and the time unit labeled below the coordinate axis to ensure clear time scale and uniform intervals; the vertical axis representing correlation degree, evenly divided into a range of 0-1. The scale is divided into small increments of 0.1 and large increments of 0.2, each labeled with a corresponding value. The correlation strength level is also indicated on the right side of the vertical axis: 0.8-1.0 strong correlation, 0.5-0.8 moderate correlation, and 0-0.5 weak correlation. Different levels are distinguished by different text labels for easy visual identification. Based on the generated correlation sequence, the timestamp and overall correlation score data pairs for each time point are extracted and converted into coordinate points in a coordinate system. The timestamp corresponds to the horizontal axis scale, and the overall correlation score corresponds to the vertical axis scale. Each coordinate point is labeled with a solid circle of uniform diameter for easy identification. The corresponding time is also labeled next to each circle. To avoid confusion between coordinate points, plot the points and their correlation values. After plotting, connect all coordinate points using a smooth curve, following a chronological order. The curve should closely reflect the changing trends of the coordinate points. If the correlation between two adjacent coordinate points changes abruptly (the difference exceeds 0.3), connect them using a broken line (rather than forced smoothing) to accurately reflect drastic changes in correlation and avoid trajectory distortion. After connection, a continuous analysis path is formed. The fluctuations and direction of the path directly correspond to the evolution of the correlation. As the path ascends, the vertical axis value increases from left to right, indicating a gradually strengthening correlation between the core and reference units, reflecting the state of their elements. As the paths become closer, the synchronicity of disaster impact trends increases; a downward shift in the path (decreasing vertical axis value) indicates a gradual weakening of the correlation, widening differences in the states of the two elements, and a divergence in disaster impact trends; a horizontal extension of the path with vertical axis value fluctuations within 0.05 indicates stable correlation and consistent trends in element state changes. Furthermore, key change nodes are marked on the path, such as points of abrupt changes in correlation, peak points, and trough points, with brief notes on the meteorological background of each node, such as peak points corresponding to afternoon high temperatures and trough points corresponding to nighttime cooling. This allows the analysis path to not only visually present the evolution trajectory of the correlation but also explain the reasons for changes in conjunction with actual scenarios.

[0021] Step 227: Extract the curvature, direction, and trend features of the analysis path, and fuse them with the spatial structure features reflected by eccentricity to generate environmental parameter correction coefficients. Specifically, this includes: first, comprehensively extracting the three core features of the analysis path, determining the extraction logic and quantification standards one by one to ensure the features are quantifiable and computable. One aspect is curvature extraction: for each coordinate point on the path, except for the first and last nodes (since the first and last nodes lack complete line segments before and after), select the node and its two adjacent coordinate points to form three continuous line segments (front segment, middle segment, and rear segment). By measuring the angle between the middle segment and the front segment, and the angle between the middle segment and the rear segment, the average of the two angles is used as the curvature judgment criterion for that node. The smaller the angle, the more pronounced the curvature of the path. The larger the rate, the more drastic the change in correlation at that time point, such as a rapid decrease from strong correlation to weak correlation. The larger the angle, the smoother the path, and the smaller the curvature, the smoother the change in correlation. Curvature is quantified into a numerical range of 0-0.5, with a 90° angle corresponding to a curvature of 0.5 (the most drastic bending), and a 180° angle corresponding to a curvature of 0 (completely smooth). For every 18° decrease in the angle, the curvature increases by 0.05, and so on, ensuring that the curvature value accurately corresponds to the degree of bending. Secondly, direction extraction: for each coordinate point and the direction of the line connecting to the next coordinate point, the path direction is determined. If the line is upward, the correlation of the next node is higher than that of the current node, and the difference exceeds 0.05, it is recorded as positive and assigned a value of 1; if the line is downward, the correlation of the next node is lower than that of the current node, and the difference exceeds 0.05, it is recorded as positive and assigned a value of 1. A value exceeding 0.05 is considered negative and assigned a value of -1; horizontal connections with a correlation difference within 0.05 are considered neutral and assigned a value of 0. Each node corresponds to a directional value, with the first and last nodes only labeled with direction and not participating in subsequent fusion calculations; thirdly, trend extraction is performed based on the changing patterns of the entire analysis path, calculating the correlation difference between the first and last nodes. A positive difference indicates that the correlation of the last node is higher than that of the first node, with a difference exceeding 0.1, indicating an overall upward trend, and is assigned a value of 1; a negative difference indicates that the correlation of the last node is lower than that of the first node, with a difference exceeding 0.1, indicating an overall downward trend, and is assigned a value of -1; a difference within 0.1 indicates a stable trend, and is assigned a value of 0. The entire path corresponds to a trend value; subsequently, the eccentricity values ​​of corresponding nodes on the same time series are retrieved and compared with... The analysis path time nodes are matched one-to-one, and feature fusion calculations are carried out to determine the calculation steps. First, the curvature value, direction value, and corresponding eccentricity value of each node are added to obtain a preliminary fusion value. The trend value is a unified value for the entire path, and this value is used for calculation at each node. Second, based on the influence weights of each feature on environmental parameters and combined with disaster monitoring practices, curvature accounts for 30%, direction for 20%, trend for 20%, and eccentricity for 30%. The preliminary fusion value is multiplied by the corresponding weights to obtain four weighted fusion values. Third, the four weighted fusion values ​​are added to obtain the environmental parameter correction coefficient for that node. The result is rounded to two decimal places, with the value range controlled between 0.8 and 1.2. If the calculated result is lower than 0.8, it is taken as 0.8; if the value is higher than 1.2, use 1.2 to avoid over-correction leading to data distortion; each time point corresponds to an environmental parameter correction coefficient, forming a continuous sequence of correction coefficients to ensure that the calibration operation can adapt to the spatiotemporal characteristics of different time points.

[0022] Step 228: Based on the environmental parameter correction coefficient, calibrate the values ​​of key monitoring elements such as temperature, precipitation, and soil moisture in the feature analysis area of ​​the fused dataset within the corresponding time period to generate a calibrated fused dataset. Specifically, this includes: first, determining the calibration range and calibration objects; the calibration period is consistent with the previous time series, being a continuous 72 hours covering 12 time nodes; the calibration spatial range covers all spatial units within the feature analysis area to avoid data imbalance caused by calibrating only the core and reference units; the calibration objects are three key monitoring elements: temperature includes daily average temperature, daily maximum temperature, and real-time temperature; precipitation includes 24-hour precipitation and real-time precipitation; and soil moisture includes stratified humidity at 0-10cm and 10-20cm depths and the mean value of both layers. To ensure full coverage of core elements, calibration calculations are performed on each key element data point for each spatial unit and each time node. The calculation logic and calibration rules are determined, and the original monitoring value of the element is multiplied by the environmental parameter correction coefficient for the corresponding time node to obtain the calibrated element value. The calculation process maintains the same number of decimal places for the original and calibrated data. The calibration logic corresponds to the data deviation correction. If the correction coefficient is greater than 1 (1.01-1.20), it indicates that the original monitoring data underestimated the actual state of the element, possibly due to obstruction or signal attenuation caused by the equipment installation location. The calibrated value is correspondingly amplified to better reflect the actual disaster impact. If the correction coefficient is less than 1 (0.80-0.99), it indicates that the original monitoring data overestimated the actual state of the element. The original data deviation may be due to slight drift caused by equipment malfunction or localized abnormal interference. The calibrated values ​​will be correspondingly reduced to correct the data deviation. If the correction factor is between 0.95 and 1.05, it indicates that the original data deviation is small, and the calibrated values ​​will remain essentially unchanged, requiring only minor adjustments to balance data accuracy and stability. After calibration, data verification and anomaly handling are conducted to ensure the calibrated data is reasonable and continuous. The calibrated data for each spatial unit and each time point are checked one by one, and abnormal data exceeding the reasonable range of the elements are removed. For example, temperatures exceeding 45℃ (the upper limit of extreme high temperatures in plains areas) or below -5℃ (unreasonably low temperatures in high-temperature and drought scenarios) after calibration, and soil moisture exceeding 100% (the upper limit of saturated humidity) or below 10% (extremely dry) after calibration. (Drought lower limit), precipitation exceeding 200mm per day after calibration (unreasonable heavy precipitation in high temperature and drought scenarios); for such abnormal data, the average of the calibrated data at two adjacent time points is used as a replacement, and the reason for the replacement and the original data are recorded to ensure data traceability; then the calibrated data is integrated with the original dataset, all calibrated key element data are associated and integrated with the uncalibrated auxiliary data in the fused dataset, the corresponding entries in the original fused dataset are updated, the original key element data are replaced, and a separate calibration record ledger is established. The ledger includes spatial unit code, time point, original data, correction coefficient, calibrated data, and anomaly handling instructions to ensure that every calibration data is verifiable; finally, a complete calibrated fused dataset is formed.

[0023] Interval definition and time-series dynamic analysis of eccentricity capture the spatial distribution heterogeneity and evolution patterns of high-temperature and drought factors, thereby improving the accuracy and reliability of fused datasets.

[0024] In a preferred embodiment of the present invention, step 3 above, based on the calibrated fused dataset, constructs a time-series dynamic analysis structure, extracts time-series dynamic feature vectors of key indicators of high temperature and drought; analyzes the change trajectory of key indicators, calculates the torsion angle of evolution over time, quantifies trend turning points and fluctuation characteristics, and fuses the time-series dynamic feature vectors and torsion angle features to form a comprehensive feature matrix, which may include: In this embodiment of the invention, step 330 involves selecting a set of key indicators for characterizing high-temperature events and a set of key indicators for characterizing drought events from the calibrated fusion dataset. Specifically, this includes: firstly, considering the bidirectional interaction between high-temperature and drought-related disasters (high temperatures exacerbate drought, and drought intensifies high temperatures), and secondly, taking into account the completeness of elements in the calibrated dataset, selecting indicators specifically for the characterization needs of different terrains and different disaster stages. Simultaneously, by comparing historical monitoring data, redundant indicators with a correlation exceeding 80% are eliminated to ensure that each indicator reflects unique disaster characteristics and that the data is traceable and verifiable. The set of key indicators for high-temperature events selects six core indicators. Taking into account intensity, duration, spatial spread, and diurnal differences, the following criteria are used: 1) Daily average temperature, representing the overall high-temperature benchmark for the region, adaptable to all terrains; 2) Daily maximum temperature, representing the intensity of extreme high temperatures, with a focus on plains areas; 3) Duration of high temperatures, setting thresholds according to terrain and counting the cumulative number of hours the daily temperature exceeds the threshold, distinguishing between day and night; 4) Cumulative high-temperature intensity, calculating the difference between the daily maximum temperature and the corresponding threshold, with a negative difference set to 0, and no accumulation when there are no high temperatures, then summing the differences over 30 days to represent the cumulative impact of high temperatures; 5) Spatial clustering of high temperatures, statistically analyzing the number of spatial unit clusters in the region where high temperatures occur for 3 consecutive days or more, within a 1km × 1km radius. The 1km unit is considered a cluster, representing the spatial spread of high temperatures; sixth, the diurnal temperature variation, calculated as the difference between the highest and lowest temperatures of the day, reflects the intensity of diurnal temperature fluctuations; the key indicators set for drought events selects six core indicators, taking into account water replenishment, soil drought, and ecological response. These include: 1) 7-day cumulative precipitation, calculated by adding the calibrated daily precipitation of the past 7 days, representing short-term water replenishment capacity; 2) 20-day precipitation anomaly percentage, calculated by first comparing the calibrated cumulative precipitation of the past 20 days with the historical average for the same period, then dividing this difference by the historical average for the same period, representing the degree of medium-term water deficit; 3) 0-10cm surface soil moisture, extracted daily from the calibrated average, directly... The indicators reflect the drought status of crops; fourth, the soil moisture at a depth of 10-20cm, indicating whether the drought is spreading to deeper layers; fifth, the soil moisture decay rate, calculated by comparing the current day's deep soil moisture calibration value with the previous day's value, with a negative difference indicating decay and a positive difference indicating mitigation, representing the speed of drought spread; and sixth, the vegetation drought response index, calculated by comparing the current vegetation coverage calibration value with the historical average for the same period, with a negative difference indicating vegetation degradation due to drought. All indicators are extracted from the calibrated fusion dataset and must meet the conditions of no more than 3 consecutive days of missing data in the past 30 days and complete calibration records. Finally, the names, definitions, applicable scenarios, data sources, and selection criteria of the two types of indicator sets are determined.

[0025] Step 331: Based on the key indicators of high temperature and drought, extract the observation values ​​of each key indicator within a preset time window from the calibrated fusion dataset, and arrange them in chronological order to generate a corresponding time series data for each key indicator. Specifically, this includes: first, setting a preset time window, considering the characteristics of short-term abrupt changes in high temperature (2-3 days) and medium-to-long-term gradual evolution of drought (15-30 days), uniformly setting it to 30 consecutive days, extracting observation values ​​at 24-hour intervals, and supplementing with instantaneous values ​​at four key time periods each day (0:00, 6:00, 12:00, 18:00) to ensure that the time granularity can capture both short-term fluctuations in high temperature and the gradual development of drought; for each indicator in the set of key indicators of high temperature, ... The following steps are performed: 1. Extract calibrated observations for each day within the corresponding time window. 2. Extract the average daily temperature from the 24-hour calibrated temperature average, sum the hourly temperatures over 24 hours, and divide by 24. 3. Extract the maximum daily temperature from the instantaneous values ​​of the four key time periods and the maximum value among other hourly values. 4. Verify the duration of high temperatures hourly against the calibrated temperature, accumulating the number of hours exceeding the corresponding terrain threshold, distinguishing between afternoon (12-18) and other time periods, and recording the afternoon high temperature duration separately. 5. Calculate the cumulative high temperature intensity daily by comparing the maximum temperature with the corresponding threshold; if the difference is negative, set it to 0. Then, sum the differences over 30 days to form a cumulative sequence. 6. Analyze the spatial clustering of high temperatures daily by statistically analyzing the number of clusters meeting the criteria within the region and marking the core location of the clusters. 7. Analyze diurnal and nighttime high temperatures. The difference calculation is performed by taking the difference between the highest and lowest temperatures of the day, with the lowest temperature being the minimum of the instantaneous values ​​at 0:00 and 6:00. For the key drought indicators, the daily cumulative precipitation over 7 days is added to the calibrated daily precipitation over the past 7 days, and the current day's precipitation is added to the sum of the precipitation over the previous 6 days, calculated daily on a rolling basis. The 20-day precipitation anomaly percentage is calculated by first taking the cumulative calibrated precipitation over the past 20 days, subtracting the historical average cumulative precipitation over the same period for the same period to obtain the precipitation difference, and then dividing the precipitation difference by the historical average for the same period. The daily calibrated average values ​​of surface and deep soil moisture are extracted, and the instantaneous value at 12:00 (the period when soil moisture is most stable) is also recorded. The soil moisture decay rate is calculated daily by subtracting the calibrated value of deep soil moisture from the previous day's deep soil moisture; a negative difference indicates a negative value. The drought intensifies, and a positive expression indicates mitigation. The vegetation drought response index is calculated by subtracting the historical average value from the current vegetation cover calibration value, and negative values ​​are labeled with the degree of vegetation degradation. After extraction, for each indicator, the 30-day observation values ​​are arranged chronologically from day 1 to day 30 to construct an independent time series for each indicator. Each series is bound with a correspondence between timestamp, calibrated observation value, adapted terrain, and time period label. If data is missing at individual time points, it is filled by the average of the observation values ​​of the corresponding time periods of the two adjacent days, by adding the observation value of the previous day to the observation value of the next day and dividing by 2. If there are no more than 2 consecutive missing days, it is filled by the average of the observation values ​​of the previous 3 days, ensuring that each time series data is continuous, complete, and without breaks.

[0026] Step 332: Based on the time series data, calculate the mean, amplitude, fluctuation frequency, and stage trend slope for each time series data, and extract and generate the statistical features corresponding to each time series. Specifically, for each indicator time series (12 in total, 6 for high temperature and 6 for drought), calculate four statistical features: First, calculate the mean by summing all observations over 30 days in the series, then dividing the sum by 30 (the number of days in the time window) to obtain the mean of the indicator time series. The mean reflects the overall average state of the indicator, such as average high temperature intensity and average soil moisture level, and also indicates the terrain adaptability corresponding to the mean; Second... The calculation involves three steps: First, calculating the amplitude. The maximum and minimum values ​​of the observed values ​​are selected from the sequence. The maximum value is subtracted from the minimum value to obtain the amplitude. The amplitude reflects the extreme fluctuation range of the indicator, such as the maximum difference in high temperature intensity or the extreme fluctuation range of soil moisture. For indicators with significant diurnal differences, the diurnal amplitude is calculated separately: the maximum value in the afternoon is subtracted from the minimum value at night. Second, calculating the fluctuation frequency. Normal fluctuation thresholds are first set according to indicator type and terrain. The daily average temperature fluctuation threshold is set to 2℃ for plains and 1.5℃ for hills. The soil moisture fluctuation threshold is uniformly set to 5%, the precipitation fluctuation threshold is set to 10mm, and the diurnal high temperature difference fluctuation threshold is set to 3℃. The daily values ​​in the sequence are then compared day by day. The absolute value of the difference between the observed value and the previous day's observed value is recorded as a valid fluctuation if the absolute value of the difference exceeds a set threshold. The total number of valid fluctuations within 30 days is counted, and then the total number of fluctuations is divided by 30 to obtain the fluctuation frequency. A higher frequency indicates more drastic changes in the indicator, such as frequent fluctuations in high temperature intensity or rapid fluctuations in soil moisture. The periods of concentrated fluctuations are also marked. Fourth, the slope of the phased trend is calculated. The 30-day time window is divided into three phases, each lasting 10 days (days 1-10, 11-20, and 21-30). For each phase, the observed values ​​on the first and last days of the phase are extracted. The observation value on the last day is subtracted from the observation value on the first day to obtain the intra-phase observation value. The difference between the measured values ​​is then divided by 10 (the number of days in the stage) to obtain the trend slope for that stage. A positive slope indicates an upward trend, while a negative slope indicates a downward trend. The larger the absolute value of the slope, the more obvious the trend. A slope close to 0 indicates a stable trend. For drought indicators, the disaster development trend corresponding to the positive and negative slopes is marked separately. An increase in the absolute value of a negative slope indicates an accelerated spread of drought. After the calculation is completed, a statistical feature set containing the mean, amplitude, fluctuation frequency, slope of stage 1, slope of stage 2, and slope of stage 3 is generated for each time series. The calculation basis, terrain adaptation description, and time period characteristics of each feature are marked to ensure traceability and repeatability.

[0027] Step 333: Combine the statistical features corresponding to all indicators in the high-temperature key indicator set in a preset order to generate a high-temperature time-series dynamic feature vector; combine the statistical features corresponding to all indicators in the drought key indicator set in a preset order to generate a drought time-series dynamic feature vector. Specifically, this includes: first, setting a unified preset order to ensure consistent and unambiguous feature combination logic, adapting to fusion operations, and facilitating reverse tracing of corresponding indicators; high-temperature key indicators are arranged in a fixed order: daily average temperature - daily maximum temperature - high-temperature duration - high-temperature intensity cumulative value - high-temperature spatial concentration - diurnal temperature difference; drought key indicators are arranged in a fixed order: 7-day cumulative precipitation - 20-day precipitation anomaly percentage - surface soil moisture - deep soil moisture - soil moisture decay rate - vegetation drought response index; the statistical features corresponding to each indicator are arranged in the order of mean - amplitude - fluctuation frequency - first-stage slope - second-stage slope - third-stage slope, and the feature belongs to which the indicator is labeled. The system identifies the statistical types of features. For the high-temperature time-series dynamic feature vector, six statistical features of the daily average temperature are first extracted and arranged in the order described above. Then, six statistical features of the daily maximum temperature, duration of high temperature, cumulative value of high temperature intensity, spatial concentration of high temperature, and diurnal temperature difference are extracted sequentially and linked one by one after the daily average temperature feature to form a continuous feature sequence. This sequence is the high-temperature time-series dynamic feature vector, with a total dimension of 6 indicators multiplied by 6 features, resulting in 36 feature dimensions. For the drought time-series dynamic feature vector, six statistical features of each drought indicator are extracted sequentially according to a preset order and combined into a continuous feature sequence using the same linking method, forming a drought time-series dynamic feature vector, also with 36 feature dimensions. After combination, detailed attributes are labeled for each feature dimension in the vector to avoid feature misalignment or confusion. At the same time, the vector is checked for completeness to verify that each feature dimension has a corresponding value and is not missing, ensuring that the structure and number of dimensions of the two types of vectors are consistent.

[0028] Step 334: Based on the indicator data in the time-series dynamic feature vector, select a continuous segment of key indicator time series with a preset time length. Map the multiple indicator values ​​corresponding to each time point of the continuous segment of key indicator time series to a coordinate point in a multi-dimensional space, forming a spatial trajectory representing the change of the indicators. Specifically, this includes: first determining the preset time length, and combining the short-term coordinated change law of high temperature and drought elements, selecting a continuous 12-hour period as the preset time length, dividing it into 6 time nodes (0:00, 2:00, 4:00, 6:00, 8:00, and 10:00) with 2-hour intervals, covering the night, early morning, and morning periods, and capturing short-term dynamic changes. The study investigated the synergistic patterns of indicators during ecological changes and diurnal transitions. From the calibrated time series of indicators, all key indicator observations corresponding to each time node were extracted: 6 high-temperature indicators and 6 drought indicators, totaling 12 indicators. This ensured that each time node corresponded to complete calibrated observations of all 12 indicators without missing values. For nodes lacking nighttime vegetation cover data, the mean of the observations at 18:00 of the previous day and 6:00 of the current day was used to fill in the gaps, ensuring data continuity. A multi-dimensional space was constructed, with the number of spatial dimensions matching the number of indicators, totaling 12 dimensions. Each dimension corresponds to a key indicator, such as dimension 1 corresponding to daily average temperature, dimension 2 corresponding to daily maximum temperature, etc. …Dimension 12 corresponds to the vegetation drought response index. Each dimension has a range set according to the calibrated numerical range of the corresponding indicator. For example, the daily average temperature dimension is set to 25℃-40℃, the soil moisture dimension to 20%-70%, and the precipitation dimension to 0-50mm. The upper and lower limits of the ranges exceed the extreme values ​​of the calibrated data by 10% to ensure that the coordinate points can be accurately mapped without overflow. For each time node, the observed values ​​of the 12 indicators at that node are mapped to the corresponding values ​​of the 12 dimensions. That is, each indicator value corresponds to a specific location within its dimension, forming a unique coordinate point in 12-dimensional space. Each coordinate point is bound to a time node… Corresponding indicator values, terrain attributes, and time periods are labeled; in chronological order, six coordinate points are marked sequentially in 12-dimensional space, and then adjacent coordinate points are connected one by one with smooth line segments to form a continuous spatial trajectory; the direction of the trajectory reflects the overall changing trend of the 12 indicators, and the curvature reflects the turning point of the coordinated changes of the indicators. If a certain indicator value changes abruptly at a certain node, such as a sudden increase in precipitation index and a synchronous increase in soil moisture due to short-term precipitation in the early morning, while the high temperature index decreases slightly, the trajectory will show a significant turning point, accurately depicting the coordinated dynamic change process of the 12 indicators within 12 hours, while also marking the time period and meteorological background corresponding to the trajectory turning point.

[0029] Step 335: Based on the spatial trajectory, calculate the vector difference between coordinate points corresponding to adjacent time points to obtain the direction vector between adjacent time points; based on the direction vector between adjacent time points, calculate the angle between every two adjacent direction vectors in sequence to generate a sequence of torsion angles that change with time; specifically, this includes: first calculating the vector difference between adjacent coordinate points; for the 6 coordinate points on the spatial trajectory, select two adjacent sets of coordinate points in chronological order: 1st and 2nd, 2nd and 3rd, 3rd and 4th, 4th and 5th, 5th and 6th, forming a total of 5 sets of adjacent coordinate points; for each set of adjacent coordinate points, subtract the previous coordinate point from the 12 dimension values ​​of the next coordinate point. The numerical values ​​corresponding to the punctuation marks are used to obtain the numerical differences across 12 dimensions. These 12 numerical differences are combined sequentially according to the dimension to form a vector difference, which represents the direction vector between adjacent time points. A positive value in the direction vector indicates an upward trend in the corresponding indicator, while a negative value indicates a downward trend. The larger the absolute value, the greater the magnitude of the change. The direction of change for each dimension's numerical difference is also marked; for example, a positive numerical difference in dimension 3 indicates an upward trend in the duration of high temperatures. Subsequently, the angles between adjacent direction vectors are calculated. For the five generated direction vectors, two adjacent direction vectors are selected sequentially: the first and second, the second and third, the third and fourth... The 4th and 5th angles together form 4 angle calculation groups. When calculating each angle, first compare the numerical trends of the two direction vectors dimension by dimension to determine if they are in the same or opposite direction. In the same dimension, if both vectors are positive, both are negative, or both are 0, they are considered to have the same trend; if one is positive and the other is negative, they are considered to have opposite trends. Count the number of dimensions with the same trend across the 12 dimensions, divide the number of dimensions with the same trend by 12 (total number of dimensions) to obtain the percentage of dimensions with the same trend; then multiply 180° by (1 - percentage of dimensions with the same trend) to obtain the angle between the two direction vectors. The angle value is strictly controlled between 0° and 180°. For example, if 9 out of 12 dimensions have the same trend, then... The proportion of the direction is 0.75, so the included angle is 180° multiplied by 0.25, which equals 45°. If the trends of the 12 dimensions are completely opposite, the included angle is 180°. If they are all in the same direction, the included angle is 0°. If 6 are in the same direction and 6 are in opposite directions, the included angle is 90°. The larger the included angle, the more drastic the change in the trend of the indicator between two time points, and the more obvious the change in the disaster status. The four included angles are arranged in chronological order to form a torsion angle sequence. Each angle corresponds to the turning relationship between two adjacent directional vectors. The time points and trend change characteristics corresponding to the angles are marked, which fully reflects the bending and turning rules of the spatial trajectory and indirectly quantifies the sudden change characteristics of the indicator trend.

[0030] Step 336: Based on the torsion angle sequence, calculate the maximum value of the torsion angle sequence as the maximum torsion angle, calculate the arithmetic mean of the torsion angle sequence as the average torsion angle, and count the number of times the torsion angle exceeds a preset threshold per unit time as the torsion frequency. Specifically, this includes: first, determining the calculation logic for each statistical item; combining practical experience in judging the trend reversal of high-temperature and drought disasters to ensure that the statistical results accurately represent the fluctuation characteristics; and simultaneously marking the statistical basis. First, calculate the maximum torsion angle: from the torsion angle sequence, comparing all angle values ​​one by one, selecting the angle with the largest value as the maximum torsion angle; the maximum torsion angle reflects the most drastic reversal of the spatial trajectory, corresponding to the most significant abrupt change in the trend of indicator changes. Simultaneously, mark the two adjacent time nodes corresponding to this angle and the details of indicator changes, such as a node where short-term precipitation causes a sudden increase in precipitation and soil moisture indicators, while the high-temperature indicator simultaneously decreases, forming a large-angle reversal; second, calculate the average torsion angle: add all four angle values ​​in the torsion angle sequence to obtain the angle sum, then divide the sum by 4 (sequence length) to obtain the arithmetic mean. The average torsion angle reflects the smoothness of the overall turning point of the spatial trajectory. The smaller the average value, the more stable the trend of the indicator change, without drastic fluctuations. The larger the average value, the more unstable the coordinated change of the indicator, and the more likely the disaster state is to recur. The third step is to calculate the torsion frequency. First, a preset threshold for the torsion angle is set. Combined with historical disaster monitoring data, when the torsion angle exceeds 60°, it is often accompanied by a significant change in the disaster trend, such as the high temperature changing from strengthening to weakening, or the drought changing from spreading to alleviating. Therefore, the threshold is set to 60°. The number of angles exceeding 60° in the torsion angle sequence is counted to obtain the number of times the threshold is exceeded. The preset time length is 12 hours. The number of times the threshold is exceeded is divided by 12 to obtain the torsion frequency per unit time. The higher the torsion frequency, the more frequent the turning point of the indicator change trend, and the more unstable the disaster state. The subsequent development trend needs to be closely monitored. After the statistics are completed, the specific angle values ​​of the maximum torsion angle and the average torsion angle, as well as the calculation results of the torsion frequency, are recorded. The basis for the threshold setting and the time nodes corresponding to the number of times the threshold is exceeded are marked to ensure that all three features can quantitatively represent the turning and fluctuation characteristics of the spatial trajectory, providing a supplementary dimension for subsequent fusion operations.

[0031] Step 337 involves fusing the time-series dynamic feature vector with the maximum torsion angle, average torsion angle, and torsion frequency features to form a comprehensive feature matrix. Specifically, this includes: first determining the fusion rules and matrix structure to ensure comprehensive feature coverage and orderly arrangement, taking into account the needs of disaster probability calculation while facilitating data traceability and verification. The fusion objects include the high-temperature time-series dynamic feature vector (36 dimensions), the drought time-series dynamic feature vector (36 dimensions), and torsion angle features (maximum torsion angle, average torsion angle, and torsion frequency), for a total feature dimension of 36+36+3=75 dimensions. The matrix is ​​constructed using a row-column structure, with rows divided into three categories: high-temperature time-series feature rows, drought time-series feature rows, and torsion angle feature rows. Columns represent feature dimensions, arranged in a fixed order: high-temperature time-series feature - drought time-series feature - torsion angle feature. Each column corresponds to a feature dimension, labeled with the feature name, associated indicator, statistical type, and applicable terrain and time period attributes. The specific fusion operation begins by fusing the high-temperature time-series dynamic feature vector with the drought time-series dynamic feature vector. The 36 feature dimensions of the eigenvector are filled into the first 36 columns of the high-temperature time series feature row in the order of the original indicators and statistical features. Then, the 36 feature dimensions of the drought time series dynamic feature vector are filled into the 37th to 72nd columns of the drought time series feature row in the corresponding order. Finally, the three features of maximum torsion angle, average torsion angle, and torsion frequency are filled into the 73rd to 75th columns of the torsion angle feature row, respectively. After filling, the matrix data is checked column by column to verify whether the values ​​of each feature dimension are correct, without missing or misaligned values. At the same time, the matrix description is supplemented to determine the feature details, fusion logic, time window, and terrain adaptation information of each row and column. A separate matrix index table is established to associate the feature dimensions with the original indicators and time series data to ensure that each feature can be traced back to the original observation value. The final comprehensive feature matrix not only covers the time series statistical features of high temperature and drought indicators, but also incorporates the turning and fluctuation features of the coordinated changes of indicators, realizing the deep integration of time series dynamic information and spatial trajectory features.

[0032] By setting a unified time window and multiple granular time nodes, coupled with the calculation of dimensional statistical features and the filling of missing data, the system comprehensively captures the overall state, extreme fluctuations, phased evolution, and time-segment differences of high temperature and drought indicators. It clearly depicts the complete process of disasters from gradual development to short-term abrupt changes, making up for the shortcomings of traditional static statistics in covering dynamic patterns and time-segment characteristics. Through the construction of multi-dimensional spatial trajectories and the quantitative analysis of torsion angles, the system realizes a concrete representation of the trend reversal characteristics of coordinated changes in indicators through dimensional trend comparison and angle calculation, capturing key signals of abrupt changes in disaster status.

[0033] In a preferred embodiment of the present invention, step 4 above, which involves inputting the comprehensive feature matrix into a pre-trained machine learning classification model to calculate the probability of occurrence of a combined high-temperature and drought disaster and to determine the warning level, may include: In this embodiment of the invention, step 440 involves inputting the comprehensive feature matrix into a machine learning classification model to generate a probability value representing the likelihood of a combined high-temperature and drought disaster. Specifically, this includes: firstly, constructing and training the machine learning classification model, based entirely on historical monitoring data of high-temperature and drought conditions, to ensure the model adapts to the disaster's evolution. The construction process is as follows: Step 1, data preparation: collecting high-temperature and drought monitoring data from different terrains (hills, plains) and seasons over the past 10 years, covering the calibrated element data, time-series feature vectors, torsion angle features, and actual disaster occurrence records for the corresponding periods mentioned above. Samples with missing data exceeding 3 days or outliers that cannot be corrected are removed, ultimately forming a valid dataset. Simultaneously, the data... The dataset is divided into training and validation sets based on time series, with the training set comprising 70% (data from the previous 7 years) and the validation set comprising 30% (data from the next 3 years). This avoids future data leakage due to random partitioning and ensures that the model's generalization ability is adapted to the actual monitoring scenario. The second step involves the selection and structural design of the machine learning classification model. A random forest classification model is chosen because it is adaptable to multi-dimensional features and has strong anti-interference capabilities, fitting the multi-dimensional attributes of the comprehensive feature matrix. The core structure of the random forest classification model is an ensemble of multiple decision trees, with the number of decision trees set to 50. This number is determined based on the dataset size; too few trees can lead to underfitting, while too many can lead to redundancy. The depth of each decision tree is controlled within 8 layers to avoid overfitting to historical data. During tree construction, 70% of the samples in the training set and 60% of the feature dimensions in the comprehensive feature matrix are randomly selected to ensure the independence and diversity of each tree. In the third step, the comprehensive feature matrix of the training set and its corresponding labels (labeled according to the actual disaster occurrence, with 1 for occurrence and 0 for non-occurrence) are input into the random forest classification model. Decision trees are trained one by one. During training of each decision tree, split nodes are selected based on the information gain of the feature dimensions. This means calculating the discriminative power of each feature for the disaster occurrence label and selecting the feature with the highest information gain as the basis for splitting the current node. This process is recursively repeated until a preset depth is reached or the number of node samples is less than 5. After all decision trees are trained, the performance of the random forest classification model is tested using a validation set, and predictions are calculated. Accuracy and error rate: If the accuracy is below 85%, adjust the number and depth of the decision tree and retrain; at the same time, pruning is used to remove redundant split nodes in the decision tree to reduce model complexity; after training, save the optimized model parameters to form a pre-trained model that can be directly called; the fourth step is feature matrix preprocessing and probability generation. The comprehensive feature matrix generated above is normalized to eliminate the difference in the numerical range of different feature dimensions. The processing logic is as follows: for all values ​​of each feature dimension in the matrix, subtract the minimum value of the dimension from the actual value of the dimension to obtain the feature difference, and then divide the difference by the difference between the maximum and minimum values ​​of the dimension to obtain the normalized feature value, ensuring that all feature values ​​are in the 0-1 range;The preprocessed comprehensive feature matrix is ​​input into a pre-trained random forest classification model. Each of the 50 decision trees in the random forest model independently predicts the probability of disaster occurrence, with each tree outputting a predicted probability value between 0 and 1. The predicted probability values ​​from all 50 decision trees are then summed to obtain a total probability. This total probability is divided by 50 (the total number of decision trees) to obtain the average probability value. This average probability value is the final probability value representing the likelihood of a combined high-temperature and drought disaster. The result is rounded to two decimal places, strictly controlled within the range of 0-1. If any individual decision tree's predicted probability exceeds the 0-1 range, that value is discarded, and the average of the remaining decision tree probabilities is calculated to ensure a reasonable result.

[0034] Step 441: Match the probability value with multiple preset probability threshold intervals. Based on the matching results, map the probability value to the corresponding warning level. Specifically, this includes: First, combining historical disaster data, disaster prevention and mitigation response capabilities, and the risk characteristics of different terrains, setting probability threshold intervals to ensure that the interval division fits the actual application scenario and corresponds to the degree of disaster risk. The correspondence between the threshold intervals and the warning levels and the setting basis are as follows: The first interval is where the probability value is less than the critical value one, corresponding to a blue warning (low risk). The setting basis is that the actual occurrence rate of disasters in this interval has been extremely low in the past 10 years, requiring only routine monitoring, with a focus on the stability of high temperature and drought indicators; The second interval is where the probability value is between the critical value one and... The threshold values ​​are as follows: Between two threshold values ​​corresponds to a yellow alert (lower risk), based on the occasional occurrence of disasters within this range, requiring increased monitoring frequency and a focus on tracking changes in soil moisture and the duration of high temperatures; the third threshold is between two and three threshold values, corresponding to an orange alert (higher risk), based on a relatively high probability of disasters occurring within this range, requiring the initiation of preventative measures, such as advance irrigation planning and stockpiling of heatstroke prevention supplies; the fourth threshold is greater than three threshold values, corresponding to a red alert (high risk), based on a high probability of disasters occurring within this range, requiring immediate activation of emergency response and implementation of measures such as crop irrigation and high-temperature protection; the threshold values ​​should be adapted to terrain differences, especially in plains areas. High temperatures tend to accumulate, while drought spreads relatively slowly. Therefore, the threshold values ​​are set as follows: Threshold 1: 0.2; Threshold 2: 0.4; Threshold 3: 0.7. In hilly areas, soil water retention is weak, and drought spreads rapidly. Therefore, Threshold 2 is lowered to 0.35, and Threshold 3 to 0.65, ensuring the risk assessment aligns with regional characteristics. After setting the threshold ranges, probability value matching and level mapping are performed. First, the disaster occurrence probability values ​​generated in step 440 are extracted and compared one by one with the corresponding terrain threshold range threshold values ​​to determine the range within which the probability value falls. If the probability value is exactly equal to the threshold value, it is mapped to a higher warning level to avoid insufficient warnings under critical conditions and ensure no risk is overlooked. After matching is complete... The probability values ​​are mapped to the corresponding warning levels, and detailed warning descriptions are generated. The risk characteristics, monitoring priorities, and response suggestions corresponding to each level are determined. Blue warnings require daily recording of changes in core indicators, yellow warnings require monitoring of soil moisture and temperature every 6 hours, orange warnings require advance coordination of irrigation resources, and red warnings require organizing personnel to carry out emergency protection. If the probability value exceeds the 0-1 range (extreme abnormal situation), it is mapped to a yellow warning by default. At the same time, the root cause of the problem is investigated, with a focus on verifying the parameter settings of the comprehensive feature matrix preprocessing link and the prediction system. After correction, the probability values ​​are recalculated and the levels are mapped. Finally, the warning level, warning description, and probability value are output together to form a complete warning result.

[0035] The training and pruning optimization process of the prediction system ensures a balance between accuracy and efficiency in the judgment results. Normalization eliminates the weight imbalance caused by differences in feature dimensions, allowing the probability value to truly reflect the likelihood of disaster occurrence. The probability threshold range is set based on historical disaster patterns and regional risk characteristics, and can be flexibly adapted to different terrains. By adjusting the critical value and the rules for raising the level of the critical state, risk classification is achieved, avoiding insufficient or excessive early warning.

[0036] In a preferred embodiment of the present invention, step 5 above, which involves issuing warning information according to the warning level and collecting actual disaster information from external feedback as new training data to periodically retrain the machine learning classification model to achieve adaptive updates, may include: In this embodiment of the invention, step 550 involves generating early warning information based on the warning level, including disaster intensity, impact range, and defense recommendations. Within a set period after the warning information is issued, actual disaster information is collected from meteorological stations, remote sensing monitoring platforms, and on-site manual reports. Specifically, this includes: firstly, combining the warning levels (blue, yellow, orange, red) determined above, linking them with the disaster risk characteristics of the corresponding terrain (plains, hills), and generating targeted early warning information to ensure complete and operable information dimensions; the disaster intensity description is refined according to the warning level, with a blue warning (low risk) indicating extremely low risk of high temperature and drought, and no obvious disaster intensity; a yellow warning (relatively low risk) indicating that there may be mild disaster. High temperatures and drought are expected, with some areas experiencing slight crop damage. An orange alert (higher risk) indicates a high probability of moderate high temperatures and drought, with some areas experiencing crop damage and a continued decline in soil moisture. A red alert (high risk) indicates a very high probability of severe high temperatures and drought, with widespread crop damage and deep soil drought. Additional key data, such as the duration of high temperatures and the trend of soil moisture decline, are also provided. The scope of impact is defined by combining regional administrative divisions and topography. Blue and yellow alerts are divided by township level, specifying the names of affected townships and terrain types. Orange and red alerts are divided by county level, indicating the core areas of affected counties and the estimated percentage of affected area. Defense recommendations are provided at different levels. The system adapts to different terrains, with plains areas focusing on irrigation planning and high-temperature protection, while hilly areas emphasize water conservation measures and vegetation protection. For example, a blue alert recommends daily monitoring of key indicators and maintaining regular irrigation frequency; a yellow alert recommends monitoring soil moisture and temperature every 6 hours and increasing irrigation in some areas; an orange alert recommends advance allocation of irrigation equipment, stockpiling of heatstroke prevention materials, and straw mulching for water conservation on hillsides; and a red alert recommends fully activating the irrigation emergency response, suspending outdoor high-temperature operations, and providing supplemental water protection for drought-stricken crops. After the alert information is generated, it is simultaneously released through the meteorological early warning platform, government notification channels, and township agricultural technology extension terminals. Subsequently, a collection period of 7 days is set to collect actual data through multiple channels. For disaster information, meteorological stations collect calibrated temperature, precipitation, and soil moisture data for the corresponding time period to supplement the element change curves during the disaster period; remote sensing monitoring platforms acquire vegetation coverage and surface temperature inversion data of the affected areas to capture the large-scale spread of the disaster; on-site manual reporting is completed by township agricultural technicians and disaster prevention monitors, recording the degree of crop drought, affected area, soil drought depth, and facility damage caused by high temperatures, while simultaneously taking on-site photos as evidence; after collection, information from multiple channels is cross-checked, duplicate information and contradictory data are eliminated, and a structured disaster information table is formed, marking the information source, collection time, corresponding warning level, and terrain attributes.

[0037] Step 551 involves matching and associating the actual disaster information with the corresponding comprehensive feature matrix to form new training samples labeled with disaster occurrence. Specifically, this includes: first, determining the core dimensions for matching and association to ensure spatiotemporal consistency and attribute correspondence. Matching dimensions are divided into three categories: time, space, and warning association. For time-based matching, the actual disaster is matched with the comprehensive feature matrix corresponding to the warning, based on the warning information release time, ensuring that the disaster information and the feature matrix belong to the same disaster evolution cycle and avoiding time misalignment. For spatial-based matching, based on the feature analysis area corresponding to the comprehensive feature matrix, the affected administrative divisions and terrain types in the actual disaster information are associated, ensuring that the feature matrix for plain areas only matches disasters in plain areas, and the same applies to hilly areas. Simultaneously, the spatial unit encoding is checked to ensure consistent coverage. For warning association matching, the warning level corresponding to the actual disaster information is matched with the warning level generated by the warning. The comprehensive feature matrix is ​​bound to form a correlation chain of warning level - feature matrix - actual disaster situation. Then, labels are marked, and the label type is determined according to the actual disaster situation information. If a combined high temperature and drought disaster actually occurs (marked as 1), the label suffix is ​​further subdivided according to the disaster intensity (mild, moderate, severe), corresponding to the actual disaster situation of orange and red warnings. If no disaster occurs (marked as 0), it corresponds to the actual disaster situation of blue and yellow warnings. After labeling, the key elements in the actual disaster situation information are associated with the features of each dimension of the comprehensive feature matrix one by one, and supplemented to the remarks column of the feature matrix to form a complete training sample of comprehensive feature matrix - actual disaster elements - disaster occurrence label. After the association is completed, the samples are checked for completeness to see if there are any problems such as missing features, incorrect labels, spatiotemporal misalignment. Samples that fail the verification are corrected, and those that cannot be corrected are removed, finally forming a valid new training sample.

[0038] Step 552: Monitor the cumulative number of new training samples. When it reaches a preset threshold, trigger the update process, merging the new training samples with the existing training samples to form an expanded training dataset. Use this expanded training dataset to retrain the machine learning classification model, achieving adaptive updates for the early warning system. Specifically, this includes: first, setting a sample accumulation threshold based on the historical training sample size, the frequency of regional disasters, and terrain complexity, while reserving adjustment space to ensure the threshold meets the model update requirements while avoiding insufficient samples leading to poor update performance; then, monitoring the cumulative number of new training samples in real time, counting the number of valid samples weekly, and automatically triggering the update process when the cumulative number reaches the preset threshold. After the update process starts, sample merging is performed, organizing the new training samples and the existing training samples according to time series, sorting them from morning to evening, while removing duplicate and abnormal samples. After merging, the total number of samples in the expanded training dataset is calculated and split into a new training set and a new validation set in a 7:3 ratio as described above, ensuring the splitting logic is consistent with the original training. Retraining is then carried out, using the classification model integrated from multiple decision trees. The system parameters were re-optimized. First, the number of decision trees was adjusted, while the depth of each decision tree was kept within 8 layers to avoid overfitting. Each decision tree was then retrained individually, with the comprehensive feature matrix of the new training set and corresponding labels input. Split nodes were selected based on feature discrimination, recursively splitting nodes until a preset depth was reached or the number of samples per node was less than 5. After all decision trees were trained, the performance was tested using a new validation set, and the prediction accuracy (number of correctly predicted samples divided by the total number of validation samples) was calculated. If the accuracy did not meet expectations, the number and depth of decision trees were adjusted, and retraining was performed. Simultaneously, pruning optimization was carried out, removing redundant split nodes to reduce system complexity and balance accuracy and efficiency. After retraining, the performance (accuracy, error rate) of the machine learning classification model before and after the update was compared. If the performance improved after the update, the new decision tree parameters were saved and replaced with the original parameters to complete the adaptive update. If the performance did not improve or even decreased, the original parameters were retained, and the reasons were analyzed. After optimization, retraining was performed. After the update, an update log was recorded, noting the number of new samples, parameter adjustments, and performance changes, achieving the adaptive capability of the early warning system to dynamically adapt to actual disaster situations.

[0039] Periodic updates triggered by a cumulative sample threshold avoid the waste of resources caused by frequent updates and ensure that the machine learning classification model can adapt to changes in the disaster situation in a timely manner. The optimized decision tree integration system can better fit the latest disaster characteristics, improve the adaptability of the early warning system to disaster changes in different terrains and time periods, and avoid early warning deviations caused by the evolution of the disaster situation.

[0040] like Figure 2 As shown, embodiments of the present invention also provide a monitoring and early warning system for combined high-temperature and drought disasters, including: The fusion module is used to collect multi-source monitoring data in real time, process the multi-source monitoring data, and generate a fused dataset. The calibration module is used to extract a subset of data from the fused dataset, construct a feature analysis region based on a preset reference relationship, select a core analysis unit as one focus of an ellipse within the feature analysis region, and select an external reference unit in the adjacent spatial range as the other focus of the ellipse. It calculates the eccentricity of the ellipse to quantify the spatial morphology and dynamic changes of the region. Based on the correlation between the core analysis unit and the external reference unit over time, it generates an analysis path and calculates environmental parameter correction coefficients to calibrate relevant elements in the fused dataset, thus obtaining the calibrated fused dataset. The feature module is used to construct a time-series dynamic analysis structure based on the calibrated fusion dataset, extract time-series dynamic feature vectors of key indicators of high temperature and drought, analyze the change trajectory of key indicators, calculate the torsional angle of evolution over time, quantify trend turning points and fluctuation characteristics, and fuse the time-series dynamic feature vectors and torsional angle features to form a comprehensive feature matrix. The early warning module is used to input the comprehensive feature matrix into a pre-trained machine learning classification model, calculate the probability of occurrence of the combined high temperature and drought disaster, and determine the early warning level. The update module is used to issue early warning information according to the warning level, collect actual disaster information from external feedback as new training data, and periodically retrain the machine learning classification model to achieve adaptive updates.

[0041] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0042] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning of combined high-temperature and drought disasters, characterized in that, The method includes: Step 1: Collect multi-source monitoring data in real time, process the multi-source monitoring data, and generate a fused dataset; Step 2: Extract a subset of data from the fused dataset, construct a feature analysis region based on a preset reference relationship, select a core analysis unit as one focus of an ellipse within the feature analysis region, and select an external reference unit in the adjacent spatial range as the other focus of the ellipse. Calculate the eccentricity of the ellipse to quantify the spatial morphology and dynamic changes of the region. Based on the correlation between the core analysis unit and the external reference unit over time, generate an analysis path, and calculate environmental parameter correction coefficients to calibrate the relevant elements in the fused dataset, obtaining the calibrated fused dataset. Step 3: Based on the calibrated fusion dataset, construct a time-series dynamic analysis structure, extract time-series dynamic feature vectors of key indicators of high temperature and drought; analyze the change trajectory of key indicators, calculate the torsion angle of evolution over time, quantify trend turning points and fluctuation characteristics, and fuse the time-series dynamic feature vectors and torsion angle features to form a comprehensive feature matrix. Step 4: Input the comprehensive feature matrix into the pre-trained machine learning classification model to calculate the probability of occurrence of the combined high temperature and drought disaster and determine the warning level; Step 5: Issue warning information according to the warning level, collect actual disaster information from external feedback as new training data, and periodically retrain the machine learning classification model to achieve adaptive updates.

2. The monitoring and early warning method for combined high-temperature and drought disasters according to claim 1, characterized in that, A subset of data is extracted from the fused dataset. A feature analysis region is constructed based on a predefined reference relationship. Within the feature analysis region, a core analysis unit is selected as one focus of an ellipse, and an external reference unit is selected in the adjacent spatial range as the other focus of the ellipse. The spatial morphology and dynamic changes of the eccentricity quantification region of the ellipse are calculated, including: Based on the geospatial distribution information in the fused dataset, target areas affected by high temperature or drought are selected, and these target areas are determined as feature analysis areas. Within the feature analysis area, based on the outliers of meteorological and environmental elements in the fused dataset, spatial units with prominent outliers are identified and determined as core analysis units. Based on the spatial correlation characteristics of the elements in the fused dataset, an external reference unit is selected within the vicinity of the core analysis unit, with the core analysis unit as the center. Using the core analysis unit and a selected external reference unit as the two foci of the ellipse, the eccentricity is calculated based on the geometric relationship between the distance between the two foci and the length of the major axis of the ellipse. Based on eccentricity, the spatial structure of high-temperature or arid elements in the region is quantitatively characterized and analyzed; and by analyzing the changes in eccentricity over a continuous time series, the evolution of spatial structure over time is dynamically described.

3. The method for monitoring and early warning of combined high-temperature and drought disasters according to claim 2, characterized in that, Based on the correlation between the core analysis unit and the external reference unit over time, an analysis path is generated, and environmental parameter correction coefficients are calculated to calibrate relevant elements in the fused dataset, resulting in a calibrated fused dataset, including: Based on the feature values ​​of the core analysis unit and the external reference unit in multiple time series, the correlation degree between the core analysis unit and the external reference unit at each time point is calculated; based on the correlation degree, a correlation relationship sequence describing the change of the relationship between the two over time is generated. Based on the correlation sequence, in the spatiotemporal coordinate system, points are plotted with time as the horizontal axis and correlation degree as the vertical axis, and the coordinate points corresponding to each time point are connected to generate an analysis path for characterizing the evolution trajectory of the correlation. The curvature, direction, and trend characteristics of the analysis path are extracted and fused with the spatial structure characteristics reflected by the eccentricity to generate environmental parameter correction coefficients. Based on the environmental parameter correction coefficient, the values ​​of key monitoring elements such as temperature, precipitation and soil moisture in the feature analysis area of ​​the fused dataset are calibrated in the corresponding time period to generate a calibrated fused dataset.

4. The monitoring and early warning method for combined high-temperature and drought disasters according to claim 3, characterized in that, Based on the calibrated fused dataset, a time-series dynamic analysis structure is constructed to extract time-series dynamic feature vectors of key indicators of high temperature and drought, including: From the calibrated fusion dataset, we selected the sets of key indicators for characterizing high-temperature events and the sets of key indicators for characterizing drought events, respectively. Based on key indicators of high temperature and drought, the observation values ​​of each key indicator within a preset time window are extracted from the calibrated fusion dataset and arranged in chronological order to generate a corresponding time series data for each key indicator. Based on time series data, the mean, amplitude, fluctuation frequency and stage trend slope of each time series data are calculated, and the statistical features corresponding to each time series are extracted and generated. The statistical features corresponding to all indicators in the set of key high-temperature indicators are combined in a preset order to generate a dynamic feature vector of high-temperature time series; the statistical features corresponding to all indicators in the set of key drought indicators are combined in a preset order to generate a dynamic feature vector of drought time series.

5. The method for monitoring and early warning of combined high-temperature and drought disasters according to claim 4, characterized in that, Analyzing the changing trajectories of key indicators, calculating the torsion angle over time, quantifying trend reversals and fluctuation characteristics, and fusing the time-series dynamic feature vector with the torsion angle characteristics to form a comprehensive feature matrix, including: Based on the indicator data in the time-series dynamic feature vector, select a continuous segment of the key indicator time series with a preset time length, and map the multiple indicator values ​​corresponding to each time point of the continuous segment of the key indicator time series to a coordinate point in a multi-dimensional space to form a spatial trajectory representing the change of the indicator. Based on the spatial trajectory, the vector difference between the coordinate points corresponding to adjacent time points is calculated to obtain the direction vector between adjacent time points; based on the direction vector between adjacent time points, the angle between every two adjacent direction vectors is calculated in turn to generate a sequence of torsion angles that change with time. Based on the torsion angle sequence, the maximum value of the torsion angle sequence is calculated as the maximum torsion angle, the arithmetic mean of the torsion angle sequence is calculated as the average torsion angle, and the number of times the torsion angle exceeds a preset threshold per unit time is counted as the torsion frequency. The time-series dynamic feature vector is fused with the maximum torsion angle, average torsion angle, and torsion frequency features to form a comprehensive feature matrix.

6. The method for monitoring and early warning of combined high-temperature and drought disasters according to claim 5, characterized in that, The comprehensive feature matrix includes original time-series statistical features and trend reversal and fluctuation information.

7. The method for monitoring and early warning of combined high-temperature and drought disasters according to claim 6, characterized in that, The comprehensive feature matrix is ​​input into a pre-trained machine learning classification model to calculate the probability of occurrence of combined high-temperature and drought disasters and determine the warning level, including: The comprehensive feature matrix is ​​input into a machine learning classification model to generate a probability value representing the likelihood of a combined high-temperature and drought disaster. The probability value is matched with multiple preset probability threshold ranges, and based on the matching results, the probability value is mapped to the corresponding warning level.

8. The method for monitoring and early warning of combined high-temperature and drought disasters according to claim 7, characterized in that, The warning levels include at least no warning, attention, warning, and emergency warning.

9. The method for monitoring and early warning of combined high-temperature and drought disasters according to claim 8, characterized in that, Warning information is issued based on the warning level, and actual disaster information from external feedback is collected as new training data to periodically retrain the machine learning classification model to achieve adaptive updates, including: Based on the warning level, early warning information including disaster intensity, scope of impact and defense recommendations is generated. Within a set period after the early warning information is issued, actual disaster information from meteorological stations, remote sensing monitoring platforms and on-site manual reports is collected. The actual disaster information is matched and associated with the corresponding comprehensive feature matrix to form new training samples with disaster occurrence labels; The system monitors the cumulative number of new training samples. When the number reaches a preset threshold, it triggers an update process, merging the new training samples with the existing training samples to form an expanded training dataset. The expanded training dataset is then used to retrain the machine learning classification model, enabling adaptive updates for early warnings.

10. A monitoring and early warning system for combined high-temperature and drought disasters, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The fusion module is used to collect multi-source monitoring data in real time, process the multi-source monitoring data, and generate a fused dataset. The calibration module is used to extract a subset of data from the fused dataset, construct a feature analysis region based on a preset reference relationship, select a core analysis unit as one focus of an ellipse within the feature analysis region, and select an external reference unit in the adjacent spatial range as the other focus of the ellipse. It calculates the eccentricity of the ellipse to quantify the spatial morphology and dynamic changes of the region. Based on the correlation between the core analysis unit and the external reference unit over time, it generates an analysis path and calculates environmental parameter correction coefficients to calibrate relevant elements in the fused dataset, thus obtaining the calibrated fused dataset. The feature module is used to construct a time-series dynamic analysis structure based on the calibrated fusion dataset, extract time-series dynamic feature vectors of key indicators of high temperature and drought, analyze the change trajectory of key indicators, calculate the torsional angle of evolution over time, quantify trend turning points and fluctuation characteristics, and fuse the time-series dynamic feature vectors and torsional angle features to form a comprehensive feature matrix. The early warning module is used to input the comprehensive feature matrix into a pre-trained machine learning classification model, calculate the probability of occurrence of the combined high temperature and drought disaster, and determine the early warning level. The update module is used to issue early warning information according to the warning level, collect actual disaster information from external feedback as new training data, and periodically retrain the machine learning classification model to achieve adaptive updates.

Images