A meteorological disaster risk early warning grading inspection method based on a square penalty matrix

Abstract

The application provides a meteorological disaster risk warning grading test method based on a square penalty matrix. The method comprises the following steps: constructing a warning-disaster sample set and dividing it into a calibration set and an independent test set; determining the normalization parameters and entropy weights of each disaster index in the calibration set, calculating a comprehensive disaster index, determining a disaster grade boundary threshold according to the true positive rate and false positive rate of a candidate threshold, dividing a four-grade disaster grade, and constructing a square penalty matrix based on the grade center value; in the independent test set, determining the disaster grade using the normalization parameters, index weights and grade boundary threshold, and looking up the penalty value in the square penalty matrix according to the warning grade and disaster grade, and calculating the grading test score. The application solves the problem that the meteorological disaster risk warning test lacks grade bias quantitative evaluation rules, and improves the objectivity, reproducibility and business applicability of the test results.

Description

Technical Field

[0001] This invention relates to the field of meteorological disaster risk early warning technology, and in particular to a method for classifying and verifying meteorological disaster risk early warning based on a squared penalty matrix. Background Technology

[0002] Meteorological disaster risk early warning products are crucial for supporting disaster prevention and mitigation, emergency dispatch, and risk control. Objectively evaluating these products helps assess the rationality of early warning level settings, identify types of early warning deviations, and provide a basis for subsequent early warning model optimization and operational rule adjustments.

[0003] Currently, existing meteorological disaster risk warning verification methods mostly focus on a binary comparison between "whether a warning has been issued" and "whether a disaster has occurred." A fundamental issue in conducting meteorological disaster risk warning classification verification is how to determine the disaster severity level. Directly setting disaster severity classification thresholds arbitrarily introduces a high degree of subjectivity; simply using quantile methods or clustering methods only reflects the distribution of disaster data itself, making it difficult to guarantee comparability between disaster severity levels and risk warning levels.

[0004] Furthermore, once the disaster severity level is determined, if only a matching method that states "the warning level must be completely consistent with the disaster severity level to be considered correct" is used, all levels will be treated as independent discrete categories, failing to reflect the orderliness, progressive relationship, and differences in the degree of deviation between levels. For example, the degree of deviation between the disaster severity level corresponding to a low-risk area and the disaster severity level corresponding to a low-risk area is different, and they should not be assigned the same evaluation result.

[0005] Therefore, there is an urgent need for a meteorological disaster risk warning classification verification method that can both objectively determine the disaster level and reflect the degree of level deviation. Summary of the Invention

[0006] This invention provides a method for classifying and verifying meteorological disaster risk warnings based on a squared penalty matrix, in order to solve the problem of the lack of quantitative evaluation rules for the deviation of meteorological disaster risk warnings, and to achieve effective classification and verification of meteorological disaster risk warning products.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] A method for classifying meteorological disaster risk warnings based on a squared penalty matrix includes:

[0009] Acquire meteorological disaster risk warning product data and disaster data, construct a warning-disaster sample set using meteorological disaster processes and administrative regions as basic sample units, and divide the warning-disaster sample set into a calibration set and an independent test set, with the independent test set being the data to be graded and tested;

[0010] Based on the disaster indexes of each sample in the calibration set, determine the normalization parameters and weights of each disaster index, and calculate the comprehensive disaster index of each sample.

[0011] The disaster severity level boundary threshold is determined based on the true positive rate (TPR) and false positive rate (FPR) of each candidate threshold on the calibration set. Based on the disaster severity level boundary threshold, the comprehensive disaster index of the samples in the calibration set is divided into no disaster, low disaster, medium disaster and high disaster. A squared penalty matrix is ​​constructed based on the center values ​​of the four disaster severity levels of no disaster, low disaster, medium disaster and high disaster.

[0012] For the samples of data to be graded in the independent test set, the comprehensive disaster index of each sample is calculated using the normalization parameters and weights determined by the calibration set. The disaster level of each sample is determined using the disaster level boundary threshold. The penalty value is retrieved from the squared penalty matrix according to the warning level and disaster level of each sample. The grading test score of the data to be graded is obtained according to the average penalty value of all samples.

[0013] Preferably, the acquisition of meteorological disaster risk early warning product data and disaster data, the construction of an early warning-disaster sample set, and its division into a calibration set and an independent test set, includes:

[0014] Acquire meteorological disaster risk warning product data and disaster data, and construct a warning-disaster sample set using meteorological disaster processes and administrative regions as basic sample units. Each sample in this warning-disaster sample set includes: warning information, disaster indicators, and matching information.

[0015] A sample correspondence relationship is established according to a unified spatial unit and time window. The spatial unit includes the business areas of cities, counties and townships, and the time window is determined according to the disaster process, the time of early warning release and the time of disaster occurrence.

[0016] The early warning-disaster sample set is divided into a calibration set and an independent test set. The calibration set is used to determine the normalization parameters and weights of each disaster indicator, the disaster level boundary threshold, the disaster level center value, and the squared penalty matrix. The independent test set is used to perform graded tests using the parameters, thresholds, and squared penalty matrix determined by the calibration set.

[0017] Preferably, the step of determining the normalization parameters and weights of each disaster index based on the disaster indexes of each sample in the calibration set, and calculating the comprehensive disaster index of each sample, includes:

[0018] Let the first one in the calibration set be... Each sample contains Disaster-related indicators, the original disaster indicators are as follows: The various disaster indicators were normalized.

[0019]

[0020] in, and They represent the first The minimum and maximum values ​​of disaster-related indicators in the calibration set;

[0021] The weights of each disaster index were determined using the entropy method. Calculate the comprehensive disaster index for each sample. :

[0022]

[0023] Comprehensive Disaster Index Continuous values ​​are used to represent samples. The overall severity of the disaster.

[0024] Preferably, the step of determining the disaster severity level threshold based on the true positive rate (TPR) and false positive rate (FPR) of each candidate threshold on the calibration set, and classifying the comprehensive disaster index of the samples in the calibration set into no disaster, low disaster, medium disaster, and high disaster based on the disaster severity level threshold, includes:

[0025] Assuming risk warning levels are coded as 0, 1, 2, and 3 according to severity, corresponding to none, low, medium, and high levels respectively, then the following three level boundaries can be constructed: Boundary This serves as the boundary between no level and levels below and above. The boundary between low and below levels and medium and above levels. It serves as the dividing line between the middle and lower levels and the high level;

[0026] Based on the comprehensive disaster index As the discrimination score, the true positive rate was calculated under different candidate thresholds. and false positive rate True positive rate Indicates at the candidate threshold Below, the proportion of positive samples correctly identified; false positive rate. Indicates at the candidate threshold Below, the proportion of negative samples that are incorrectly identified as positive samples;

[0027] For the There are several level boundaries, and the candidate threshold is set as follows: ,in: True positive rate and false positive rate Calculate according to the following formulas respectively:

[0028]

[0029]

[0030] in, This represents the number of true positive samples. Indicates the number of false negative samples. Indicates the number of false positive samples. This represents the number of true negative samples;

[0031] Calculate the difference corresponding to each candidate threshold. That is, the Kolmogorov-Smirnov statistic under the candidate threshold. , will make The candidate threshold that reaches the maximum value is determined as the optimal threshold for the corresponding disaster level boundary. ;

[0032]

[0033]

[0034] get , and Three disaster severity classification thresholds;

[0035] when The comprehensive disaster index will then be used. Classified as having no disaster situation;

[0036] when The comprehensive disaster index will then be used. Classified as low-risk;

[0037] when The comprehensive disaster index will then be used. Classified as a moderate disaster;

[0038] when The comprehensive disaster index will then be used. It is classified as a high-risk disaster area.

[0039] Preferably, the construction of the squared penalty matrix based on the central values ​​of four disaster levels—no disaster, low disaster, medium disaster, and high disaster—includes:

[0040] Based on the disaster severity level thresholds, the calibrated sample set is divided into four levels: no disaster, low disaster, medium disaster, and high disaster, and the center value of each level is determined. The central value of the disaster level is determined by the mean, median, quantile center or robust mean of the comprehensive disaster index of the samples within the corresponding disaster level.

[0041] The squared penalty matrix is ​​set as a 4×4 matrix, where the rows of the squared penalty matrix correspond to the actual disaster level and the columns correspond to the warning level, where the first column corresponds to the warning level. Line number The elements of a column are denoted as This indicates that when the actual disaster level is level 1... Level 1, warning level is No. The penalty value corresponding to the level;

[0042] The warning level and the disaster level are mapped to the center value of the four disaster levels according to the same severity order. The elements in the squared penalty matrix are determined by the following formula:

[0043]

[0044] in, , Indicates the actual level of disaster The corresponding level center value, Indicates the warning level The rank center values ​​are obtained by mapping the rank according to the same severity order. and These represent the maximum and minimum values ​​among the central values ​​of the four-level disaster severity levels, respectively.

[0045] The squared penalty matrix, once determined on the calibration set, remains fixed and is used for hierarchical testing of the independent test set.

[0046] Preferably, the method further includes: when the warning level is lower than the disaster level, it is determined to be an underestimation, and the full penalty is applied; when the warning level is higher than the disaster level, it is determined to be an overestimation, and the penalty value is reduced. Multiply by the overestimation weight ,when At that time, there was no distinction between underestimation and overestimation;

[0047] , like ;

[0048] , like

[0049] Where i represents the actual disaster level, and j represents the warning level. The aforementioned asymmetric weights... .

[0050] Preferably, the process of calculating the comprehensive disaster index for each sample of data to be graded in the independent test set using the normalization parameters and weights determined by the calibration set, determining the disaster level of each sample using the disaster level boundary threshold, retrieving the penalty value from the squared penalty matrix based on the warning level and disaster level of each sample, and obtaining the grading test score of the data to be graded based on the average penalty value of all samples includes:

[0051] The disaster index of each sample in the independent test set is normalized by using the normalization parameters of each disaster index determined by the calibration set, and the comprehensive disaster index is calculated by using the weights of each disaster index determined by the calibration set.

[0052] The disaster severity level of each sample of data to be graded in the independent test set is determined by using the disaster severity level boundary threshold determined by the calibration set;

[0053] Early warning level for samples of data to be graded in the independent test set and disaster level The corresponding penalty value is retrieved from the squared penalty matrix determined by the calibration set, and then the average penalty value of all samples is calculated. .

[0054]

[0055] This represents the total number of samples of data to be graded in the independent testing cluster that participate in the grading test.

[0056] Calculate the GradeScore of the data to be graded based on the average penalty value:

[0057] .

[0058] As can be seen from the technical solution provided by the present invention, the present invention first divides the early warning-disaster sample set into a calibration set and an independent test set. In the calibration set, a comprehensive disaster index is constructed based on multi-dimensional disaster indicators, and the correspondence between the risk warning level and the comprehensive disaster index is used to determine the disaster level boundary threshold, improving the correspondence between disaster level classification and risk warning level. Simultaneously, the degree of deviation between the warning level and the disaster level is quantified by a squared penalty matrix based on the level center distance, distinguishing between adjacent level deviations and severe cross-level deviations. Furthermore, asymmetric weights can be combined to characterize underestimation and overestimation, achieving a refined evaluation of early warning grading deviations. In the independent test set, the normalized parameters, indicator weights, disaster level boundary thresholds, and squared penalty matrix determined by the calibration set are used for grading tests, avoiding the involvement of independent test data in the determination of grading rules, thus improving the objectivity, reproducibility, and operational applicability of the test results.

[0059] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 The flowchart illustrates a meteorological disaster risk early warning classification verification method based on a squared penalty matrix, as provided in this embodiment of the invention.

[0062] Figure 2 This invention provides a flowchart for determining the thresholds for no, low, medium, and high disaster levels based on the ROC (receiver operating characteristic) curve, as an embodiment of the invention.

[0063] Figure 3 A flowchart of a method for constructing a squared penalty matrix provided in an embodiment of the present invention. Detailed Implementation

[0064] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0065] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0066] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0067] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0068] The flowchart of a meteorological disaster risk early warning classification verification method based on a squared penalty matrix provided in this embodiment of the invention is as follows: Figure 1 As shown, the processing steps include the following:

[0069] Step S10: Obtain meteorological disaster risk warning product data and disaster data, and construct a warning-disaster sample set using meteorological disaster processes and administrative regions as basic sample units.

[0070] Each sample should include at least the following information:

[0071] Warning information: type of meteorological disaster, warning issuance time, affected area, risk level (no risk, low risk, medium risk, high risk);

[0072] Disaster indicators: Disaster-affected area, time of occurrence or statistical time, direct economic losses, affected population, death toll, agricultural area affected, collapsed houses, and other disaster indicators;

[0073] Matching information includes: disaster event number, administrative division code, time window, disaster attribution information, etc.

[0074] The sample correspondence is established according to a unified spatial unit and time window. The spatial unit can be a city, county, township, or other business area; the time window can be determined according to the disaster process, the time of early warning issuance, and the time of disaster occurrence.

[0075] In cases where there are multiple early warning results for the same region and the same disaster event, the early warning result with the highest risk level and the closest time to the disaster occurrence, or the representative early warning result stipulated in the business regulations, can be used as the early warning result for that sample.

[0076] After constructing the early warning-disaster sample set, the sample set is divided into a calibration set and an independent test set. The calibration set is used to determine the normalization parameters and entropy weights of each disaster indicator, the disaster level boundary threshold, the central value of the four disaster levels, and to construct the squared penalty matrix. The independent test set is used to perform graded tests using the normalization parameters, indicator weights, disaster level boundary thresholds, and squared penalty matrix determined by the calibration set.

[0077] Step S20: Determine the normalization parameters and entropy weights of the disaster index based on the calibration set, and calculate the comprehensive disaster index.

[0078] Let the first Each sample contains Disaster-related indicators, the original disaster indicators are as follows: ,Should This includes direct economic losses, deaths, affected population, agricultural damage area, and collapsed houses.

[0079] First, the various disaster indicators were normalized:

[0080]

[0081] in, and They represent the first The minimum and maximum values ​​of disaster-related indicators in the calibration set.

[0082] Subsequently, the entropy method was used to determine the weights of each disaster index. And calculate the comprehensive disaster index. :

[0083]

[0084] Comprehensive Disaster Index Continuous values ​​are used to represent samples. The overall disaster severity index corresponds to the comprehensive severity of the disaster. The central value is the mean, median, quantile center, or robust mean of the comprehensive disaster index within the corresponding disaster level.

[0085] The comprehensive disaster index is obtained by normalizing and weighting at least two types of disaster indicators from direct economic losses, affected population, death population, agricultural disaster area and collapsed houses.

[0086] Step S30: Determine the disaster severity level threshold in the calibration set.

[0087] Based on the true positive rate of each candidate threshold on the threshold calibration set and false positive rate Determine the disaster severity level threshold and classify the comprehensive disaster index of the samples into no disaster, low disaster, medium disaster and high disaster based on the disaster severity level threshold.

[0088] Figure 2 A method based on the embodiments of the present invention is provided. A flowchart for determining the thresholds for no, low, medium, and high disaster levels using curves. (Example) Figure 2 As shown, in the calibration set, the level boundaries of the four-level risk warning levels are transformed into three ordered binary classification problems. The comprehensive disaster index obtained in step S20 is used as the discrimination score to calculate the risk levels under different candidate thresholds. and And select to make The candidate threshold that reaches the maximum value is taken as the optimal threshold for the corresponding disaster level boundary.

[0089] If risk warning levels are coded as 0, 1, 2, and 3 according to severity, corresponding to none, low, medium, and high levels respectively, then the following three level boundaries can be constructed:

[0090] boundary This represents the boundary between no risk level and low or higher risk levels. No-risk samples can be marked as 0, while low-risk, medium-risk, and high-risk samples can be marked as 1, used to determine the threshold between no disaster and low or higher risk levels.

[0091] boundary This defines the boundary between low-risk (no risk), medium-risk (low risk), and high-risk (medium risk). No-risk and low-risk samples can be marked as 0, while medium-risk and high-risk samples can be marked as 1, used to determine the threshold between low and medium disaster levels.

[0092] boundary This defines the boundaries between no, low, medium, and high risk levels. No-risk, low-risk, and medium-risk samples can be marked as 0, while high-risk samples are marked as 1, used to determine the threshold between medium and high risk levels.

[0093] The comprehensive disaster index of each sample in the calibration set is used as the discrimination score. For the first... There are several level boundaries, and the candidate threshold is set as follows: ,in: True positive rate and false positive rate Calculate according to the following formulas respectively:

[0094]

[0095]

[0096] in, This represents the number of true positive samples. Indicates the number of false negative samples. Indicates the number of false positive samples. This represents the number of true negative samples.

[0097] Furthermore, calculate the threshold corresponding to each candidate threshold. - The difference is used as the corresponding candidate threshold. value( (Kolmogorov-Smirnov statistics).

[0098]

[0099] The larger the value, the stronger the ability of the candidate threshold to distinguish between samples on both sides of the corresponding level boundary.

[0100] Therefore, it will make The candidate threshold that reaches the maximum value is determined as the optimal threshold for the corresponding disaster level boundary. .

[0101]

[0102] This leads to the three disaster severity level thresholds. , and .

[0103] Based on the three disaster severity level thresholds, the comprehensive disaster index of the calibrated sample is divided into:

[0104] < No disaster reported; ≤ DI < Low disaster level; ≤ < Disaster situation in China; ≥ High disaster situation

[0105] If all disaster indicators for some samples are 0 or there are no disaster records, their comprehensive disaster index can be set to 0, and they can be used as samples without disaster records for threshold calibration and independent testing. If passed... The thresholds obtained from the curves do not meet the requirements. < < The monotonicity of the fourth-level threshold can be ensured through sorting correction or ordered constraint optimization.

[0106] Step S40: Construct a squared penalty matrix based on the center value of the four disaster levels in the calibration set.

[0107] Based on the three disaster level thresholds determined in step S30, the calibrated sample is divided into four levels: no disaster, low disaster, medium disaster, and high disaster.

[0108] The central values ​​of the four disaster levels in the calibration set are calculated separately and used as representative values ​​of the severity of each disaster level.

[0109] The central values ​​of the four levels of disaster severity are respectively denoted as:

[0110] The central value of the level can be determined by the mean, median, quantile center or robust mean of the comprehensive disaster index of the samples within the corresponding disaster level.

[0111] Based on this, a squared penalty matrix is ​​constructed according to the distance between the central values ​​of the four disaster levels.

[0112] Figure 3 This is a flowchart for constructing a squared penalty matrix and calculating a graded test score, provided as an embodiment of the present invention.

[0113] The squared penalty matrix is ​​a 4×4 matrix, where the rows correspond to the actual disaster level and the columns correspond to the warning level.

[0114] The element in the i-th row and j-th column of the matrix is ​​denoted as .

[0115] This indicates that when the actual disaster level is level 1 Level 1, warning level is No. At level 1, the penalty value resulting from the inconsistency between the two is, where, .

[0116] Since both the warning level and the disaster level are coded as 0, 1, 2, and 3 from low to high severity, the warning level and the disaster level are mapped to the central value of the fourth-level disaster level according to the same severity sequence.

[0117] The elements in the squared penalty matrix are determined by the following formula:

[0118]

[0119] in, Indicates the actual level of disaster The corresponding level center value, Indicates the warning level The rank center values ​​are obtained by mapping the rank according to the same severity order. and These represent the maximum and minimum values ​​among the central values ​​of the four-level disaster severity levels, respectively.

[0120] pass The distance between the grade center values ​​is normalized so that the penalty value is between 0 and 1; the square operation is used to strengthen the penalty for distant grade deviations relative to near grade deviations.

[0121] When the warning level matches the disaster level: The corresponding penalty value is 0.

[0122] The greater the difference between the warning level and the disaster level, the greater the distance between their corresponding level center values, and the greater the penalty value.

[0123]

[0124] In an alternative implementation, underestimation / overestimation of asymmetric weights can also be introduced.

[0125] When the warning level is lower than the disaster level, it is considered an underestimation, and the full penalty is applied; when the warning level is higher than the disaster level, it is considered an overestimation, and the penalty value can be reduced. Multiply by an overestimation weight less than 1 This is to reflect the asymmetry between underestimating and overestimating early warning in terms of business risks, emergency costs, and social impact.

[0126] , like ;

[0127] , like

[0128] The asymmetric weights The tolerance level for underreporting and false reporting can be determined based on the business department's relative tolerance, or it can be determined by iterating through candidate parameters in a calibration sample. If it is not necessary to distinguish between underreporting and overreporting, then let... = 1.

[0129] Step S50: Calculate the graded test score using the calibration results in the independent test set.

[0130] The independent test set includes graded early warning data and ungraded disaster data.

[0131] For each sample of data to be graded in the independent test set, the warning level is obtained directly from the graded warning data, and the disaster level is determined according to the following process.

[0132] First, the disaster indicators of each sample in the independent test set are normalized using the normalization parameters of each disaster indicator determined in the calibration set in step S20.

[0133] For the k-th type of disaster index of the nth sample in the independent test set of data to be graded and tested... The calibration set is used to determine and Normalize:

[0134]

[0135] Subsequently, the disaster index weights determined by the calibration set were adopted. Calculate the comprehensive disaster index of the nth sample in the independent test set of data to be graded:

[0136]

[0137] Furthermore, the three disaster severity level thresholds determined in the calibration set in step S30 are used. , and The comprehensive disaster index of each sample is divided into four levels: no disaster, low disaster, medium disaster, and high disaster, thus obtaining the disaster level of each sample.

[0138] For the nth sample of data to be graded and tested, let its disaster level be i and its warning level be j.

[0139] Based on the disaster level i and the warning level j, the corresponding penalty value is retrieved from the squared penalty matrix constructed in step S40:

[0140]

[0141] If underestimation or overestimation of asymmetric weights are used, the penalty value after adjustment of the asymmetric weights is retrieved and calculated.

[0142] Let N be the total number of samples participating in the tiered test in the independent test set. Calculate the average penalty value for all samples:

[0143]

[0144] Further define the grading test score for the data to be graded. Subtract the average penalty value of all samples from 1:

[0145]

[0146] in, The closer the value is to 1, the more consistent the warning level is with the disaster level; The lower the value, the worse the early warning grading effect.

[0147] In addition to the graded inspection score, this invention can also output the following auxiliary indicators to enable business departments to analyze the effectiveness of early warning products from multiple perspectives:

[0148] Strict level accuracy: The proportion of samples where the warning level and the disaster level are completely consistent;

[0149] Adjacent level deviation rate: The proportion of samples where the warning level and the disaster level differ by one level;

[0150] Severe cross-level error rate: The proportion of samples where the warning level differs from the disaster level by two or more levels;

[0151] Underestimation rate: The proportion of samples where the warning level is lower than the disaster level;

[0152] Overestimation rate: The proportion of samples where the warning level is higher than the disaster level;

[0153] Traditional confusion matrix metrics such as Precision, Recall, F1, FAR, FNR, and CSI at various levels.

[0154] Example 1: Graded Inspection of Meteorological Disaster Risk Early Warning Products

[0155] Taking a certain type of meteorological disaster risk warning product as an example, the risk warning product is divided into no risk, low risk, medium risk and high risk, and the disaster data includes indicators such as direct economic losses, affected population, agricultural disaster area, collapsed houses and deaths.

[0156] The implementation steps are as follows:

[0157] Using "meteorological disaster process - administrative region" as the basic sample unit, a set of early warning-disaster samples corresponding to risk warning levels and disaster indicators is constructed.

[0158] The aforementioned early warning-disaster sample set is divided into a calibration set and an independent test set.

[0159] In the calibration set, the minimum and maximum values ​​of each disaster index are determined as disaster index normalization parameters; the disaster indexes in the calibration set are normalized using the normalization parameters, and the weight of each disaster index is determined using the entropy method, and the comprehensive disaster index of each sample in the calibration set is calculated.

[0160] Three ROC binary classification problems are constructed in the calibration set, corresponding to the boundary between no risk and low or above risk, the boundary between low or below risk and medium or above risk, and the boundary between medium or below risk and high risk, respectively.

[0161] For each level boundary, calculate the true positive rate, false positive rate, and KS value corresponding to different candidate thresholds. The candidate threshold that maximizes the KS value is determined as the corresponding disaster level boundary threshold. , and .

[0162] Based on the three disaster level thresholds, the calibrated sample set is divided into no disaster, low disaster, medium disaster, and high disaster.

[0163] Calculate the median of the comprehensive disaster index for the four disaster levels in the calibration set to obtain the central values ​​for the four disaster levels. , and .

[0164] A 4×4 squared penalty matrix is ​​constructed based on the central values ​​of the four disaster levels and the squared penalty formula.

[0165] In the independent test set, the disaster index is normalized using the disaster index normalization parameters determined by the calibration set, and the comprehensive disaster index of each sample is calculated using the entropy weights determined by the calibration set.

[0166] Determined using the calibration set , and The samples in the independent test set were divided into four categories: no disaster, low disaster, medium disaster, and high disaster.

[0167] Based on the warning level and disaster level of each sample in the independent test set, the corresponding penalty value is directly retrieved from the squared penalty matrix determined in the calibration set.

[0168] Calculate the average penalty value and graded test score for all samples in the independent test set, and output auxiliary indicators such as the four-level confusion matrix, the accuracy of the strict grade, the adjacent grade bias rate, the serious cross-grade error rate, the underestimation rate, and the overestimation rate.

[0169] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0170] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0171] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0172] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for classifying meteorological disaster risk early warning based on a squared penalty matrix, characterized in that, include: Acquire meteorological disaster risk warning product data and disaster data, construct a warning-disaster sample set using meteorological disaster processes and administrative regions as basic sample units, and divide the warning-disaster sample set into a calibration set and an independent test set, with the independent test set being the data to be graded and tested; Based on the disaster indexes of each sample in the calibration set, determine the normalization parameters and weights of each disaster index, and calculate the comprehensive disaster index of each sample. The disaster severity level boundary threshold is determined based on the true positive rate (TPR) and false positive rate (FPR) of each candidate threshold on the calibration set. Based on the disaster severity level boundary threshold, the comprehensive disaster index of the samples in the calibration set is divided into no disaster, low disaster, medium disaster and high disaster. A squared penalty matrix is ​​constructed based on the center values ​​of the four disaster severity levels of no disaster, low disaster, medium disaster and high disaster. For the samples of data to be graded in the independent test set, the comprehensive disaster index of each sample is calculated using the normalization parameters and weights determined by the calibration set. The disaster level of each sample is determined using the disaster level boundary threshold. The penalty value is retrieved from the squared penalty matrix according to the warning level and disaster level of each sample. The grading test score of the data to be graded is obtained according to the average penalty value of all samples.

2. The method according to claim 1, characterized in that, The process of acquiring meteorological disaster risk early warning product data and disaster data, constructing an early warning-disaster sample set, and dividing it into a calibration set and an independent test set includes: Acquire meteorological disaster risk warning product data and disaster data, and construct a warning-disaster sample set using meteorological disaster processes and administrative regions as basic sample units. Each sample in this warning-disaster sample set includes: warning information, disaster indicators, and matching information. A sample correspondence relationship is established according to a unified spatial unit and time window. The spatial unit includes the business areas of cities, counties and townships, and the time window is determined according to the disaster process, the time of early warning release and the time of disaster occurrence. The early warning-disaster sample set is divided into a calibration set and an independent test set. The calibration set is used to determine the normalization parameters and weights of each disaster indicator, the disaster level boundary threshold, the disaster level center value, and the squared penalty matrix. The independent test set is used to perform graded tests using the parameters, thresholds, and squared penalty matrix determined by the calibration set.

3. The method according to claim 1, characterized in that, The process of determining the normalization parameters and weights of each disaster index based on the disaster indexes of each sample in the calibration set, and calculating the comprehensive disaster index for each sample, includes: Let the first one in the calibration set be... Each sample contains Disaster-related indicators, the original disaster indicators are as follows: The various disaster indicators were normalized. in, and They represent the first The minimum and maximum values ​​of disaster-related indicators in the calibration set; The weights of each disaster index were determined using the entropy method. Calculate the comprehensive disaster index for each sample. : Comprehensive Disaster Index Continuous values ​​are used to represent samples. The overall severity of the disaster.

4. The method according to claim 3, characterized in that, The process of determining the disaster severity level threshold based on the true positive rate (TPR) and false positive rate (FPR) of each candidate threshold on the calibration set, and classifying the comprehensive disaster index of the samples in the calibration set into no disaster, low disaster, medium disaster, and high disaster based on the disaster severity level threshold, includes: Assuming risk warning levels are coded as 0, 1, 2, and 3 according to severity, corresponding to none, low, medium, and high levels respectively, then the following three level boundaries can be constructed: Boundary This serves as the boundary between no level and levels below and above. The boundary between low and below levels and medium and above levels. It serves as the dividing line between the middle and lower levels and the high level; Based on the comprehensive disaster index As the discrimination score, the true positive rate was calculated under different candidate thresholds. and false positive rate True positive rate Indicates at the candidate threshold Below, the proportion of positive samples correctly identified; false positive rate. Indicates at the candidate threshold Below, the proportion of negative samples that are incorrectly identified as positive samples; For the There are several level boundaries, and the candidate threshold is set as follows: ,in: True positive rate and false positive rate Calculate according to the following formulas respectively: in, This represents the number of true positive samples. Indicates the number of false negative samples. Indicates the number of false positive samples. This represents the number of true negative samples; Calculate the difference corresponding to each candidate threshold. That is, the Kolmogorov-Smirnov statistic under the candidate threshold. , will make The candidate threshold that reaches the maximum value is determined as the optimal threshold for the corresponding disaster level boundary. ; get , and Three disaster severity classification thresholds; when The comprehensive disaster index will then be used. Classified as having no disaster situation; when The comprehensive disaster index will then be used. Classified as low-risk; when The comprehensive disaster index will then be used. Classified as a moderate disaster; when The comprehensive disaster index will then be used. It is classified as a high-risk disaster area.

5. The method according to claim 4, characterized in that, The method of constructing a squared penalty matrix based on the central values ​​of four disaster levels (no disaster, low disaster, medium disaster, and high disaster) includes: Based on the disaster severity level thresholds, the calibrated sample set is divided into four levels: no disaster, low disaster, medium disaster, and high disaster, and the center value of each level is determined. The central value of the disaster level is determined by the mean, median, quantile center or robust mean of the comprehensive disaster index of the samples within the corresponding disaster level. The squared penalty matrix is ​​set as a 4×4 matrix, where the rows of the squared penalty matrix correspond to the actual disaster level and the columns correspond to the warning level, where the first column corresponds to the warning level. Line 1 The elements of a column are denoted as This indicates that when the actual disaster level is level 1... Level 1, warning level is No. The penalty value corresponding to the level; The warning level and the disaster level are mapped to the center value of the four disaster levels according to the same severity order. The elements in the squared penalty matrix are determined by the following formula: in, , Indicates the actual level of disaster The corresponding level center value, Indicates the warning level The rank center values ​​are obtained by mapping the rank according to the same severity order. and These represent the maximum and minimum values ​​among the central values ​​of the four-level disaster severity levels; The squared penalty matrix, once determined on the calibration set, remains fixed and is used for hierarchical testing of the independent test set.

6. The method according to claim 5, characterized in that, The method further includes: when the warning level is lower than the disaster level, it is judged as an underestimation and the full penalty is applied; when the warning level is higher than the disaster level, it is judged as an overestimation and the penalty value is reduced. Multiply by the overestimation weight ,when At that time, there was no distinction between underestimation and overestimation; , like ; , like ; Where i represents the actual disaster level, j represents the warning level, and the asymmetric weights... .

7. The method according to any one of claims 1 to 6, characterized in that, For the samples of data to be graded in the independent test set, the comprehensive disaster index of each sample is calculated using the normalization parameters and weights determined by the calibration set. The disaster level threshold is used to determine the disaster level of each sample. The penalty value is retrieved from the squared penalty matrix according to the warning level and disaster level of each sample. The grading test score of the data to be graded is obtained based on the average penalty value of all samples, including: The disaster index of each sample in the independent test set is normalized by using the normalization parameters of each disaster index determined by the calibration set, and the comprehensive disaster index is calculated by using the weights of each disaster index determined by the calibration set. The disaster severity level of each sample of data to be graded in the independent test set is determined by using the disaster severity level boundary threshold determined by the calibration set; Early warning level for samples of data to be graded in the independent test set and disaster level The corresponding penalty value is retrieved from the squared penalty matrix determined by the calibration set, and then the average penalty value of all samples is calculated. ; This represents the total number of samples of data to be graded in the independent testing cluster that participate in the grading test. Calculate the GradeScore of the data to be graded based on the average penalty value: 。

Images