Tropical Cyclone Damage Assessment using a Projection Pursuit Dynamic Cluster Model


 Using data from 62 tropical cyclones (TCs) that landed in Guangdong Province in China between 2000 and 2019, we calculated six indices—minimum central pressure, maximum wind speed, maximum rainstorm ratio, cumulative surface rainfall, cyclone track length and lifetime—and constructed a projection pursuit dynamic cluster (PPDC) model to assess TC damage risk. Although a single index may provide correct information on the intensity of certain types of damage, a comprehensive damage risk assessment cannot be obtained from individual indices alone. The PPDC model is a stable tool for TC damage risk assessment, especially in terms of economic loss, agricultural disaster area and disaster-affected population. Model validation improved the correlation of each of the indices. Output from the PPDC model for disaster-affected population and agricultural disaster-affected area also improved after model validation. We examined the limitations of the single indices using data from three TCs. Output from the PPDC model can closely reflect the intensity of the damage caused by the cyclones. Projection pursuit dynamic clustering is a new and objective method for typhoon damage risk assessment, and provides the scientific basis to support disaster prevention and mitigation.

constructed a projection pursuit dynamic cluster (PPDC) model to assess TC damage risk. Although  Many damagerisk assessments focus on TC intensity, which is quantified by minimum sea 45 level pressure or maximum surface wind speed and is strongly correlated with TC destructiveness.  Zong and Chen (1999)  proposed in the 1970s; it is a set of data analytic techniques used to identify structures in large 80 multivariate datasets, and has been widely used in hydrology and meteorology. Wang and Ni (2008) 81 combined dynamic cluster with projection pursuit, and developed a projection pursuit dynamic 82 cluster (PPDC) model; their results show that the PPDC model is a powerful multi-factor cluster 83 analysis tool, which can be applied to the regional division of water resources in China. This model 84 has a number of advantages: it does not require the input of artificial parameters, it is stable and 85 easy to operate, and dynamic cluster produces objective and clear results.

86
Damagerisk assessments provide the scientific basis for disaster prevention. In this study, we 87 developed a landfalling TC classification index system using PPDC, TC intensity, track, duration  defined as the ratio between daily rainfall that is greater than or equal to 50 mm and daily rainfall 102 that is greater than or equal to 10 mm. The maximum rainstorm ratio is defined as the maximum 103 rainstorm ratio during a TC. During TCs, light rainfall is relatively rare. Therefore, we excluded 104 daily rainfall that is less than 10 mm from our analysis.  To retain the maximum amount of information and characteristics contained in the data, we 117 analyzed the data from different angles to find the optimal projection direction. Projecting 118 high-dimensional data on to a low-dimensional space allows for conventional and intuitive image 119 analyses. Let the j th index of the i th sample be 0 ij where n is sample 120 size, and m is the number of indicators selected), and the PPDC model was constructed as 121 follows:

122
Step 1: Data standardization 123 The data were standardized to minimize differences in magnitude. For a large evaluation index,

124
Eq. 1 was used: For a small index, Eq. 2 was used: Step 2: Linear projection , then the ij x projection eigenvalue i z can be expressed as: Step 3: Construction of projection indicators 136 To identify the optimal projection direction, high-dimensional data projections in sequences of all samples. The set was dynamically clustered into p ( p n  ) classes as follows:

139
(1) Randomly select p points and use them as p convergent nuclei, which are denoted as

141
(2) Divide the points in Ω by 0 L ; the points are classified into p categories. The results are 142 denoted as 0

147
(4) Repeat the steps above to obtain a classification result sequence ( , ) error range is sufficiently small, and the algorithm is terminated. This algorithm is convergent.

150
The set of projected eigenvalues of all samples belonging to the h th category is represented by ; the absolute value of the distance between any two projected eigenvalues is represents the proximity of the samples within the class; between maximizing inter-class sample distances and minimizing intra-class samples distances.

164
Step 4: Model optimization 165 We applied the immune evolutionary algorithm ( Tables S1 and S2.

191
Using the PPDC model, the data were divided into two clusters using the following settings: m=6, which is higher than that in Fig. 1 (0.437). For direct economic loss, the average correlation 270 coefficient is 0.413, which is lower than that in Fig. 1 (0.454). The PPDC model can adequately parameters is 0.220, which is lower than the corresponding value in Fig. 1. In Fig. 2   hPa, which is 13 hPa lower than the sample mean; the maximum rainstorm ratio was 0.2643, which 305 is 130.84% that of the sample mean; the surface rainfall was 3.6×10 5 mm, which is equivalent to the 306 sample mean; track length was 1,688 km, which is 89.44% that of the sample mean; lifetime was 84 307 hours, which is 17 hours below than the sample mean. Khanun caused direct economic losses of minimum central pressure was 905 hPa, which is lower than the sample mean by 58.0 hPa; 320 maximum rainstorm ratio was 0.2645, which is higher than the sample mean; cumulative surface 321 rainfall was 1.4×10 6 mm, which is 281% that of the sample mean; track length was 6,549 km, 322 which is 247% that of the sample mean; lifetime was 234 hours, which is 124% that of the sample cumulative rainfall, maximum rainstorm ratio, and track length overestimate disaster intensity.

329
Output from the PPDC model is 0.8251 (Fig. 6), which is 124% that of the sample mean, which 330 indicates a very strong TC and closely reflects the intensity of the damage caused by Mangkhut.         Average landfalling TC classi cation indices for three TCs.

Figure 6
Average output from the projection pursuit dynamic cluster (PPDC) model for three TCs.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.