Development of a Storm Surge Prediction Model Using Typhoon Characteristics and Multiple Linear Regression

Abstract

Storm surges pose a significant threat to coastal regions worldwide, particularly as sea levels continue to rise due to climate change. This study aims to develop a storm surge height prediction model for the southeastern coast of Korea using a multiple linear regression (MLR) approach. Typhoon characteristics, including location and intensity derived from best-track data, were used as independent variables, while observed storm surge heights served as the dependent variable. The model’s predictive performance was assessed using the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Squared Error (MSE) and the coefficient of determination (R²). To enhance model accuracy and interpretability, a threshold-based model configuration strategy was implemented by categorizing data according to (1) the distance between the typhoon center and the observation point, and (2) the magnitude of the observed storm surge height. The results indicate that restricting typhoon events to within 900–1000 km of the observation site and segmenting surge heights into low and high ranges significantly improves predictive skill, especially for extreme surge events. For example, at Masan station, the model achieved an R² of 0.82 for high storm surge height (>0.2 m), and Gwangyang station showed an R² of 0.57 at a 500 km distance threshold, demonstrating substantial skill in predicting extreme surges. However, limitations remain in capturing the variability of lower-magnitude surges, suggesting the need for future research incorporating nonlinear and ensemble methods. This study provides a foundation for improving coastal hazard prediction and contributes to the development of more effective early warning systems and risk management strategies.

1. Introduction

Sea level rise exacerbates the risk of coastal disasters by amplifying flooding, inundation and shoreline erosion [1]. From a coastal engineering perspective, sea level can be broadly decomposed into several components: the mean sea level, the astronomical tide, the meteorological tide (i.e., storm surge), and other residual effects [2,3,4,5,6]. Among these, storm surges have emerged as a major threat to coastal regions worldwide due to their capacity to cause extensive inundation and damage [7,8,9,10,11,12,13,14,15]. In addition, the risk of coastal inundation from storm surges has been observed to increase with the ongoing rise in sea levels attributable to climate change [16,17,18,19].

Storm surges are abnormal rises in sea level primarily driven by meteorological factors—particularly low atmospheric pressure and strong winds [2,3,4,5,6]. Its research is often conducted in relation to tropical cyclones, which are accompanied by both low atmospheric pressure and strong winds [15,18,19]. In general, two main approaches have been adopted for predicting storm surge: numerical models and statistical models. While numerical hydrodynamic models offer detailed physical representation, they are computationally expensive and time-consuming (e.g., [16,17,18,19]). On the other hand, statistical models—particularly regression-based approaches—enable rapid predictions and are more suitable for operational forecasting (e.g., [20,21,22,23,24,25]). Given the importance of speed and reliability in early warning systems for coastal hazards, statistical models provide a practical alternative for storm surge prediction.

To date, various statistical methods, including machine learning techniques, have been applied to the development of storm surge height prediction models [26,27,28,29,30,31,32,33,34,35]. However, Multiple Linear Regression (MLR) remains the most widely used approach, demonstrating strengths in predictive accuracy, interpretability, and computational efficiency. Roberts et al. [26,27] developed an MLR-based storm surge prediction model for the New York-New Jersey coastal region. They used meteorological inputs from the NARR and CFSR reanalysis datasets as independent variables and tide gauge observations as the dependent variable. Their model achieved a high level of accuracy—within 0.1 m of the observed peak water level—for extreme events such as Hurricane Sandy (2012), performing comparably to the physics-based numerical model, NYHOPS (New York Harbor Observing and Prediction System). Notably, due to its low computational cost, the MLR model could be used to operate dozens of ensemble forecasts, making it advantageous for both real-time and long-term scenario-based predictions.

However, some studies utilizing MLR have tended to focus primarily on estimating optimal regression coefficients and achieving predictive performance, without applying systematic data partitioning methods or establishing a clear model development strategy. This tendency is particularly evident in studies that use synthetic typhoon scenarios or data derived from numerical models. Al Kajbaf and Bensi [31] developed an MLR-based surrogate model for storm surge prediction along the U.S. East Coast, using thousands of synthetic typhoon scenarios generated by the ADCIRC (Advanced Circulation Model for Oceanic, Coastal and Estuarine Waters). The input variables included central pressure deficit, radius to maximum winds, forward speed, and track direction of the typhoon, while the output variable was the peak storm surge height. Although the model enabled rapid surge estimation through a simple linear regression formula—without requiring complex hydrodynamic simulations—the absence of a well-defined data splitting and modeling strategy may lead to overfitting, reduced model robustness, and limited applicability to real-world storm events (e.g., [36]).

In recent study, numerical model-based datasets such as ERA5 (European Centre for Medium-Range Weather Forecasts Reanalysis v5), GTSR (Global Tide and Surge Reanalysis), and ADCIRC have been increasingly used to overcome the limitations of observational data, including spatial sparsity and the lack of extreme event records. For example, Tadesse and Wahl [35] developed a global storm surge prediction model using ERA-Interim, satellite data, and GTSR—a reanalysis dataset based on numerical modeling—for coastal regions worldwide. They applied and compared several statistical methods, including MLR, KNN (K-Nearest Neighbors) and Random Forest, and found that atmospheric pressure—particularly lagged pressure—was the most significant predictor in more than 70% of the study areas. While their study demonstrated the feasibility of global-scale surge prediction using only model-based data, the direct applications of such an approach to coastal disaster risk planning at the national level may be limited. This is because the performance of the global model represents an average over broad spatial domains, which makes it difficult to optimize the underlying numerical models for every individual coastline. As a result, model-driven errors are inevitable (e.g., [15]), and certain regions within a given country may not be suitable for the application of such generalized methods.

In addition, compared to the extensive body of research conducted in the United States and Europe, studies focusing on the Korean coastline remain extremely limited. Choo et al. [37] applied logistic regression and MLR techniques to assess sea level anomalies at three sites along the southeastern coast of Korea—Busan, Geoje, and Gadeokdo. Using meteorological and oceanographic observation variables as inputs, they developed a statistical model to predict anomalous sea level events. Their results showed that the MLR-based model improved predictive performance by approximately 4.9% in terms of the coefficient of determination (R²) compared to an existing empirical approach. Their study is one of the few cases in which MLR has been applied to Korean coastal area, highlighting both the potential for region-specific statistical model development and the relative scarcity of such research in the region.

Taken together, these previous studies underscore the need for storm surge prediction models that not only utilize reliable statistical techniques such as multiple linear regression (MLR), but also incorporate physically meaningful variables, observational data, and region-specific characteristic. Addressing these gaps, the present study aims to develop a storm surge height prediction model tailored to the southeastern coast of Korea, using multiple linear regression technique.

Figure 1 illustrates the overall workflow of this study. The storm surge height prediction model was developed using typhoon characteristics such as typhoon location and intensity, extracted from best-track data as independent variables, while observed storm surge heights were used as the dependent variable. The model’s predictive performance was evaluated using the Root Mean Square Error (RMSE) and the coefficient of determination (R²). To enhance model reliability and interpretability, this study implemented a structures model configuration strategy that distinguishes it from previous studies. Specifically, models were separately developed based on threshold criteria, including (1) the distance between the typhoon center and the observation point, and (2) the magnitude of the observed storm surge height. This approach contributes to both improved model performance and structural simplicity while addressing key limitations of prior research. In general, higher storm surges occur at locations situated to the right side of the typhoon track. However, because this study developed an MLR model applicable to 11 stations, the relative position between the typhoon and each observation station varies for the same typhoon. Therefore, the variation in storm surge height according to the relative position was not considered in this study.

2. Materials and Methods

2.1. Research Area

As shown in Figure 2a, the southeastern coast of the Korea Peninsula (KP) was selected as the research site. This region has frequently experienced typhoon-related damages associated with high storm surges in the past [15]. The southeastern coastline is characterized by complex coastal topography and is relatively less affected by astronomical tides and wind waves compared to the western and eastern coasts of the KP. In contrast, the southwestern coast exhibits a large tidal range, necessitating an analysis of the nonlinear interactions between tide and storm surges. The eastern coast of the KP, adjacent to the deep waters of the East Sea, is subject to significant wave transformations, requiring a detailed examination of the nonlinear interactions between wind waves and storm surges.

Eleven locations along the southeastern coast of the KP were selected as the points of interest for the development of storm surge prediction models using the multiple linear regression (MLR) technique. The selected locations include Geomundo, Goheung, Yeosu, Gwangyang, Tongyeong, Masan, Geojedo, Gadeokdo, Busan, Ulsan and Pohang. Figure 2b presents the locations of tide gauge stations installed within the area of interest for this study, and

Table 1. The geographic coordinates of these points of interest.

Point Name	Longitude	Latitude
Gadeokdo	128°48′39″ E	35°01′27″ N
Geomundo	127°18′32″ E	34°01′42″ N
Geojedo	128°41′57″ E	34°48′05″ N
Goheung	127°20′34″ E	34°28′52″ N
Gwangyang	127°45′17″ E	34°54′13″ N
Masan	128°35′20″ E	35°12′36″ N
Busan	129°02′07″ E	35°05′47″ N
Yeosu	129°23′14″ E	35°30′07″ N
Ulsan	127°45′57″ E	34°44′50″ N
Tongyeong	128°26′05″ E	34°49′40″ N
Pohang	129°23′02″ E	36°02′50″ N

provides their detailed coordinates.

2.2. Data

2.2.1. Independent Variable (Predictors)

In this study, typhoons that passed through the region spanning from 32° N to 40° N and from 122° E to 132° E over the period from 1979 to 2020, as shown in Figure 3, were defined as typhoons that affected the KP. A total of 155 typhoons were considered in this study, and their key characteristics are presented in

Table A1. Lists of Typhoons that affected Korea and were defined as such based on their passage through the region from 32° N to 40° N and from 122° E to 132° E.

No.	Typhoon Name	Pmin	Umax	Typhoon Lifetime
1	IRVING	958	75	1979	8	7	-	1979	8	20
2	JUDY	980	50	1979	8	15	-	1979	8	27
3	KEN	991	43	1979	8	30	-	1979	9	10
4	IDA	996	NaN	1980	7	5	-	1980	7	15
5	NORRIS	1002	NaN	1980	8	23	-	1980	8	31
6	ORCHID	967	70	1980	9	1	-	1980	9	16
7	IKE	1006	NaN	1981	6	7	-	1981	6	17
8	JUNE	990	45	1981	6	15	-	1981	6	26
9	OGDEN	983	NaN	1981	7	26	-	1981	8	1
10	AGNES	970	55	1981	8	25	-	1981	9	6
11	CLARA	1004	NaN	1981	9	13	-	1981	10	2
12	CECIL	975	55	1982	8	1	-	1982	8	19
13	ELLIS	955	70	1982	8	17	-	1982	9	4
14	FORREST	968	70	1983	9	16	-	1983	9	30
15	ALEX	1004	NaN	1984	6	28	-	1984	7	6
16	HOLLY	965	70	1984	8	12	-	1984	8	23
17	GERALD	1002	NaN	1984	8	14	-	1984	8	24
18	JUNE	1002	NaN	1984	8	25	-	1984	9	3
19	HAL	996	NaN	1985	6	11	-	1985	6	28
20	JEFF	992	45	1985	7	18	-	1985	8	3
21	KIT	970	70	1985	7	30	-	1985	8	17
22	LEE	980	60	1985	8	8	-	1985	8	16
23	ODESSA	985	55	1985	8	19	-	1985	9	2
24	PAT	965	70	1985	8	24	-	1985	9	2
25	BRENDAN	980	70	1985	9	25	-	1985	10	8
26	NANCY	994	45	1986	6	18	-	1986	6	27
27	VERA	960	70	1986	8	13	-	1986	9	2
28	ABBY	996	NaN	1986	9	9	-	1986	9	24
29	THELMA	960	78	1987	7	6	-	1987	7	18
30	ALEX	994	NaN	1987	7	21	-	1987	8	2
31	DINAH	940	85	1987	8	19	-	1987	9	3
32	ELLIS	990	40	1989	6	18	-	1989	6	25
33	JUDY	970	65	1989	7	20	-	1989	7	29
34	VERA	1002	NaN	1989	9	11	-	1989	9	19
35	OFELIA	996	NaN	1990	6	15	-	1990	6	26
36	ROBYN	992	40	1990	6	29	-	1990	7	14
37	ABE	996	NaN	1990	8	22	-	1990	9	3
38	CAITLIN	945	80	1991	7	18	-	1991	7	30
39	GLADYS	975	50	1991	8	13	-	1991	8	24
40	UNNAMED	994	35	1991	8	21	-	1991	8	31
41	KINNA	965	70	1991	9	8	-	1991	9	16
42	MIREILLE	935	95	1991	9	13	-	1991	10	1
43	JANIS	965	70	1992	7	30	-	1992	8	13
44	IRVING	994	40	1992	7	30	-	1992	8	5
45	KENT	980	50	1992	8	3	-	1992	8	20
46	POLLY	1000	NaN	1992	8	23	-	1992	9	4
47	TED	992	45	1992	9	14	-	1992	9	27
48	OFELIA	990	40	1993	7	24	-	1993	7	29
49	PERCY	980	55	1993	7	25	-	1993	8	1
50	ROBYN	945	85	1993	7	30	-	1993	8	14
51	YANCY	955	75	1993	8	27	-	1993	9	7
52	RUSS	1004	NaN	1994	6	2	-	1994	6	12
53	WALT	992	40	1994	7	11	-	1994	7	28
54	BRENDAN	992	45	1994	7	25	-	1994	8	3
55	DOUG	985	48	1994	7	30	-	1994	8	13
56	ELLIE	970	65	1994	8	3	-	1994	8	19
57	FRED	1004	NaN	1994	8	12	-	1994	8	26
58	SETH	975	55	1994	9	30	-	1994	10	16
59	FAYE	950	75	1995	7	12	-	1995	7	25
60	JANIS	990	NaN	1995	8	17	-	1995	8	30
61	RYAN	985	60	1995	9	14	-	1995	9	25
62	EVE	980	60	1996	7	10	-	1996	7	27
63	KIRK	960	75	1996	7	28	-	1996	8	18
64	PETER	975	60	1997	6	15	-	1997	7	4
65	TINA	975	60	1997	7	21	-	1997	8	10
66	OLIWA	970	65	1997	8	28	-	1997	9	19
67	YANNI	975	55	1998	9	24	-	1998	10	2
68	NEIL	980	50	1999	7	22	-	1999	7	28
69	OLGA	975	60	1999	7	26	-	1999	8	5
70	PAUL	992	35	1999	7	31	-	1999	8	9
71	RACHEL	1000	NaN	1999	8	5	-	1999	8	11
72	SAM	1004	NaN	1999	8	17	-	1999	8	27
73	WENDY	1006	NaN	1999	8	29	-	1999	9	7
74	ZIA	990	40	1999	9	11	-	1999	9	17
75	ANN	994	38	1999	9	14	-	1999	9	20
76	BART	940	85	1999	9	17	-	1999	9	29
77	DAN	1012	NaN	1999	10	1	-	1999	10	12
78	KAI-TAK	994	35	2000	7	2	-	2000	7	12
79	BOLAVEN	985	40	2000	7	19	-	2000	8	2
80	BILIS	1001	NaN	2000	8	17	-	2000	8	27
81	PRAPIROON	965	70	2000	8	24	-	2000	9	4
82	SAOMAI	970	60	2000	8	31	-	2000	9	19
83	XANGSANE	1003	NaN	2000	10	24	-	2000	11	2
84	CHEBI	1000	NaN	2001	6	19	-	2001	6	25
85	RAMMASUN	965	65	2002	6	26	-	2002	7	7
86	NAKRI	996	NaN	2002	7	7	-	2002	7	13
87	FENGSHEN	980	50	2002	7	13	-	2002	7	28
88	RUSA	960	70	2002	8	22	-	2002	9	3
89	KUJIRA	1000	NaN	2003	4	8	-	2003	4	25
90	SOUDELOR	975	60	2003	6	7	-	2003	6	24
91	MAEMI	935	90	2003	9	4	-	2003	9	16
92	MINDULLE	984	45	2004	6	21	-	2004	7	5
93	NAMTHEUN	996	40	2004	7	24	-	2004	8	3
94	MEGI	970	65	2004	8	13	-	2004	8	22
95	CHABA	955	80	2004	8	17	-	2004	9	5
96	SONGDA	945	75	2004	8	26	-	2004	9	10
97	MEARI	975	60	2004	9	18	-	2004	10	2
98	MATSA	998	NaN	2005	7	29	-	2005	8	9
99	NABI	955	75	2005	8	28	-	2005	9	9
100	KHANUN	1000	NaN	2005	9	5	-	2005	9	13
101	CHANCHU	996	NaN	2006	5	7	-	2006	5	19
102	EWINIAR	975	60	2006	6	29	-	2006	7	12
103	WUKONG	980	45	2006	8	12	-	2006	8	21
104	SHANSHAN	950	80	2006	9	9	-	2006	9	19
105	MAN-YI	955	70	2007	7	6	-	2007	7	23
106	USAGI	960	80	2007	7	27	-	2007	8	4
107	PABUK	995	NaN	2007	8	4	-	2007	8	15
108	NARI	960	75	2007	9	11	-	2007	9	18
109	WIPHA	1005	NaN	2007	9	14	-	2007	9	20
110	KROSA	1010	NaN	2007	10	1	-	2007	10	14
111	KALMAEGI	994	NaN	2008	7	11	-	2008	7	24
112	LINFA	998	NaN	2009	6	13	-	2009	6	30
113	MORAKOT	998	NaN	2009	8	2	-	2009	8	13
114	DIANMU	985	50	2010	8	6	-	2010	8	13
115	KOMPASU	970	70	2010	8	27	-	2010	9	6
116	MALOU	992	50	2010	8	31	-	2010	9	10
117	MERANTI	1003	NaN	2010	9	6	-	2010	9	14
118	MEARI	980	55	2011	6	20	-	2011	6	27
119	MUIFA	973	63	2011	7	26	-	2011	8	15
120	KULAP	1012	NaN	2011	9	5	-	2011	9	11
121	KHANUN	991	43	2012	7	13	-	2012	7	20
122	DAMREY	965	70	2012	7	27	-	2012	8	4
123	TEMBIN	980	55	2012	8	17	-	2012	9	1
124	BOLAVEN	960	65	2012	8	18	-	2012	9	1
125	SANBA	940	85	2012	9	10	-	2012	9	18
126	LEEPI	1002	NaN	2013	6	16	-	2013	6	23
127	DANAS	965	65	2013	10	1	-	2013	10	9
128	NEOGURI	975	50	2014	7	2	-	2014	7	13
129	MATMO	994	NaN	2014	7	16	-	2014	7	26
130	NAKRI	980	50	2014	7	27	-	2014	8	4
131	FUNG-WONG	998	35	2014	9	17	-	2014	9	25
132	VONGFONG	975	60	2014	10	1	-	2014	10	16
133	CHAN-HOM	973	58	2015	6	29	-	2015	7	13
134	HALOLA	994	45	2015	7	6	-	2015	7	26
135	SOUDELOR	998	35	2015	7	29	-	2015	8	12
136	GONI	945	85	2015	8	13	-	2015	8	30
137	NAMTHEUN	994	45	2016	8	30	-	2016	9	5
138	MERANTI	1004	NaN	2016	9	8	-	2016	9	17
139	CHABA	965	70	2016	9	24	-	2016	10	7
140	NANMADOL	985	55	2017	7	1	-	2017	7	8
141	PRAPIROON	965	60	2018	6	27	-	2018	7	5
142	JONGDARI	992	45	2018	7	23	-	2018	8	4
143	LEEPI	998	40	2018	8	10	-	2018	8	15
144	SOULIK	963	73	2018	8	15	-	2018	8	30
145	KONG-REY	975	65	2018	9	27	-	2018	10	7
146	DANAS	985	43	2019	7	14	-	2019	7	23
147	FRANCISCO	975	65	2019	8	1	-	2019	8	11
148	LINGLING	963	73	2019	8	30	-	2019	9	12
149	TAPAH	975	60	2019	9	17	-	2019	9	23
150	MITAG	988	50	2019	9	24	-	2019	10	5
151	HAGUPIT	996	NaN	2020	7	30	-	2020	8	12
152	JANGMI	996	40	2020	8	6	-	2020	8	14
153	BAVI	950	85	2020	8	20	-	2020	8	29
154	MAYSAK	950	80	2020	8	26	-	2020	9	7
155	HAISHEN	945	85	2020	8	30	-	2020	9	10

. The characteristics of typhoons were used as input variables. Typhoon data were obtained from the best track dataset provided by IBTrACS.

IBTrACS is an international database that integrates best track data of tropical cyclones worldwide [38]. It is a project developed by the National Centers for Environmental Information (NCEI) of National Oceanic and Atmospheric Administration (NOAA), initiated to unify tropical cyclone track records that were previously managed independently by Regional Specialized Meteorological Centers (RSMCs) and Tropical Cyclone Warning Centers (TCWCs) across different ocean basins [38,39]. The primary objective of IBTrACS is to standardize these records—originally archived using different formats and criteria by each agency—into a consistent dataset that is easily accessible and usable by researchers. The IBTrACS dataset [40] covers the global ocean region from 70° N to 70° S and from 180° W to 180° E, and classifies data into seven basins: North Atlantic (NA), Eastern Pacific (EP), Western Pacific (WP), North Indian Ocean (NI), South Indian Ocean (SI), South Pacific (SP), and South Atlantic (SA). IBTrACS has been continuously updated since 1842, and as of 24 July 2025, the most recent version is v04r01. The dataset generally provides data at 3-h intervals and focuses on storm center position, maximum sustained wind speed, and minimum central pressure. Additionally, depending on the source agency, it may include other tropical cyclone-related parameters such as radius of maximum winds, environmental pressure, storm classification.

In this study, among the various tropical cyclone characteristics provided by IBTrACS, the independent variables for the multiple linear regression model were selected based on the “TOKYO” data, which are issued by the Worle Meteorological Organization Regional Specialized Meteorological Center in Tokyo, operated by the Japan Meteorological Agency, the official forecasting authority for typhoons in the western North Pacific, with records available from 1951 to the present. The selected variables include the latitude and longitude of the typhoon center, the maximum sustained wind speed and central pressure at the typhoon center, the translational speed of the typhoon, the distance between the typhoon center and a specific location, and the angle of approach of the typhoon relative to that location. These variables were selected based on their known influence on storm surge dynamics and their availability from typhoon track datasets [29,30,31,32,33,34]. The angle of approach was calculated clockwise from true north. For instance, if the point of interest is located in the upper-right quadrant relative to the typhoon center, the angle of approach falls within the range of 0° to 90°.

2.2.2. Dependent Variable (Predictand)

To better understand and predict various oceanographic phenomena occurring along the Korean coast—such as tides, storm surges, and sea level rise—a nationwide tide gauge network consisting of approximately 50 stations has been established and is operated primarily by the Korea Hydrographic and Oceanographic Agency (KHOA) [41,42]. The temporal resolution of tide observations varies by station, with data available at intervals of 1 min, 10 min, or 1 h. KHOA provides two types of tide observation datasets with a temporal resolution of 1 h, both of which have undergone quality control procedures, including raw time series processing, gap filling, and outlier removal [43].

In this study, as mentioned in Section 2.1 Research area, the dependent variable for the multiple linear regression model was the storm surge height observed at eleven tide gauge stations located along the southeastern coast of the KP, the designated study area. The storm surge height was calculated by removing the astronomical tide components and the annual mean sea level from the quality-controlled water level records with one-hour interval at each tide gauge station. The astronomical tide components were estimated using the default setting of the T_tide MATLAB (R2025a) toolbox [44].

2.3. Multiple Linear Regression

Multiple linear regression (MLR) is a widely adopted statistical approach for quantifying the relationship between a set of independent (explanatory) variables and a dependent variable. By constructing a mathematical model based on observed data, MLR enables the prediction of the dependent variable’s behavior as a function of several explanatory factors. The general structure of an MLR model is expressed as follows:

$$Y={\beta }_0+{\beta }_1{X}_1+{\beta }_2{X}_2+\dots +{\beta }_n{X}_n$$

(1)

where $Y$ denotes the dependent variable, $X_i \left(i=1, \dots , n\right)$ represent the independent variables, and ${\beta }_i$ are the regression coefficients estimated by the least squares method. Although MLR assumes linear and additive relationships among variables, it has demonstrated robust applicability in modeling complex real-world phenomena [45,46,47].

In this study, the dependent variable is the observed storm surge height (SSH) at each station. The set of independent variables used in the regression analysis comprises the typhoon’s latitude (TOKYO_LAT), longitude (TOKYO_LON), the angle (ang) and distance (dis) between the typhoon center and the observation station, typhoon speed (speed_km), and central pressure (TOKYO_PRES).

Prior to model development, the dataset was rigorously preprocessed to ensure statistical validity and robustness. Missing values and storm surge heights less than zero were excluded from the dataset and the corresponding values of the dependent and independent variables at those timestamps were excluded from further analysis. This exclusion is to avoid phenomenon in which storm surge height becomes negative as a typhoon approaches, known as a “negative storm surge” or “reverse storm surge”, which occurs through a mechanism opposite to that of a typical storm surge [48,49,50]. It should be noted that due to the linear formulation of MLR, the regression occasionally yielded slightly negative predicted SSH values. These outputs are physically unrealistic because negative storm surges (reverse surges) were excluded from the dataset and are not the focus of this study. In practical applications, such negative predictions can be safely truncated to zero to ensure physical consistency and to prevent false positives.

Two different data splitting strategies were applied. In the first approach, all available time series data, regardless of typhoon event, were pooled and randomly split into training and testing sets at a 7:3 ratio. In the second approach, data were grouped by individual typhoon events, and the list of typhoons was split into training and testing groups (7:3 ratio), after which each station’s time series data were merged accordingly. This dual strategy was designed to test whether including all-time series typhoon events together influences the regression results, allowing for the evaluation of model performance both within and across typhoon events.

All input variables were standardized using z-score scaling prior to regression modeling to ensure comparability among predictors and improve numerical stability. To address potential multicollinearity among explanatory variables, variance inflation factors (VIFs) were computed for all predictors. The VIF is a commonly used diagnostic that quantifies the extent to which the variance of a regression coefficient is inflated due to multicollinearity with other variables. Generally, a VIF greater than 5 or 10 is considered indicative of problematic multicollinearity. Accordingly, in this study, only independent variables with a VIF value less than 5 were used in the regression analysis, thereby ensuring that multicollinearity does not pose a significant problem.

All analyses, including data preprocessing, scaling, VIF-based feature selection, model training, and performance evaluation were performed using Python 3.9. Executable code and the corresponding datasets have been provided as supplementary data to ensure transparency and reproducibility.

2.4. Objective Functions

To objectively evaluate the predictive performance of the MLR models developed in this study, four widely used statistical criteria were employed: mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and coefficient of determination (R²).

MAE measures the average magnitude of the errors in a set of predictions, without considering their direction. It is given by:

$$MAE={\displaystyle \frac{1}{n}}{\Sigma }_{i=1}^n\left|O_i-S_i\right|$$

(2)

where $O_i$ and $S_i$ are the observed and simulated (predicted) values, respectively, and $n$ is the number of data points. Lower MAE values indicate higher model accuracy. MAE is widely used because it provides a straightforward interpretation of the average prediction error in the original units of the target variable.

MSE represents the average of the squared differences between observed and predicted values:

$$MSE={\displaystyle \frac{1}{n}}{\Sigma }_{i=1}^n{\left(O_i-S_i\right)}^2$$

(3)

MSE penalizes larger errors more than smaller ones due to the squaring of the residuals, making it particularly sensitive to outliers. It is commonly used in regression analysis for model calibration and comparison.

RMSE is the square root of the mean squared error:

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\Sigma }_{i=1}^n{\left(O_i-S_i\right)}^2}$$

(4)

This criterion expresses the model error in the same units as the observed variable, aiding in interpretation and practical assessment of the model’s predictive performance. A lower RMSE value indicates better model performance.

R² quantifies the proportion of the variance in the observed data that is explained by the model. It is defined as:

$$R^2={\left({\displaystyle \frac{{\Sigma }_{i=1}^n\left(O_i-\overline{O}\right)\left(S_i-\overline{S}\right)}{\sqrt{{\Sigma }_{i=1}^n{\left(O_i-\overline{O}\right)}^2}\sqrt{{\Sigma }_{i=1}^n{\left(S_i-\overline{S}\right)}^2}}}\right)}^2$$

(5)

where $\overline{O}$ and $\overline{S}$ are the mean observed and predicted values, respectively. R² values range from 0 to 1, with higher values indicating better model fit. In general, R² values greater than 0.5 are considered acceptable for applications [49,50,51].

3. Results

3.1. Effect of Typhoon Event Grouping on Model Performance

3.1.1. Without Consideration of Individual Typhoon Events

To evaluate the impact of typhoon–station distance on regression model performance without explicit consideration of individual typhoon events, all available station data were pooled and randomly split into training and testing sets at a 7:3 ratio. The regression models were then constructed and evaluated across a range of distance thresholds, from 2000 km down to 500 km (

Table 2. Regression model test R² values for each station by typhoon–station distance. The “Total” row represents the R² value derived from regression using the combined dataset from all stations.

Station	Distance (km)
	2000	1500	1000	900	800	700	600	500
Gadeokdo	0.1396	0.1542	0.2132	0.2152	0.2087	0.1995	0.2104	0.2057
Geomundo	0.2303	0.2289	0.2494	0.2556	0.2599	0.2590	0.2528	0.2590
Geojedo	0.2425	0.2662	0.3205	0.3292	0.4012	0.3956	0.3968	0.4220
Goheung	0.1919	0.2361	0.1972	0.2724	0.2870	0.3354	0.3743	0.3722
Gwangyang	0.3627	0.2460	0.5359	0.5317	0.5264	0.5300	0.5440	0.5588
Masan	0.1438	0.1985	0.1546	0.2394	0.2307	0.2245	0.2918	0.3012
Busan	0.1461	0.1969	0.2267	0.2137	0.2156	0.2130	0.2126	0.2098
Yeosu	0.1423	0.1581	0.2106	0.2108	0.2053	0.2013	0.2165	0.2304
Ulsan	0.1659	0.1820	0.2404	0.2366	0.2289	0.2373	0.2276	0.1920
Tongyeong	0.1359	0.1615	0.2037	0.2017	0.1972	0.2000	0.2202	0.2155
Pohang	0.1333	0.1573	0.2462	0.2428	0.2360	0.2363	0.2230	0.2011
Total	0.1330	0.1651	0.1930	0.1930	0.1897	0.1866	0.1864	0.1736

).

The results show that model predictive performance, as measured by test R², is lowest when the distance threshold is set to 2000 km, with a general improvement as the threshold is reduced. The highest predictive performance is typically achieved when the distance threshold is set at either 1000 km or 900 km, after which further reductions in the threshold do not lead to substantial improvement and may even slightly decrease model accuracy. For example, among all stations, Gwangyang exhibited the highest R² value (0.5588) at a threshold of 500 km, while Geojedo station recorded the second-highest (0.4220) at the same distance. When all stations were combined into a single dataset (“Total” row), the best R² values (0.19) were observed at 1000 km and 900 km thresholds.

This trend suggests that the inclusion of data from distant typhoon events (i.e., those with typhoon centers located more than 1000 km from the observation site) introduces substantial noise to the regression model, likely because such events do not meaningfully affect local storm surge heights. By excluding these data points, the model focuses on physically relevant typhoon events, thereby increasing explanatory power and prediction accuracy. This finding aligns with the physical expectation that storm surge responses are most pronounced when a typhoon is within a certain proximity to the observation site.

The results further indicate spatial variability in model performance across stations. Notably, the Gwangyang and Geojedo stations consistently display higher R² values at shorter distance thresholds compared to other locations. This may be attributed to local geographic and bathymetric conditions—such as coastline orientation, shallow continental shelves, or exposure to typhoon tracks—that amplify the sensitivity of these sites to nearby typhoon-induced surges [52,53,54].

Figure 4 illustrates the predictive performance of the regression model using all station data at each distance threshold. In panels (a) through (c) (2000 km to 1000 km), red ellipses highlight regions where predicted values remain near zero despite variation in observed values. These points, which are gradually eliminated as the threshold decreases, do not contribute to the predictive capability of the model. In panels (e) through (h) (800 km to 500 km), blue ellipses indicate areas where the observed values exhibit sharp increases, but predicted values fail to exceed approximately 0.2, reflecting the model’s limited ability to capture extreme observed SSHs in these cases.

In summary, the distance-based thresholding of input data was found to be a critical factor in optimizing regression model performance, and the observed spatial differences among stations underscore the importance of local factors in storm surge prediction.

3.1.2. With Typhoon Event Grouping

To assess whether grouping data by individual typhoon events can improve regression model performance, we applied a data partitioning strategy in which the entire set of 76 typhoon events was randomly split into training and testing groups in a 7:3 ratio. Specifically, 53 typhoon events were used for model training, while the remaining 23 events were reserved for independent testing (

Table 3. Regression model test R² values for each station by typhoon–station distance according to typhoon event-based data grouping. The “Total” row represents the R² value derived from regression using the combined dataset from all stations.

Station	Distance
	2000	1500	1000	900	800	700	600	500
Gadeokdo	0.1527	0.1990	0.2411	0.2268	0.2208	0.2103	0.2228	0.2179
Geomundo	0.2145	0.2666	0.2781	0.2778	0.2809	0.2851	0.2899	0.2992
Geojedo	0.2757	0.2548	0.3823	0.3817	0.3539	0.3751	0.3608	0.3235
Goheung	0.1060	0.1865	0.1607	0.2939	0.3127	0.4089	0.4039	0.3632
Gwangyang	0.1620	0.1741	0.5339	0.5313	0.5265	0.5041	0.5335	0.5696
Masan	0.0962	0.1490	0.2179	0.1966	0.1870	0.1858	0.1735	0.2339
Busan	0.1590	0.2514	0.2445	0.2263	0.2244	0.2192	0.2219	0.2098
Yeosu	0.1636	0.2159	0.2622	0.2610	0.2540	0.2558	0.2779	0.2867
Ulsan	0.1720	0.2270	0.2503	0.2477	0.2315	0.2285	0.2096	0.1638
Tongyeong	0.1523	0.2036	0.2362	0.2334	0.2300	0.2337	0.2544	0.2480
Pohang	0.1591	0.2632	0.3025	0.2879	0.2706	0.2577	0.2201	0.1815
Total	0.1316	0.2110	0.2261	0.2189	0.2119	0.2066	0.2050	0.1863

).

Overall, the event-based grouping approach led to modest improvements in predictive performance compared to the results without grouping (Section 3.1.1, Table 2). For example, at Goheung station, the R² value exceeded 0.4 at the 700–600 km threshold, which was not observed in the non-grouped analysis. Similarly, stations such as Gadeokdo, Geomundo, Yeosu, Tongyeong, and Pohang showed R² increases of approximately 0.02–0.05 under event-based grouping. However, the effect was not uniformly positive. In some cases, such as Goheung at 2000–1000 km thresholds, R² decreased by 0.03–0.08, and Masan and Geojedo stations exhibited both improvements and reductions in R² values depending on the distance threshold.

These findings indicate that event-based grouping helps preserve the temporal coherence and intra-event variability inherent in typhoon time series data, Reducing potential information leakage between training and testing sets. This leads to a more realistic estimate of the model’s generalization ability for future, unseen typhoon events. The observed improvements are particularly notable at stations more exposed to typhoon impacts and at shorter distance thresholds, which aligns with the expectation that local bathymetry, coastline orientation, and storm track proximity all play significant roles in storm surge predictability.

Direct comparison of Table 2 and Table 3 reveals that the greatest R² improvements from event grouping were achieved at Goheung (up to +0.07 at a 700 km threshold) and Gwangyang (up to +0.03 at a 500 km threshold). For some stations, such as Masan and Geojedo, R² values fluctuated, reflecting the complex interaction of local factors and data partitioning method.

From an operational standpoint, the event-based grouping approach is advantageous because it better simulates real-world forecasting scenarios—where the model must predict storm surge heights for entirely new typhoon events—thereby supporting more robust early warning system development. However, it should be noted that grouping by event reduces the effective sample size in both training and testing datasets, which may affect the statistical stability of regression results, especially at higher distance thresholds where sample sizes are inherently smaller.

Figure 5 presents the regression model test results for all stations under event-based grouping across various typhoon–station distance thresholds. In each panel, most data points with observed SSH values below 0.3 cluster densely, with close correspondence between observed and predicted values. For observed SSH values exceeding 0.3, however, the model persistently underestimates, with predicted values rarely surpassing 0.2. This limitation remains even as the distance threshold decreases, although the range of predicted values becomes somewhat broader at lower thresholds. This suggests that while event-based grouping enhances performance for moderate SSH events, the regression model continues to struggle with predicting extreme storm surge events.

In summary, grouping by typhoon events can yield slight improvements in regression model accuracy and generalizability, particularly for stations and thresholds most affected by typhoon surges. Nevertheless, the model’s limitations in capturing the highest observed SSH values persist. To address this issue, the following section presents an additional analysis applying threshold values to the observed SSH data in order to further investigate and potentially improve the model’s predictive capability.

3.2. Model Performance According to SSH Threshold Values

Since the regression analysis in Section 3.1 indicated that a distance threshold of 1000 km yielded the best model performance, the distance was fixed at 1000 km for subsequent analyses. Consistent with the approach in Section 3.1.2, the dataset was grouped by individual typhoon events prior to partitioning.

For the SSH thresholds, values were empirically determined based on the distribution of observed SSHs shown in Figure 5. Specifically, the threshold was estimated by identifying the point at which data began to cluster along the lower bound (red line), yielding an initial cut-off of approximately 0.2 m. The dataset was then divided into two groups: a low SSH range (0 ≤ SSH ≤ threshold) and a high SSH range (threshold < SSH), and performing regression modeling separately for each interval. To further assess model sensitivity, the threshold was subsequently increased in increments of 0.05 m, and regression modeling was performed separately for each interval.

3.2.1. Model Performance in the Low SSH Range

For the low SSH range, the results showed that as the threshold increased—thereby broadening the low SSH interval—the test R² values generally improved across most stations (

Table 4. Test R² values of regression models for each station in the low SSH range (0 ≤ SSH ≤ threshold) according to observed SSH thresholds. The “Total” row represents the R² value derived from regression using the combined dataset from all stations.

Station	Threshold
	0.2	0.25	0.3	0.35	0.4	0.45	0.5	0.55	0.6
Gadeokdo	0.1646	0.1857	0.2153	0.2250	0.2325	0.2355	0.2368	0.2392	0.2412
Geomundo	0.1649	0.1916	0.2199	0.2461	0.2608	0.2708	0.2759	0.2765	0.2781
Geojedo	0.4361	0.4013	0.3948	0.3853	0.4033	0.4095	0.4095	0.4095	0.4064
Goheung	0.2992	0.2125	0.1692	0.1670	0.1670	0.1606	0.1607	0.1607	0.1607
Gwangyang	0.4355	0.4796	0.4945	0.4941	0.5034	0.5283	0.5241	0.5339	0.5339
Masan	0.1354	0.1299	0.1332	0.1244	0.1313	0.1458	0.1535	0.1637	0.1637
Busan	0.1644	0.1869	0.2255	0.2334	0.2378	0.2421	0.2440	0.2469	0.2458
Yeosu	0.1819	0.2055	0.2302	0.2598	0.2634	0.2664	0.2743	0.2758	0.2776
Ulsan	0.1615	0.2089	0.2291	0.2469	0.2497	0.2503	0.2503	0.2503	0.2503
Tongyeong	0.1407	0.1726	0.1997	0.2158	0.2245	0.2288	0.2334	0.2344	0.2371
Pohang	0.2038	0.2390	0.2645	0.2845	0.2973	0.3012	0.3012	0.3025	0.3025
Total	0.1461	0.1759	0.2009	0.2167	0.2241	0.2275	0.2299	0.2306	0.2310

). This trend suggests that expanding the low range to include a wider range of SSH values, not only the smallest but also those approaching the threshold, enables the regression model to better capture the underlying linear relationship between storm surge height and the explanatory variables. In very narrow low ranges (e.g., threshold = 0.2 or 0.25), the regression outcomes are often more scattered, and linearity is not clearly observed, due to statistical noise and limited variability in the observed SSH values. As the threshold increases, the inclusion of more SSH cases closer to the threshold leads to a clearer linear structure and stronger model fit, as reflected in the increasing R².

It is noteworthy that a few stations, such as Geojedo and Goheung, exhibited higher R² values when the low range was more restricted. This phenomenon may be associated with a limited number of data points or the presence of a more pronounced linear relationship within a narrower SSH interval at these locations. In such cases, the calculated model performance becomes highly sensitive to both data range and distribution, which may yield higher R² values for small but well-structured datasets.

Figure 6 presents the scatter plots of predicted versus observed SSH for all stations across various threshold values. For lower thresholds, predicted SSH values are largely confined within a narrow range, exhibiting minimal sensitivity to variations in the observed SSH. As the threshold increases, this pattern persists: the predicted values remain concentrated within a limited interval, even as the observed SSH demonstrates greater variability. Consequently, the regression model systematically underestimates higher observed SSH values within the low range, resulting in a plateau effect rather than improved alignment with the 1:1 reference line. These findings indicate that increasing the threshold does not substantively enhance the model’s ability to reproduce the full spectrum of observed SSH values. The improvement in R² at higher thresholds is therefore primarily attributable to the statistical effects of expanding the data range, rather than to genuine advancements in predictive performance.

Overall, these results highlight the limitations of the regression model in accurately capturing the variability of SSH in the low range, regardless of the threshold setting, and suggest that alternative modeling strategies or the inclusion of additional explanatory variables may be required to improve predictive skill.

3.2.2. Model Performance in the High SSH Range

For the high SSH range (threshold < SSH), the regression model’s predictive performance generally improved as the threshold value increased, indicating that more selective data curation led to enhanced model skill (

Table 5. Test R² values of regression models for each station in the high SSH range (threshold < SSH) according to observed SSH thresholds. Empty cells indicate cases where regression analysis was not performed due to insufficient data for model construction. The “Total” row represents the R² value derived from regression using the combined dataset from all stations.

Station	Threshold
	0.2	0.25	0.3	0.35	0.4	0.45
Gadeokdo	0.4131	0.4331	0.5048	0.5417	-	-
Geomundo	0.3373	0.3175	0.3370	0.4959	0.5828	0.5191
Geojedo	0.5892	-	-	-	-	-
Goheung	0.1621	0.3508	-	-	-	-
Gwangyang	-	-	-	-	-	-
Masan	0.8215	0.7534	0.5696	-	-	-
Busan	0.3810	0.4250	0.3740	0.5088	-	-
Yeosu	0.2767	0.3408	0.3875	0.4463	0.4396	0.3248
Ulsan	0.3732	0.4002	-	-	-	-
Tongyeong	0.2946	0.3024	0.3246	0.4575	0.2631	0.4640
Pohang	0.3604	0.5284	0.3442	-	-	-
Total	0.2299	0.2550	0.2476	0.2513	0.2586	0.1335

). With the exception of Masan and Pohang, all stations exhibited increasing R² values with higher thresholds. This trend reflects that, for more extreme storm surge events, the model is better able to capture linear relationships among the input variables, possibly because the variance and dynamic range of SSH are higher in this subset. This observation is supported by the calculated variance and range of storm surge height values across different threshold datasets (see Table S1 in the Supplementary Materials), which demonstrate that both variance and range increase with higher thresholds. This finding confirms that more selective data curation in the high storm surge height range enhances the linear relationship between predictors and storm surge heights.

However, it is important to note that as the threshold increases, the number of available data points within the high SSH range decreases. In several cases, particularly for the height thresholds, the sample size became insufficient to conduct regression analysis, as indicated by empty cells in Table 5. This reduction in sample size may introduce risks of overfitting and statistical instability, as regression estimates become more sensitive to individual data points when the dataset is small. Although in this study both p-values and variance inflation factors (VIFs) were carefully checked to mitigate these risks, the potential for overfitting remains a consideration that warrants caution in the interpretation of these results.

Table 5 summarizes the test R² values for each station at varying SSH thresholds in the high range. Figure 7 presents the scatter plots for each station at the threshold yielding the highest R² value. Compared to the low SSH range, the high range results demonstrate clearer linearity and improved model performance. For example, the Masan station achieved the highest R² value of 0.82. With the exception of Goheung, Yeosu, Ulsan, and Tongyeong, most stations exhibited R² values exceeding 0.5, indicating a substantial gain in explanatory power within this range.

Although Figure 7c,e shows relatively high R² values, the predicted SSH values are more widely scattered when the observed values cluster around 0.2–0.3 m. This dispersion arises because global linear regression attempts to preserve a single linear trend across the entire dataset, leading to systematic deviations in ranges where the true relationship is nearly flat. Consequently, small variations in predictors translate into disproportionately large differences in fitted SSH. Such behavior reflects the inherent limitation of global linear regression, which cannot fully capture localized nonlinear relationships, and the increased residual sensitivity near thresholds with high sample density, potentially inducing heteroscedasticity [55]. In addition, nonlinear tide–surge interactions and site-specific coastal geometry may contribute to the variability in this range [52,53,54]. In future research, we plan to incorporate geographical factors such as the relative position of observation stations to typhoon tracks, coastline configuration, and the presence of bays to reduce such dispersion and improve model performance.

Overall, these findings suggest that, although the regression model shows limited performance for smaller storm surge heights, it is able to capture the relationships among typhoon characteristics and SSH more effectively for higher surge events, provided sufficient data are available.

4. Discussion

This study demonstrated that storm surge prediction using MLR can be significantly improved by careful data selection and preprocessing, particularly with respect to the typhoon–station distance and the application of SSH thresholds. Although Tadesse and Wahl [36] reported that atmospheric pressure, particularly lagged pressure, was a dominant predictor in many regions globally, typhoons approaching the Korean Peninsula generally weaken in intensity, making pressure- or wind-based thresholds less effective. Instead, distance- and SSH-based thresholds were adopted, as they better capture the regional characteristics of storm surges along the Korean Peninsula. Yang et al. [15] also demonstrated in their study that storm surge height are strongly influenced by the distance between the typhoon center and the point of interest, supporting the validity of the present findings. Our results indicated that when the distance between a typhoon center and the observation station was restricted to within 900–1000 km, and when the dataset was partitioned according to SSH ranges, the overall predictive performance of the regression models was enhanced. This improvement can be attributed to the exclusion of data points corresponding to events with negligible storm surge response, thereby reducing noise and focusing model training on physically meaningful cases.

Despite these improvements, the model’s predictive skill in the low SSH range remained relatively limited. The results suggest that storm surge heights in this regime may be influenced by factors or interactions not fully captured by the linear and additive assumptions of MLR. This is further supported by the observed plateauing of predicted values and the limited alignment with the 1:1 reference line in scatter plots, indicating that the linear model may not adequately capture the underlying dynamics of storm surge events, particularly for lower-magnitude occurrences.

In contrast, in the high SSH range, the regression models achieved notably higher R² values, with several stations exhibiting values greater than 0.5. This can be ascribed to the increased variance and dynamic range of SSH in this interval, which facilitates the identification of linear relationships among predictors. Nevertheless, it must be acknowledged that as the SSH threshold increases, the number of data points in the high range decreases, potentially resulting in overfitting and reduced statistical robustness. Although this study implemented rigorous checks of statistical significance (p-values) and multicollinearity (VIFs), these limitations should be considered when interpreting the results.

Given these findings, future research should explore the use of nonlinear or ensemble modeling approaches—such as random forest regression, gradient boosting, or neural networks—which may better accommodate the inherent nonlinearity and complex interactions in storm surge processes. In addition, expanding the range of predictor variables to include real-time sea level observations, additional meteorological parameters, and local bathymetric features may further enhance predictive performance [15,35,36]. The integration of physical-based numerical models with data-driven statistical approaches also warrants investigation as a potential pathway for achieving greater accuracy and reliability. In this context, storm surge height is affected not only by typhoon characteristics but also by the site-specific conditions of the location of interest (e.g., [15,56,57,58,59]). As the threshold was defined based on the distance between the typhoon and the observation site, it may be difficult to ensure the same model performance in other regions, even if the threshold value is identical. Therefore, when applying the model development in this study to other regions with the same threshold values, the predictive performance should be examined separately for those regions.

From an operational perspective, the methodology and findings presented here offer valuable insights for the development of storm surge early warning systems and coastal risk management. Nevertheless, additional validation and calibration across a broader range of sites, as well as under diverse climatic and typhoon scenarios, will be required to generalize the applicability of these results.

5. Conclusions

This study evaluated the performance of multiple linear regression models for storm surge prediction, focusing on the influence of typhoon–station distance and SSH thresholds on model accuracy. The results demonstrated that careful data partitioning—restricting typhoon events to within 900–1000 km of the observation site and separating data into low and high SSH ranges—significantly improved predictive skill, especially for high surge events. However, the linear regression model exhibited limitations in accurately capturing the variability of lower-magnitude storm surges, highlighting the complexity and inherent nonlinearity of storm surge processes.

The methodology and findings presented herein provide valuable insights for the development of coastal risk management strategies and early warning systems. Nevertheless, further research is needed to address the identified limitations, particularly by applying nonlinear and ensemble modeling approaches, expanding the set of predictor variables, and validating the models across broader spatial and climatic contexts. Continued advancements in this field will support more reliable storm surge forecasting and enhanced coastal resilience in the face of extreme weather events.