I. Introduction

Air pollution has received attention as a public health and economic concern. A large empirical literature suggests that exposure to particulate matter and related pollutants is associated with higher mortality risk, higher health care costs, and lower labor productivity (Deryugina et al., 2019;Graff Zivin and Neidell, 2012;Greenstone and Hanna, 2014). In South Korea, air pollution is often discussed through two broad channels. One is domestic emissions, including power generation, industrial activity, and road traffic (An and Heshmati, 2019; Kim, 2021; Lee et al., 2011). The other is transboundary transport, in which dust and pollution episodes originating outside the country contribute to elevated particulate matter concentrations under certain meteorological conditions (Heo et al., 2025;Kim, 2019). Both channels are likely to contribute to observed concentrations.

An insight of the earlier literature is the role of heterogeneity. Air pollution is not driven by the same sources everywhere. Local economic activity, transport infrastructure, and geography can generate variation in which sources matter and when, even within the same country and period. A large applied literature accounts for this heterogeneity by using high-frequency data, rich fixed effects, and related designs to estimate aggregate relationships between pollution and outcomes at broader geographical scales (Deryugina et al., 2019;Moretti and Neidell, 2011;Schlenker and Walker, 2016;Schwartz et al., 2017;Zhao et al., 2023)

This heterogeneity is salient in coastal port cities. Vessel traffic and port operations are spatially concentrated near shoreline neighborhoods, and short-run fluctuations in maritime activity may be reflected in short-run variation in ambient concentrations measured at nearby monitors. These settings therefore provide a natural environment to document a port-related exposure-concentration relationship using high-frequency data.

In South Korea, a few studies examined port and coastal-related externalities through coastal environmental restoration and welfare effects, port development and fisheries compensation, and environment-fisheries trade linkage (Pyo, 2009;Moon and Kang, 2006;Lin and Kim, 2015). By contrast, there is limited evidence on how vessel movements near ports map into contemporaneous ambient air pollution concentrations.

Motivated by this gap, this study uses administrative port-call records to measure hourly vessel arrivals and departures as a high-frequency proxy for maritime activity. The analysis links these records to hourly monitor-level air-quality readings and meteorology data, and asks whether short-run variation in vessel activity is associated with contemporaneous PM 10 and PM 2.5 at nearby monitors, conditional on meteorology and rich fixed effects.

The empirical setting is Busan, Korea's primary container port and a major fish-landing and seafood logistics hub. Port activities and fisheries-related industries are spatially proximate, and the city plays an important role in maritime logistics, coastal production, and distribution systems supporting transportation, fisheries, and seafood markets. While the analysis is based on Busan, the setting shares key features with many large port cities-high maritime throughput and shoreline-adjacent urban activity-so the estimates are intended to be informative for broader discussions of local pollution margins in port-intensive coastal hubs such as Los Angeles, Guangzhou, Zhoushan, and Callao/Lima.

This study contributes to the literature in three ways. First, this study documents a port-related local pollution margin in a setting where national discussions often emphasize transboundary episodes, suggesting that coastal logistics activity may be relevant for urban particulate concentrations. Second, this study leverages high-frequency administrative port-call records and monitor-level air-quality data within a panel framework, using rich fixed effects and meteorological controls to better account for unobserved monitor-specific differences and common time shocks. Third, this study quantifies an exposure-concentration relationship in a large port city, providing a practical input for subsequent work that evaluates port-related policies or links port-associated pollution to health outcomes.

The remainder of the paper proceeds as follows. Section 2 describes the data sources and construction of the port exposure index. Section 3 presents the empirical framework and fixed-effects specifications. Section 4 reports the main results. Section 5 concludes with implications for Korea's port cities and directions for future research.

Ⅱ. Data

1. Data sources and sample construction

This study constructs an hourly monitor-time panel for Busan, South Korea, spanning from January 2016 through December 20191). The analysis integrates three data sources: (i) administrative port-call records, (ii) hourly air-quality readings from regulatory monitors, and (iii) hourly meteorological observations.

For port-call data, event-level information on vessel arrivals and departures is obtained from the Public Data Portal, which provides Ministry of Oceans and Fisheries' public Vessel Operation Information (VOIS) system. Each record contains vessel identifiers, gross tonnage, origin and destination ports, a pier/anchoring-site, and timestamps for entry and exit. Because VOIS records are event-based, vessel activity is sparse at the hourly level: many hours have no recorded arrivals or departures at a given pier/anchorage. To construct an hourly activity measure, port-call events are aggregated to hourly gross tonnage by location. Events are then mapped to a complete hourly time grid; hours without recorded movements are assigned zeros by construction. Blue dots in <Fig. 1> show piers and anchoring-sites.

For air pollution data, hourly concentrations of PM 2.5, PM 10, SO 2, NO 2, O 3, and CO are obtained from AirKorea. The sample is restricted to the 23 monitoring stations located within Busan's administrative boundary. Monitor coordinates reported by AirKorea are used to locate each station spatially - i.e., red dots in <Fig. 1> PM concentrations are measured in µg/m³.

For meteorological data, hourly weather covariates-temperature, wind speed, wind direction, and precipitation-are collected from the Public Data Portal, which provides the Korea Meteorological Administration's Automated Weather System (AWS). AWS observations originate from high-frequency measurements (e.g., 1-minute or 10-minute intervals) that are aggregated and quality-controlled to the hourly level by KMA. All measurements follow Korean Standard Time (KST). Wind direction is converted into four directional indicators (NE, SE, SW, NW) to summarize prevailing dispersion conditions. Meteorological data are merged to the monitor-hour panel.

2. Data management: Port-exposure index

The empirical strategy requires linking hourly vessel activity to the ambient pollution recorded at each air-quality monitor. Vessel activity is linked directly to monitor coordinates to preserve the spatial granularity of the monitoring network and to reduce measurement error that could arise from assigning uniform exposure within administrative units.

Two geospatial layers are assembled: (i) the coordinates of Busan's 23 air-quality monitors and (ii) the coordinates of port locations where arrivals and departures occur, including piers and anchoring sites (91 locations during the study period). Euclidean distances between each monitor m and each pier location p are computed to form a time-invariant monitor-pier distance matrix d m p. These distances provide a simple proxy for the idea that vessel activity occurring closer to a monitor is more likely to be reflected in the monitor's readings than activity occurring farther away.

Port-call records are first aggregated to hourly gross tonnage at each port location. Let T p t denote the total gross tonnage associated with recorded arrivals and departures at location p during hour t. Using the distance matrix, monitor-level exposure is defined as an inverse-distance-weighted tonnage of contemporaneous activity:

In equation (1), E m t reflects both the scale of vessel activity and proximity to active piers. This formulation is consistent with the intuition that nearby vessel activity contributes more strongly to measured concentrations than distant activity. For interpretability, regression tables report the coefficient on E m t per 1,000 tons (i.e., E m t scaled by 10 3)

Because port-call records are event-based, many location-hours have no arrivals or departures. The port-activity series is therefore mapped to a complete hourly time grid at the location level, and hours without recorded movements are assigned zeros by construction. When activity is concentrated in only a few locations in a given hour, most T p t terms in the sum are zero. In such cases, E m t remains positive but relatively low because it is driven by the small number of locations with T p t > 0. E m t would be mechanically zero only if all locations had T p t = 0 in the same hour; in practice, many locations are inactive in a given hour, while a subset has positive activity.

<Table 1> reports descriptive statistics for vessel activity, the port exposure index, air pollutants, and meteorological covariates. The VOIS port-call dataset contains 339,618 event records, each corresponding to an individual vessel arrival or departure. In contrast, the exposure index E m t is constructed at the hourly monitor-time level. Accordingly, the monitor–hour panel includes 703,464 observations, determined by the number of monitors multiplied by the number of hourly periods in the study window. Missing observations in the effective estimation sample primarily reflect missing readings in the air-quality and/or meteorological series. When E m t is missing, this reflects data availability or matching (e.g., missing coordinate metadata or merge limitations), rather than "no port activity," which is coded as zero by construction.

Ⅲ. Empirical strategy

The analysis leverages an hourly monitor-level panel, rather than a single time series, to exploit cross-monitor variation while controlling for monitor fixed effects and common time shocks. Rich fixed effects and meteorological covariates are included to limit confounding from persistent monitor differences and citywide temporal patterns, providing a transparent characterization of how vessel activity maps into monitor-level particulate concentrations in Busan. The empirical specification relates monitor-hour pollution to vessel activity exposure as shown in equation (2).

Pollution m t denotes the ambient concentration of PM 10 or PM 2.5 measured at monitor m in hour t, and E m t is the port exposure index defined in Section 2.2. The vector X m t ′ includes meteorological controls - temperature, wind speed, and wind-direction indicators - and ϵ m t is an idiosyncratic error term. Standard errors are clustered at the monitor level to allow for serial correlation within monitoring stations over time. The coefficient of interest, β, summarizes how contemporaneous pollution varies with a one-unit increase in exposure (i.e., an increase in the inverse-distance-weighted tonnage of vessel arrivals and departures).

In addition, γ m are monitor fixed effects, and τ t are rich time fixed effects (e.g., year-by-month-by-day-by-hour or monitor-specific year fixed effects). The analysis uses two fixed-effects specifications. Both are designed to absorb persistent spatial differences across monitors and common temporal shocks, but they differ in how flexibly they control for time-varying confounders.

In the results, Model 1 includes monitor fixed effects and separate fixed effects for year, month, day, and hour. Monitor fixed effects absorb time-invariant differences across stations, such as local land use, elevation, proximity to major roads, and persistent neighborhood characteristics. The year, month, day, and hour fixed effects absorb aggregate temporal variation common to all monitors, including seasonality, day-specific shocks, and diurnal patterns (e.g., typical morning/evening pollution cycles).

Model 2 includes a more demanding fixed-effects structure: monitor-by-month-by-day-by-hour and monitor-by-year fixed effects. The monitor-by-month-by-day-by-hour fixed effects allow each monitor to have its own highly flexible baseline that varies at the month-day-hour level, absorbing granular time patterns that may differ across locations (e.g., monitor-specific diurnal profiles, local microclimate patterns, and recurring location-specific activity schedules). The monitor-by-year fixed effects allow each monitor to have its own year-specific level, absorbing slow-moving local changes (e.g., gradual neighborhood development, long-run shifts in nearby traffic or construction) that could otherwise confound the relationship between vessel activity and pollution. With these controls, identification relies on within-monitor variation that remains after removing monitor-specific baseline patterns at a fine time scale.

Ⅳ. Results

<Table 2> reports the particulate matter estimates. Model 1 (columns 1 and 3) includes monitor, year, month, day, and hour fixed effects. Model 2 (columns 2 and 4) applies a more demanding fixed-effects structure: monitor-by-month-by-day-by-hour and monitor-by-year fixed effects. Across specifications, higher port exposure is associated with higher PM 10 and PM 2.5, with larger point estimates under Model 2.

In model 2, a 1,000-ton increase in hourly vessel activity (as summarized by the port exposure index) is associated with approximately 0.076µg/m³ higher PM 10 and 0.078µg/m³ higher PM 2.5. The corresponding point estimates in Model 1 are smaller but remain positive. One interpretation is that the more demanding fixed effects in Model 2 absorb additional monitor-specific temporal patterns that may be correlated with both vessel activity and pollution, so the remaining within-monitor variation yields larger coefficients. At the same time, the two specifications rely on different residual variation, so differences in magnitudes should be interpreted cautiously2).

A back-of-the-envelope calculation helps interpret magnitudes. In Model 2, a 1,000-unit increase in the exposure index is associated with 0.076µg/m³ higher PM 10 and 0.078µg/m³ higher PM 2.5. Relative to the sample means (41.28 for PM 10 and 23.94 for PM 2.5), these changes correspond to approximately 0.18% and 0.33%, respectively.

Consistent with prior atmospheric studies for Korea (Heo et al., 2025;Kim, 2019), wind direction is included as a control and is informative in magnitude and sign. Relative to southeasterly winds, northwesterly and southwesterly winds are associated with higher PM 10 (2.11µg/m³ and 1.69µg/m³, respectively, in Model 2), and northwesterly winds are associated with higher PM 2.5. (1.10 µg/m³). On the other hand, northeasterly winds are associated with lower PM 10 and PM 2.5 (-2.39 µg/m³ and -1.76 µg/m³, respectively). These patterns are consistent with Busan's coastal setting, in which winds from inland directions may coincide with higher pollutant loads than winds originating over open sea. Consistent with the paper's scope, wind direction is treated as a covariate rather than a structural dispersion mechanism.

Temperature is positively associated with PM 10 and PM 2.5 in this monitor-hour panel, whereas wind speed is negatively associated with particulate concentrations. Because the specifications include rich time fixed effects that absorb seasonal patterns, the positive temperature coefficient is interpreted as a within-season association. This pattern is consistent with higher temperatures being correlated with conditions conducive to higher particulate concentrations, including stronger photochemical activity, while higher wind speeds are consistent with greater dispersion.

As a sensitivity check, the same specifications are estimated for SO 2, NO 2, O 3, and CO as shown in <Table 3>. The estimated coefficients on port exposure are small and statistically indistinguishable from zero at conventional levels. This is possibly because O 3 is largely secondary, and SO 2 and NO 2 are key precursors to secondary particulate formation, so contemporaneous associations with a short-run activity proxy need not be pronounced at the monitor-hour scale3). The evidence is consistent with the exposure index mapping more directly into particulate concentrations than into gaseous measures (Deryugina et al., 2019; Schlenker & Walker, 2016).

Ⅴ. Conclusion

To quantify negative externalities due to coastal vessel activity, this study links high-frequency vessel activity to monitor-level air quality in Busan during 2016~2019. Using an hourly port-exposure index constructed from administrative port-call records and panel regressions with meteorological controls and rich fixed effects, the estimates suggest a positive relationship between port exposure and both PM 10 and PM 2.5. The results are intended to document an exposure–concentration mapping for a coastal logistics hub, providing evidence on a local port-related margin that may complement broader discussions of regional transport in Korea and may be informative for other large port cities where maritime operations and urban activity co-locate.

From a fisheries and coastal-management perspective, the results highlight that maritime activity in port-adjacent areas may be a non-negligible local environmental margin, alongside transboundary pollution. This is potentially relevant for coastal regions where port and fisheries-related activities cluster near the shoreline. The results highlight that deteriorating air quality affects the health and livelihoods of local fisheries workers and local communities. The estimates also provide a quantitative segue for future studies examining how fisheries-related emissions affect fisheries productivity, long-term health impacts, and the effectiveness of pollution-reduction policies.

This study has several limitations. First, the study could not disentangle fisheries-related traffic from other maritime activity, nor did it incorporate records of coastal small fishing vessels. Accordingly, the estimates reflect overall port-related vessel activity; future work could combine vessel-type information to construct fisheries-specific exposure measures.

Second, the analysis ends in 2019. Extending the panel into 2020 and beyond would coincide with large, overlapping shocks-most notably COVID-19 disruptions to shipping and local activity—and major maritime regulatory changes, including the IMO 2020 global sulfur cap and Busan's Vessel Speed Reduction (VSR) program introduced in December 2019. These concurrent changes complicate separating policy effects from broader demand and routing shifts within the same empirical design. A natural extension is to assemble a longer post-2020 panel and implement a policy-focused design (e.g., difference-in-differences exploiting VSR eligibility and timing), potentially complemented by alternative exposure constructions and additional validation exercises.

Third, the paper focuses on a single, consistently measured high-frequency signal of port activity-arrivals and departures. Other port-related activities, such as cargo-handling equipment use or trucking flow, may also be associated with vessel activity and air pollution, but are not observed at comparable frequency in the available data. Accordingly, the estimates provide a reduced-form exposure–concentration mapping based on vessel movements, rather than a decomposition of total port-related emissions by source.

Lastly, the analysis controls for wind direction as a contemporaneous dispersion condition but does not explicitly model port-to-monitor downwind alignment. Future work could construct a downwind measures and estimate E m t × W i n d d i r e c t i o n m t.