Severity-Based Assessment of Maritime Accidents in Indonesian Inland Waterway and Offshore Shipping Environments

Vessels were grouped into analytically meaningful categories aligned with the operational structure of Indonesian fleets in Table 1. To improve interpretability, abbreviations were replaced with explicit vessel-type labels: general cargo ship, ferry/Ro-Ro, fishing vessel, high-speed craft/speedboat, small passenger craft, service/special-purpose vessel, tanker, tug-barge combination, and unknown/not-specified (UNK-V). Accident modes were coded into standard operational categories consistent with the narrative fields and dominant failure mechanisms in each environment (collision/allision, grounding, capsizing/sinking, fire/explosion, machinery/technical failure, person overboard, and other/unknown). The UNK-V category constitutes 38% of RIV records and 27% of SEA records in the cross-tabulations, a distributional asymmetry that is acknowledged as a limitation; a sensitivity analysis excluding UNK-V is reported alongside primary results to confirm that the zone-severity association is not driven by this category. This classification approach supports an interpretable comparison of how socio-technical and organisational controls interact with environment-specific operating constraints, consistent with systems-oriented reasoning that emphasises context, control structures, and failure propagation rather than single-factor event chains [14-16].

2.3 Statistical analysis

The analytical strategy integrates descriptive statistics and inferential modelling to assess whether observed differences between RIV and SEA persist after accounting for compositional variation in vessel types and accident modes. First, accident frequency was summarised by year and operational zone (2018-2021) to establish the temporal pattern of incidents and the relative contribution of each environment. Second, cross-tabulations were used to characterise how vessel categories and accident modes differ across RIV and SEA, and to identify dominant operational profiles that explain environment-specific accident patterns. In addition, spatial concentration was assessed descriptively by mapping dominant vessel-type accident clusters across major Indonesian waters and rivers, using the vessel-category counts and dominant risk profiles reported in the updated hotspot table, thereby linking national patterns to geographically concentrated operational risk consistent with spatial analytical practice in maritime safety research [14, 15].

Third, severity indicators were computed by operational zone using the cleaned dataset (n = 1,523). These included the proportion of accidents with at least one death or missing person and the distribution of per-event casualty counts, to capture both fatality likelihood and casualty intensity. A chi-square test of independence was applied to confirm whether the difference in severe-outcome proportions between RIV and SEA was statistically significant, providing an inferential check on the zone-severity contrast independent of the regression model. Finally, a multivariable logistic regression model was estimated to quantify zone-specific differences in the probability of severe outcomes — the binary indicator of at least one death or missing person per event — while controlling for vessel category and accident mode. The model included three categorical predictors: (i) operational zone (RIV = reference; SEA = comparator); (ii) vessel category, with fishing vessel (FISH) as the reference category given its numerical predominance across both zones; all remaining categories, including UNK-V, were entered as separate binary indicators; (iii) accident mode, with capsizing/sinking (CAPS) as the reference category; all other modes entered as binary indicators. Because UNK-V is distributed asymmetrically across zones (38% of RIV vs. 27% of SEA cross-tabulation records), both a primary model including UNK-V and a sensitivity model excluding UNK-V were estimated; consistency of the zone effect across both models is reported. Model fit was assessed using the Hosmer-Lemeshow goodness-of-fit test; variance inflation factors (VIF) were computed for all predictors to check multicollinearity. The model was estimated on the cleaned dataset (n = 1,523). RIV was treated as the reference environment throughout; all estimates are reported with 95% confidence intervals and p-values.

The results indicate that Indonesia's maritime risk landscape has a dual structure: open-water/coastal routes account for the majority of recorded accidents, while river/inland waterways contribute a smaller but persistent share. This distribution is consistent with evidence that accident counts concentrate along high-traffic corridors [17, 18] and should be interpreted as an exposure-driven signal — reflecting the scale and diversity of maritime activity across Indonesia's major straits and seas — rather than as direct evidence of higher per-incident severity. The distinction carries practical importance for engineering intervention prioritisation: high absolute counts in SEA indicate the scale of fleet exposure and point to a need for broad-coverage monitoring and vessel-safety standards across offshore fleets, while the lower but persistent RIV counts signal a structural risk that warrants targeted waterway-specific engineering controls.

Accident-mode differences reinforce this interpretation. River/inland incidents are more strongly characterised by close-quarters navigation outcomes—collisions/allisions and capsizing/sinking—consistent with constrained channels and frequent interactions with fixed structures [22]. In open-sea/coastal environments, a stronger presence of stability- and machinery-related failures and fire/explosion events reflects longer voyage durations and higher mechanical reliability demands under wave-climate variability [24, 27]. These contrasts are consistent with a context-mechanism interpretation in which the operational environment shapes which failure pathways are most likely to manifest — although causal attribution cannot be established from accident records alone, as the data do not capture exposure rates, vessel condition, or crew experience. The mode-specific patterns do, however, translate into concrete engineering control priorities: collision-prevention systems and capsizing-resistance measures for riverine vessels, and machinery-reliability standards and fire-suppression systems for offshore fleets.

An important contribution of this study is operationalising severity as two distinct dimensions: (i) the probability that an accident results in at least one fatality or missing person, and (ii) the casualty burden when casualties occur. Using the cleaned dataset (n = 1,523), RIV shows a higher proportion of severe outcomes (44.4%) than SEA (31.6%) despite SEA recording more accidents overall. This pattern is consistent with the interpretation that constrained waterways may amplify the consequences of operational failures, producing a higher observed likelihood of fatal outcomes per incident [20]. At the same time, offshore accidents show a higher mean casualty burden per event (2.29 vs. 1.14), a pattern consistent with the higher vessel capacity and offshore exposure conditions of open-water operations [23, 25]. The combined evidence argues against a single "more dangerous environment" narrative. These two severity dimensions have different engineering and governance implications: reducing fatal-outcome probability in RIV calls for vessel-stability improvements, collision-prevention controls, and passenger-survivability measures at the point of incident; reducing casualty burden in SEA points to vessel capacity management, search-and-rescue response time, and emergency equipment provision.

The logistic regression provides inferential support that zone differences in severe outcomes persist after controlling for vessel category and accident mode (adjusted OR 0.48; 95% CI 0.34-0.68; p < 0.001), consistent with an independent effect of operational environment on fatality likelihood beyond fleet and mode composition [20, 26]. The sensitivity model excluding UNK-V records yielded a directionally consistent result (adjusted OR 0.46; 95% CI 0.33-0.66; p < 0.001), confirming that the unknown vessel-type category does not drive the main finding. This does not contradict the higher mean casualty burden offshore; rather, it clarifies that operational environment exerts an independent influence on fatality likelihood over and above fleet composition. The engineering and governance implication is specific: if the goal is to reduce the number of incidents that result in fatalities, riverine settings require more targeted per-incident intervention than offshore settings — even after accounting for differences in vessel type and accident mode — while offshore settings require measures focused on limiting casualty escalation when incidents do occur.

The vessel-specific concentration patterns translate zone-level findings into spatially targeted risk governance. Offshore regions — the Java Sea, South China Sea, Makassar Strait, and eastern Indian Ocean — show the highest concentration of fishing-vessel incidents; the Java Sea additionally shows elevated cargo and passenger vessel involvement in Table 5 and Figure 1, consistent with evidence from high-density corridors [20, 26]. Inland hotspots on the Mahakam and Musi rivers reflect tug-barge and passenger-speedboat concentrations, consistent with restricted manoeuvring and infrastructure interactions [22]. These spatial signatures have direct engineering implications: the Java Sea multi-vessel hotspot warrants AIS-based vessel-separation monitoring and fishing-vessel stability enforcement; the Makassar Strait and South China Sea fishing concentrations indicate priorities in weather-based departure restrictions and life-saving equipment compliance; and the Mahakam and Musi river profiles point to tug-barge load regulation, speed limits for passenger speedboats, and improved channel marking in constrained waterways. Prioritising these spatially concentrated risk pathways over uniform national measures represents a more efficient allocation of safety-engineering resources.

The study has three primary limitations. First, the 2021 records represent a partial-year observation, as data collection did not cover all twelve calendar months uniformly; they are included in the severity analysis, and a sensitivity analysis restricted to 2018-2020 confirmed that all directional findings remain stable. Second, the analysis relies on accident records without exposure denominators such as traffic volume or vessel-days; consequently, the results reflect accident counts and proportions rather than absolute risk rates normalised by activity levels. Third, the hotspot analysis uses a descriptive concentration framework rather than probabilistic spatial modelling; future studies should apply kernel density estimation or Bayesian spatial methods to quantify clustering significance. Explanations offered in this discussion regarding environmental exposure differences, institutional enforcement, and emergency response constraints are interpretive inferences consistent with the observed patterns, rather than conclusions directly measurable from accident records alone. Future research should integrate AIS-derived exposure metrics to test whether zone effects persist when normalised by activity levels and to strengthen causal inference [20, 26].