HOTSPOTS FOR VESSEL-TO-VESSEL AND VESSEL-TO-FIX OBJECT ACCIDENTS ALONG THE GREAT LAKES SEAWAY

Transcription

1 0 0 HOTSPOTS FOR VESSEL-TO-VESSEL AND VESSEL-TO-FIX OBJECT ACCIDENTS ALONG THE GREAT LAKES SEAWAY Bircan Arslannur* MASc. Candidate Department of Civil and Environmental Engineering, University of Waterloo 0 University Avenue West, Waterloo, Ontario, Canada NL G Tel: + () 00-; barslann@uwaterloo.ca Frank F. Saccomanno Professor Department of Civil and Environmental Engineering, University of Waterloo 0 University Avenue West, Waterloo, Ontario, Canada NL G Tel: + () -0; saccoman@uwaterloo.ca Word count: words text + tables/figures x 0 words (each) =, words November th, Corresponding author

2 Arslannur, Saccomanno ABSTRACT Over the past decade, an average of. vessel accidents per year have been reported along the Great Lakes Seaway (GLS), the vast majority of which tend to be clustered at specific unsafe locations or sites along the route. Sites with unacceptably high potential for accidents are referred to as hotspots, and these hotspots become prime candidates for future safety intervention. Given the rare random nature of accidents, the identification of hotspots must be based on robust sitespecific prediction models of accident expectation. This paper presents an empirical Bayes prediction model developed for GLS that considers four types of accident scenarios: vessel-tovessel (VV) and vessel-to-fix objects (VF) for river and canal/lock sections. Hotspot sites are determined using two risk tolerance thresholds: th % exceedance (high risk sites) and th % exceedance (moderate and high risk sites). For the th % threshold and VV accidents, a total of hotspots were identified over the 00 km length of GLS being studied (excluding lake or port areas). Of the designated hotspots sections, km (0.%) were located along natural river courses and the rest at canals/locks. For VF accidents, all the high risk hotspots were located at canal/lock sections (. km). Reducing the threshold to th % resulted in a.% increase in seaway length that is designated as a hotspot. The locations of these hotspot sections along the GLS were consistent for both thresholds. Keyword: Vessel Accidents, Empirical Bayes, Hotspot Identification, Great Lakes Seaway

3 Arslannur, Saccomanno INTRODUCTION The GLS is a deep draft waterway extending for,00 km (,0 miles) from the Atlantic to the headwaters of Lake Superior. Since, more than. billion tonnes of cargo estimated at US$ billion have been transported to and from Canada, the U.S., and nearly fifty other nations. About % of this traffic travels to and from overseas ports, especially Europe, with the remainder being internal GLS movements (). In, a total of,,00 gross tonne were carried on the St. Lawrence, Montreal-Lake Ontario section of the seaway by vessels (). As illustrated in Figure the portion of the seaway considered in this paper extends for about 00 Km from the Gulf of the St. Lawrence to the Sault St. Marie canal/lock sections between Lake Huron to Lake Superior. This stretch of seaway consists of approximately 0 km of canal/lock sections and 0 km of river. + Canal Sections Rive Sud ( km) Beaharnois ( km) Wiley Dondero ( km) Iroquois Lock ( km) Welland ( km) West Neebish ( km) Soo Locks ( km) FIGURE GLS from Rimouski to Sault Ste. Marie with mileage offsets. While there are sections of the seaway near the Gulf of the St. Lawrence with widths in excess of 0 km, much of the St. Lawrence (00 km of length) consists of narrower river sections and canal/lock sections with widths ranging from 0 m to km. Restrictions to vessel movement caused by reduced channel width can be a major cause of vessel accidents along the GLS. For this study, vessel accidents that occurred within inland ports and on the open lakes will not considered in this study. GLS ACCIDENT HISTORY A total of 0 vessel accidents and safety occurrences were reported throughout the entire system in Canada and the U.S. for the period of 0- (). Most of the marine accidents were minor in nature with % of the accidents reported in the U.S. waters resulting in no damage to the vessel and.% of accidents reported in Canadian waters classified at the minimum severity ranking.

4 Accidents Arslannur, Saccomanno 0 Figure illustrates the frequency of more serious vessel accidents per year over a year period from 00 to. Over the past years the distribution of accidents has varied from a low of in to a high of in 0, with an average rate of. accidents per year. Each year, the GLS is operational for about an month period between March and December. From Figure we note that no appreciable reduction in the frequencies can be ascertained, despite several improvement dressy safety operations and standards that meet or exceed international water transport protocols Year FIGURE Number of freight vessel accidents occurred in GLS per year. The reduction of vessel accidents needs to be an integral part of cost-effective safety management. In this paper, we argue that vessel accidents can best be reduced by providing answers to two fundamental safety questions: ) Which sites are unsafe such that some form of intervention or countermeasure is warranted? and ) What form should this intervention take, such that accidents are reduced in a cost-effective and practicable manner? The focus of this paper is on the first question, i.e. to identify unsafe sites for intervention. We refer to these targeted sites (or GLS sections) as hotspots. The first step in identifying hotspots is to develop and calibrate a robust vessel accident prediction model for the GLS. This model serves to provide estimates of the expected number of vessel accidents (or accident potential) at different locations of the seaway for different types of accidents. The expected number of accidents obtained from this model can then be compared to suitable risk tolerance criteria (as specified independently) to highlight locations of unacceptable risk (hotspots). (, ). Once hotspots have been identified, the nature and causes of likely or historical accidents can be investigated, and appropriate countermeasures developed for accident reduction. The problem of countermeasure development and evaluation is not considered within the scope of this paper. The natural tendency in hotspot identification is to use observed accident frequencies directly to measure lack of safety and identify hotspots. This is not readily possible given the rare random nature of accidents with significant year-to-year fluctuations. At the site specific level, this year-to-year fluctuation in annual accident frequency can be more pronounced than was shown in Figure for the entire length of GLS. For a given site for example, accidents in one year can be quite high (site designated as unsafe), while for another year they could be low (site designated as safe). To account for these year-to-year fluctuations in the historical data, it becomes necessary to develop site-specific prediction models that provide a measure of the potential or expectation for accidents at specific locations ().

5 Arslannur, Saccomanno 0 In this paper, several empirical Bayesian (EB) accident prediction models have been developed and evaluated using GLS data. These models are used to designate hotspots for two types of accidents: vessel-to-vessel (VV) and vessel-to-fixed objects (VF) and two types of route: canals/locks and river channels. ACCIDENT DATA USED MODEL CALIBRATION In this research, only freight vessel accidents as reported by the TSB for canals/locks and river channels have been considered. Between 00 and there were 0 such accidents (about 0% of total GLS). Figure illustrates the breakdown of these accidents. VV accidents involve a collision between two or more vessels in-transit (head-on and side impact), while VF accidents include, running aground, or hitting a fixed object within the channel (either natural or man-made). Accident Occurrence Percentage Chart 0% 0% 0% 0% % 0% Vessel-Vessel Vessel-Fixed Object Vessel-Vessel Vessel-Fixed Object per Km per Km River Canal 0.. Canal River FIGURE Percentage of marine accident occurrences in the GLS between 00 and. Over the year period, about 0% of VV accidents were found to take place at canalsand locks, with the remainder occurring along river sections. Of VF accidents, took place at canals and locks, while took place along river sections (roughly equally split). However, river sections of the GLS under consideration comprises a total length of,0 km, while canals and locks represent a total length of 0 km. Hence, on a route Km basis most VF accidents tend to be concentrated at canals and locks (approx. %). This is expected given the narrower width involved and restrictions on vessel manoeuvrability increasing the likelihood of collision with the sides of the navigation channel. In this analysis, to obtain the length and the width of the GLS, one week of freight vessel Automatic Identification System (AIS) data was used. Historic routes (longitudes and latitudes) of freight vessels were obtained from one week of AIS data on the GLS. AIS data gives the specific routes that are typically followed by vessels. This dataset was used to pinpoint the location of the vessel as it progresses along the seaway. Information on length and width of GLS sections was obtained from ArcGIS software () for the appropriate AIS vessel location. In this way, channels that are not navigable were not included in the GLS route determination. The full length of the seaway was divided into geographic sections with uniform physical features using ArcGIS software, sections of canals and locks, and natural river courses. Individual sections were determined based on similar geometric features. (For example, a given section could reflect a curve with the same radius from its beginning to its endpoint.) This was carried out using physical inspection of satellite GLS images.

6 Arslannur, Saccomanno 0 0 Canal shipment data were extracted from GLS Management Corporation annual reports for the period 00 to. River vessel traffic along the seaway between Montreal and Kingston was assumed to be similar to canal traffic in the same vicinity, since separate data on river traffic volumes was not available. For the Rimouski-Montreal, Detroit and St. Clair River segments, we obtained traffic data from the Canadian Coast Guard for years between 0 and. For the St. Mary River segment, traffic data were obtained from the US Coast Guard. EB ACCIDENT PREDICTION MODEL An EB model approach has been adopted in this study as the basis for the vessel accident prediction model. The EB approach has been examined and explored by several researchers and was found to provide reliable site-specific results (with reduced over-dispersion error and regression-to-themean bias) (-). In the EB approach the best estimate of expected vessel accidents at a specific site is obtained by combining two sources of inference:. Historical observed accidents (y) for a specific site (data likelihood). Expected accident frequency for similar sites obtained from a calibrated safety performance function (or prior). The expected number of vessel accidents ( ) or posterior is obtained, such that: i E( y ) ( ) y [] i i i i i i i with, E( ), and Var( ) ( ) E( ) [] i i i i i The timeframe for prediction can be a year (as per these expressions) or any specified period. where i = EB expected annual accident frequency at site i, = expected annual accident frequency at similar sites (i.e., from safety performance i functions), yi = observed crash frequency in n years at site i, and = weight factor. i The weight factor i is estimated from the mean and variance of the safety performance function (SPF), such that: Ey ( i) i [] Var( yi) E( yi) A number of SPF expressions were investigated for the GLS data for different geometric and operational input factors. The expressions that yielded the best statistical results included: section width (km), section length (km) and annual vessel passages. A separate regional affiliation term is meant to explain for the discrepancy that we found in the number of VF accidents between areas. For instance, we would expect to have more VF accidents in higher traffic areas such as the Welland Canal compared the lower traffic areas of the St. Lawrence River canals. However, this is not the case. One of the factors that may contribute to the increase of accidents in the St. Lawrence River canals despite less traffic in this area compared

7 Arslannur, Saccomanno 0 0 to the other canal sections are the number of canal entrance points. The bulk of VF accidents in a canal occurred at its entrance. Due to the geometric nature of the St. Lawrence River there canal sections separated by portions of river. As a result, there are eight canal entrance points along this seaway. On the other hand, other GLS canals consist of a continuous uniform nature (same width and geometry through their length). And, therefore, only two entrances to these sections. In the paper, this difference between the St. Lawrence and Great Lakes was accounted by a regional input variable. In the model, this region effect is designated as for the St. Lawrence canals and 0 for the other canals in the GLS. R-software () was used to fit various Generalized Linear Models (Poisson and Negative Binominal) for VV and VF impact types. These expressions are of the form: T( x z) Li ln( E(y i)) ln( ) [] Wi T( x z) Li ln( E(y i)) ln( ) Dvar(i) [] Wi where yi = observed crash frequency in n years at section i T(x-z) = Total number of freight vessel traffic the years between x and z. Li = length of the section i Wi = average width of the section i Dvar(i) = region effect at the section i The Generalized Linear Model-Negative Binominal (GLM-NB) distribution was found to yield better results for VF accidents on river sections, while a GLM-Poisson distribution was found to yield the better results for the VF canal and all VV models. From the investigated accident reports and accident history data, it was concluded that other variables such as, current, wind and time had no significant effect on the vessel accident frequency along the GLS. Table summarizes the organization of the GLS vessel accident/traffic data for input into the SPF and the EB posterior. The EB posterior makes use of the SPF or prior input data and separate observational data for data likelihood. This Table illustrates the data inputs for the two types of vessel impact model and two GLS section types. For instance, years of data between 0 and 0 were used to fit various GLM for VV accidents on the river sections. These data were used to establish the EB prior or SPF function. The best fit GLM for the posterior was established using years of data between and for the data likelihood or observational component. TABLE Data for SPF (Prior) and Data Likelihood Inputs by Type of Accident and Section VV VF Section Type Prior Data likelihood Prior Data likelihood River Canal/Lock SPF AND EB PREDICTION MODEL RESULTS Table and summarize the SPF results for river and canal/lock sections, respectively. Separate models were calibrated for VV and VF impact types. The ratio of the residual deviance to the

8 Arslannur, Saccomanno 0 degrees of freedom yielded a dispersion parameter, which measures the degree of over-dispersion or unexplained variation in the site accident experience as compared to what the model predicts. A value of for this parameter suggest no over-dispersion and the model is assumed to fully capture the site-specific variation in accident frequency (). For the river section in Table, the degree of dispersion for VV and VF accidents was found to be 0. and.0, respectively, and for these models both Poisson and NB distributions were considered acceptable. In this paper, we selected a NB link function for the VF river, and a Poisson link function for the VV river model based on the significance of the input factors in Equation and. TABLE Prior Model Results for River Sections Outputs VV River VF River coefficient p-value coefficient p-value β β Residual Deviance 0.0. Degrees of freedom AIC.. Dispersion parameter 0..0 For the canal-lock, the SPF VV and VF expressions yielded dispersion parameters of.0 and., respectively, suggesting that for VF accidents the data is slightly over-dispersed. Analysis of both Poisson and NB link functions, however, yielded statistically more robust inputs for the Poisson link function. Hence in this paper a Poisson link function was used for canal/lock accident prediction for both VV and VF accident types. From Table and Table, we can conclude that all the variable parameters in the fitted expressions were highly significant at the % level. TABLE Prior Model Results for Canal-Lock Sections Outputs VV Canal/Lock VF Canal/Lock coefficient p-value coefficient p-value β β β Residual Deviance. 0.0 Degrees of freedom AIC..0 Dispersion parameter.0. The data likelihood component of the EB posterior makes use of data for a year period (-) for both VV and VF accidents at rivers and canal/lock sections (Table ). The four year period for data likelihood is considered sufficient to capture long term year-to-year variations in the observed accident frequencies (). In total there are river and canal/lock sections over the entire length of the GLS. For demonstration purposes a 0 Km stretch of GLS along the St. Lawrence River was selected to report the posterior estimates of expected frequencies for both VV and VF accident types. The results are summarized in Table. For these results, the prior µ and the posterior E(µ/y) have are reported for a combined year period, as is the number of observed accidents. VAR(y) was calculated based on the section s dispersion parameter using the summation of years data. To

9 Arslannur, Saccomanno obtain the -year expected value, each year s expectation was summed ( E(µ/y)) for VV and VF accidents. TABLE Posterior Model Results for the St. Lawrence River VF accident results on St. Lawrence River VV accident results on St. Lawrence River Section # y µ Var(y) α E(µ/y) y µ Var(y) α E(µ/y) ANALYSIS OF HOTSPOTS The th and th percentiles were obtained based on the distribution of years total expected accidents per section divided by the section length and the year total volume and multiplied by 0,000 for scaling. The th and th percentile represents the expected year total number of accidents that are exceeded by % and % of sections along the GLS. Basically, these expectations reflect higher than average values for seaway conditions. For this paper, sections with expected number of accidents exceeding the th and th percentiles are deemed to be hotspots. The th and th year percentiles for VV accidents were estimated to be 0. and 0., respectively. For VF accidents, the th and th percentiles were estimated to be. and., respectively. These assumed thresholds are juxtaposed on the expected accident distributions for VV and VF types along the GLS, as illustrated in Figure. FIGURE Distribution of years total expected number of accidents.

10 Arslannur, Saccomanno Table provides a list of GLS hotspots for the th percentile (dark color) and th percentile (light color) thresholds for VV and VF accidents. The Table includes information on mileage offset, section type and E(µ/y)/ yr. For the higher risk th percentile threshold a total of sections (. km,.0% of the seaway) were identified as hotspots for VV accidents, and sections (. km, 0.% of the seaway) for VF accidents. These hotspots are identified out of a total of sections considered (00 km) along the GLS. A total of sections ( km,.% of the seaway) were identified as hotspots when both types of accidents (VV and VF) were combined. For the lower risk threshold of th percentile, the number of hotspots increased to sections for VV (. km or.0% of GLS length) and sections for VF (0. km,.% of GLS length). A total of sections (. km,.% of the seaway) were identified as hotspots when both VV and VF accidents were combined into a single model. Reducing the threshold by % has had the effect of adding sections to the list of hotspots for all accident types. Seven of these hotspots were isolated sections (. Km), while hotspots (. km) were found to be extensions of the th percentile sites. TABLE Hotspots for the GLS Section # Mileage Offsets (km) Region Section Type VV Accidents E(µ/y) per km* % value VF Accidents E(µ/y) per km* % value 00+ St. Lawrence River Canal St. Lawrence River Canal-lock St. Lawrence River Canal St. Lawrence River Canal St. Lawrence River Canal-lock St. Lawrence River Canal-lock St. Lawrence River Canal-lock St. Lawrence River Canal-lock St. Lawrence River River St. Lawrence River Canal-lock Welland Canal Canal Welland Canal Canal-lock Welland Canal Canal-lock Welland Canal Canal-lock Welland Canal Canal-lock Welland Canal Canal Welland Canal Canal Welland Canal Canal-lock Welland Canal Canal Detroit River River Detroit River River Detroit River River St. Mary's River River St. Mary's River Canal St. Mary's River River For the th percentile threshold, canal and lock sections are decidedly more unsafe than river sections, although the actual Km designated as hotspots is significantly greater for the th percentile. The table represents the total linear km distance from the start point for both VV and VF hotspots. For the th percentile threshold, hotspot sections were found to be located on canals

11 0.. Arslannur, Saccomanno and one hotspot on a river course. For th percentile value, the number of hotspots increased to for canals and for river sections. Although there are more hotspot segments on canals than rivers for the th percentile, the total route length of river hotspots is greater than for canals (see Figure ). Km of hotspots Canal River th percentile FIGURE Km of hotspots on canal and river. A factor that significantly affects the expected number of accidents (VV and VF) on the GLS is navigation channel width. The average width of navigation channels along the GLS is. km, while the average width of the hotspot sections for the th percentile is 0. km and only 0. km for the th percentile. This shows when the width of the section is decreased, the risk of accident increases significantly. Table indicates the number of hotspot sections and their length in Km for four different regions of the GLS. Both th and th percentile thresholds are shown. The highest risk sites tend to be located in the Detroit River-St. Clair River sections followed closely by the St. Lawrence River. This is especially true if adjustments are made to consider the section length and the concertation of risk. The highest expected number of hotspots are found along the St. Lawrence for both th and th percentiles. TABLE Number and Km of Hotspots on Each Region th percentile Number of Hotspots Km of Hotspots Region th percentile value th percentile value th percentile value th percentile value St. Lawrence River. Welland Canal.. Detroit River-St. Clair River 0. St. Mary's River 0. 0 Figure and Figure show the hotspot locations on Welland Canal and the north entrance of Rive Sud Canal, Montreal. The thin line corresponds to hotspots based on the th percentile, while the thicker line corresponds to hotspots for the th percentile for selected sections of the GLS. These figures illustrate that a % reduction in threshold ( to th percentile) yields an extension of the hotspots zones along the segments, and that hotspots identified for the th percentile tend to be located within the length of channel hotspot for the th percentile. For canal sections most of high risk sites (at th percentile) are found at the canal entrance/exit points; whereas for the lower th percentile, these hotspot zones tend to be extended over the length of the canal.

12 Upbound Traffic Downbound Traffic Upbound Traffic Downbound Traffic Arslannur, Saccomanno Legend Route th Percentile th Percentile FIGURE Hotspot locations on Rive Sud Canal, Montreal. Legend Route th Percentile th Percentile FIGURE Hotspot locations on Welland Canal.

13 Arslannur, Saccomanno 0 0 CONCLUSIONS The EB approach adopted in this paper has provided a robust objective basis for predicting vessel accidents along the GLS. types of prediction models were developed for different types of accidents and navigation channel features: namely vessel-to-vessel collisions for river and canal/lock sections and vessel-to-fix objects collisions for river and canal/lock sections. Major confounding factors that were found to be statistically significant in these models were: vessel volume (number of vessels per year traversing each section), length of section (in Km) to account for exposure, navigation channel width (in Km), and for vessel-to-fix object accidents a dummy variable to account for regional affiliation. The application of the model for GLS hotspot identification necessitated the introduction of risk tolerance thresholds to guide decisions as to when safety intervention is warranted. In the absence of objective risk tolerance criteria, this paper made use of the distribution of expected number of vessel accidents to establish the th and th percentiles. These values serve as high and moderate tolerance levels for hotspot identification. The application of the EB accident prediction models yielded a total of VV (. km) and VF (. km) hotspots for the th percentile threshold and VV (. km) and VF (0. km) for the th percentile threshold. For the th percentile threshold, canals and locks were found to be more unsafe than river sections for VV and VF accident types combined. A total of hotspot sections were found to be located on canals and one hotspot on a river course. For th percentile value, the number of hotspots increased to for canals and for river sections. The model suggested that a reduction in risk threshold from the th to th percentile resulted in a more extensive length of GLS channel designated as a hotspot, although the general locations of these hotspots along the route were similar in nature. Essentially the th percentile hotspot sections were included as a subset in the th percentile hotspots. The model introduced in this paper provides a reliable cost effective method for allocating scarce safety budgets to those GLS locations where these funds have a greater potential for higher safety dividends. REFERENCES. The Economic Impacts of the Great Lakes-St. Lawrence Seaway System. October,. Martin Associates, Lancaster, Pennsylvania. Accessed July,.. The St. Lawrence Seaway Traffic Report Navigation Season. The St. Lawrence Seaway Management Corporation and the Saint Lawrence Seaway Development Corporation, St. Lambert, Quebec. en.pdf. Accessed July,.. Safety Profile of the Great Lakes-St. Lawrence Seaway System. March,. Research and Traffic Group, Glenburnie, Ontario. Profile-ExSum.pdf. Accessed November,.. Montella, Alfonso. "A comparative analysis of hotspot identification methods." Accident Analysis & Prevention, Vol.,, pp. -.. Saccomanno, F., Fu, L., & Miranda-Moreno, L. Risk-based model for identifying highwayrail grade crossing blackspots. In Transportation Research Record: Journal of the

14 Arslannur, Saccomanno Transportation Research Board, No., Transportation Research Board of the National Academies, Washington, D.C., 0, pp. -.. ESRI (). ArcGIS Desktop: Release.. Environmental Systems Research Institute, Redlands, CA. Accessed July,. Persaud, Bhagwant, Dominique Lord, and Joseph Palmisano. Calibration and Transferability of Accident Prediction Models for Urban Intersections. In Transportation Research Record, No., Transportation Research Board of the National Academies, Washington, D.C., 0, pp. -.. Hauer, Ezra, Douglas W. Harwood, Forrest M. Council, and Michael S. Griffith. Estimating Safety by the Empirical Bayes Method: A Tutorial. In Transportation Research Record, No., Transportation Research Board of the National Academies, Washington, D.C., 0, pp. -. Miranda-Moreno, L., Fu, L., Saccomanno, F., & Labbe, A. Alternative risk models for ranking locations for safety improvement. In Transportation Research Record: Journal of the Transportation Research Board No. 0, Transportation Research Board of the National Academies, Washington, D.C., 0, pp. -.. R Core Team (). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN , Accessed July,.. McCullagh, P. and A.J. Nelder. Generalised linear models. nd edition. Chapman and Hall/CRC, Boca Raton, Florida,.