AI-Driven Estimation of Vessel Sailing Times and Underwater Acoustic Pressure for Optimizing Maritime Logistics ^†

Abstract

This paper presents an innovative AI-based approach to estimate vessel sailing times in port surroundings. Leveraging historical vessel data, including ship characteristics and weather conditions, the model employs preprocessing techniques to enhance accuracy. Additionally, an underwater acoustic propagation model is used to study underwater noise pressure, aligning with environmental goals. The dataset, covering January to December 2022 in the Port of Cartagena, Spain, undergoes analysis, revealing intriguing patterns in ship routes. Employing various ML models, the study selects Random Forest as the most accurate, achieving an R2 of 0.85 and MSE of 0.145. The research showcases promising accuracy, aiding port optimization and environmental impact reduction, advancing maritime logistics with AI.

1. Introduction

Today, a large number of ships and nautical elements are active at sea. According to UNCTAD [1], approximately 80% of world trade is transported by sea, and this number is expected to further increase in the coming years [1]. In addition, shipping companies have been reporting over the years that disruptions and deviations from the initial plan occur frequently, resulting in most cases in delays [2]. These delays contribute to poor port optimization, disruptions in the market chain, and increased pollution, mainly greenhouse gas emissions and underwater radiated noise, due to prolonged idle times of vessels awaiting port calls. In fact, in April 2018, the IMO adopted the Initial Strategy for the reduction of GHG emissions from shipping which sets key ambitions, including cutting annual greenhouse gas emissions from international shipping by at least half by 2050 compared with their level in 2008.

This strategy goes in line with the Zero-Emission Waterborne Transport, the Horizon Europe partnership that aims to deliver and demonstrate zero-emission solutions for all major ship types and services before 2030. Ports over the world are starting to implement R&D tools to optimize their own performance and to be able to partner with these strategies. One example is the Port of Rotterdam, which has developed and implemented tools based on the Just In Time (JIT) arrival criterion, optimizing the speed of each vessel throughout its journey and reducing CO₂ emissions by 14%.

In this paper, we present an innovative approach that incorporates artificial intelligence (AI) models, specifically machine learning (ML), and preprocessing techniques, to estimate the sailing time of vessels in port surroundings. All of this is accomplished by leveraging historical vessel data, such as ship characteristics, movement patterns, weather conditions, and port-specific factors (docks and areas of action). Also, by implementing an underwater acoustic propagation model in each ship in its route, direct aspects related to the underwater noise pressure in the port context are studied. This study aligns with the MSFD, in particular regarding Descriptor 11 [3], searching for a balance between optimizing economical marine activities with good environmental status.

2. Study Area: Cartagena Port

This study encompasses two port docks of Cartagena Port (sited in Murcia, Spain) specialized in different traffic: the Cartagena Dock (sports marinas, cruise ships, and container cargo) and the Escombreras Dock (specialized in liquid and solid bulk traffic and storage activities) which was recently expanded, resulting in the port performing a hegemonic role in the management of this traffic throughout the Spanish Southeast. The port receives important flows that cross the Mediterranean, having dense networks with the Maghreb, the French, and the Italian coasts. The waters of the area are also furrowed by the local professional fishing boats and by the maritime traffic that connects the Atlantic Ocean and the Mediterranean Sea. Moreover, Cartagena receives part of the maritime passenger traffic that connects the Peninsula with the Balearic Islands. Additionally, more and more cruise lines are calling at the port (240,000 cruise tourists on 170 ships in 2019), being one of the national ports that is growing the most in this sense.

3. Methodology

Hence, in the present document, we propose a methodological focus on data analysis, emphasizing preprocessing, to enhance further predictions with machine learning models and underwater acoustic propagation models to assess the underwater radiated noise of ships in the port surroundings with the aim of reducing their impact.

3.1. Data Analysis

The dataset was derived from a Shiplocus (a multi-application platform for port management and maritime traffic exploitation (GMV)) account provided by the APC (Port Authority of Cartagena) through LIFE PortSounds (LIFE PortSounds. Reducing the impact of underwater noise on the marine environment of the Port of Cartagena (LIFE2020)) project. It consisted of 472 MB (1,585,941 × 32) of vessels and trajectories’ relevant parameters of the selected area (

Table 1. Impact Zone meshgrid coordinates.

Points	Lat	Lon
1	37°37,930′ N	01°10,613′ O
2	37°37,930′ N	00°33,988′ O
3	37°21,783′ N	01°10,613′ O
4	37°21,783′ N	00°33,988′ O

), which corresponds to the Impact Zone (IZ) of the APC. A preliminary data analysis was conducted for computational purposes, and thus undefined and incomplete data were removed, as well as irrelevant columns, maintaining ‘Latitude’, ‘Longitude’, ‘MMSI’, ‘Name’, ‘Date’, ‘Vessel type’, ‘SOG’, ‘COG’, ‘Length’, ‘Cargo’, and ‘Registered Owner’ parameters. In addition, ‘SOG’ (speed over the ground) was used to remove data coming from vessels moving at abnormal speeds, such as very low speeds (1.5 knots) or physically impossible speeds, given by the expression (1). Hence, the dataset obtained after the preliminary analysis consisted of 113 MB (58,632 × 11).

$$v_{max}=2.8\sqrt{L}$$

(1)

where L refers to the vessel’s length.

A route was defined as the union of successive AIS messages from a vessel, where successive messages are defined as those between which no more than 5 h have elapsed. Therefore, AIS messages remaining on the dataset after the data processing steps were used in the crafting of routes.

Given the dataset (consisting of a concatenation of AIS points), routes were transformed into the following features: ‘MMSI’, ‘Time spent’, ‘Vessel type’, ‘Length’, ‘Mean SOG’, ‘Cargo’, ‘Owner’, ‘Arrival date’, ‘Start point’, ‘End point’, and ‘Passing through’, where the new columns were:

• ‘Time spent’: The duration of the whole route.
• ‘Arrival date’: The timestamp where the route begins.
• ‘Start point’: A sectorization of the area was performed and key areas were defined, so the start point defines the key area where the route begins.
• ‘End point’: As with the start point, the end point is given by the key area where the route ends.
• ‘Passing through’: Coded as ‘YES’ or ‘NO’ based on whether a vessel is ending its itinerary in the port area or is just passing through the area but will not end its itinerary.

As can be seen in Figure 1, vessels show great similarities in their routes over the year, except for the tugs, which are always moving close to the docks and show significant uncertainty in the duration of the routes, ranging from 1 h to 22 h. This is due to tugs being vessels that reside in the port and are designed primarily for towing and pushing other vessels in harbors, canals, and other confined waterways. As tugs are vessels from the port, they were excluded from the route’s dataset. Also, in Figure 1d, anomalous routes can be observed. These routes were filtered to avoid confusions in the model.

With this, a thorough exploratory analysis was conducted to analyze route differences and similarities among vessel types to understand their behavior, as well as to sectorize areas depending on the traffic density. Thus, the sectorization made it possible to understand the key areas (areas with the higher density of AIS points) and to filter out those routes that did not start or end in a key area. Finally, a curated dataset (Figure 2) was obtained to be implemented in underwater acoustic pressure and machine learning analysis.

3.2. Underwater Acoustic Pressure

To assess the noise emitted by the vessels, the Ross model [4] was applied to each route. The Ross model considers physical vessel parameters, such as the speed and the length, and also parameters like the frequency to estimate the Source Level (SL), which is essentially the Sound Pressure Level (SPL) at 1 m distance of the acoustic source.

After the evaluation of the SL, a spherical loss model dependent only on the distance (for the sake of simplicity) was used to obtain a first approximation of the noise distribution over the area. Two cases were studied: one with a high number of vessels in the impact zone and one with just one vessel in the test site.

For the test with a high number of vessels, four ships (one cruise ship and three cargo ships) were selected from the data set. For these four ships, the SL was obtained at frequencies 62.5 Hz and 125 Hz, which are affected by cavitation noise. Once the SL was obtained, the spherical model was run in the area and the transmission losses were obtained. Finally, the SPL field that would be generated by each of the vessels was calculated and added coherently and 100% additively in linear units to obtain the worst case that could not occur. The physical properties of the 4 ships selected are shown in

Table 2. Physical properties of the vessels modelled as sources.

Type	Speed (Knots)	Length
Cruise Ship	15	247
General Cargo Ship 1	10.7	108
General Cargo Ship 2	10.7	108
General Cargo Ship 3	7	90

.

For the one-vessel case, the cruise ship from the first case was selected as the source to be modelled. The methodology is the same but this time no SPL maps need to be added, as there is only a single source in the area.

3.3. ML Models

Several machine learning (ML) models were tested to predict the time spent for a vessel knowing its arrival date, starting and ending point, and the passing through field. These models were Gradient Boosting and Random Forest Regressor. Gradient Boosting (GB) is a machine learning algorithm that uses an ensemble technique to create a more accurate prediction model from multiple simpler models. The main idea behind GB is to combine several weak models to form one strong model [5]. Random Forest Regressor (RFR) is a meta estimator that fits a number of classifying decision trees on various sub-samples of the dataset and uses averaging to improve the predictive accuracy and control over-fitting [6]. For each model, hyperparameter optimization techniques were conducted and the results were compared to select the best model.

4. Results

4.1. Underwater Acoustic Pressure Assessment

As for the noise assessment, the Source Level of every route was studied and classified for every type of vessel.

As depicted in Figure 3a, the SL generated by the vessels at the 62.5 Hz frequency mostly ranged between 150 and 165 dB. Also, for the 125 Hz component, the SL values were between 110 and 170 dB, where most values were distributed between 140 and 155 dB.

In Figure 4, an SL breakdown per type of vessel and frequency is shown. As can be seen, the lowest values found are from the FSV (Fishing Support Vessel) type. Also, the biggest values are from the cruise ship and the LNG tanker types, whose median values are up to 165 dB in (a) and up to 155 dB in (b).

Using the methodology described in Section 3.2, the SPL maps at 5 m depth for 63 and 125 Hz were obtained, both for the one-vessel case and the four-vessels case in the IZ of the Cartagena port (see Figure 5).

The highest SPL levels were found close to the source and ranged between 140 dB and 130 dB. In Figure 5, a slight difference between the two cases can be appreciated, as the highest SPL in the distance was found for the four-vessel case. For the four-vessel case, mean values of SPL for 62.5 Hz and 125 Hz, are, respectively, 75.8 $\mathrm{d}\mathrm{B}/\mathrm{k}{\mathrm{m}}^2$ and 65.3 $\mathrm{d}\mathrm{B}/\mathrm{k}{\mathrm{m}}^2$. For the one-vessel case, the values found are 72.6 $\mathrm{d}\mathrm{B}/\mathrm{k}{\mathrm{m}}^2$ and 61.7 $\mathrm{d}\mathrm{B}/\mathrm{k}{\mathrm{m}}^2$.

Between the one-vessel case and four-vessel case, the SPL differences found were close to 4 $\mathrm{d}\mathrm{B}/\mathrm{k}{\mathrm{m}}^2$. Also, it should be noted that the SPL levels found for the 125 Hz component are lower than for the 62.5 Hz due to the SL being lower for the 125 Hz.

It should be noted that these models were computed as a first approximation to understand underwater acoustic pressure, and that the 100% coherent and additive addition of the acoustic waves carried out for the four-vessel case will never occur in real terms.

4.2. ML Models

In

Table 3. Model summary and parameters.

Model	${\mathit{R}}_{\mathit{t}\mathit{r}\mathit{a}\mathit{i}\mathit{n}}^2$	${\mathit{R}}_{\mathit{t}\mathit{e}\mathit{s}\mathit{t}}^2$	MSE
Gradient Boosting	$0.884$	$0.8$	$0.198$
Random Forest Regressor	0.974	0.85	0.145

we can see the model metrics after the fine-tuning of the hyperparameters using a GridSearchCV algorithm, where it can be observed that RFR performs the best.

5. Conclusions

This paper presents the study and descriptive analysis of an AIS dataset with the aim of creating an ML-driven tool for the optimization of waiting times in the Port of Cartagena. In addition, a first approach to the visualization of the impact generated by the traffic from an acoustic point of view has been carried out. The machine learning model used was able to predict the transit time of vessels in the defined area with an MSE of 0.145. The acoustic models, although built as a first approximation, showed differences between different frequencies and different numbers of coherent vessels.

Future work will focus on the use of more complex models such as the MMPE (Miami–Monterrey Parabolic Equation) for a better estimation of transmission losses in the area. In addition, these more complex models will be used over a time range of several hours to see the noise signature left by the traffic. On the other hand, more complex models such as NN will be used for a better prediction of the sailing time.