Random Forest Classifier-based Ship Type Prediction with Limited Ship Information of AIS and V-Pass

Abstract

Identifying ship types is an important process to prevent illegal activities on territorial waters and assess marine traffic of Vessel Traffic Services Officer (VTSO). However, the Terrestrial Automatic Identification System (T-AIS) collected at the ground station has over 50% of vessels that do not contain the ship type information. Therefore, this study proposes a method of identifying ship types through the Random Forest Classifier (RFC) from dynamic and static data of AIS and V-Pass for one year and the Ulsan waters. With the hypothesis that six features, the speed, course, length, breadth, time, and location, enable to estimate of the ship type, four classification models were generated depending on length or breadth information since 81.9% of ships fully contain the two information. The accuracy were average 96.4% and 77.4% in the presence and absence of size information. The result shows that the proposed method is adaptable to identifying ship types.

1. Introduction

Marine accidents related to safety and the environment have increased since marine use has become complex and diverse. There have been 14,042 maritime accidents with an increasing rate of 0.8/year, and 1,459 kL (approximately 292 kL per year) oil spillage in recent – five years (2017 2021) in the Republic of Korea (hereafter referred as to Korea). Thus, the interest in maritime surveillance has steadily increased to secure the marine environment and resources and prevent maritime accidents and illegal activities (MOF, 2022).

Coast guards of countries have utilized marine traffic reporting systems such as the Automatic Identification System (AIS) and Vessel Monitoring System (VMS). In addition, Korea has a unique fishing boat tracking system called V-Pass. Furthermore, space-born remote sensing-based marine surveillance also has been developing in countries (Pappas and Achim, 2018; Lee and Lee, 2022).

AIS data contain dynamic and static data. Since dynamic information such as longitude and latitude, speed and course are reported within a minute scale, the movement of a ship can be specified (Park et al.,    2018). And the static data, including the identification number, name, length, width, type, and navigation status, and thus the identification of a ship is feasible. AIS communication is made between ships as well as between ship and land (MOF, 2014), and it has been broadly utilized to prevent ship collisions at sea (Park, 2015). Although the equipment and the use of AIS are regulated by the International Convention as well as domestic law for every country, the obligation is only applicable to ocean-going ships over 300 gross tons, non-ocean-going ships over 500 gross tons, coastal ships over 50 gross tons, and all kinds of passenger ships. Thus, ships less than 50 gross tons are not responsible for equipping the AIS, and they are fishing boats in most.

On the other hand, V-Pass is the ocean wireless communication system to secure fishing ships’ safety and automate the record of their entering and leaving a harbor in Korea (Han, 2021). Around 94% of registered fishing ships are less than ten gross tons, and the Korean government enforces the ships to equip the V- Pass for safety reasons (MOF, 2019), and is regulated by domestic law. V-Pass also has dynamic and static information like AIS contains.

Among the dynamic and static information, ship type is significant to estimating a ship’s maneuvering characteristics, such as its submerged depth (draft), its accordance maneuverability, and its route in the near future. For this reason, the ship type information has been conventionally used by navigators and Vessel Traffic Services Officers (VTSOs) since the main navigation route depends on the type (Jeon and Jung, 2018; Lee and Cho, 2022), and the geometrical relation between ships in the near future also be predictable (Lee and Song, 2017). Furthermore, ship type is the first factor to be considered to prevent maritime accidents since the damage amount from collisions and oil spills also highly depends on the type (MOF, 2022).

Despite that, 25–60% of AIS data do not contain ship type (Hong and Yang, 2014), and the problem is from equipment failure (Kazimierski and Stateczny, 2015; Emmens et al., 2021), the failure of message transmission (Harati-Mokhtari et al., 2007; Greidanus et al., 2015; Emmens et al., 2021), and unfamiliarity of seafarers to handling AIS equipment (Harati-Mokhtrai et al., 2007; Johansson et al., 2013; Emmens et al., 2021). Obviously, the insight of identifying ship type can be from long-term and repeated experience in monitoring vessel traffic. However, the insight is limited to sharing with other members.

On the other hand, there have been researches to predict ship types using the length and width information. Here, length and width indicate Length Overall (LOA) and breadth, respectively. The method has been mainly used in classifying the type of ships detected in satellite imagery in the remote sensing field, and the method is categorized into three types of use cases; length and width (Heiseberg, 2016; Liu et al., 2017), dimension (Deng et al., 2013), the ratio of length to width (Jubelin and Khenchaf, 2014; Xiaoyang et al., 2016; Jeon and Yang, 2021). However, the AIS is applicable to the method above only if the length and width information is provided.

This paper, therefore, suggests a method of identifying ship types using the Random Forest Classifier (RFC) after additionally defining the main dynamic features, time, location, speed, and course. Cargo, tanker, and passenger ships have releatively regular schedules and routes, and they are reflected in the number of recorded points at a specific time, course, and location. Fishing ships have various speeds and courses at sea, but they also make clusters with low speeds where a fishing resource is abundant depending on the seasons. Thus, this study begins with the hypothesis that ship type can be determined if the four features are given and aim to reveal the performance, and the power of explainability of length and width is also confirmed through the RFC model. Unlike RNN-based models, which require time-series data to predict ship type (Park et al., 2021a; Park et al., 2021b), this paper focuses on ship type identification with given dynamic information at a timestamp.

RFC is an ensemble machine learning model that makes a vote on multiple estimation results of the decision tree classifier (DTC) and determines the most as the final result. DTC is intuitive and strong at nonlinear classification and thus explainable of the determination process. Thus, the decision from multiple DTC is equivalent to that of multiple navigators and VTSOs in identifying ship types.

The remainder of this paper consists as follows. Section 2 introduces the coverage and feature of the study area in Ulsan, and descriptive statistics of AIS and V-Pass in the area. The overall methodology, including preprocessing and generation of training and testing of the RFC is described in section 3. Section 4 discusses the result of ship type prediction. Finally, the conclusions are drawn, and the future study is introduced in section 6.

2. Study Area and Data

The study area is Ulsan adjacent sea, ranging from 129.3 to 129.57° longitudinally, and from 35.3 to 35.6° latitudinally. The prediction target is limited to navigational ships not entering/departing/berthing the harbor, and not anchoring in the adjacent water. Thus, AIS and V-Pass data 2km offshore only used in this study (Fig. 1).

Fig. 1. Study area. 2 km off shoreline (blue line) is made to eliminate the vessels entering/departing/berthing the harbor and anchoring in the adjacent water.

Ulsan port is the major harbor of tankers in South Korea (Kim et al., 2011). The type of tanker has the highest proportion (44,513 ships and 34.6%) of overall ships (128,626 ships). Among the total ships in the AIS and V-Pass integrated data (hereafter integrated data), 81% of ships have information on their type, and 73% of ships have their size information (Table 1).

Table 1. Presence ratio of size information according to ship type information presence (Aug. 1, 2018)

L: length overall, B: breadth, None: none of L&B

3. Methodology

1) Extraction of ship data outside of 2 km off shoreline

2 km off shoreline is determined to eliminate ship data that are entering/departing/berthing in Ulsan por since ships having speed less than 1.0 knot can be confused with anchoring ships or fishing boats in the port adjacent water. That is, this paper copes with the ships over 2 km off from the shoreline.

To set the 2 km off the shoreline, a Global Self- consistent, Hierarchical, High-resolution Geography Database (GSHHG) was used. The GSHHG has been developed and operated by Hawaii University and the National Oceanic and Atmospheric Administration of the USA (Wessel andSmith, 1996). The 2km buffer is computed by using QGIS 3.0 (QGIS Development Team, 2019), and ship data extraction was made by function “cover” in R package {raster} (Hijmans and Etten, 2012).

Table 2. The categories of ship type with common nature

2) Recategorizing of ship types

AIS categorizes ship types through code numbers from 0 to 99 in integer format. Because the one hundred ship types are not efficient for classification, the ship types are rearranged following the two simple rules. Firstly, the types of passenger, tanker, wing-in-ground (WIG), high-speed craft (HSC), other, reserved, and sparse are unified by their type name since they have multiple codes. Secondly, the ship types with similar characteristics are also unified following their common nature (Table 2).

Accordingly, there are 13 ship types; cargo, tanker, passenger, fishing, pilot, pleasure craft, special, tug & tow, warship, for public, HSC, other, and spare & reserved.

3) Descriptive statistics

The basic statistics of 583 ships were recorded in the AIS and V-Pass integration data on August 1, 2018, the date when no typhoon warn (Table 3). In the proportion of ship types, the top three are fishing boats (33.8%), tankers (24.0%), and cargo ships (10.0%), and the bottom three types are high-speed boats (0.2%), pilot ships (0.3%), and warships (0.5%). Depending on the presence of size information (length and breadth), the top five types are pilot ships (100%), warships (100%), cargo ships (98.0%), tankers (97.1%), and fishing vessels (95.9%). On the other hand, the bottom three types are high-speed boats (0%), leisure boats (20.9%), and unknown types of boats (21.1%).

From the proportion of retaining size information, cargo ships and tankers that cause enormous damage in a marine accident, and pilot ships that support the entry and departure of these ships, equip AIS with size information. Although warships (100%) are known not to report their location through AIS for security reasons, if they provide, they show their size clearly. Non mainstream pleasure (20.9%), other (62.5%), and unknown types (20.6%) did not have size information properly.

Table 3. Presence ratio of length information dependent on ship type

Fig. 2. Distribution of vessels’ traces. Dataset is interpolated 5 minutes in equal and grids are set approx. 0.05° in equals 5 km. Vessels 2 km off from shoreline (blue line) are only used as datasets.

DTC chooses the most frequent variable in the classification process. Prior to calculating the frequencies, it is required to prepare standard spatial and temporal units. Firstly, AIS and V-Pass data have irregular temporal intervals. Thus, data should be interpolated with five minutes intervals. Secondly, it is necessary to consider the main routes on the ship types as well as the standard spatial unit. Thus, the study area is divided at approximately 0.05° (approximately 5.5 km) at equal intervals, and the vertical and horizontal orders of zone X and Y are assigned. (Fig. 2).

Subsequently, the AIS and V-Pass data are organized with the following variables; time, speed over ground (SOG), course over ground (COG), length, breadth, and area coordinates X and Y.

4) Random Forest Classifier (RFC)

RFC is a model that presents the final judgment by voting the primary class from the decision from multiple DTCs (Ho, 1995). It has shown excellent performance among many classification and prediction models (Park et al., 2018; You et al., 2019; Lee et al., 2019).

The features and labels have to be defined and they should be divided into training data and test data; training features, training labels, test features, and test labels.

As the features, time, SOG, COG, length, breadth, and zone coordinates X and Y are defined, and ship types are defined as labels. On the other hand, the test features and labels should be independent of the target date of classification. In addition, typhoons may affect the ship’s speed and course; thus, vessel type identification is performed on August 1, 2018, when no typhoon is in Ulsan Port.

RFC is implemented using the Random Forest Classifier module of the Python Library {Scikit-learn} (Pedregosa et al., 2011). RFC is composed of multiple DTCs, and the DTCs have multiple child nodes from the root node, and they lead to a conclusion by making positive/negative decisions in every child node. Here, the criterion for determining whether it is purely positive (or negative) is the Gini coefficient (Eq. 1). A node is called “pure” if all samples belonging to a node are in an identical label.

\(G _ { i } = 1 - \sum _ { k = 1 } ^ { n } p _ { i , k } ^ { 2 }\)       (1)

i: No.of node

k: feature

The {Scikit-learn} uses the cost function of the Classification and Regression Tree (CART) algorithm that assigns weights depending on the number of samples to obtain the Gini coefficients (Eq. 2, Fig. 3).

\(\left. \begin{array} { c } { J ( k , t _ { k } ) = \frac { N _ { \text { left } } } { N } G _ { \text { left } } + \frac { N _ { r i g h t } } { N } G _ { \_ r i g h t } } \end{array} \right.\)       (2)

J: cost function

j: feature

t_k: threshold for feature k

The “GridSearchCV” function of {scikit-learn} is used (Pedregosa et al., 2011) to determine the hyperparameters for optimum learning. Hyperparameter is defined as the parameter closely related to model optimization and cannot be set by the model itself.

Four hyperparameters are set; Bootstrap, the number of estimators, maximum features, and maximum depth. Bootstrap is the sampling method with replacement or not, the number of estimators means the number of DTCs consisting of the RF, and the maximum depth is how many generations of child nodes will be made.

Fig. 3. Structure of Decision Tree and Random Forest Classifier. (a) Example of Nodes of DTC, (b) Schematic diagram of RFC.

However, the ‘GridSearchCV’ requires long time to handle large data, and the computation duration depends on hyperparameter options. The ‘GridSearchCV’ receives the input variables such as the type of hyperparameters, options, and data, and performs predictions through every combination of the hyperparameter type and option value.

Due to the reason above, it is necessary to determine the amount of training data. Inoue (1973) proposed a seven-day survey period as a sample instead of one year for representing the population for efficiency in analyzing maritime traffic. Thus, seven days is set as the initial value, and the training data is composed of July 25-31, 2018, and test data is set on August 1, 2018. Both data have no typhoon warning. The optimum amount of training days will be found in Chapter 4.2 by increasing the training days and comparing the performance.

Then, Table 3 shows the hyperparameter type and option value and the number of features that are cross tested from one to seven when they have length and breadth while one to five when they do not include length and breadth; subsequently, the former has 224 combinations (=2×4×7×4), and the latter has 160 combinations (=2×4×5×4).

Table 3. Hyperparameter options for GridSearchCV

* Case in presence of length and breadth is 1 to 7 while absence case has 1 to 5.

4. Results and Discussion

1) Optimum hyperparameter

Optimal hyperparameters were determined through seven-day learning data for August 1, 2018. With the size information, the number of predictors is two times lower than otherwise. On the other hand, the maximum number of features and the maximum depth were higher than those in the absence (Table 4).

Table 4. Optimum hyperparameter

2) Optimum training period

The period of training data was gradually increased from 1 to 20 earlier days before the prediction date of August 1, 2018, and the four performance indicators, f1-score, recall, and precision, were computed in every gradual step.

Fig. 4. Prediction scores according to training period. (a) presence of L&B, (b) absence of L&B.

Table 5. Accuracy upon ship types (Aug. 1, 2018)

With the size information, all four performance indicators showed an increasing trend with the increment of the training period and showed a sharp decrease in 19-day training (Fig. 4(a)). Accuracy was stable from the five-day training, and the rest of the indicators were stable from 12-day training and then decreased sharply on 19-day training. In the absence of size information, accuracy was equally stable from 5-day training, while the rest of the indicators had a score of 0.5 or less and were not stable overall (Fig. 4(b)). From the result, it can be considered that predicting ship type is highly dependent on the presence or absence of size information. Subsequently, the optimal learning day was set to ten day training with the given results.

The prediction accuracy of each ship type for January 1, 2018, is presented (Table 5). The accuracy was high in the presence of size information since the primary ship type were tankers (0.96), fishing boats (0.98), and cargo ships (0.93). On the other hand, the type of public has low accuracy (0.59) even the size was given. In the case of absence size, the accuracies of all other ships were less than 0.7 except for tanker ships (0.83), fishing ships (0.81), pilot ships (0.77), and leisure ships (0.73). In addition, the following types have abrupt decreases in accuracy; battleships (–1.00), other ships (–0.88), public – – ships ( 0.56), and cargo ships ( 0.45). Subsequently, the weighted mean of presence and absence are 0.96 and 0.77.

3) Daily forecasting performance

Daily ship type prediction was made for January 10 to December 31, 2018 (355 days) with the presence and absence of size information, and the performance was recorded (Fig. 5(a)).

Fig. 5. Prediction scores inpresence of LB.(a)times series, (b) distribution.

Fig. 6. Prediction scores in absence of LB.(a)times series, (b) distribution.

Table 6. Explainability of Features

*m represents median. Q1 and Q3 denote the first and third quarters. Y and X represent zone coordinates Y and X.

In the model with size, the first quarter (Q1) and the third quarter (Q3) were higher than 0.8 and the inter quantile range (IQR) was stable at 0.1 or less (Fig. 5(b)). In the case of other f1-score, recall, and precision, Q1 and Q3 were located between 0.6 and 0.9, and the IQR size was less than 0.2, which is less stable than accuracy, but the median value exceeded 0.7, indicating high predictive performance (Fig. 5(b)).

In contrast, the model without size has accuracy higher than 0.6 in most, and the rest three indicators are less than 0.4 in most (Fig. 6). This is the result of the unbalanced number of samples upon ship types provided in Table 5.

4) Explainability of feature

The trained model is suitable to predict ship type if ship data is given for only a specific time, like remote sensing data, and not like time-series data. The four dynamic features, time, SOG, COG, location (zone Y & X), and two static features, length and breadth, have their explanatory power for the type identification. For instance, fishing ships have both high and low speeds, therefore, they can be distinguished from cargo ships by their high speed, but there is no distinction with low speed. On the other hand, the fishing ships have lower speed at specific times and locations for spread and heave nets (Park et al., 2021), and thus the time and location have more power of explanatory than cargo ships.

The explanatory power of each feature was derived for the entire prediction period (Table 6, Fig. 7). In the presence of size, the total length (.367) and breadth (.326) exceeded half of the total explanatory power based on the median value. In the absence of size, SOG (.308) was the highest, and the second was COG (.213), and they exceeded half of the total explanatory power.

The features of time and zone coordinates had lower explanatory power in both models. Through this, it can be considered that the SOG and COG become the main feature if the size is not given. The time can be a meaningful feature since it has a smaller IQR and a higher median than the zone coordinates XY in the absence of size.

Then, why does the zone coordinate X have higher explanatory power than Y in both models? In Ulsan Port, there are many vessels moving north and south, and thus it is challenging to have a discriminatory power on the type of vessel upon latitude. On the other hand, longitude refers to the distance from the port. In other words, vessels close to the port are more likely to be tankers heading to/leaving from Ulsan port and the vessels farther are cargo ships and tankers passing by Ulsan Port, and vessels farther away are fishing vessels without major routes (Fig. 2)

Fig. 7. Explainability of features. (a) presence of size, (b) absence of size.

5. Conclusions and future study

RFC can be easily implemented through the Python library {scikit-learn}. It is essential to refine and process data and set appropriate hyperparameters to improve classification performance prior to implementing machine learning-based classification. In this study, the preprocessing of AIS and V-Pass data was introduced, the optimum hyperparameters were found through GridSearchCV, and the optimal learning period was set by performing the training models.

This study has limitations in the use of AIS and V- Pass data. Firstly, although the AIS and V-Pass have been linearly interpolated, that is only for longitude and latitude. Thus, it requires SOG and COG interpolation using a kinetic interpolation that is able to correct and predict the ship’s states such as position, SOG, and COG. Secondly, the model without size information showed low classification performance. Major routes depending on ship types can be considered as another feature to overcome this issue.

In addition, maritime surveillance through satellite imagery also requires not only detection but also classification. The proposed method can be considered as a method to predict ship type from the detected ship from satellite images.