Multiple aspect trajectories: a case study on fishing vessels in the Northern Adriatic sea Giulia Rovinelli Stan Matwin Fabio Pranovi 867381@stud.unive.it stan@dal.ca Elisabetta Russo Università Ca’ Foscari Venezia Institute for Big Data Analytics, Claudio Silvestri Venezia, Italy Dalhousie University Marta Simeoni Halifax, Canada Alessandra Raffaetà Institute of Computer Science fpranovi@unive.it Polish Academy of Sciences elisabetta.russo@unive.it Warsaw, Poland silvestri@unive.it simeoni@unive.it raffaeta@unive.it Università Ca’ Foscari Venezia Venezia, Italy ABSTRACT Our aim in this paper is to build, implement and analyze a spatio- In this paper we build, implement and analyze a spatio-temporal temporal database for gaining a sound knowledge about the database describing the fishing activities in the Northern Adri- fishing activities in the Northern Adriatic basin and trying to atic Sea over four years. The database results from the fusion address the above matters. To accomplish this task, we start from of two complementary data sources: trajectories from fishing two complementary data sources covering a time period of four vessels (obtained from terrestrial Automatic Identification Sys- years, from January 2015 to December 2018. The first data source tem, or AIS, data feed) and the corresponding fish catch reports is the set of terrestrial Automatic Identification System (AIS) data, (i.e., the quantity and type of fish caught). We present all the i.e., the AIS data sent by ships and received by ground stations on phases of the dataset creation, starting from the raw data and the Italian coast of Northern Adriatic sea. In particular, we focus proceeding through data exploration, data cleaning, trajectory on the AIS data of the fishing vessels. The second data source reconstruction and semantic enrichment. Moreover, we formalise is the fish catch reports of the Chioggia fish market, which is and compare different techniques to distribute the fish caught the primary market of the Northern Adriatic basin. Such reports by the fishing vessels along their trajectories. We implement the contain the quantity and type of fish caught by all vessels selling database with MobilityDB, an open source geospatial trajectory their landings at the Chioggia fish market. data management and analysis platform. Subsequently, guided Similar data have been used in [1] to develop early results by our ecological experts, we perform some analyses on the re- on the use of machine learning techniques to predict the future sulting spatio-temporal database, with the goal of mapping the Catch Per Unit Effort (CPUE), an indicator of fishing resources fishing activities on some key species, highlighting all the in- exploitation, from the past data of the Northern Adriatic sea. The teresting information and inferring new knowledge that will be mentioned work had some limitations, mainly related to the short useful for fishery management. temporal horizon – only two years, 2015 and 2016 – of the landing and AIS data. This, in fact, turned out to be a serious problem for the application of prediction methods: using the first year for 1 INTRODUCTION training and the second one for testing, was not sufficient. The The Northern Adriatic Sea area is one of the most exploited areas novel database that originates from the present work, thanks to of the Mediterranean Sea, causing an over-exploitation of the the availability of the data sources for two further years, puts fish resources. Having a clear representation and understanding the basis to overcome such limitations and paves the way for a of the main factors driving such phenomenon is of paramount subsequent favorable application of prediction methods. importance both for ecologists and for local policy makers that, We present all the phases of the database creation. Trajecto- together, could use such information for the development of ef- ries are reconstructed by linear interpolation of the raw AIS data. fective fishery management plans, able to make fishing activities We first clean the data, then we detect the trips performed by sustainable and ensure a productive and healthy ecosystem. the fishing vessels and we enrich the resulting trajectories with Some interesting objectives relative to the sea monitoring in additional information concerning the activities and anomalies the Adriatic Sea are: occurring during their trips. Moreover, relying on the landing • improving the knowledge of the fishing activities in the reports of the Chioggia fish market, we add a further, valuable Northern and Central Adriatic Sea, semantic aspect to the trajectories which denotes the quantity • evaluating the effectiveness of the current fishery man- of fish caught in each trajectory segment of the fishing vessel. agements, and In order to distribute the total catches along the trajectories we • detecting the spatial distribution of commercial fishery define two different approaches, which are subsequently put into catches. action and compared through specific analyses. First, we adopt the uniform distribution where the catch of a given species is © 2021 Copyright for this paper by its author(s). Published in the Workshop Proceed- ings of the EDBT/ICDT 2021 Joint Conference (March 23–26, 2021, Nicosia, Cyprus) uniformly distributed along the fishing segments of the corre- on CEUR-WS.org. Use permitted under Creative Commons License Attribution 4.0 sponding trajectory: each fishing segment is associated with a International (CC BY 4.0) portion of the total amount of fish, proportional to its length. The uniform distribution is clearly a simplification of the reality. We The analyses show that spatialising the distribution of catches refined it by considering a weighted distribution, whose underly- allows one to single out the fishing grounds and their seasonal ing idea is that the areas where more vessels are fishing during a and annual variation. This can be useful for the explanation of the given time period are more likely to have higher catch rates. fishermen behaviour, as well as to better understand the seasonal In [4], the authors review current research challenges and migration of the target species. trends tied to the integration, management, analysis, and visu- In summary, the main contributions of the paper are: (i) the alization of objects moving at sea. Several strategies have been use of MobilityDB, a database platform which is proved to be proposed to deal with the fusion of heterogeneous ocean data particularly suited for creating and analysing semantic trajec- properly. For example, the paper [13] shows a platform in the tories; (ii) the development of a case study concerning fishing maritime vessel traffic domain for discovering real-time traffic activities based on real and heterogeneous datasets, with the alerts by querying and reasoning across numerous streams (e.g., proposal of different approaches for distributing catches along AIS, weather, ice). The authors use semantic web technologies to the trajectories; (iii) the execution of various qualitative analyses, integrate heterogeneous data sources. In [3], the authors propose proposed and assessed by our ecological experts, for detecting a model for integration and analysis of data for vessel movement spatial and temporal patterns of the fishing activities. in a real-time maritime situation awareness system, also using The paper is organised as follows: Section 2 describes the semantic web techniques and tools. Unlike the previous meth- trajectory reconstruction and enrichment and the creation of ods, we represent our trajectory data with a semantic model. By a spatio-temporal database by means of MobilityDB. Section 3 considering data sources such as AIS and landing reports, the reports and discusses the results of some specific analyses per- trajectory of every fishing vessel becomes a complex object with formed with MobilityDB on the obtained database. In particular, several data dimensions that are contextual to the movement. we show the usefulness of recording and visualizing possible Several semantic models for trajectory data have been pro- anomalies of the trajectories as well as how to take advantage posed, such as the stops and moves [7], CONSTANT [2], and of the catches distribution for gaining new knowledge on key recently MASTER [6]. In this paper we follow the MASTER se- species in the area. Finally, we draw some concluding remarks in mantic model which, among the three proposals, is the more Section 4. flexible and expressive since it allows for enriching trajectories with complex objects. We represent the trajectory of fishing 2 FROM RAW DATA TO MULTIPLE ASPECT vessels as a multiple aspect trajectory. The AIS data constitute TRAJECTORIES the sequence of spatio-temporal points. Moreover, the MASTER In this section we illustrate the various ingredients and steps model introduces the concept of aspect which consists of “a real- we followed to produce a spatio-temporal database of fishing world fact that is relevant for the trajectory data analysis” [6]. It vessels’ trajectories in the Northern Adriatic sea, enriched with distinguishes between volatile aspects — usually associated with landing data from the Chioggia market. We start by describing the trajectory points, since they vary during the object movement the data sources of our case study, that is, the terrestrial AIS — and long term aspects — which do not change during an entire data of the Northern Adriatic sea, and the landing reports of trajectory, and hence they are associated with the whole trajec- the Chioggia fish market. Next, we explain how trajectories can tory. For instance, for vessel trajectories, the speed is a volatile be reconstructed by linear interpolation of the raw AIS data. In aspect, whereas the fishing gear type is a long term aspect. this step, we clean the data, we detect the trips performed by We also provide a prototype implementation of our spatio- the fishing vessels and we enrich the resulting trajectories with temporal database based on MobilityDB [15], an open source additional information concerning the activities and anomalies geospatial trajectory data management and analysis platform, occurring during the trips. Then we illustrate how to assign specifically developed to support the representation and the anal- landing reports to trajectories and we formalise the two different ysis of moving objects. On the one hand, the implementation in techniques to distribute the fish catches along the trajectories. MobilityDB allows us to perform various analyses on the dataset Finally, we give some details of the implementation by showing and assess the appropriateness of the conceptual framework. On the advantages of using MobilityBD as database to store and the other hand, it reveals the potentialities of MobilityDB for the analyse trajectories. reconstruction and management of semantic trajectories. In fact, The overall view of the process is depicted in Fig. 1: Starting the system offers temporal types that are suited to model points from the raw terrestrial AIS data of the fishing vessels and from changing their position along a time period. It also provides a lot the landing reports of the Chioggia’s market, we build up on of spatio-temporal operators to handle trajectories, e.g for getting top of MobilityDB a spatio-temporal database of multiple aspect the position and the associated annotations of a trajectory at a trajectories that enables us to perform analyses on the spatio- certain time instant, or checking topological relations or comput- temporal and semantic features of the trajectories. ing the distance between trajectories. It supports also the GiST (Generalized Search Tree) and SP-GiST (Space-Partition GiST) indexes, which can be used for accelerating spatial, temporal and 2.1 Data sources spatio-temporal queries. Finally, trajectories can be visualized Automatic Identification System (AIS). AIS raw data, provided by traditional tools such as QGIS [11], an Open Source GIS that by the Italian Coast Guard, were obtained for the trawl fishing supports viewing, editing, and analysis of geospatial data. vessels operating in the Northern Adriatic Sea from January 2015 The spatio-temporal database is used for analysing some phe- until December 2018. A total of 70 (2015), 77 (2016), 82 (2017) nomena of interest. First, we check the AIS coverage and we de- and 81 (2018) trawlers, with a length overall above 15m, were tect areas where there are transmission problems. Then, guided taken into consideration in this study: in particular, small and by our ecological experts, we map the fishing activities on some large bottom otter trawl (SOTB and LOTB), Rapido, one specific key species, highlighting all the interesting information and infer- kind of beam trawl (RAP), and midwater pair trawl (PTM). The ring new knowledge that will be useful for fishery management. identification of the vessels was performed by matching the data Figure 1: Bird’s eye view of the process: data sources, reconstruction and enrichment of trajectories and data analysis Year Number of vessels Number of transactions Year AIS data Number of trajectories 2015 71 64180 2015 29757601 11280 2016 79 70017 2016 38519864 11130 2017 80 71716 2017 21247207 35335 2018 76 72165 2018 25098120 9549 Table 1: Dimension of the Landing dataset Table 2: Raw AIS data vs trajectories present in the AIS (MMSI code, vessel name and the call sign) datum outside a port area and the immediate previous AIS datum with those of the European Fleet Register, which supplies specific is inside a port area and the time period between the two AIS information on the vessels (i.e., primary and secondary gear, data is greater than 20 minutes. The first condition corresponds length overall, gross tonnage, etc.). All the data given by the to the fact that the vessel ends a trip, it switches off the AIS, it AIS (i.e., data position, speed, time, MMSI) were used to identify is docked at the port and after a while it starts a new trip. The the fishing tracks and analyze the fishing activities (fishing, not second one corresponds to a situation in which a vessel leaves fishing). out of the port and then it starts transmitting when it is outside Daily landing reports. Landing dataset was obtained from the the port (20 minutes is the minimum time a vessel takes to leave Chioggia’s Fish Market, whose harbor hosts one of the main the port). A detailed analysis reveals that some fishing vessels, fishery fleets of the Adriatic Sea. This dataset consists of daily after entering the port area at the end of a trip, continue to trans- landings (catch amounts in kilogram) for 104 commercial species mit their position. In this way, none of the above criteria is met. caught during four years, from January 2015 to December 2018 in This causes a wrong trip reconstruction in which two or more the Northern Adriatic Sea. The records pertain around 80 fishing trips are considered as a unique trip with a duration of several vessels, and contains a total of 278078 transactions over the four days. Hence, to avoid this phenomena we remove the AIS data years, as detailed in Table 1. transmitted inside the port when the vessel returns to a port. In Table 2 we report the dimension of the original AIS datasets and 2.2 Trajectory reconstruction and the resulting number of trajectories. enrichment A trajectory, resulting from the reconstruction, is a sequence Trajectories are reconstructed by linear interpolation of the raw of segments, obtained by connecting consecutive AIS points. It AIS data. While performing the reconstruction raw data are is enriched with the following information: cleaned: all the points implying movements that are not physi- • MMSI, boat identifier; cally feasible due to a maximum possible boat speed are removed. • trip duration (in hours); Next, in order to organize the data into distinct trajectories fol- • trip length (in meters); lowed by the fishing boats, we apply two criteria: a new trip • total time of fishing activity (in hours); begins 𝑎) when the vessel is inside a port area and there is no • total length of the fishing activity (in meters); transmission for longer than a fixed time, or 𝑏) there is an AIS • date and time of the trip departure and conclusion; ID Description Gear description ID min. speed max. speed 0 normal trip Small bottom otter trawl SOTB 3.704 8.334 1 no transmission for more than 30 minutes outside a port area Large bottom otter trawl LOTB 3.704 8.334 2 trip always inside a port area Pelagic pair trawl PTM 3.704 10.186 3 trip duration exceeds the 24 hours. Rapido RAP 7.408 12.964 Table 3: Values of the anomaly attribute Table 5: Gears and their minimum and maximum fishing speed (in km/h) ID Activity description 0 in port 1 exiting from port spatio-temporal point as in the original MASTER model, but 2 entering to port segments. This is motivated by the fact that we want to high- 3 fishing light the presence of homogeneous trajectory portions, which 4 navigation. are the appropriate granularity level for our analyses. According Table 4: Values of the activity attribute to the MASTER model classification, the information listed above can be classified as long-term aspects, (those associated with the full trajectory), and volatile aspects (those associated with the segments). • total number of segments with more than 30 minutes be- By using the MASTER model we are able to represent different tween two consecutive AIS transmissions; aspects of our trajectories in a uniform and simple way. Moreover, • anomaly, a code specifying whether the trip presents an this representation allows us to perform complex queries merging anomaly or not and the kind of anomaly. spatial, temporal and semantic features. In the rest of the paper, The anomaly attribute highlights some strange behaviour of the we denote by 𝑇 the resulting set of multiple aspect trajectories. fishing vessel. Possible anomalies are: • the time interval between two consecutive AIS data is 2.3 Catch distribution longer than 30 minutes outside the port, suggesting some We next describe how to merge the trajectories of the fishing points could be missing (anomaly is set to 1); vessels with the daily landing reports provided by the Chioggia • a boat remains inside a port area for the whole trip (anom- fish market. The latter dataset contains information about each aly is set to 2); trading transaction, including the landing date, MMSI of the • the duration of the trip exceeds the 24 hours (anomaly is seller, the species, and the quantity of fish. Note that we work on set to 3); a subset of the set of reconstructed multiple aspect trajectories. If none of the above cases occurs, the trip is considered as normal In fact, we exclude from our analysis, fishing vessels that do not and anomaly is set to 0. Table 3 summarizes the possible values sell their fish in Chioggia, trajectories with anomaly 2, i.e., the of the anomaly attribute. ones that do not leave the port area, and trajectories that do not It is worth noting that through the MMSI, we can obtain further have any fishing activity. information on the vessel, such as its name and the fishing gear. In order to perform the merge we need to associate each fish Each segment in the trajectory is in turn annotated with: market transaction with a trajectory of the vessel having the • speed; specified MMSI. To accomplish this task, for each transaction, we • position of the segment with respect to the port areas; select the vessel trip with the most recent arrival in the port (be- • activity of the boat within the segment; fore 4 PM of the landing date). Arrivals after 4 PM are associated • length of the segment; with transactions occurring the next day. The quantity (weight) • time spent in the segment; of fish assigned to a trajectory is called a catch. • transmission. In order to distribute the fish associated with a trajectory over its fishing segments we follow two different approaches: The activity attribute describes what the vessel is doing ac- cording to Table 4. The in port, exiting from port and entering to • uniform distribution, and port situations can be deduced from the position of the extremes • weighted distribution. of the segment w.r.t. the port area. If none of the previous cases In the first case, the catch is uniformly distributed along the applies, the fishing or navigation activities are established on fishing segments of the corresponding trajectory. Each fishing the basis of the average speed of the boat. More precisely, if the segment of the trajectory is associated with a fraction of the total average speed is in the range of the fishing speed of the gear amount of fish, proportional to its length. We consider separately the boat is equipped with, the boat is assumed to be in a fishing each species that the fishing vessel caught. phase; otherwise, it is assumed to be in a navigation phase. The Definition 2.1 (Uniform distribution). Let 𝑡𝑟 be a trajectory and considered gears and their minimum and maximum speed during let catch the record containing the quantities of the different the fishing activity are reported in Table 5. species associated with the trajectory 𝑡𝑟 . Given a segment 𝑠 be- The attribute transmission records whether the end points of longing to 𝑡𝑟 with activity set as fishing and a species 𝑠𝑝, the the segment have a time distance greater than 30 minutes. If this uniform catch for segment 𝑠 and species 𝑠𝑝 is defined as happens the attribute is set to 1, otherwise to 0. As explained 𝑠.𝑙𝑒𝑛 above, the presence of segments with transmission set to 1 allows 𝑑𝑈 (𝑠, 𝑠𝑝) = ∗ catch.𝑠𝑝 (1) for the detection of an anomalous behaviour of the trajectory. 𝑡𝑟 .𝑙𝑒𝑛_𝑓 𝑖𝑠ℎ𝑖𝑛𝑔 These trajectories are modeled as a multiple aspect trajectory, where following MASTER model [6]. Actually, as minimum granular- • tr.len_fishing is the attribute storing the total length of ity to attach semantic information, we do not consider a single the fishing activity for the trajectory 𝑡𝑟 ; • s.len is the length of the segment; receives a weight which is proportional not only to the length • catch.𝑠𝑝 selects the quantity of a certain species 𝑠𝑝. s.len of the segment but also to the fishing coefficient 𝛼 (s.cell, 𝑠𝑝) of the cell the segment belongs to. Clearly the assumption of uniform catch distribution is a sim- plification of reality. We consider also a refinement based on a so called weighted distribution. The idea is that the areas where 2.4 Implementation more vessels are fishing, during a given time period, are more To construct and store the set of multiple aspect trajectories, likely to have higher catch rates. we used MobilityDB [15], an open source extension to the Post- In order to implement this technique, we need to suitably par- greSQL database system [10] and its spatial extension PostGIS [9]. tition the fishing area because it becomes crucial to evaluate the It provides temporal types and spatio-temporal operators that number of fishing vessels present in a certain zone. We decided ease the management of moving objects. to divide the Northern Adriatic sea into a square grid with 3×3 One main feature of MobilityDB is that it offers a construct for km cell size. The size has been chosen in agreement with the representing the evolution of a value during a sequence of time environmental scientists based on the dimension of the fishing instants. The values between successive instants are interpolated vessels and their behaviour during the fishing activity. using a linear function. Clearly, this construct perfectly suites The introduction of the grid leads to a further segmentation of the representation of trajectories, which are reconstructed from a the trajectories. In fact, each segment that spatially crosses one sequence of spatio-temporal data. In our case, the spatio-temporal or more cells of the grid needs to be split into smaller segments points are the AIS data aggregated on the basis of the trajectory in such a way that each portion is completely inside a single cell. 𝑖𝑑. We created a set of objects of type tgeompoint, which is a Moreover, since we deal with a spatio-temporal grid, all segments temporal type modelling a point changing its position along a spanning over two days are split into two smaller segments by time period. taking as extra point the interpolated position at midnight. Next, the function trajectory is applied to these objects, and In order to compute the weighted distribution, we associate a a geometry value is returned. In this way the trajectory can coefficient to each spatio-temporal cell of the grid. be visualized. In our work, for visualizing trajectories and the Definition 2.2 (fishing coefficient). Let 𝑐 be a spatio-temporal result of our analyses, we used QGIS [11], an Open Source GIS cell and 𝑠𝑝 a species. The fishing coefficient of cell 𝑐 for the species that supports viewing, editing, and analysis of geospatial data. 𝑠𝑝 is defined as follows: For instance, Figure 2(left) shows the sequence of AIS data, i.e., the sequence of spatio-temporal points, related to the trip of a 𝛼 (𝑐, 𝑠𝑝) = |{𝑡𝑟 ∈ 𝑇 ↓ 𝑠𝑝 | 𝑡𝑟 ∩ 𝑐 ≠ ∅}|∗ (2) fishing boat, whereas Figure 2(center) illustrates the continuous Σ𝑡𝑟 ∈𝑇 ↓𝑠𝑝 Σ𝑠 ∈𝑡𝑟 ∩𝑐∧s.activiy=fishing s.len representation of the same trip obtained by using the MobilityDB where construct. The interpolation is internally implemented by the • 𝑇 ↓ 𝑠𝑝 is the set of trajectories having a landing report system, with the dual advantage of raising the user from this task with the species 𝑠𝑝; and simplifying queries and analyses. • 𝑡𝑟 ∩ 𝑐 returns the intersection between the trajectory 𝑡𝑟 MobilityDB provides a lot of spatio-temporal operators to and the cell 𝑐; handle trajectories. For instance, startTimestamp and endTime- • s.activity and s.len are respectively the attributes of seg- stamp return respectively the first and last time instant among a ment 𝑠 storing the activity and the length of the segment. set of time instants and this can be useful to extract the beginning and ending points of a trajectory; getValue returns a value at The coefficient 𝛼 (𝑐, 𝑠𝑝) combines the number of fishing vessels a particular time instant. There are operators to check topologi- and the amount of fishing activity they perform in the cell, hence cal relations between trajectories, like tintersects, tdisjoint, it provides a measure of the fishing activity in the cell. Note that and others to compute distances. Interestingly the results of these the coefficient depends on the species. Hence, for each species operators are values changing in time. In fact, it can happen that 𝑠𝑝, we select only the trajectories having a landing report for the at certain time periods trajectories enjoy the relations whereas at given species 𝑠𝑝. other ones they do not, and the distance between the objects can Since it is natural to expect that vessels will mostly concen- vary depending on the movement of the objects themselves. For trate in fishy areas, the intuition is that cells where the fishing instance, the user can check whether a fishing vessel respects the coefficient is higher will have higher catch rates. This leads to rule that it can fish only at a distance greater than three nautical the idea, formalised below, of using such coefficient as a weight miles from the coast and eventually detect where and when the when distributing catches over a trajectory. ban has not been observed. Definition 2.3 (Weighted distribution). Let 𝑡𝑟 be a trajectory MobilityBD allows for an easy representation of semantic tra- and let catch the record containing the quantities of the differ- jectories where semantic attributes can be modelled as temporal ent species associated with the trajectory 𝑡𝑟 . Given a segment 𝑠 types. This means that we can model in a single table both the belonging to 𝑡𝑟 with activity set as fishing and a species 𝑠𝑝, the sequence of spatio-temporal points forming a trajectory and in- weighted catch for segment 𝑠 and species 𝑠𝑝 is defined as formation associated with the whole trajectory itself, such as the MMSI of the vessel, the duration and length of the trajectory. 𝛼 (s.cell, 𝑠𝑝) ∗ s.len Moreover, a trajectory can be segmented and each segment can 𝑑𝑊 (𝑠, 𝑠𝑝) = ∗catch.sp Σ𝑠 ′ ∈𝑡𝑟 ∧s′ .activity=fishing (𝛼 (s ′ .cell, 𝑠𝑝) ∗ s ′ .len) be stored as a temporal type. Even in this case we can add other (3) attributes modelling features of the segment itself, such as the where s.cell is the unique cell the segment 𝑠 belongs to. speed, the activity, the transmission and the quantity of caught When distributing the catch over the segments of the trajec- fish. In Figure 2(right) the different colours describe the activities tory 𝑡𝑟 , again only segments which are classified as fishing are of the fishing vessel. They allow the user to immediately detect considered. The difference is that in this case each segment 𝑠 where the vessel is fishing and also the shape of the movement. Figure 2: Trajectory visualisation as a sequence of spatio-temporal points (left), as a continuous function (center), and as a semantic object where the activity attribute is highlighted (right) For instance, the figure highlights several circular movements the area. In fact, spatializing the distribution of catches has sev- and the experts have confirmed that they are typical of this kind eral important applications. For instance, it allows us to obtain of fishing activity. knowledge about the seasonal variation of the fishing grounds Finally, MobilityDB provides support for the GiST (Generalized and this, in turn, is useful for explaining the fisherman behaviour, Search Tree) and SP-GiST (Space-Partition GiST) indexes, which as well as to better understand the seasonal migration of a tar- can be created for table columns of temporal types. We used such get species. Figure 4 reports the seasonal spatial distribution of indexes for accelerating spatial, temporal and spatio-temporal cuttlefish, Sepia officinalis, aggregated by fishing gears (SOTB, queries. LOTB and RAP) in 2018. Cuttlefish is one of the main target species of the Adriatic Sea, hence it is an ideal case study for showing a seasonal migratory behaviour. It is worth noting that 3 DATA ANALYSIS AND DISCUSSION the most productive seasons were autumn and winter, with two In this section we present some analyses performed with Mobili- high density areas, one nearer the coast and the other one more tyDB on the obtained spatio-temporal database of the Northern offshore, at the border with the Croatian waters. In spring the Adriatic sea. catches resulted more scattered, while in summer the catch area The first analysis aims at visualizing the regions where there was more defined and localized closer to the Italian coast. This are transmission problems. We exploit the anomaly attribute is in line with the general ecological knowledge about the be- and in particular we investigate trajectories having this attribute haviour of the species, hence, the catches data correctly reflect set to 1. In Figure 3 we show for each cell, the percentage of cuttlefish seasonal spatial distribution behaviour. Figure 4 reports trajectories that got disconnected from the AIS for a time period also the comparison between the uniform (A) and the weighted greater than 30 minutes while crossing that cell with respect to (B) distribution maps of cuttlefish Sepia officinalis in 2018. It is the total number of trajectories passing through the cell. Looking evident that the maps obtained with the weighted distribution at Figure 3, it is evident that the no-transmission anomaly has (B) result more defined, allowing to better identify the fishing decreased a lot from 2015 to 2018. In fact, in 2015 the area where grounds of cuttlefish. this percentage is over 50% is very large and it covers almost the Another important application of the spatial distribution of whole fishing zone. Instead, in 2018 this phenomenon is localized catches is the detection of different fishing grounds among years. in few areas, i.e., close to the coasts and along the territorial As an example, the catches of anchovy, Engraulis encrasicolus, waters borders. Moreover, in 2018 there are also some isolated recorded in winter 2015, 2016, 2017 and 2018 and distributed cells in the southern part. according with the weighted distribution are reported in Figure 5. The low spatial coverage of AIS is a well-known issue and the The maps clearly show how the fishing grounds, and conse- amount of missing data can vary substantially between vessels quently the distribution of anchovies, changed along the years. as discussed in [12]. Our analysis reveals that data from 2018 are In particular, a gradual reduction of the fishing grounds is ob- more reliable and can be useful for detecting areas where the AIS served from 2016 to 2018. This is clearly a relevant information signal is not received well, like the isolated cells in the southern for both ecologists and policy makers: if the fishing ground re- portion of the sea area under investigation. duction is the result of an over exploitation of the species they This analysis is an example of how the semantic knowledge can adopt appropriate countermeasures. hidden in a single attribute, such as the anomaly attribute, can To end up, we would like to point out that these are only a be useful to greatly improve the general spatio-temporal knowl- few examples of the analyses that can be performed by using the edge of the domain of interest. On one hand the progressive dataset of multiple aspect trajectories. For instance, we can focus low-coverage reduction of AIS data is per se a highly valuable in- on vessels equipped with a specific fishing gear (i.e., LOTB, SOTB, formation for ecologists and policy makers, since this ensures the RAP and PTM) and determine their fishing grounds and the reliability of the collected data. On the other hand, the proposed corresponding degree of exploitation. This fine-grained analysis implementation allows the experts to continuously monitor the could help to reveal different efficiency degrees of fisheries that, in degree of coverage and eventually decide to add further terrestrial turn, could constitute a basis to implement specific management AIS receivers. actions for these activities. Moreover, we can vary our analysis The second and third analyses take advantage of the catches according to different time periods and consider only certain sea distribution and try to infer some knowledge on key species in Figure 3: Spatial distribution of the no-transmission anomaly, years 2015, 2016, 2017 and 2018 (from left to right) Sepia Officinalis, 2018, Uniform distribution (A) Sepia Officinalis, 2018, Weighted distribution (B) Figure 4: Comparison between uniform (A) and weighted (B) distribution of cuttlefish Sepia officinalis, aggregated by seasons (winter, spring, summer and autumn 2018) areas. For instance one could focus on protected areas, like the 4 CONCLUSIONS Pomo Pit or the Sole Sanctuary. We can also select the behaviour In this paper we built, implemented and analysed a spatio-tempo- of single trajectories satisfying complex conditions concerning ral database of the vessels trajectories in the Northern Adriatic both their movements and their semantic annotations by using sea. We started from the terrestrial AIS data of the area of interest the operators available in MobilityDB. and the fish reports of the main fish market, Chioggia, for the years 2015, 2016, 2017, 2018. We determined the trajectories and introduced semantic attributes able to unveil interesting infor- mation and aspects of the original data themselves. Moreover, Figure 5: Spatial distribution of anchovy Engraulis encrasicolus in winter, years 2015, 2016, 2017 and 2018 (from left to right) we gave a formal definition of two different catch distribution Multiple-Aspect Analysis of Semantic Trajectories - First International Workshop, techniques, the uniform and weighted, with the aim of putting MASTER 2019 (Lecture Notes in Computer Science), Vol. 11889. Springer, 83–99. [2] Vania Bogorny, Chiara Renso, Artur Ribeiro de Aquino, Fernando de them at work and comparing their behavior. Lucca Siqueira, and Luis Otavio Alvares. 2014. Constant–a conceptual data Additionally, we implemented the spatio-temporal database model for semantic trajectories of moving objects. Transactions in GIS 18, 1 (2014), 66–88. using MobilityDB, thus ensuring a suitable environment for stor- [3] Stefan Brüggemann, Konstantina Bereta, Guohui Xiao, and Manolis ing, querying and visualizing trajectories of moving objects. Koubarakis. 2016. Ontology-based data access for maritime security. In Euro- The ecological experts proposed some analyses on the ob- pean Semantic Web Conference. Springer, 741–757. [4] Christophe Claramunt, Cyril Ray, Elena Camossi, Anne-Laure Jousselme, tained database. We started with the analysis of the transmission Melita Hadzagic, Gennady L. Andrienko, Natalia V. Andrienko, Yannis anomalies – stored as a new semantic feature – that allowed us Theodoridis, George A. Vouros, and Loïc Salmon. 2017. Maritime data integra- to acknowledge a concrete and progressive improvement of the tion and analysis: recent progressand research challenges. In Proceedings of the 20th International Conference on Extending Database Technology. 192–197. data completeness in the years 2015-2018, thanks to the growing [5] Ioannis Kontopoulos, Konstantinos Chatzikokolakis, Dimitris Zissis, Kon- use of the AIS transmission systems in the fishing vessels and to stantinos Tserpes, and Giannis Spiliopoulos. 2020. Real-time maritime anom- aly detection: detecting intentional AIS switch-off. Int. J. Big Data Intell. 7, 2 the increasing AIS data receiving coverage. (2020), 85–96. We proceeded then with the analysis of the two proposed dis- [6] Ronaldo dos Santos Mello, Vania Bogorny, Luis Otavio Alvares, Luiz Hen- tribution techniques. It turned out that the weighted distribution rique Zambom Santana, Carlos Andres Ferrero, Angelo Augusto Frozza, Geo- mar Andre Schreiner, and Chiara Renso. 2019. MASTER: A multiple aspect is actually a refinement of the uniform one, able to better define view on trajectories. Transactions in GIS 23, 4 (2019), 805–822. the fishing ground of the species of interest. Besides, we showed [7] Christine Parent, Stefano Spaccapietra, Chiara Renso, Gennady Andrienko, how the use of semantic trajectories can provide an assessment Natalia Andrienko, Vania Bogorny, Maria Luisa Damiani, Aris Gkoulalas- Divanis, Jose Macedo, Nikos Pelekis, et al. 2013. Semantic trajectories modeling of the fishing activities, capturing spatial and temporal patterns. and analysis. ACM Computing Surveys (CSUR) 45, 4 (2013), 42. All these results put the best possible basis for a favorable [8] Lucas May Petry, Amílcar Soares, Vania Bogorny, Bruno Brandoli, and Stan Matwin. 2020. Challenges in Vessel Behavior and Anomaly Detection: From application of prediction methods, which is the next step to be Classical Machine Learning to Deep Learning. In Advances in Artificial Intel- done on the obtained database. In particular, first we would like ligence - 33rd Canadian Conference on Artificial Intelligence (Lecture Notes in to test whether the Random Forest prediction results reported Computer Science), Vol. 12109. Springer, 401–407. [9] PostGIS [n.d.]. PostGIS: Spatial and Geographic objects for PostgreSQL. https: in [1] improve thanks to the availability of 2017 and 2018 data. //postgis.net/ Then, we would like to experiment other prediction techniques, [10] PostgreSQL [n.d.]. PostgreSQL Open Source Relational Database. https://www. such as lag variables [14], or modern time series prediction. postgresql.org/ [11] QGIS [n.d.]. QGIS: a free and open source Geographic Information System. Finally, another interesting line of research is to extract fishing https://qgis.org/en/site/ patterns, like the circular one illustrated in Figure 2, or anomaly [12] Jennifer L Shepperson, Niels T Hintzen, Claire L Szostek, Ewen Bell, Lee G Murray, and Michel J Kaiser. 2018. A comparison of VMS and AIS data: the behaviour, as investigated in [5, 8]. effect of data coverage and vessel position recording frequency on estimates of fishing footprints. ICES Journal of Marine Science 75, 3 (2018), 988–998. [13] Amílcar Soares, Renata Dividino, Fernando Abreu, Matthew Brousseau, An- ACKNOWLEDGEMENTS thony W Isenor, Sean Webb, and Stan Matwin. 2019. CRISIS: Integrating AIS This paper is supported by the MASTER project that has received and Ocean Data Streams Using Semantic Web Standards for Event Detec- tion. In International Conference on Military Communications and Information funding from the European Union’s Horizon 2020 research and Systems. innovation programme under the Marie-Sklodowska Curie grant [14] Hristos Tyralis and Georgia Papacharalampous. 2017. Variable Selection in agreement N. 777695. Time Series Forecasting Using Random Forests. Algorithms 10 (2017), 114. [15] Esteban Zimányi, Mahmoud Sakr, Arthur Lesuisse, and Mohamed Bakli. 2019. MobilityDB: A Mainstream Moving Object Database System. In Proceedings of REFERENCES the 16th International Symposium on Spatial and Temporal Databases. 206–209. [1] Pedram Adibi, Fabio Pranovi, Alessandra Raffaetà, Elisabetta Russo, Claudio Silvestri, Marta Simeoni, Amílcar Soares, and Stan Matwin. 2019. Predicting Fishing Effort and Catch Using Semantic Trajectories and Machine Learning. In