1. Introduction
The continuing decline in global fish stocks places an increasing importance on fishing vessel surveillance to better understand individual fishing grounds and to provide a basis for enforcement of fishing regulations and counter illegal, unreported, and unregulated (IUU) fishing activities. However, there is no uniform and mandatory reporting on fishing boat activity or their catch. Log books recording date ranges, locations, and catch weights are required in certain jurisdictions. However, these records are typically not submitted until after the vessel returns to its landing site and there is no centralized system for sharing these records. Individual fishing boats can be tracked in near real time using vessel tracks from automatic identification system (AIS) and vessel monitoring systems (VMS). However, these systems provide an incomplete view of fishing boat activity and the data are generally restricted in terms of access. In terms of completeness, the requirements for AIS and VMS generally cover only the larger boats, with smaller boats able to operate “under the radar.” Indonesia has implemented a monitoring system for fishing vessels on a voluntary basis since 2003. While formal implementation of VMS attached on board of fishing vessel and fish carrier vessel for 30 gross tonnage (GT) or larger has been formalized by the Ministry Regulation number 42 in 2015, which represent approximately 10% of the domestic fishing fleet [
1]. In the USA, VMS is required on fishing boats longer than 19 meters and, in certain federal waters, all fishing boats are required to carry VMS [
2]. In other countries, such as the Philippines, there is no VMS requirement as of 2018. The International Maritime Organization (IMO) requires that all boats 300 GT and larger carry AIS [
3]. The vast majority of fishing boats fall under these weight limits and are under no requirement to broadcast their location. A second shortcoming of VMS and AIS for vessel surveillance is that operators are able to evade detection by disabling their devices, reporting a false identification, or reporting a series of false locations. Amongst maritime enforcement agencies, there is substantial interest in identifying “dark vessels” that lack an accurate AIS or VMS signal. This is a tip-off of possible illegal fishing in certain fishing grounds. The only way to identify “dark vessels” is to combine AIS and VMS with complimentary vessels detection sources.
Certain satellite remote sensing systems can detect vessels and have potential value in fishery surveillance. High spatial resolution optical sensors are able to detect and characterize vessels under low cloud daytime conditions. Satellite synthetic aperture radar (SAR) systems are also able to detect offshore vessels and have the advantage of all-weather operations [
4,
5]. Also, high resolution optical imagery have been used to detect marine vessel activities [
6,
7,
8,
9]. The downside for these sources are that it may take several days to access the data, there are generally fees associated with the data access and global coverage is not currently available on a daily basis.
Data from meteorological satellite sensors have the favorable characteristics of providing global coverage, with data open access and near real time availability. In addition, there are long term archives for meteorological satellite data, which can be used to develop extended temporal records. The major disadvantage of meteorological satellite data for vessel detection is the spatial resolution, which is typically far too coarse for the detection of vessels using normal spectral bands. The exception to this is the detection of electric lighting on boats at night.
It has been known since the 1970s that lights from fishing boats can be detected with low light imaging data collected at night by sensors flown on weather satellites [
10]. The low light imaging on these sensors is designed for the detection of moonlit clouds in the visible, but also enables the detection of electric lights present at the Earth’s surface [
11]. One of the lighting sources detected are fishing boats that deploy lights to attract catch. More recently, boat detection has been demonstrated using low light imaging data collected by the NASA/NOAA visible infrared imaging radiometer suite (VIIRS). Multiple groups have utilized VIIRS boat detection data in their research [
12,
13,
14,
15,
16]. The first VIIRS was launched in 2011 and the second in 2017 [
17], which will help to provide more coverage per day. In fact, the Earth Observation Group operates a near real-time VIIRS boat detection system which produces a nightly global mapping of VIIRS boat detections (VBD), and are available for open access download [
18].
The purpose of this paper is to define a VBD cross-matching algorithm suitable for vessel track records like AIS or VMS, and to use this to characterize the VBD product. Studies had been carried out to match vessel track record with satellite based boat detection [
19,
20,
21,
22,
23]. Most of these studies use vessel track records to verify the vessel detection result from selected scenes of satellite imagery. To aid satellite based vessel detection like VBD, whose records lack specific information on the boat detected, track data like VMS and AIS have to be further analyzed to provide information beyond vessel name and type. It is particularly important for fishery management to be able to answer questions like: is a detection an indication of fishing activity? Is it possible to identify “dark vessels” detected by satellite but lacking track records? What percentage of fishing vessels are detected by satellite?
To investigate these questions, we conducted a cross comparison analysis of VBD and VMS records for more than 3600 vessels spanning 32 months for Indonesia. The VMS records include unique identifiers for individual vessels and gear type registrations. By cross matching the VMS tracks with VBD it is possible to calculate the match rates and average VIIRS radiance for individual vessels. These results are aggregated for all the vessels having a specific gear type registration to answer the question on the types of fishing vessels detected by VIIRS.
In addition to cross matching, we also analyze VMS tracks for the status of the vessel at the time of recording. The VMS tracks are classified into four activity types by their location, speed and heading changes: landing, transit, stationary, and maneuvering. Such efforts help to answer the question on whether a VBD can be interpreted as “fishing.” Separate match rate summaries are calculated for fishing and transit activity types.
Clearly, any new data source on vessel detection needs to be thoroughly examined and compared to other available data sources. Otherwise the users will not know how to interpret the data or develop standard operating procedures for their use. Our intention is to provide results on VBD in reference to a widely recognized vessel track data source.
2. Methodology
2.1. Data Collection
2.1.1. VMS
Under a cooperative research agreement with the Global Fishing Watch (GFW), we obtained 32 months of VMS data collected by the Indonesian Ministry of Marine Affairs and Fisheries (MMAF a.k.a. KKP). This consists of records for Indonesian fishing vessels larger than 30 GT from January 2014 to August 2016. The VMS records comprise 15 gear types with more than 3600 distinct vessels. The typical gap between vessel track records is about an hour. Here the term “track record” refers to the location report of the vessel at the time of reporting. The summary of gear types included in the database is shown in
Table 1. The records include transmitter number, timestamp in UTC, latitude/longitude, registered gear type, vessel size and other useful information. There are nearly 29 million VMS records, with roughly 20 million track-hours. For easier reference to the gear types, abbreviations are devised as shown in
Table 1 and used throughout this work.
2.1.2. VBD
The VBD data used in this study were downloaded from the EOG website [
18] for the corresponding time range, i.e., from January 2014 to August 2016. The VBD records were filtered to a rectangular area covering the Indonesian Fishery Management Zones (Wilayah Pengelolaan Perikanan, WPP) as shown in
Figure 1. The VBD data used in this study were produced by the V23 algorithm. VBD records were also imported to the same database as the VMS records for further processing. The VBD data were provided in the form of CSV files, with data from columns listed in
Table 2 used in the study.
VBD from 2014–2016 being a dataset derived from observations of a single polar orbit satellite with nominal swath of 3000 km, can cover the globe daily, with minimal coverage of 1 at the equator, and increased coverage closer to both poles. Indonesia usually gets 1 coverage, and 2 coverages in the orbit overlap area. The band used for detection in VBD is VIIRS day night band (DNB), with a nominal pixel footprint of 742 m by 742 m. VBD also adopts data from VIIRS Nightfire which was also a product of EOG [
25], to prevent gas flares of oil platforms from being recognized as lit vessels [
16].
2.2. Vessel Location Prediction
To match VMS to VBD, the vessel location at the time of satellite overpass has to be known. This is achieved by interpolating the vessel location between neighboring records at the time of anticipated satellite overpass time. The location interpolation is shown as
where
and
are coordinates of the vessel at the satellite overpass time
.
and
are vessel coordinates in VMS records immediately prior to
. Likewise,
and
are vessel coordinates in VMS records immediately after
. This can also be re-written to do extrapolation of vessel location at the beginning and ending of vessel tracks. Such process are called vessel location prediction in this study. The overall process is shown in
Figure 2.
Before performing vessel location prediction, the VMS track record needed to be cleaned to remove records with erroneous timestamps or coordinates. Records that indicate moving speed larger than 30 knots were also neglected.
VMS track series were processed by UTC date. The record that was nearest to local midnight was selected, whose coordinate and timestamp were used as seed to begin the search for the nearest satellite overpass time, here referred as the predicted time.
A program, VIIRS overpass predictor, was developed to find the closest satellite overpass time for a given location and time. The orbital location of satellite at any given time can be propagated with SGP4 model [
26] and proper two line element (TLE) [
27]. TLE provides all the necessary parameters describing the satellite orbit at a specific time. TLE is retrieved from CelesTrak which usually being updated daily [
28].
Since TLE was frequently updated, the program finds the TLE closest to the seeding time, and it propagates satellite location with 10 min interval until the seeding location passed the scanning plane and is visible, i.e., the shortest distance from target to satellite path is within the scanning swath. Then it reverses the direction of propagation with 1/10 of the previous time step until the seeding location passed the scanning plane again. The time step is then reversed with 1/10 of its current setting. The process is repeated until the change of angle between the satellite to the visible observer and the scanning plan is less than 0.0001 degrees for the current time step. The time step is then reset and the next loop is initiated with seed time advanced by 60 min from the current predict time to search for the next possible overpass. When the given coordinate is within the orbital overlap zones, it is possible to find multiple predict overpass times. The process is continued until the seed time is 24 h advanced from the first seed time. By verifying the predicted time with VBD records, the precision of the VIIRS overpass predictor is determined to be 2 s.
The predicted VIIRS overpass time is then used to interpolate the vessel location between immediate neighboring VMS track before and after it. If the predicted VIIRS overpass time is beyond the ending or starting record of the VMS track series, extrapolation is adopted. The process is skipped if the predicted VIIRS overpass time is 2 h away from any nearest VMS track records. The calculated record with predicted VIIRS overpass time and interpolated vessel location is referred as predicted record. Such records are inserted into the VMS database and treated as part of the VMS track series.
2.3. VMS Record Status Classification
VMS tracks with predicted records inserted are then classified by their type of activity. In this study we tried to separate the type of activity into maneuvering (M), transit (T), stationary (S), and landing (L). Briefly speaking, M records were those with lower speed and larger heading change, while T records were those with higher speed and smaller heading change. S and L records were those not moving for a period of time, with L records were found close to recognized landing sites. We specifically call vessels in M and S as being fishing (F). There are a number of different methods developed to perform status classification of track records [
29]. Here the method proposed by de Souza et al. [
30] for long line vessels using an R package called adehabitatLT [
31] was chosen.
Status classification as shown in
Figure 3 was applied on one vessel within a calendar year at a time. If the number of track records within a calendar year was less than 50, the track will not be subjected to classification. The adehabitatLT package requires the track to be evenly separated temporally. The VMS records were not strictly 1 h apart, hence it had to be normalized to a regular hourly interval before classification.
The coordinates were reprojected to corresponding universal transverse mercator (UTM) zones in accordance to the average latitude and longitude of the selected annual track to have the calculation carried out in meters instead of degrees. Although vessels can cross multiple UTM zones, the errors caused were assumed negligible in the scope of this study. This was because fishing vessels usually take regular routes and are unlikely to cross more than two UTM zones.
The process of applying adehabitatLT can be simplified in two stages. First, the annual track was cut into segments. The maximum likelihood method gives the most likely number of segments in the track, which describes how many times a vessel changed its pattern of maneuvering in the given track. Then for each segment, we found the most likely model from the predefined velocity model for each segment [
31].
The velocity model in this study was defined as a series of Gaussian distribution with mean = 2 (km/h) , and standard deviation as 2.5 (km/h). For our purpose, segments with velocity model lower than or equal to model 5, i.e., velocity normal distribution of mean velocity of 10 km/h with standard deviation of 2.5 km/h, are determined to be maneuvering (M). Fast moving segments were deemed in transit (T). If slow moving segments have an average cosine of turning angle between records larger than 0.8 or smaller than −0.8, i.e., angle smaller than approximately 36.8 degrees, then the segment was also considered as in transit.
Besides checking by segment, the track was also checked record-wise for stationary records. First, the track was scanned for stationary records. If any of the four consecutive records moved less than 100 m, their status was set to stationary (S). If the whole track was stationary, no further process was applied, and the whole track was marked as stationary. The distance to the nearest coast of these stationary records are then determined by querying the NASA coastal distance database [
32]. The status of records which are within 2 km from the nearest coast were overridden to landing (L), and a possible landing site was written the the record by referencing the list of known landing sites, which is extracted from the MMAF website [
33].
Segment-wise and record-wise results were then consolidated and written into normalized track record, which was then de-normalized back to its original time stamp and written into the database. The same algorithm was applied to all gear types, as we found the output to be satisfactory within the scope of this study.
2.4. VBD Cross Matching
The threshold of matching predicted VMS record to VBD was set to be within 700 m and 5 s. The reason to set the distance threshold at 700 m was in consideration of the VIIRS-DNB pixel footprint size, while a temporal threshold of 5 s was because the precision of overpass prediction was 2 s. Such a threshold was considered to provide enough tolerance. If a VBD match was found for any predicted record, the matched VBD record ID along with the distance and temporal difference was recorded.
There were cases when a VBD record was found with similar coordinate in neighboring orbits. Being able to add temporal discrimination in matching VMS and VBD gave us the ability to match the exact VBD record within orbit overlap areas.
4. Discussion
Our study has four objectives: (1) to identify the gear types that are commonly detected by VIIRS in Indonesia, (2) to rate the probability that a boat is fishing if it is detected by VIIRS, (3) to determine the probable gear type for a VBD in Indonesia fishery management zones (WPP), and (4) to determine if different styles of lighting can be discerned. To address these objectives we combined two types of data. The reference data are 32 months of VMS data supplied by the Indonesia Ministry of Marine Affairs and Fisheries (MMAF). The subject data are VIIRS boat detections from the same 32 months. Out of 3683 VMS equipped vessels, 2632 had at least one VBD match.
The methodology can be divided into two parts. The first is an algorithm for cross-matching VIIRS boat detections with VMS tracks. The process involves predicting the probable location of a VMS equipped vessel at the time of each VIIRS data collection using an orbital model. The probable vessel location is interpolated between the two immediate neighboring VMS records found before and after the predicted VIIRS overpass time. Matches are confirmed if the VIIRS has a vessel detection within 700 m and 5 s of the predicted location. The second part involves algorithm to segments and classifies VMS records into landing, transit and fishing activity types based on their location, velocity and change in heading.
VMS/VBD match rates were calculated separately for the transit and fishing activity types. The gear types show two different types in VBD match rates for the fishing activity type. Our interpretation is that gear type with low match rates are not using lights to attract catch and are only occasionally detected when nighttime operations call for extra deck lighting. This is the group with fishing match rates less than 11%: basic longline (RD), longline tuna (RT), pole and line (H), carrier (P), gill net (JIO/JLB), and trawlers (PI/PUD). Gear types that are routinely using lighting to attract catch, with match rates in excess of 30% include two types of squid boats (jigging and stick dipnet), handline, handline tuna, and two types of purse seiners. The match rates for transit activity lacks this two-typed distribution and is consistently in the 0–10% range, comparable to the range found for gear types that are not using lights to attract catch. Based on these results we conclude that if a vessel is detected by VIIRS, there is a 96.42% probability that it is fishing.
The difference in light usage during fishing may also associate with the use of fish aggregating devices (FADs). FADs are known to be a common practice, enabling fishermen to use minimal effort for maximal catch. FADs are extensively used to attract tuna, for they tend to be attracted by floating objects. The exact number and distribution of FADs is unclear, while licensed numbers in Indonesia in 2006–2008 counts near 100 [
36], some say there are 3858 or more FADs existed in Indonesian water [
40]. Deployment of deep-sea FADs are reported to be found in provinces like North/West Sumatra, Lampung, east/west Java, Celebes, Maluku and Papua [
36], which has immediate access to deep waters as shown in
Figure 1. In these waters, we found smaller numbers of VBD/VMS matches as shown in
Figure 7, and with very little to no vessel lighting observed as shown in
Figure 14. This connection implies that the two-typed distribution on match rate for purse seiners operating in Indonesia is due to the difference in how they operate. Those operating in WPP 715, 716, and 717 are utilizing deep-water FADs while those in WPP 712 and 713 are not. Likewise, large pelagic purse seine with one ship (PCOB) operating in WPP 572 and 573 also more likely to rely on deep-water FADs hence they used less lighting compared to those operating in shallow waters. Overall, as
Figure 14 suggests, all bright lighting is observed in shallow waters including WPP 711, 712, and 718, while deep water regions like WPP 572, 273, 714, 715, 716, and 717 seldom detects bright lights.
During the 32 month data period, there were 2.1 million VBD records and the match rate to VMS vessels was 6.72%. Thus 93% of the VBD (2 million) records were not represented in the VMS record. We believe the vast majority of the VBD lacking VMS are from vessels using lights to attract catch that are under 30 GT level that triggers the VMS requirement. Other possible reasons why vessels may be detected by VIIRS but lack VMS are the possibilities that the vessel has turned off their VMS or is an illegal foreign fishing vessel. The spatial context of the VBD lacking VMS may be used to guide the interpretation. For instance, VBD lacking VMS in the far northern part of the Natuna Sea are suspect as foreign fishing vessels.
By taking the gear type match rates as a sample representing the gear types engaged in fishing with lights it is possible to calculate the probable gear types for VBD (
Table 8). Across all of Indonesia it is most likely to find a VBD representing a small pelagic purse seiner (PCK). This is the most likely gear type for VBD in 6 of the eleven WPP zones (571, 572, 711, 713, 715, and 716). For the remaining five WPP the probable gear type for a VBD ranges from large pelagic purse seine with one ship (PCOB) in WPP 573, squid dipnet (BA) in WPP 712, longline tuna (RT) in WPP 717 and 714, and squid jigging in WPP 718.
Consideration of the match rates versus average radiance indicates that most gear types have a largely consistent style of lighting. The exception is small pelagic purse seiner (PCK), which shows two distinct populations of vessel lighting styles in
Figure 10c. There is one loose cluster with a center of mass near 60% raw match rate and an average radiance of 50 nW/sr/cm
2. The other population forms a largely vertical column in
Figure 10c, with raw match rates less than 40% and average radiances less than 5 nW/sr/cm
2.
Table 4 and
Table 5 offers a geographical clue in that PCK has high match rates with brighter lighting in the Natuna Sea and Makassar Straits and extremely low match rates in the Maluku Sea with dimmer lighting. In tracing the geography of the vertical column cluster of
Figure 10c we found these vessels are operating in the Maluku Sea. The second author conducted lighting surveys on PCK vessels at landing sites on Java Island and in the Maluku Sea region (Ambon and Bitung). The PCK at Java landing sites invariably carried string of bare metal halide bulbs above the deck edges, in some cases reaching 100,000 watts in total.
Besides the using of FADs, in Ambon and Bitung the PCK still use lights to attract catch, but deploy small numbers of submersible lights. Field survey of lighting used on PCK vessels operating out of Ambon and Bitung indicates the boats have 1000 to 3000 watts of deck lighting and use 1000 to 8000 watts of submersible lighting as shown in
Figure 11. Our interpretation is the vertical column on
Figure 10c is an indication for the use of submersible lighting. Certainly the detection of submersible lighting on fishing boats by satellite is quite challenging. Not only is the wattage used quite small but also the water is opaque in the near infrared, which covers half of the DNB bandpass, which straddles the visible and NIR. This study lays the foundation for further study to improve VBD regarding this issue.
While no field data has been collected on PCOB, it is clear from
Figure 10d that he primary cluster is a vertical column quite similar to the one seen on
Figure 10c. This suggests that PCOB vessels are commonly using submersible lights. Another gear type with a similar vertical column cluster is longline tuna (RT). Submersible LED lights are widely used to attract catch on RT vessels [
41], with on board lighting also becoming more popular in recent years as shown in
Table 5. Based on these new understandings, we can further investigate the impact on VBD resulted from difference lighting methods.
5. Conclusions
A methodology has been developed for cross-matching VIIRS boat detections with status-classified VMS tracks. The process involves predicting the probable location of a VMS equipped vessel at the time of each VIIRS data collection. An orbital model is used to calculate VIIRS overpass times. VMS tracks are segmented and classified by their location, speed, and heading change to determine the status of each record. The algorithm can also be applied to AIS tracks.
Given an extended temporal record, it would be possible to develop DNB lighting profiles for individual VMS vessels, providing additional information for more detailed analyses of vessel behavior. Such vessel specific profiles could include the overall match rate, spatial distribution of matches, and radiance levels. It is also possible to derive group profiles for all vessels of a specific gear type or associated with a specific landing site.
There are several applications for cross-matching VIIRS boat detections with GPS based vessel tracks. If the matching is done in near real time as VBD and VMS data becomes available, it will be a great help to the authororties by identifying “dark vessels” that lack an operating VMS or AIS system. In some areas, this is a tip-off for potentially illegal fishing. Examples may include fishery closures, restricted waters, or Exclusive Economic Zone (EEZ) boundary zones. It is reasonable to assume that most AIS and VMS vessels operate legally most of the time. It has been reported that vessels turn off their AIS or VMS devices while engaged in illegal fishing. Thus a VBD lacking AIS or VMS in fishing grounds where all vessels should reasonably have one or the other is instantly suspect in terms of IUU fishing. Another potential application is the identification of offshore transshipment events. Several organizations analyze AIS and VMS tracks to identify transshipment events [
42,
43,
44]. The typical positive identification of transshipment is two or more stationary vessels in an offshore location. The conclusion is particularly strong if one of the vessels is registered as a carrier or transporter. But what about the cases where an AIS or VMS vessel is “loitering” with no other AIS or VMS vessel present? If the loitering period extends after midnight, our methods make it possible to check for VBD matches that may indicate a transshipment event. This evidence could be interpreted relative to the DNB match rate and radiance profile of the subject AIS or VMS vessel. Common questions from the VBD users are what type of vessels are detected by VIIRS and can the VBD be interpreted as fishing activity, as opposed to transit? The results from this study confirmed that in Indonesia, five gear types including squid fishing, lift net, and purse seiners commonly deploy and operate lights to attract catch. This includes what are the detection limits in terms of light output. The results from this study show that given an extended period of observation, it is possible to develop match rate profiles for individual vessels that can be used with some advantage.
The vast numbers of VBD that lack VMS testify to the value of the VBD data in monitoring fishing grounds where lighting is used to attract catch. In fishing grounds where lights are used to attract catch, VBD provide the most complete near real time data source for locating fishing activity. EOG posts the latest VBD output within 4 to 6 h from satellite overpass. The time includes data gathering and processing.
Further development of this study can be used to estimate unreported catch and improve stock assessments as well as other related information relevant for decision makers. Moreover, improved monitoring programs can benefit the assessment of efficient harvest strategies through improved estimates of maximum sustainable yield (MSY), biomass and carrying capacity of the fishery [
45]; while the problems with catch under-reporting appear to be particularly serious in Indonesia, especially for catches of tuna and tuna like species [
46].
Verifying the reported logbooks against landing data using VBD and VMS data proposed by this research, could address the problem of uncertainty in the level of reporting from fishers to the fishing port authority. The urgency of such verification using VBD and VMS data is due to the current trip length which was estimated by the number of fishing days by departure and arrival dates of vessels, while some purse-seine vessels, however, transfer their catch to a carrier vessel that brings the catch ashore without any proper catch reporting to the authority.