Fig 1.
Presentation of raw S-AIS tracks for three individual vessels using different fishing gear types in global overview (A) and more fine-scale representations of potential fishing behavior for a trawler (green, B), longliner (red, C) and purse seiner (blue, D).
Dots represent individual S-AIS signal detections, lines interpolated tracks. Note the global-range behavior of longliners, and the more regional basin-wide operations of purse seine and trawl vessels. Map data by Natural Earth.
Table 1.
Performance measures for the worldwide trawl dataset.
NF stands for probable non-fishing and F for probable fishing events. Sensitivity is related with non-fishing detection, and specificity with fishing detection. The column Stat. Diff. Fish Effort shows the t-test statistical comparison (p-value) between the predicted fishing effort time calculated from the algorithm’s labels and the expert’s labels. The asterisk indicates a significant difference.
Table 2.
Performance measures for the 16 longliner vessels in different oceans.
NF stands for probable non-fishing and F for probable fishing events. Sensitivity is related with non-fishing detection, and specificity with fishing detection. The column Stat. Diff. Fish Effort shows the t-test statistical comparison (p-value) between the predicted fishing effort time calculated from the algorithm’s labels and the expert’s labels. Two of the vessels could not be measured because they did not have any labeled fishing activity. The asterisk indicates a significant difference.
Fig 2.
Speed distributions for trawlers during fishing and non-fishing activities.
The red line represents probable non-fishing activity, and the black line probable fishing activity.
Fig 3.
Speed distributions for longliners during fishing and non-fishing activities.
The red line represents probable non-fishing activity, and the black line probable fishing activity.
Fig 4.
Speed distribution for Purse Seiners.
The red line indicates the probable non-fishing activity labeled by the expert, while the black line is the probable fishing activity. This distribution considers only the speeds reported by the vessels 10 km from shore, and during day time.
Fig 5.
Accuracy/Recall measured for trawlers with HMM using Monte Carlo Simulation.
Results do not consider the 10 km coastal distance threshold.
Fig 6.
Comparison of the Hidden Markov Model algorithm results to the expert labels.
Matching results for fishing activity presented in blue, expert labels in green and the algorithm’s fishing activity predictions in red. Empty circles represent non-fishing activity as identified by algorithm and expert. The track corresponds to vessel number 2 in Table 1. Map data by Natural Earth.
Fig 7.
Results for longliner number 8 from Table 2(Accuracy: 89%).
Matching results for fishing activity presented in blue, expert labels in green and the algorithm’s fishing activity predictions in red. Empty circles represent non-fishing activity as identified by algorithm and expert. Map data by Natural Earth.
Fig 8.
Results for longliner number 1 from Table 2 (Accuracy: 46%).
Matching results for fishing activity presented in blue, expert labels in green and the algorithm’s fishing activity predictions in red. Empty circles represent non-fishing activity as identified by algorithm and expert. Map data by Natural Earth.
Table 3.
Result for the purse seiner filtering approach.
The seven vessels were randomly chosen from multiple parts of the world. NF stands for probable non-fishing and F for probable fishing events. Sensitivity is related with non-fishing detection, and specificity with fishing detection.
Fig 9.
Comparison of the purse seiner algorithm results to the expert labels.
Matching results for fishing activity presented in blue, expert labels in green and the algorithm’s fishing activity predictions in red. Empty circles represent non-fishing activity as identified by algorithm and expert. The track corresponds to vessel number 6 in Table 3. Map data by Natural Earth.