1 Introduction
-
A new database6 containing more than nine thousand trajectories recorded from 828 fishing vessels with a sampling period of 5 min, to overcome some limitations of previous study based on hourly sampling periods. This database reduce by more than 10 the Nyquist band-limit of existing databases.
-
We present comprehensive experiments including spatio-temporal features. These features were extracted using two different approaches: one based on local analysis and the other on global analysis of the trajectories.
-
A novel method based on the fusion of global and local features to classify the trajectories of vessels according to their fishing gear with high reliability.
-
We present a model specifically trained to detect Trawl Fishing Gear, achieving classification accuracies of over \(99\%\).
-
We extend the Tragsatec database by increasing the number of fishing vessels. Whilst 357 different fishing vessels were included in [32], here we present information on 828, an increase of almost 2.5 times the previous database.. Furthermore, compared to the 5 fishing gear classes of [32], the database presented here spans 7 different classes. Nevertheless, only 5 classes are included in the multi-classification experimental section due to the limited number of samples for the two new classes.
-
We expand our experiments (see Sect. 4.3) by training and evaluating specific trawl detection models, an useful application to prevent IUU fishing. We also provide and ablation study to understand how factors such as data availability or sampling period influence the performance.
2 Related works
3 Database and methodology
3.1 Database
Attributes | Ours | Indonesian VMS [6] | Thai VMS/AIS [7] |
---|---|---|---|
Fishing vessels | 828 | 1227 | 32 |
GPS positions | 1.43 M | 5.26 M | 184.5 K |
Trajectories | 9.37 K | – | 771 |
Observed days | 66 | 365 | – |
Sampling period (min) | \(5 \pm 0.83\) | \(60\pm 15\) | 120 |
Nyquist Band-limit (Hz) | 1/600 | 1/7,200 | 1/14,400 |
Classes (Fishing Gears) | Trawls, Purse seines, Trammel, Longline, Gillnets, Dredges, Pots and traps | Trawls, Longlines, Purse seines, Pole and line | Trawls, Purse Seines, Longlines, Reefer |
Info record | Description | Samples | Fields |
---|---|---|---|
AIS messages | Messages issued by the vessel’s AIS beacon | 5 M | Geom. position, date, hour, speed, course, vessel id |
Vessels | Basic data of a vessel | 1647 | Vessel id, usual fishing gears |
Diary statements | Info on when and where the vessels start/end their navigation | 31.8 K | Diary id, vessel id, departure date, return date, departure port id, return port id |
Fishing gears carried | Fishing gears carried on board | 33.1 K | Record id, diary id, fishing gear id |
Fishing gears | Information about fishing gears | 157 | Fishing gear id, name, details, code |
Ports | Information about ports | 11.6 K | Port id, geometric outline of the port, name |
-
Trawls: A fishing method that involves dragging a cone-shaped net, usually known as trawl, along the ocean floor to capture the target species.
-
Purse seines and surrounding nets: This technique consists in encircling an entire area or school of fish with a surrounding wall of net (i.e. the seine) that hang vertically. Then, the bottom is pulled close to trap the fish inside.
-
Gillnets: A fishing method that hangs a wall of net, typically made of nylon, vertically in a water column. Fish swimming into the net are entangled, with a backward structure that prevents their escape.
-
Trammel: A variation of the gillnets which employs up to three layers of nets.
-
Longline: This technique consists in attaching a long main line with bated hooks behind the boat. The bated hooks are attached at intervals to attract the different species of fish.
-
Dredges: This technique involves the use of a rigid structure called dredge to collect shellfish by dragging the dredge along the seafloor.
-
Pots and traps: This is a stationary method of capturing sea animals, in which pots and traps are deployed for a period of time (e.g. 24 h) and then hauled aboard to harvest the trapped fish.
3.2 Data curation
3.3 Feature extraction
3.3.1 Global features
-
Time: 25 features related to the duration of the trajectory, events such as raising the digital pen, or local maximums/minimums.
-
Velocity and acceleration: 25 features obtained from the first and second order temporal derivatives of position-temporal functions, such as the standard deviation of these.
-
Direction: 18 features extracted from the trajectory, for instance the starting direction, or direction histograms.
-
Geometry: 32 features associated with the line or aspect of the dynamic trajectory.
3.3.2 Local features
3.4 Classification models
4 Experiments and results
4.1 Experimental protocol
-
SVM. Two different hyper-parameters are tuned for the SVM, the complexity C, and the \(\gamma\) value. The complexity controls the trade-off between correctly classifying all training samples (i.e., low values of C) and maximizing the margin of the classifier (i.e., high values of C). On the other hand, \(\gamma\) controls the curvature of the decision boundary through the RBF function, with high values of \(\gamma\) representing more curvature. We will explore values of \(C\in\)[1, 10, 100], and \(\gamma \in\)[0.1, 0.01, 0.001].
-
Random Forests. For the RF model, we will only explore the number of estimators N, which is the number of decision trees included in the forest. In this work, the consider values of \(N\in\)[101, 1, 10 K].
-
Neural Network. Two different hyper-parameters are considered for the NN classifier, namely the number of units L in the hidden layer, and the learning rate of the network \(\alpha\). Note that the output layer of the Neural Network contains the same number of units as classes in the multi-class configuration (i.e., 5 output units) and uses softmax, while only 1 output unit with sigmoid activation is used for binary classification. As for the hyper-parameters, we will explore the values of \(L\in\)[100, 1, 10 K], and \(\alpha \in\)[\(1e-3\), \(1e-4\), \(1e-5\)].
4.2 Multi-class fishing gear classification
Classifier | Features | Hyperparameters | Mean acc. (95% CI) [%] |
---|---|---|---|
BiGRU | Local | – | \(75.60\,{\pm\, 3.62}\) |
MLP | Global | \(L=1\),000; \(\alpha =0.0001\) | \(82.69\,{\pm\, 2.00}\) |
SVM | Global | \(C=100\); \(\gamma =0.01\) | \(83.16\,{\pm\, 1.77}\) |
RF | Global | \(N=101\) | \(86.22\,{\pm\, 2.66}\) |
RF + BiGRU | Fusion | \(w_{RF} = 0.8\); \(w_{BiGRU} = 0.2\) | \(\mathbf {90.13\,{\pm\, 3.13}}\) |
4.3 Binary fishing gear classification: trawls detection
Classifier | Features | Hyperparameters | mAP (95% CI) [%] |
---|---|---|---|
BiGRU | Local | – | \(99.81\,{\pm\, 0.04}\) |
MLP | Global | \(L=1\),000; \(\alpha =0.0001\) | \(99.82\,{\pm\, 0.08}\) |
SVM | Global | \(C=10\); \(\gamma =0.01\) | \(99.83\,{\pm\, 0.03}\) |
RF | Global | \(N=1\),001 | \(99.97\,{\pm\, 0.01}\) |
MLP + BiGRU | Fusion | \(w_{MLP}=0.49\); \(w_{BiGRU}=0.51\) | \(99.91\,{\pm\, 0.01}\) |
SVM + BiGRU | Fusion | \(w_{SVM}=0.52\); \(w_{BiGRU}=0.48\) | \(99.91\,{\pm\, 0.01}\) |
RF + BiGRU | Fusion | \(w_{RF}=0.75\); \(w_{BiGRU}=0.25\) | \(\mathbf {99.98\,{\pm\, 0.01}}\) |
Classifier | Features | Threshold (%) | EER (%) |
---|---|---|---|
BiGRU | Local | 45.11 | 1.32 |
MLP | Global | 88.93 | 0.99 |
SVM | Global | 70.32 | 0.89 |
RF | Global | 53.15 | 0.86 |
MLP + BiGRU | Fusion | 52.23 | 0.56 |
SVM + BiGRU | Fusion | 51.06 | \(\mathbf {0.43}\) |
RF + BiGRU | Fusion | 57.96 | 0.64 |
4.3.1 Ablation study: effect of number of training samples and sampling period
\(S_t\) | Classifier | Features | Hyperparameters | Mean acc. (95% CI) [%] |
---|---|---|---|---|
200 | BiGRU | Local | – | \(84.31\,{\pm\, 0.87}\) |
MLP | Global | \(L=1\),000; \(\alpha =0.001\) | \(95.13\,{\pm\, 0.16}\) | |
SVM | Global | \(C=10\); \(\gamma =0.01\) | \(95.65\,{\pm\, 0.27}\) | |
RF | Global | \(N=1\),001 | \(95.85\,{\pm\, 0.23}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}=0.63\); \(w_{BiGRU}=0.37\) | \(94.85\,{\pm\, 0.30}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}=0.85\); \(w_{BiGRU}=0.15\) | \(95.48\,{\pm\, 0.26}\) | |
RF + BiGRU | Fusion | \(w_{RF}=0.98\); \(w_{BiGRU}=0.02\) | \(\mathbf {95.93\,{\pm\, 0.24}}\) | |
400 | BiGRU | Local | – | \(88.41\,{\pm\, 0.79}\) |
MLP | Global | \(L=1\),000; \(\alpha =0.0001\) | \(96.44\,{\pm\, 0.23}\) | |
SVM | Global | \(C=10\); \(\gamma =0.001\) | \(96.06\,{\pm\, 0.25}\) | |
RF | Global | \(N=10\),001 | \(97.20\,{\pm\, 0.16}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}=0.72\); \(w_{BiGRU}=0.28\) | \(96.20\,{\pm\, 0.29}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}=0.57\); \(w_{BiGRU}=0.43\) | \(95.74\,{\pm\, 0.18}\) | |
RF + BiGRU | Fusion | \(w_{RF}=0.95\); \(w_{BiGRU}=0.05\) | \(\mathbf {97.24\,{\pm\, 0.18}}\) | |
1 K | BiGRU | Local | – | \(94.13\,{\pm\, 0.68}\) |
MLP | Global | \(L=100\); \(\alpha =0.00001\) | \(97.44\,{\pm\, 0.23}\) | |
SVM | Global | \(C=10\); \(\gamma =0.01\) | \(97.94\,{\pm\, 0.12}\) | |
RF | Global | \(N=1\),001 | \(98.26\,{\pm\, 0.22}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}=0.55\); \(w_{BiGRU}=0.45\) | \(97.69\,{\pm\, 0.08}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}=0.81\); \(w_{BiGRU}=0.19\) | \(97.90\,{\pm\, 0.10}\) | |
RF + BiGRU | Fusion | \(w_{RF}=0.93\); \(w_{BiGRU}=0.07\) | \(\mathbf {98.38\,{\pm\, 0.18}}\) | |
2 K | BiGRU | Local | – | \(97.35{\pm 0.46}\) |
MLP | Global | \(L=1\),000; \(\alpha =0.0001\) | \(98.85\,{\pm\, 0.07}\) | |
SVM | Global | \(C=100\); \(\gamma =0.001\) | \(98.71\,{\pm\, 0.10}\) | |
RF | Global | \(N=101\) | \(97.76\,{\pm\, 0.22}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}=0.50\); \(w_{BiGRU}=0.50\) | \(\mathbf {99.23\,{\pm\, 0.16}}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}=0.50\); \(w_{BiGRU}=0.50\) | \(98.99\,{\pm\, 0.22}\) | |
RF + BiGRU | Fusion | \(w_{RF}=0.75\); \(w_{BiGRU}=0.25\) | \(98.95\,{\pm\, 0.17}\) | |
4 K | BiGRU | Local | – | \(96.91\,{\pm\,1.02}\) |
MLP | Global | \(L=10\),000; \(\alpha =0.0001\) | \(99.03\,{\pm\, 0.11}\) | |
SVM | Global | \(C=100\); \(\gamma =0.001\) | \(98.99\,{\pm\, 0.13}\) | |
RF | Global | \(N=1\),001 | \(98.71\,{\pm\, 0.09}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}=0.40\); \(w_{BiGRU}=0.60\) | \(98.09\,{\pm\, 0.65}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}=0.62\); \(w_{BiGRU}=0.38\) | \(\mathbf {99.43\,{\pm\, 0.11}}\) | |
RF + BiGRU | Fusion | \(w_{RF}=0.67\); \(w_{BiGRU}=0.33\) | \(98.53\,{\pm\, 0.35}\) | |
5.28 K | BiGRU | Local | – | \(98.10\,{\pm\, 0.26}\) |
MLP | Global | \(L=10\),000; \(\alpha =0.001\) | \(98.90\,{\pm\, 0.12}\) | |
SVM | Global | \(C=100\); \(\gamma =0.01\) | \(98.65\,{\pm\, 0.13}\) | |
RF | Global | \(N=10\),001 | \(98.99\,{\pm\, 0.10}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}=0.50\); \(w_{BiGRU}=0.50\) | \(\mathbf {99.36\,{\pm\, 0.09}}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}=0.51\); \(w_{BiGRU}=0.49\) | \(99.28\,{\pm\, 0.12}\) | |
RF + BiGRU | Fusion | \(w_{RF}=0.67\); \(w_{BiGRU}=0.33\) | \(98.99\,{\pm\, 0.10}\) |
\(T_s\) (min) | Classifier | Features | Hyperparameters | mAP (95% CI) [%] |
---|---|---|---|---|
5 | BiGRU | Local | – | \({98.64}\,{\pm\, 0.26}\) |
MLP | Global | \(L={100}\); \(\alpha ={0.001}\) | \({99.83}\,{\pm\, 0.02}\) | |
SVM | Global | \(C={10}\); \(\gamma =0.01\) | \(\mathbf {{99.93}\,{\pm\, 0.01}}\) | |
RF | Global | \(N=10\),001 | \({99.86}\,{\pm\, 0.01}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}={0.87}\); \(w_{BiGRU}={0.13}\) | \({99.90}\,{\pm\, 0.02}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}={0.93}\); \(w_{BiGRU}={0.07}\) | \(\mathbf {{99.93}\,{\pm\, 0.01}}\) | |
RF + BiGRU | Fusion | \(w_{RF}={0.84}\); \(w_{BiGRU}={0.16}\) | \({99.90}\,{\pm\, 0.02}\) | |
10 | BiGRU | Local | – | \({98.93}\,{\pm\, 0.19}\) |
MLP | Global | \(L=100\); \(\alpha ={0.001}\) | \({99.83}\,{\pm\, 0.03}\) | |
SVM | Global | \(C=10\); \(\gamma =0.01\) | \({99.90}\,{\pm\, 0.02}\) | |
RF | Global | \(N=10\),001 | \({99.77}\,{\pm\, 0.01}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}={0.77}\); \(w_{BiGRU}={0.23}\) | \(\mathbf {{99.91}\,{\pm\, 0.03}}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}={0.64}\); \(w_{BiGRU}={0.36}\) | \(\mathbf {{99.91}\,{\pm\, 0.02}}\) | |
RF + BiGRU | Fusion | \(w_{RF}={0.71}\); \(w_{BiGRU}={0.29}\) | \({99.84}\,{\pm\, 0.02}\) | |
20 | BiGRU | Local | – | \({98.80}\,{\pm\, 0.27}\) |
MLP | Global | \(L={1}\),000; \(\alpha =0.001\) | \({99.77}\,{\pm\, 0.08}\) | |
SVM | Global | \(C={100}\); \(\gamma =0.01\) | \({99.87}\,{\pm\, 0.02}\) | |
RF | Global | \(N=10\),001 | \({99.66}\,{\pm\, 0.02}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}={0.91}\); \(w_{BiGRU}={0.09}\) | \({99.87}\,{\pm\, 0.03}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}={0.85}\); \(w_{BiGRU}={0.15}\) | \(\mathbf {{99.90}\,{\pm\, 0.02}}\) | |
RF + BiGRU | Fusion | \(w_{RF}={0.79}\); \(w_{BiGRU}={0.21}\) | \({99.80}\,{\pm\, 0.02}\) | |
35 | BiGRU | Local | – | \({98.56}\,{\pm\, 0.28}\) |
MLP | Global | \(L=100\); \(\alpha ={0.0001}\) | \({99.67}\,{\pm\, 0.05}\) | |
SVM | Global | \(C=100\); \(\gamma =0.01\) | \({99.70}\,{\pm\, 0.03}\) | |
RF | Global | \(N=1\),001 | \({99.49}\,{\pm\, 0.04}\) | |
MLP + BiGRU | Fusion | \(w_{MLP}={0.71}\); \(w_{BiGRU}={0.29}\) | \(\mathbf {{99.80}\,{\pm\, 0.04}}\) | |
SVM + BiGRU | Fusion | \(w_{SVM}={0.67}\); \(w_{BiGRU}={0.33}\) | \({99.77}\,{\pm\, 0.03}\) | |
RF + BiGRU | Fusion | \(w_{RF}={0.63}\); \(w_{BiGRU}={0.37}\) | \({99.71}\,{\pm\, 0.05}\) |