TensorFlow implementation of the model proposed in "A Multi-Task Deep Learning Architecture for Maritime Surveillance Using AIS Data Streams" (https://ieeexplore.ieee.org/abstract/document/8631498) and "GeoTrackNet—A Maritime Anomaly Detector using Probabilistic Neural Network Representation of AIS Tracks and A Contrario Detection" (https://arxiv.org/abs/1912.00682). (GeoTrackNet is the anomaly detection module of MultitaskAIS).

All the codes related to the Embedding block are adapted from the source code of Filtering Variational Objectives: https://github.com/tensorflow/models/tree/master/research/fivo

The elements of the code are organized as follows:

geotracknet.py                   # script to run the model (except the A contrario detection).
runners.py                        # graph construction code for training and evaluation.
bounds.py                         # code for computing each bound.
contrario_kde.py                  # script to run the A contrario detection.
contrario_utils.py
distribution_utils.py
nested_utils.py
utils.py
data
├── datasets.py                   # reader pipelines.
├── calculate_AIS_mean.py         # calculates the mean of the AIS "four-hot" vectors.
├── dataset_preprocessing.py      # preprocesses the AIS datasets.
└── csv2pkl.py                    # parse raw AIS messages from aivdm format to csv files.
└── csv2pkl.py                    # loads AIS data from *.csv files.
models
└── vrnn.py                       # VRNN implementation.
chkpt
└── ...                           # directory to keep checkpoints and summaries in.
results
└── ...                           # directory to save results to.

See requirements.yml

The MarineC dataset is provided by MarineCadastre.gov, Bureau of Ocean Energy Management, and National Oceanic and Atmospheric Administration, (marinecadastre.gov), and availble at (https://marinecadastre.gov/ais/)

The Brittany dataset is provided by CLS-Collecte Localisation Satellites (https://www.cls.fr/en/) and Erwan Guegueniat, comprises AIS messages captured by a coastal receiving station in Ushant, from 07/2011 to 07/2019. We provide here a set of processed AIS messages (data/ct_2017010203_10_20.zip) on which readers can re-produce the results in the paper GeoTrackNet. This set comprises dynamic information of AIS tracks (LAT, LON, SOG, COG, HEADING, ROT, NAV_STT, TIMESTAMP, MMSI) of cargo and tanker vessels from 01/2017 to 03/2017, downsampled to a resolution of 5 minutes. For the full Brittany dataset, please contact CLS (G. Hajduch, [email protected]).

Converting to csv:

• MarineC dataset: we use QGIS (https://qgis.org/en/site/) to convert the original metadata format to csv files.
• Brittany dataset: we use libais (https://github.com/schwehr/libais) to parse raw AIS messages to csv files (see avidm_decoder.py).

csv2pkl.py then loads the data from csv files, selects AIS messages in the pre-defined ROI, creates AIS trajectories (keyed by the MMSI) then saves them in pickle format.

Preprocessing steps: the data are processed as described in the paper by dataset_preprocessing.py.

First we must train the Embedding layer:

python geotracknet.py \
  --mode=train \
  --dataset_dir=./data \
  --trainingset_name=ct_2017010203_10_20/ct_2017010203_10_20_train.pkl \
  --testset_name=ct_2017010203_10_20/ct_2017010203_10_20_valid.pkl \
  --lat_min=47.5 \
  --lat_max=49.5 \
  --lon_min=-7.0 \
  --lon_max=-4.0 \
  --latent_size=100 \
  --batch_size=32 \
  --num_samples=16 \
  --learning_rate=0.0003 \

A model trained on the dataset comprising AIS messages of cargo and tanker vessels, from January 01 to March 10, 2017 can be found at chkpt/elbo-ct_2017010203_10_20_train.pkl-data_dim-602-latent_size-100-batch_size-50.zip.

After the Embedding layer is trained, we can run task-specific blocks.

To avoid re-caculating for each task, we calculate them once and save the results as a .pkl file.

python geotracknet.py \
  --mode=save_logprob \
  --dataset_dir=./data \
  --trainingset_name=ct_2017010203_10_20/ct_2017010203_10_20_train.pkl \
  --testset_name=ct_2017010203_10_20/ct_2017010203_10_20_valid.pkl \
  --lat_min=47.5 \
  --lat_max=49.5 \
  --lon_min=-7.0 \
  --lon_max=-4.0 \
  --latent_size=100 \
  --batch_size=32 \
  --num_samples=16 \
  --learning_rate=0.0003 \

Similarly for the test set (testset_name=ct_2017010203_10_20/ct_2017010203_10_20_test.pkl).

log_logprob divides the ROI into small cells and saves the of AIS messages in each cell.

python geotracknet.py \
  --mode=local_logprob \
  --dataset_dir=./data \
  --trainingset_name=ct_2017010203_10_20/ct_2017010203_10_20_train.pkl \
  --testset_name=ct_2017010203_10_20/ct_2017010203_10_20_valid.pkl \
  --lat_min=47.5 \
  --lat_max=49.5 \
  --lon_min=-7.0 \
  --lon_max=-4.0 \
  --latent_size=100 \
  --batch_size=32 \
  --num_samples=16 \
  --learning_rate=0.0003 \

contrario_detection detects abnormal vessels' behaviors using a contrario detection and plots the results.

python geotracknet.py \
  --mode=contrario_detection \
  --dataset_dir=./data \
  --trainingset_name=ct_2017010203_10_20/ct_2017010203_10_20_train.pkl \
  --testset_name=ct_2017010203_10_20/ct_2017010203_10_20_test.pkl \
  --lat_min=47.5 \
  --lat_max=49.5 \
  --lon_min=-7.0 \
  --lon_max=-4.0 \
  --contrario_eps=1e-10 \
  --latent_size=100 \
  --batch_size=32 \
  --num_samples=16 \
  --learning_rate=0.0003 \

We would like to thank MarineCadastre, CLS and Erwan Guegueniat, Kurt Schwehr, Tensorflow team, QGIS and OpenStreetmap for the data and the open-source codes.

We would also like to thank Jetze Schuurmans for helping convert the code from Python2 to Python3.

For any questions, please open an issue.