(PART) Global Fishing Watch Datasets

Fishing effort and vessel presence

Data Overview

The flagship datasets of Global Fishing Watch are vessel presences and apparent fishing effort based on transmissions from vessel tracking datasets. These datasets represent the amount of time vessels are present in an area and, for fishing vessels, how much of that time is spent fishing. Using cloud computing, machine learning, and public vessel registry information, GFW analyzes tens of millions of AIS positions each day to map global apparent fishing effort.

Producing such a dataset involves two key steps:

  1. Identification of fishing vessels in the AIS data

  2. Detection of fishing activity

We combine our comprehensive database of vessels with a known vessel type from vessel registries with two convolutional neural networks (CNN)—a cutting edge form of machine learning model—to help us classify vessels and predict when they are fishing. The resulting output are a list of vessels by type (e.g. trawler, drifting_longline, cargo, etc.) and a dataset of AIS positions, where each position is labeled as a fishing/non-fishing position. These datasets can then be combined to evaluate the presence and fishing activity of individual vessels or make rasters of aggregate vessel presence or fishing activity.

Note: Every AIS position receives a fishing score (or NULL if the model can’t make a prediction), regardless of whether that MMSI is classified as a fishing vessel.

Key Tables

  • pipe_ais_v3.research_messages - AIS positions, thinned to one minute per segment, for all vessels.

  • pipe_production_vYYYYMMDD.research_segs - Summaries of vessel activity by segment (seg_id). Includes fields indicating which segments are likely noise (e.g. good_seg, good_seg2, overlapping_and_short) and should be filtered out.

  • pipe_production_vYYYYMMDD.research_segs_daily - Daily summaries of vessel activity and fishing by segment (seg_id), including by EEZ. This table can be used for non-spatial analyses that summarize vessel activity over specific time periods and/or by EEZ.

Public Fishing Effort Tables

GFW’s public fishing effort and vessel presence datasets (2012-2020) are gridded at 100th degree resolution by flag state and geartype and at 10th degree resolution by MMSI, respectively. These tables are available in a public BigQuery project (global-fishing-watch) and are inexpensive ways to quickly map fishing effort and vessel presence.

  • global-fishing-watch.gfw_public_data.fishing_effort_v2: Public daily fishing effort by flag state and geartype (2012-2020) at 0.01 degree resolution.
  • global-fishing-watch.gfw_public_data.fishing_effort_byvessel_v2: Public daily fishing effort by MMSI (2012-2020) at 0.1 degree resolution.

Data Description

is_fishing_vessel

The research_messages table includes positions for all vessels, of which ~75% are non-fishing vessels. Therefore, it can be expensive and inefficient to query the whole table for fishing analyses. To compensate for this, the table is clustered on the boolean is_fishing_vessel field, which indicates positions belonging to potential fishing vessels - any ssvid on one of the fishing lists in the vessel info tables.

For fishing analyses, include a WHERE is_fishing_vessel filter statement to reduce query size by approximately 80%

hours and fishing_hours

Each AIS position is assigned an amount of time (hours), which is calculated as the time since the vessel’s previous AIS position. Vessel presence is calculated by simply summing the hours of all AIS positions from good segments in the analysis.

AIS positions also receive two types of fishing score (0-1) - one from the GFW fishing detection algorithm (nnet_score), and one from a separate algorithm that detects likely fishing by squid jiggers (night_loitering). AIS positions where the relevant fishing score is equal to 1 are considered fishing positions and all hours assigned to that position are considered fishing effort (fishing_hours).

Fishing score (nnet_score vs. night_loitering)

The GFW fishing detection algorithm performs poorly for squid jiggers, which fish at night using lights while remaining largely stationary. To account for this, a second fishing algorithm identifies when vessels exhibit this night loitering behavior. Because the vessel class associated with an MMSI can change - either over time or as a result of improved GFW vessel classification - all AIS positions receive both fishing scores so that they are available. As a result, when calculating fishing effort, it is important to first assign a vessel class to each MMSI in order to use night_loitering for squid jiggers and nnet_score for all other fishing vessels.

Caveats & Known Issues

While AIS provides a revolutionary way to monitor global commercial fishing activity, there are several important limitations and caveats.

Many fishing vessels are not trackable via AIS

AIS data includes only a small fraction—approximately 70,000—of the world’s estimated 2.8 million fishing vessels. Coverage is much higher for larger vessels with less than 1 percent of vessels under 12 meters represented, 14-19 percent for vessels between 12-24 meters, and 52-85 percent for vessels larger than 24 meters. The International Maritime Organization mandates AIS for most vessels larger than 36 meters, and vessels broadcasting AIS are predominantly from upper and upper-middle income countries.

is_fishing_vessel outdated over time

The is_fishing_vessel field indicates positions for potential fishing vessels in order to make fishing-related queries considerably cheaper. In order to do so, however, the table is created from the list of all ssvid on one of the fishing lists using the most current vessel info table at the time the research_messages table is created (which you can see in the DETAILS tab of the research_messages table). As a result, over time as new MMSI appear in the AIS data, or vessel classifications change for existing MMSI, not all fishing vessels may have positions marked as is_fishing_vessel = True.

If you are using concerned about missing vessels, simply omit the is_fishing_vessel filter and join on the most up-to-date fishing vessel list.

GFW identifies apparent fishing effort

The GFW fishing detection algorithm is a best effort mathematically to identify “apparent fishing activity.” It is possible that some fishing activity is not identified as such by our algorithms and, conversely, we may predict fishing activity where fishing is not actually taking place. For these reasons, GFW qualifies designations of vessel fishing activity, including synonyms of the term “fishing activity,” such as “fishing” or “fishing effort,” as “apparent,” rather than certain.

Fishing activity is under-represented in areas of poor AIS reception

For various technical reasons, not every AIS message that is broadcast is recorded. Satellites must be overhead to receive AIS signals, terrestrial receivers only receive signals near shore, AIS messages can interfere with each other in areas of high vessel density, and AIS devices vary in broadcast strength and frequency. As a result, fewer AIS positions are received by vessels operating in certain parts of the world, limiting the effectiveness of our machine learning models and our ability to detect apparent fishing effort.

Example Queries

  • fishing_hours_by_position.sql: Basic query to pull AIS positions for fishing vessels and calculate fishing hours.

  • fishing_hours_gridded.sql: Query to create a raster of fishing hours, including fishing hours per square kilometer.

  • fishing_hours_by_eez.sql: Query to quickly calculate fishing hours per EEZ over a specific date range using the daily segments research table (research_segs_daily)