14 SAR object detections

Synthetic Aperture Radar (SAR) object detections consist of vessels and offshore infrastructure detected, classified, and matched to AIS/VMS (when available) from satellite radar images taken by ESA’s Sentinel-1 mission.

14.1 1st Release (2022-06-08)

What to know regarding this preliminary release:

Fixed objects like infrastructure have been removed, so only vessels ¹
Detections refer, for the most part, to fishing and non-fishing vessels
Large areas with significant sea ice and icebergs have been excluded (high latitudes)
False positives (noise) have been filtered out with ML (this is not perfect)
Detections have been matched to available vessels with AIS
Detections are classified and separated as matched and unmatched
Detections are computed with a delay of 72 hours, and tables updated daily
Detections are available globally from January 1st 2022 onward ²
Detection footprints (area of the ocean scanned) are also provided

¹ Future releases will include all offshore infrastructure
² Future releases will include detections from 2015/2017 onward

Link to Map

Link to Data

14.2 Key Concepts

Data point: individual detection of maritime object from a single SAR scene
Polygon: outline of the valid area within the SAR scene used for detection
Matched: individual detection has been paired with a specific AIS message

14.3 Key Tables

BQ: world-fishing-827.proj_sentinel1_v20210924

detect_scene_raw_YYYYMMDD - id, lon, lat, time from Earth Engine detections
detect_foot_raw_YYYYMMDD - id, time, wkt-polygon from Earth Engine scene footprints
detect_scene_pred_YYYYMMDD - id, presence, length from machine learning prediction
detect_scene_match - id, matching-info, vessel-info from matching algorithm

14.4 Key Fields

scene_id - ID of Sentinel-1 scene from where detections were extracted
detect_id - ID unique to every detection, composed as scene_id;lon;lat
detect_lon - longitude coordinate
detect_lat - latitude coordinate
detect_timestamp - time the image was taken
ssvid - vessel identifier if the detection matches to AIS/VMS
score - score cutoff used to match the detection
confidence - confidence cutoff used to match the detection
presence - probability of vessel presence from a Conv Neural Net
length_m - length estimated with a Conv Neural Net
footprint_wkt - WKT multipolygon of image area used for detection
var1/var2 - vessel info before and after the SAR image used for matching

14.5 Key Buckets

GCS: Data output from Earth Engine used to produce the BQ Tables:

Detections: gs://gfw-production-sentinel1-detection-gee/v20210924/<date>
Footprints: gs://gfw-production-sentinel1-detection-footprints/v20210924/<date>
Thumbnails: gs://gfw-production-sentinel1-detection-thumbnails/v20210924/<date>

14.6 Data Description

The data tables contain all detections from 2017-01-01 to present, updated intermittently. Imagery availability increases along data acquisition timeframe, especially from mid-2016 on. Vessels are extracted from single scenes while fixed infrastructure is extracted from 6-month median composites, constructed every month with a moving window. Detection was performed on Google Earth Engine with a Constant False Alarm Rate (CFAR) algorithm using the GRD product, Interferometric Wide Mode, VH band at 20 m resolution. Vessel classification to identify noise and regression to estimate length was performed with a Convolutional Neural Network using both the VH and VV bands.

14.7 Caveats & Known Issues:

Sentinel-1 does not take images over the open ocean, mostly near coastal waters
spatial coverage is not homogeneous, e.g. European waters are imaged more frequently
temporal coverage is not homogenous, e.g. 2017 has less area covered than 2021
some missing days, e.g. in 2016, is due to no images available over the ocean
Sentinel-1B stopped operating on 2021-12-23, reducing daily images from ~400 to ~250
sea ice and icebergs, at high latitudes, are known to introduce false positives
images prior 2018 are of lower quality, introducing potential false positives