Satellite mapping reveals extensive industrial activity at sea - training data

Data repository for the paper "Satellite mapping reveals extensive industrial activity at sea (Paolo et al. 2024, Nature)". Here we present a detailed global map of major industrial activities at sea. To detect and characterize vessels and offshore infrastructure in coastal waters around the globe, we analyzed 2 petabytes of satellite imagery spanning 2017–2021, with our analyses covering over 15% of the ocean (Extended Data Figure 1) where more than 75% of industrial activity is concentrated (see Methods). We designed and trained three deep convolutional neural networks to identify objects (>97% accuracy) and estimate their lengths (0.84 R2); to classify offshore infrastructure into oil, wind, and other objects (>98% accuracy); and to classify vessels as fishing or non-fishing (>90% accuracy). Combined, we classified over 67 million image tiles, including dual-polarization synthetic-aperture radar (SAR) imagery from Sentinel-1 and optical (red, green, blue, and near infrared) imagery from Sentinel-2. The resolution of SAR allows us to capture the majority of objects larger than 15 meters (detection rate of >70% for 25 m vessels, and >90% for 50 m and larger, Extended Data Figure 2). We also analyzed 53 billion vessel GPS positions from AIS and matched them to the satellite detections to determine whether a detected vessel was publicly tracked.