202603 - Vessel detections from Sentinel 2

Vessel detections with Sentinel-2 satellite imagery from 2019 to 14 days ago. Multiyear detection of vessels with respective length, speed and orientation estimates based on deep-learning models applied to electro-optical imagery (RGB and near-infrared) at 10-m resolution covering most exclusive economic zones and marine protected areas across the ocean. Sentinel-2 detections are based on electro-optical imagery at 10-m resolution. Compared with our published Sentinel-1 detections based on SAR imagery at 20-m resolution, Sentinel-2 allows detecting smaller vessels with all types of vessel materials, in contrast with the weakness of SAR imagery on revealing vessels made of wood and fiberglass. Sentinel-2 also images the wakes of moving vessels, which further increase the detectability of small vessels and allow us to infer the vessels’ speed and orientation.The rich information contained in the optical imagery can be used to map not only vessel presence, but also vessel activities such as vessel encounters and bottom trawling. Sentinel-2 also covers more area of the ocean than Sentinel-1, and the novel neural-net detection approach does not require the exclusion of near-shore regions (as opposed to the CFAR approach), allowing detection of vessels in highly-packed areas all the way to the shoreline where most human activity is concentrated. Overall, with Sentinel-2 imagery we are able to detect about 3 times more vessels, “see” a broader range of vessel lengths and types, and infer more information of vessel activities than in our previous mapping using Sentinel-1 imagery.基于2019年至14天前的哨兵二号（Sentinel-2）卫星影像开展船舶检测。本数据集依托深度学习模型对10米分辨率的光电影像（红-绿-蓝RGB与近红外波段）进行处理，实现多年级船舶检测，并估算对应船舶的长度、航速与航向，覆盖全球绝大多数专属经济区与海洋保护区。 哨兵二号（Sentinel-2）检测基于10米分辨率的光电影像。相较于我们已发布的基于20米分辨率合成孔径雷达（Synthetic Aperture Radar, SAR）影像的哨兵一号（Sentinel-1）检测结果，哨兵二号可检测所有材质类型的小型船舶，而合成孔径雷达影像难以识别木质与玻璃纤维材质的船舶。此外，哨兵二号可捕捉移动船舶的尾迹，进一步提升小型船舶的可检测性，并支持我们推算船舶航速与航向。光电影像蕴含的丰富信息不仅可用于绘制船舶存在分布，还可用于刻画船舶活动（如船舶会遇与海底拖网作业）。相较于哨兵一号，哨兵二号的覆盖海域范围更广；其新颖的神经网络检测方法无需像恒虚警率（Constant False Alarm Rate, CFAR）方法那样排除近岸区域，可在直至人类活动高度集中的海岸线附近的高密度船舶区域开展检测。总体而言，相较于此前基于哨兵一号影像的测绘工作，利用哨兵二号影像可多检测约3倍的船舶，“观测”到更广尺寸范围与类型的船舶，并能推算更多船舶活动相关信息。