Mbani Benson, Greinert Jens
DeepSea Monitoring Group, GEOMAR Helmholtz Center for Ocean Research Kiel, Wischhofstraße 1-3, 24148, Kiel, Germany.
Institute of Geosciences, Kiel University, Ludewig-Meyn-Str. 10-12, 24118, Kiel, Germany.
Sci Rep. 2025 Jul 28;15(1):27481. doi: 10.1038/s41598-025-12723-y.
The deep-sea comprises diverse habitats and species whose characterisation provides crucial insights into the health and resilience of our oceans. Whereas direct sampling enables investigation of the vertical variability of the seafloor at small spatial scales, optical imaging allows for multi-scale assessment of the spatial distribution of (mega)benthos and substrates. However, modern seafloor imaging surveys typically generate thousands of images that are infeasible to manual annotation. Consequently, transforming these terabyte-scale datasets into actionable insights requires automated workflows. Here, we deployed two A.I workflows to automate the annotation of substrates and megafaunal taxa in seafloor images from the tropical North Atlantic. Clustering, feature space visualisation and multivariate statistical analysis techniques were used to classify the seafloor into habitats, estimate megafaunal distribution patterns, and to identify environmental drivers that influence observed patterns. We found that the seabed here formed seven clearly distinct clusters, with visible sub-partitions observed in each cluster. Investigations revealed a gradient of sediment disturbance due to biogenic activity, with images showing little-to-no sediment disturbance mapping to one half of the feature space, whereas images exhibiting visibly vigorous sediment reworking mapping to the other half of the feature space. Also, megafaunal abundances were 14 times higher in the shallower Eastern region of the seabed, potentially due to higher Particulate Organic Carbon flux and relatively warmer temperatures. Moreover, geographic clustering of megafauna was observed in topographically complex features such as slopes of submarine canyons and on top of seamounts, where heterogeneity created diverse microhabitats and unique niches that megafauna could exploit.
深海包含各种不同的栖息地和物种，对它们的特征描述能为我们深入了解海洋的健康状况和恢复能力提供关键信息。直接采样能够在小空间尺度上研究海底的垂直变化，而光学成像则可以对（巨型）底栖生物和基质的空间分布进行多尺度评估。然而，现代海底成像调查通常会生成数千张图像，人工标注这些图像是不可行的。因此，要将这些太字节规模的数据集转化为可操作的见解，就需要自动化工作流程。在这里，我们部署了两种人工智能工作流程，以自动标注来自热带北大西洋的海底图像中的基质和大型动物类群。我们使用聚类、特征空间可视化和多元统计分析技术，将海底分类为不同的栖息地，估计大型动物的分布模式，并识别影响观测模式的环境驱动因素。我们发现，这里的海床形成了七个明显不同的聚类，每个聚类中都观察到了明显的子分区。调查揭示了生物活动导致的沉积物扰动梯度，图像显示几乎没有沉积物扰动的映射到特征空间的一半，而显示出明显剧烈沉积物重塑的图像映射到特征空间的另一半。此外，在海床较浅的东部区域，大型动物的丰度高出14倍，这可能是由于较高的颗粒有机碳通量和相对较高的温度。此外，在地形复杂的特征区域，如海底峡谷的斜坡和海山顶部，观察到了大型动物的地理聚类，这些地方的异质性创造了多样化的微生境和独特的生态位，大型动物可以利用。
