Fast and accurate mapping of fine scale abundance of a VME in the deep sea with computer vision
Fast and accurate mapping of fine scale abundance of a VME in the deep sea with computer vision
复制标题
DOI：
10.1016/j.ecoinf.2022.101786
复制
发表时间：
2022-11-01
影响因子：
5.1
通讯作者：
Howell, Kerry L.
中科院分区：
文献类型：
--
作者：
Piechaud, Nils;Howell, Kerry L.
随着人类对深海生态系统的压力越来越大，需要大量的数据来了解它们的生态，监测随着时间的推移而发生的变化，并向保护管理者提供信息。目前的图像分析方法速度太慢，不能满足这些要求。最近，生物学家更容易接触到计算机视觉，这可能有助于应对这一挑战。在这项研究中，我们演示了一种非专家可以训练YOLOV4卷积神经网络(CNN)的方法，该网络能够计算和测量单一类别的对象。我们将CV应用于从东北大西洋1200米深的AUV拍摄的58,000多张图像中提取关于异植生物Syringammina Fragilissima的密度和种群规模结构的定量数据。开发的工作流程使用了开源工具和基于云的硬件，只需要在生态学家中常见的CV方面的经验。CNN表现出色，召回率为0.84，准确率为0.91。模型预测得到的每幅图像的个体计数和大小测量结果与手动收集的数据高度相关(分别为0.96和0.92)。分析可以在不到10天的时间内完成，从而为这一脆弱海洋生态系统的种群规模结构和细微规模分布带来新的见解。结果表明，脆弱链霉菌的分布呈斑块状。平均密度为2.5微米(-2)，但可以从高达45微米(-2)不等，距离几乎没有它的地区只有几十米。平均大小为5.5厘米，最大的个体(>；15厘米)往往在低密度地区。这项研究展示了研究人员如何利用CV快速有效地生成关于底栖生态系统范围和分布的大型量化数据集数据。这一点，再加上AUV的大采样能力，可以绕过图像分析的瓶颈，极大地便利未来的深海勘探和监测。它还说明了这些新技术在实现联合国海洋十年设定的目标方面的未来潜力。
With growing anthropogenic pressure on deep-sea ecosystems, large quantities of data are needed to understand their ecology, monitor changes over time and inform conservation managers. Current methods of image analysis are too slow to meet these requirements. Recently, computer vision has become more accessible to biologists, and could help address this challenge. In this study we demonstrate a method by which non-specialists can train a YOLOV4 Convolutional Neural Network (CNN) able to count and measure a single class of objects. We apply CV to the extraction of quantitative data on the density and population size structure of the xenophyophore Syringammina fragilissima, from more than 58,000 images taken by an AUV 1200 m deep in the North-East Atlantic. The workflow developed used open-source tools, cloud-base hardware, and only required a level of experience with CV commonly found among ecologists. The CNN performed well, achieving a recall of 0.84 and precision of 0.91. Individual counts per image and size measurements resulting from model predictions were highly correlated (0.96 and 0.92, respectively) with manually collected data. The analysis could be completed in less than 10 days thus bringing novel insights into the population size structure and fine scale distribution of this Vulnerable Marine Ecosystem. It showed S. fragilissima distribution is patchy. The average density is 2.5 ind.m(-2) but can vary from up to 45 ind.m(-2) only a few tens of meter away from areas where it is almost absent. The average size is 5.5 cm and the largest individuals (>15 cm) tend to be in areas of low density. This study demonstrates how researchers could take advantage of CV to quickly and efficiently generate large quantitative datasets data on benthic ecosystems extent and distribution. This, coupled with the large sampling capacity of AUVs could bypass the bottleneck of image analysis and greatly facilitate future deep-ocean exploration and monitoring. It also illustrates the future potential of these new technologies to meet the goals set by the UN Ocean Decade.