The Interpretation of Particle Size, Shape, and Carbon Flux of Marine Particle Images Is Strongly Affected by the Choice of Particle Detection Algorithm
The Interpretation of Particle Size, Shape, and Carbon Flux of Marine Particle Images Is Strongly Affected by the Choice of Particle Detection Algorithm
复制标题
DOI：
10.3389/fmars.2020.00564
复制
发表时间：
2020-07
期刊：
影响因子：
--
通讯作者：
S. Giering;Brett Hosking;N. Briggs;M. Iversen
中科院分区：
文献类型：
--
作者：
S. Giering;Brett Hosking;N. Briggs;M. Iversen
海洋中颗粒的原位成像正在迅速成为研究海洋碳循环的强大工具，包括通过生物碳泵下沉颗粒对碳固定的作用。在分析相机图像中的颗粒时，一大挑战是确定颗粒的大小，这是计算碳含量、下沉速度和通量所必需的。图像处理的一个关键决策是用来确定图像的哪一部分形成粒子、哪一部分是背景的算法。然而，这一关键的分析步骤往往不被提及，其影响也很少被探讨。在这里，我们表明，当为单个数据集选择不同的算法时，最终的通量估计很容易发生数量级的变化。我们将一系列静态阈值和11种不同的算法(七种阈值和四种边缘检测算法)应用于LISST-Holo系统在两个不同环境中收集的颗粒轮廓。我们的结果表明，颗粒检测方法不仅影响估计的颗粒大小，而且还影响颗粒形状。当不同的颗粒检测方法混合时，例如，当来自不同研究或设备的数据集被合并时，不确定性可能会加剧。我们的结论是，显然需要更透明的方法描述和粒子检测算法的合理性，以及允许不同设备之间相互比较的校准标准。
In situ imaging of particles in the ocean are rapidly establishing themselves as powerful tools to investigate the ocean carbon cycle, including the role of sinking particles for carbon sequestration via the biological carbon pump. A big challenge when analysing particles in camera images is determining the size of the particle, which is required to calculate carbon content, sinking velocity and flux. A key image processing decision is the algorithm used to decide which part of the image forms the particle and which is the background. However, this critical analysis step is often unmentioned and its effect rarely explored. Here we show that final flux estimates can easily vary by an order of magnitude when selecting different algorithms for a single dataset. We applied a range of static threshold values and 11 different algorithms (seven threshold and four edge detection algorithms) to particle profiles collected by the LISST-Holo system in two contrasting environments. Our results demonstrate that the particle detection method does not only affect estimated particle size but also particle shape. Uncertainties are likely exacerbated when different particle detection methods are mixed, e.g., when datasets from different studies or devices are merged. We conclude that there is a clear need for more transparent method descriptions and justification for particle detection algorithms, as well as for a calibration standard that allows intercomparison between different devices.