Plankton and particles from Seaexplorer glider and UVP6 across the Ligurian Front

We targeted a permanent mesoscale front in the Ligurian Sea (NW Mediterranean) that we repeatedly sampled between January and June 2021 using a SeaExplorer glider equipped with a UVP6, a versatile in situ imager. We aimed to resolve plankton and particle distribution during the spring bloom, to assess whether the front was a location of increased concentration of zooplankton, and if it constrained the distribution of particles. During the 5 months, the glider did more than 5,000 dives and the UVP6 collected 1.1 million images. Images captured by the UVP6 during cruising (n = 785,405) were imported into the Morphocluster application to quickly detect large clusters of similar objects (e.g. marine snow aggregates). In a second step, images collected during back transects (n = 434,129, on which we focused our analyses) were imported onto the EcoTaxa web application with their Morphocluster label in order to be sorted at a finer scale into taxonomic or morphological groups (marine snow, artefact, badfocus, reflection or unidentifiable) with the help of a supervised machine learning algorithm. As sorting all 400k+ images would have required a multiple months effor, we instead decided to rely on the prediction of a Random Forest classifier fed with both handcrafted and deep features generated by a MobileNet V2 feature extractor previously finetuned on UVP6 data. We selected a RF classifier for the following reasons: RFs tend to produce good classification probability estimates (Niculescu-Mizil and Caruana 2005), they are faster to train than a full CNN stack and, when trained with deep features, they perform as well as a full CNN. The dataset thus contains the following elements: - CTD data, some collected by the glider payload, and other collected by a SMRU - particles data, exported from Ecopart - plankton data, exported from Ecotaxa. Validated objects were either individually inspected by an operator, or batch validated in the morphocluster application. Predicted classifications were not reviewed.

我们将研究目标定为利古里亚海（西北地中海）的一处永久性中尺度锋面（mesoscale front），于2021年1月至6月期间，使用搭载UVP6多功能原位成像仪的SeaExplorer滑翔机开展重复采样。本研究旨在解析春季水华期间的浮游生物与颗粒物分布，评估该锋面是否为浮游动物浓度升高的区域，以及是否对颗粒物分布构成约束。
在为期5个月的作业中，该滑翔机累计完成超过5000次下潜，UVP6共采集110万张图像。其中巡航阶段采集的785405张图像被导入Morphocluster应用程序，以快速检测相似物体的大型集群（例如海洋雪聚集体）。
第二步，我们将返程断面采集的434129张图像（为本研究分析的核心数据集）连同其Morphocluster标签一同导入EcoTaxa网页应用，借助监督式机器学习算法，将图像进一步细分为分类学或形态学组别，包括海洋雪、人工伪影、失焦、反光或无法识别样本。
由于手动标注全部40余万张图像需耗时数月，我们转而采用经UVP6数据微调的MobileNet V2特征提取器生成的手工特征与深度特征，输入随机森林分类器（Random Forest classifier）进行预测。选择随机森林分类器的原因如下：其一，随机森林可生成优质的分类概率估计结果（Niculescu-Mizil与Caruana，2005）；其二，其训练速度快于完整卷积神经网络（CNN）栈；其三，当使用深度特征训练时，其性能可媲美完整卷积神经网络。
本数据集包含以下要素：
- 温盐深仪（CTD）数据：部分由滑翔机载荷采集，其余由SMRU采集
- 颗粒物数据：从Ecopart导出
- 浮游生物数据：从EcoTaxa导出
经标注的样本均经过核验：或由操作人员逐一人工检视，或通过Morphocluster应用程序完成批量核验；而模型预测得到的分类结果未经过人工复核。