DataSheet_1_Biomass estimations of cultivated kelp using underwater RGB images from a mini-ROV and computer vision approaches.docx

Seaweed farming is the fastest-growing aquaculture sector worldwide. As farms continue to expand, automated methods for monitoring growth and biomass become increasingly important. Imaging techniques, such as Computer Vision (CV), which allow automatic object detection and segmentation can be used for rapid estimation of underwater kelp size. Here, we segmented in situ underwater RGB images of cultivated Saccharina latissima using CV techniques and explored pixel area as a tool for biomass estimations. Sampling consisted of underwater imaging of S. latissima hanging vertically from a cultivation line using a mini-ROV. In situ chlorophyll a concentrations and turbidity (proxies for phytoplankton and particle concentrations) were monitored for water visibility. We first compared manual length estimations of kelp individuals obtained from the images (through manual annotation using ImageJ software). Then, we applied CV methods to segment and calculate kelp area and investigated these measurements as a robust proxy for wet weight biomass. A strong positive linear correlation (r2 = 0.959) between length estimates from underwater image frames and manual measurements from the harvested kelp was observed. Using unsupervised learning algorithms, such as mean shift clustering, colour segmentation and adaptive thresholding from the OpenCV package in Python, kelp area was segmented and the number of individual pixels in the contour area was counted. A positive power relationship was found between length from manual measurements with CV-derived area (r2 = 0.808) estimated from underwater images. Likewise, CV-derived area had a positive power relationship with wet weight biomass (r² = 0.887). When removing data where visibility was poor due to high turbidity levels (mid-June), the power relationship was stronger between CV-derived area estimates and the field measurements (r² = 0.976 for wet weight biomass and r² = 0.979 for length). These results show that robust estimates of cultivated kelp biomass in situ are possible through kelp colour segmentation. However, we demonstrate that the quality of CV post-processing and accuracy of the model are highly dependent of environmental conditions (e.g. turbidity and chlorophyll a concentrations). The establishment of these technologies has the potential to offer scalability of production, efficient real-time monitoring of sea cultivation and improved yield predictions.

海藻养殖业是全球增长最快的水产养殖行业。随着养殖场的不断扩张，对生长和生物量的自动化监测方法变得越来越重要。成像技术，如计算机视觉（CV），能够实现自动目标检测和分割，可用于对水下海带尺寸的快速估算。在本研究中，我们利用计算机视觉技术对养殖的 Saccharina latissima 的现场水下 RGB 图像进行了分割，并探讨了像素面积作为生物量估算工具的应用。采样包括使用微型遥控潜水器（ROV）对悬挂在养殖线上的 S. latissima 进行水下成像。通过监测现场叶绿素 a 浓度和混浊度（作为浮游植物和颗粒浓度的指标）来评估水透明度。我们首先比较了从图像中获得的手动海带个体长度估算值（通过使用 ImageJ 软件进行手动标注）。然后，我们应用计算机视觉方法对海带进行分割和面积计算，并研究了这些测量值作为湿重生物量的稳健代理。观察到水下图像帧的长度估算与收获海带的手动测量之间存在着强烈的正线性相关（r² = 0.959）。通过使用无监督学习算法，如 Python 中 OpenCV 包的均值漂移聚类、颜色分割和自适应阈值，对海带面积进行了分割，并计算了轮廓面积内的单个像素数量。发现手动测量的长度与计算机视觉得出的面积之间存在正幂关系（从水下图像中估算的 r² = 0.808）。同样，计算机视觉得出的面积与湿重生物量之间存在正幂关系（r² = 0.887）。当移除由于混浊度水平高（六月中期）而导致透明度差的数据时，计算机视觉得出的面积估算值与现场测量值之间的幂关系更强（湿重生物量的 r² = 0.976 和长度的 r² = 0.979）。这些结果表明，通过海带颜色分割，可以实现对现场养殖海带生物量的稳健估算。然而，我们证明了计算机视觉后处理的质量和模型的准确性高度依赖于环境条件（例如混浊度和叶绿素 a 浓度）。这些技术的建立有可能为生产的可扩展性提供支持，实现海栽培的实时高效监测，并提高产量预测的准确性。