A method to map seafloor substrates using machine learning, based on geophysical data including multibeam bathymetry, backscatter, and side-scan sonar, is currently being developed. Results from a case study in Table Bay, southwestern South Africa, are presented here, showing a method of physical seafloor classification that uses a number of statistical algorithms and software programs. In the first step of the process, a customised tool was created within ArcGIS using python scripting language to classify seafloor bathymetry, which can be applied to areas beyond South Africa. The tool created by the authors was based on pioneering work done by the National Oceanic and Atmospheric Administration on a benthic terrain modelling toolbox. In step two, multibeam bathymetry, backscatter and side-scan sonar data processed using Qimera, Fledermaus Geocoder Toolbox, and Navlog processing software, were classified using machine learning techniques including Decision Trees, Random Forests, and k-means clustering computer algorithms. The results from these algorithms were compared to classify the seafloor substrate distribution. Our results have allowed a comparison of advantages and disadvantages of each machine learning technique and we found that the k-means clustering techniques were the simplest to implement and understand and worked best based on their seafloor segmentation capabilities in Table Bay, with all three data sets (multibeam bathymetry, backscatter and side-scan sonar). In future research, ground-truthing methods (for example underwater video and grab samples) will be used to validate the interpreted data to create accurate seafloor substrate maps. This work provides the initial steps to develop a holistic predictive tool that classifies geophysical data into substrate maps using machine-learning techniques. The maps can be used to model biological communities and to produce benthic habitat maps for use in marine science and management.

目前正在开发一种基于机器数据的海底底物测绘方法，该方法基于包括多波束测深，反向散射和侧扫声纳在内的地球物理数据。此处显示的是南非西南部Table Bay的案例研究结果，显示了使用许多统计算法和软件程序进行的物理海底分类方法。在此过程的第一步中，在ArcGIS中使用python脚本语言创建了一个自定义工具，以对海底测深法进行分类，该工具可以应用于南非以外的地区。作者创建的工具是基于美国国家海洋与大气管理局在底栖地形建模工具箱上所做的开创性工作。在第二步中，使用Qimera处理了多光束测深，反向散射和侧扫声纳数据，Fledermaus Geocoder Toolbox和Navlog处理软件使用机器学习技术（包括决策树，随机森林和k-均值聚类计算机算法）进行了分类。比较了这些算法的结果，以对海底基质分布进行分类。我们的结果可以比较每种机器学习技术的优缺点，并且我们发现k-means聚类技术是最简单的实现方法，它基于表湾中的海底分割能力（具有全部三个数据集），因此工作效果最佳。 （多光束测深，反向散射和侧扫声纳）。在未来的研究中，将使用地面处理方法（例如，水下视频和抓取样本）来验证解释后的数据，以创建准确的海底底物图。这项工作提供了开发整体预测工具的初始步骤，该工具可使用机器学习技术将地球物理数据分类为基质地图。这些地图可用于为生物群落建模并制作底栖生物栖息地地图，以用于海洋科学和管理。