Automated classification of three-dimensional reconstructions of coral reefs using convolutional neural networks
Automated classification of three-dimensional reconstructions of coral reefs using convolutional neural networks
复制标题
DOI：
10.1371/journal.pone.0230671
复制
发表时间：
2020-03-24
期刊：
影响因子：
3.7
通讯作者：
Bhandarkar, Suchendra M.
中科院分区：
文献类型：
--
作者：
Hopkinson, Brian M.;King, Andrew C.;Bhandarkar, Suchendra M.
珊瑚礁是生物多样性和结构复杂的生态系统，它们分别受到人类活动的影响。因此，需要对珊瑚礁进行快速生态评估，但目前的方法需要耗时的人工分析，无论是在潜水调查期间还是在调查期间收集的图像上。珊瑚礁结构的复杂性对生态功能是必不可少的，但衡量起来具有挑战性，往往归因于简单的衡量标准，如粗糙度。计算机视觉和机器学习的最新进展为缓解其中一些限制提供了可能性。我们开发了一种对珊瑚礁剖面的3D重建进行自动分类的方法，并评估了该方法的准确性。使用商业Structure-From-Motion软件利用从视频调查中提取的图像生成珊瑚礁切片的3D重建。为了生成3D分类地图，3D重建上的位置被映射回原始图像以提取该位置的多个视图。测试了几种方法，将一个点的多个视图的信息合并到一个分类中，所有这些方法都使用卷积神经网络来分类或从图像中提取特征，但合并信息所采用的策略不同。合并信息的方法包括投票、概率平均和学习的神经网络层。所有方法都执行了类似的操作，在大多数类别上获得了接近96%和90%的总体分类准确率。这些方法具有很高的分类精度，适合于许多生态应用。
Coral reefs are biologically diverse and structurally complex ecosystems, which have been severally affected by human actions. Consequently, there is a need for rapid ecological assessment of coral reefs, but current approaches require time consuming manual analysis, either during a dive survey or on images collected during a survey. Reef structural complexity is essential for ecological function but is challenging to measure and often relegated to simple metrics such as rugosity. Recent advances in computer vision and machine learning offer the potential to alleviate some of these limitations. We developed an approach to automatically classify 3D reconstructions of reef sections and assessed the accuracy of this approach. 3D reconstructions of reef sections were generated using commercial Structure-from-Motion software with images extracted from video surveys. To generate a 3D classified map, locations on the 3D reconstruction were mapped back into the original images to extract multiple views of the location. Several approaches were tested to merge information from multiple views of a point into a single classification, all of which used convolutional neural networks to classify or extract features from the images, but differ in the strategy employed for merging information. Approaches to merging information entailed voting, probability averaging, and a learned neural-network layer. All approaches performed similarly achieving overall classification accuracies of similar to 96% and > 90% accuracy on most classes. With this high classification accuracy, these approaches are suitable for many ecological applications.