Whale counting in satellite and aerial images with deep learning
Whale counting in satellite and aerial images with deep learning
复制标题
DOI：
10.1038/s41598-019-50795-9
复制
发表时间：
2019-10-03
影响因子：
4.6
通讯作者：
Herrera, Francisco
中科院分区：
文献类型：
--
作者：
Guirado, Emilio;Tabik, Siham;Herrera, Francisco
尽管鲸鱼的兴趣和威胁地位很高，但世界海洋中的鲸鱼数量仍然高度不确定。鲸鱼的检测通常是通过昂贵的目击调查、声学调查或高分辨率图像进行的。由于深度卷积神经网络(CNN)在一些计算机视觉任务中取得了很好的性能，本文基于开放的数据和工具，提出了一个健壮的、可推广的基于CNN的卫星和航空图像中鲸鱼的自动检测和计数系统。特别是，我们设计了一个两步鲸鱼计数方法，第一个CNN找到有鲸鱼存在的输入图像，第二个CNN定位和计数这些图像中的每一头鲸鱼。该系统在全球十个鲸鱼观赏热点的谷歌地球图像上进行了测试，获得了81%的检测性能和94%的鲸鱼计数性能。将这两个步骤结合起来，与单独使用基线检测模型相比，准确率提高了36%。在全球范围内应用这一具有成本效益的方法可能有助于对鲸鱼数量的评估，以指导保护行动。出于保护目的，全球免费获取高分辨率图像将推动这一进程。
Despite their interest and threat status, the number of whales in world's oceans remains highly uncertain. Whales detection is normally carried out from costly sighting surveys, acoustic surveys or through high-resolution images. Since deep convolutional neural networks (CNNs) are achieving great performance in several computer vision tasks, here we propose a robust and generalizable CNN-based system for automatically detecting and counting whales in satellite and aerial images based on open data and tools. In particular, we designed a two-step whale counting approach, where the first CNN finds the input images with whale presence, and the second CNN locates and counts each whale in those images. A test of the system on Google Earth images in ten global whale-watching hotspots achieved a performance (F1-measure) of 81% in detecting and 94% in counting whales. Combining these two steps increased accuracy by 36% compared to a baseline detection model alone. Applying this cost-effective method worldwide could contribute to the assessment of whale populations to guide conservation actions. Free and global access to high-resolution imagery for conservation purposes would boost this process.