Blue eyes: New tools for monitoring coastal environments using remotely piloted aircraft and machine learning

沿海环境的遥感是一个快速发展的领域，遥控飞机(RPA)显示出越来越大的潜力来测绘和跟踪沿海地物的分布、健康和动态，例如植被、贝类、鸟类、哺乳动物、地貌甚至凋落物(例如Anderson&Gaston，2013)。与传统环境监测技术相比，它们的优势是部署速度快，能够提供整个地貌及其环境特征的超高分辨率2-D和3-D图像，可用于准确绘制地图并客观评估物理环境和生物多样性的细微或早期变化。RPA监测产生的信息使管理者能够详细和准确地了解沿海生态系统的整个特征如何随着气候、水文条件和人类活动的变化而变化，因此它们代表了如何管理这些生态系统及其自然和人为危害的蜕变。对图像中的环境特征进行检测和分类所采用的方法是海洋生态系统RPA使用成功的关键。传统的图像处理技术使用基于像素的监督分类，但RPA图像的超高分辨率会使这种方法变得繁琐和无效-增加了成本，并导致数据生产的延迟。因此，必须专门开发新技术来处理RPA映像中包含的大量信息。这个项目的目标是开发用于RPA图像分析的定制工具，使用的是一种复杂的机器学习技术，最近已被证明在许多计算机视觉应用中提供了阶梯变化(例如，LeCun等人)。，2015)。特别是，我们预计现有的3-D信息与传统的RGB成像的融合将是该项目的主要焦点之一，因为它承诺显著提高性能(Gupta等人)。，2014)。算法开发将需要大量用于训练的注解图像数据集和CEFAS提供的关于图像外观的专家知识。CEFA有一个广泛的RPA图片库，涵盖海滩、泥滩和岩石海岸，这些图像是由监测海岸地貌和侵蚀、海草、盐沼、藻类和海洋无脊椎动物的项目产生的。学生将从上面提到的想法开始应用各种计算技术，以期开发识别关键沿海特征的算法，其好处是工具开发可以在博士学位开始时立即开始。随着研究的进展，学生还将为Cefas RPA航班的设计和部署做出贡献，为任何难以识别的环境特征提供有针对性的信息。开发的工具将能够更快、更有效地解释图像，极大地提高沿海变化跟踪技术的效用。

Remote sensing of coastal environments is a rapidly evolving field, with remotely piloted aircraft (RPA) showing growing potential for mapping and tracking the distribution, health and dynamics of coastal features such as vegetation, shellfish, birds, mammals, geomorphology and even litter (e.g. Anderson & Gaston, 2013). Their advantages over traditionalenvironmental monitoring techniques is in their speed of deployment and ability to deliver ultra-high resolution 2-D and 3-D images of entire landscapes and the environmental features within them, which can be used to accurately map and objectively assess even subtle or early-day changes in the physical environment and in biodiversity. The informationgenerated from RPA monitoring allows managers to gain a detailed and accurate understanding of how entire features of coastal ecosystems change in response to climate, hydrographic conditions and anthropogenic activities and so they represent a metamorphosis in how these ecosystems and their natural and anthropogenic hazards can be managed. The methods employed to detect and classify the environmental features in the images are pivotal to the success of RPAuse in marine ecosystems. Traditional image processing techniques use pixel-based supervised classification, but the super-high resolution of RPA images can render such methods cumbersome and ineffective - increasing costs and causing delays in data production. Hence, new techniques must be developed specifically to deal with the significant amounts of information contained in RPA images. This studentship aims to develop bespoke tools for RPA image analysis using `deeplearning', a complex machine learning technique that has recently proven to provide a step-change in a number of computer vision applications (e.g. LeCun et al. ,2015). In particular, we envisage that fusion of the now available 3-D information with the traditional RGB imaging will be one of the main focal points of the project as it promises significant improvements in performance (Gupta et al. ,2014). The algorithm development will require a large dataset of annotatedimagery for training and the expert knowledge on the image appearance which are available in Cefas. Cefas have an extensive library of RPA images covering beaches, mudflats and rocky shores generated from projects monitoring coastal geomorphology and erosion, seagrass, saltmarsh, algae and marine invertebrates. The student will apply various computational techniques starting with the ideas mentioned above with a view to develop algorthims for identifying the key coastal features, with the benefit that tool development can begin immediately on commencement of the PhD. The student will also contribute to the design and deployment of Cefas RPA flights as the research progresses, to provide targeted information for any environmental features that provde difficult to identify. The developed tools will allow a faster and more effective interpretation of images, vastly improving the utility of the coastal change tracking technology.