Improved pattern recognition in environmental signals using machine learning

识别不同类型的环境信号的模式和分类对于理解产生它们的过程至关重要。机器学习可以显著提高我们从噪音中识别信号的能力，并增加我们挖掘大量环境数据的能力，这意味着它可以提取比传统方法更多的信息。该项目最初将使用在婆罗洲、土耳其和法罗群岛记录的连续宽带地震数据，这些数据来自之前由NERC资助或由NERC地球物理设备设施支持的项目。这些地震数据集记录了自然发生的环境过程的连续时间序列测量，是测试和开发机器学习技术的理想训练数据。应用于地震数据的检测和分类算法通常侧重于地震信号，因为它们构成了危险。然而，其他可以检测到的自然现象，如山体滑坡和岩石坠落，也可能代表着重大的风险。由于世界某些地区气候变化导致降雨量增加，这些现象在未来几年可能会变得更加普遍。改进我们对这类现象的检测，更好地了解造成这些现象的机制，对于减轻它们所造成的危害至关重要。通过使用地震数据，我们的理解可能会加深，然而，目前的一个挑战是，与地震目录相比，这些信号的目录相对较少。利用婆罗洲、土耳其和法罗群岛的数据集，学生将生成不同类型信号的注释数据集，如岩崩、滑坡、采石场爆炸、运输和动物运动。这些带注释的数据集将与有监督的机器学习算法一起使用，生成在不同位置检测到的不同类型的自然和人为地震信号的目录。考虑到这三个地点有不同的噪声分布、不同的仪器和可能不同的信号源，这将提供一个机会，让人们了解如何在全球范围内最好地将这种算法应用于公开可用的地震数据。新的、扩大的环境信号目录将用于研究产生这些信号的过程。

The identification of patterns in, and the classification of, different types of environmental signals is crucial for understanding the processes that generate them. Machine learning can significantly improve our ability to identify signal from noise and increase our capacity to mine large volumes of environmental data, meaning it is possible to extract more information than is possible with traditional methodologies. This project will initially use continuous broadband seismic data recorded in Borneo, Turkey and the Faroe Islands from projects that have previously been funded by NERC or supported through the NERC Geophysical Equipment Facility. These seismological datasets record continuous time series measurements of naturally occurring environmental processes and are ideal training data for testing and developing machine learning techniques.Detection and classification algorithms applied to seismological data typically focus on earthquake signals due to the hazards that they pose. However, other natural phenomena that can be detected, such as landslides and rock falls, can also represent a significant risk. With increased rainfall due to a changing climate in some parts of the world, these are phenomena that are likely to become more common in the coming years. Improving our detection of such phenomena, and better understanding of the mechanisms that cause them, is vital to mitigate against the hazards they pose. Our understanding may be deepened through the use of seismological data, however one challenge at present is that catalogues of such signals are relatively sparse in comparison to catalogues of earthquakes. Using the datasets from Borneo, Turkey and the Faroes, the student will generate annotated datasets of different types of signals, such as rockfall, landslides, quarry blasts, transport, and animal movement. These annotated datasets will then be used with supervised machine learning algorithms to generate catalogues of different types of natural and anthropogenic signals detected seismically in different locations. Given that the three locations have different noise profiles, different instrumentation, and potentially different sources of signals, this will provide an opportunity to develop understanding of how best to apply such algorithms on a global scale to publicly-available seismological data. The new, expanded, catalogues of environmental signals will then be used to investigate the processes that generate them.