MistNet: Measuring historical bird migration in the US using archived weather radar data and convolutional neural networks
MistNet: Measuring historical bird migration in the US using archived weather radar data and convolutional neural networks
复制标题
DOI：
10.1111/2041-210x.13280
复制
发表时间：
2019-11-01
影响因子：
6.6
通讯作者：
Sheldon, Daniel
中科院分区：
文献类型：
--
作者：
Lin, Tsung-Yu;Winner, Kevin;Sheldon, Daniel
大型气象雷达网络是研究鸟类迁徙的综合工具。例如，美国的WSR-88 D网络覆盖了整个美国大陆，自20世纪90年代以来一直在存档数据。这些数据可以量化广泛和精细规模的鸟类运动，以解决一系列迁徙生态学问题。然而，自动区分降水与生物的问题大大限制了利用历史雷达数据进行生物分析的能力。我们开发了MistNet，这是一种深度卷积神经网络，可以在雷达扫描中区分降水和生物。与之前的机器学习方法不同，MistNet可以进行精细的预测，并可以从雷达扫描中收集生物信息，这些信息也包含降水。MistNet基于图像神经网络，包括几个针对雷达数据独特特征量身定制的架构组件。为了避免大量的人工标记工作，我们使用从双极化雷达数据中获得的大量噪声标签来训练MistNet。在历史和当代WSR-88 D数据中，MistNet识别了至少95.9%的生物量，错误发现率为1.3%。双极化训练数据和我们的雷达专用架构组件是有效的。通过在单次雷达扫描中保留与降水同时发生的生物量，MistNet比传统的全扫描筛选方法多保留15%的生物量。MistNet是完全自动化的，可以应用于数百万次雷达扫描的数据集，以产生细粒度的预测，从而实现一系列应用，从大陆规模的测绘到空域使用的本地分析。雷达鸟类学正在迅速发展，并导致有关大陆规模的鸟类运动模式的重大发现。通用和经验验证的方法来量化雷达数据中的生物信号是必不可少的，这一领域的未来发展。MistNet可以对整个迁移系统进行大规模、长期和可重复的测量。
Large networks of weather radars are comprehensive instruments for studying bird migration. For example, the US WSR-88D network covers the entire continental US and has archived data since the 1990s. The data can quantify both broad and fine-scale bird movements to address a range of migration ecology questions. However, the problem of automatically discriminating precipitation from biology has significantly limited the ability to conduct biological analyses with historical radar data. We develop MistNet, a deep convolutional neural network to discriminate precipitation from biology in radar scans. Unlike prior machine learning approaches, MistNet makes fine-scaled predictions and can collect biological information from radar scans that also contain precipitation. MistNet is based on neural networks for images, and includes several architecture components tailored to the unique characteristics of radar data. To avoid a massive human labelling effort, we train MistNet using abundant noisy labels obtained from dual polarization radar data. In historical and contemporary WSR-88D data, MistNet identifies at least 95.9% of all biomass with a false discovery rate of 1.3%. Dual polarization training data and our radar-specific architecture components are effective. By retaining biomass that co-occurs with precipitation in a single radar scan, MistNet retains 15% more biomass than traditional whole-scan approaches to screening. MistNet is fully automated and can be applied to datasets of millions of radar scans to produce fine-grained predictions that enable a range of applications, from continent-scale mapping to local analysis of airspace usage. Radar ornithology is advancing rapidly and leading to significant discoveries about continent-scale patterns of bird movements. General-purpose and empirically validated methods to quantify biological signals in radar data are essential to the future development of this field. MistNet can enable large-scale, long-term, and reproducible measurements of whole migration systems.