Identification and Classification of Relativistic Electron Precipitation at Earth Using Supervised Deep Learning
Identification and Classification of Relativistic Electron Precipitation at Earth Using Supervised Deep Learning
复制标题
DOI：
10.3389/fspas.2022.858990
复制
发表时间：
2022-03
期刊：
影响因子：
--
通讯作者：
L. Capannolo;Wen Li;Sheng Huang
中科院分区：
文献类型：
--
作者：
L. Capannolo;Wen Li;Sheng Huang
我们展示了在太空科学中的深入学习的应用。但是，我们的大气层尚未完全理解电子降水的详细特性和驱动因素，我们的目标是建立一个深度学习模型，以确定相对论的降水事件及其相关的驱动程序（WAVE或CSS）。依赖能量的降水模式。可以接受的事件可接受，并将其适当地分类为代表或CSS，因为将深度学习用于此任务的优势是有意义的，因为驱动器将其驱动器分类为相当廉价的降水事件，并且通常涉及人类的限制，因此可以通过统计范围驱动和CSS的属性。在所有L壳和MLT领域的过程及其相对作用，因此对于改善辐射带模型是有用的。
We show an application of supervised deep learning in space sciences. We focus on the relativistic electron precipitation into Earth’s atmosphere that occurs when magnetospheric processes (wave-particle interactions or current sheet scattering, CSS) violate the first adiabatic invariant of trapped radiation belt electrons leading to electron loss. Electron precipitation is a key mechanism of radiation belt loss and can lead to several space weather effects due to its interaction with the Earth’s atmosphere. However, the detailed properties and drivers of electron precipitation are currently not fully understood yet. Here, we aim to build a deep learning model that identifies relativistic precipitation events and their associated driver (waves or CSS). We use a list of precipitation events visually categorized into wave-driven events (REPs, showing spatially isolated precipitation) and CSS-driven events (CSSs, showing an energy-dependent precipitation pattern). We elaborate the ensemble of events to obtain a dataset of randomly stacked events made of a fixed window of data points that includes the precipitation interval. We assign a label to each data point: 0 is for no-events, 1 is for REPs and 2 is for CSSs. Only the data points during the precipitation are labeled as 1 or 2. By adopting a long short-term memory (LSTM) deep learning architecture, we developed a model that acceptably identifies the events and appropriately categorizes them into REPs or CSSs. The advantage of using deep learning for this task is meaningful given that classifying precipitation events by its drivers is rather time-expensive and typically must involve a human. After post-processing, this model is helpful to obtain statistically large datasets of REP and CSS events that will reveal the location and properties of the precipitation driven by these two processes at all L shells and MLT sectors as well as their relative role, thus is useful to improve radiation belt models. Additionally, the datasets of REPs and CSSs can provide a quantification of the energy input into the atmosphere due to relativistic electron precipitation, thus offering valuable information to space weather and atmospheric communities.