• Home
  • About
    • Kira's Blog Site

      Doctor X's blog site

    • Learn More
    • Email
    • Twitter
    • Github
    • Steam
    • Weibo
  • Posts
    • All Posts
    • All Tags
  • Data Analysis
  • Climate Projects
  • Videos
  • Life
  • Notes
  • Algorithm
  • Papers

Deep learning and process understanding for data-driven Earth system science

17 Feb 2020

关于DL应用于地球科学的文章:

  • Deep learning for multi-year ENSO forecasts
  • AI COPERNICUS ‘DISCOVERS’ THAT EARTH ORBITS THE SUN
  • Complex networks reveal global pattern of extreme-rainfall teleconnections (还没想好放到那个目录,姑且当作是深度学习的引用)

Deep learning and process understanding for data-driven Earth system science is a review published in 《Nature》 as a Perspectives

In the Paper, author review the development of machine learning in the geoscientific context, and highlight how deep learning—that is, the automatic extraction of abstract (spatio-temporal) features—has the potential to overcome many of the limitations that have, until now, hindered a more wide-spread adoption of machine learning. Author further lay out the most promising but also challenging approaches in combining machine learning with physical modelling.

Big data challenges

Big data challenges in the geoscientific context.

首先文章介绍了地球系统的资料属于标准的 “Four V’s” of “Big Data”:Volume(深度), velocity(更新速度), Variety(多样性), Veracity(真实性)。

紧接着文章提到了两个:“事实”:

  1. 我们收集、创造数据的能力远远的超过我们能够明智的选择数据的能力,更别说理解它们。
  2. 在过去几十年中,预测能力并没有随着数据的可用性的增加而增加。

Four examples of typical deep learning applications

Four examples of typical deep learning applications.

作者介绍了深度学习的四个引用例子:

  • Object recognition 对应极端事件识别
  • Super-resolution applications 对应分辨率处理
  • Video prediction 对应短临预测
  • Language translation 对应时间序列预测

我是个初学者,对于上面作者提到的四个例子,我个人觉得其实主要就是一个是CNN和Sequence Model的应用。 比较容易的是可以采用YOLO

Integration with physical modelling

在模式方面,一般来说物理模式和机器学习被认为是两种不同领域;一个是根据理论,一个是根据数据。 但是这两个方式其实应该是互补的;因此作者在这里阐述了如何将数据科学应用到理论科学中。

Linkages between physic al models and machine learning.

Interpretation of hybrid modelling



Like Tweet +1