Year: 2020
Type: article
Abstract: The Earth system is considered to possess certain modes - preferred patterns of variability that can represent the latent structure of the climate system, also known as teleconnections. There are approaches to discover these patterns, Principal Components Analysis and Empirical Orthogonal Teleconnection (EOT) analysis. However, while the latter is very effective, it is computationally intensive. An autoencoder is an unsupervised neural network that learns an efficient neural representation of input data. It is considered as a dimensionality reduction tool that is highly similar to PCA and EOT. The hidden layer of an autoencoder represents the most significant information of the input, which can extract crucial latent structures. When applying Earth Observation data in a linear autoencoder, the information presented in the nodes is teleconnections. However, with an increased number of nodes in the hidden layer, teleconnections spread across nodes and the patterns become mixed. This research presents a sequential autoencoder (SA) with only one single node in the hidden layer based on subsequent residual series analysis. It has a major finding that the hidden nodes of SA are identical to the corresponding PCA components.
Author: Jiena He
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W3049064027