INVESTIGATING THE LIMITED SAMPLE SIZE OF SUBSEASONAL CLIMATE FORECASTING DATASETS IN DEEP LEARNING APPLICATIONS

CHMIELOWIEC, Philip

Permalink

https://hdl.handle.net/2142/124855 Copy

Description

Title

INVESTIGATING THE LIMITED SAMPLE SIZE OF SUBSEASONAL CLIMATE FORECASTING DATASETS IN DEEP LEARNING APPLICATIONS

Author(s)

CHMIELOWIEC, Philip

Issue Date

2023-05-01

Keyword(s)

Deep Learning, Subseasonal Climate Forecasting, Climate Models, Transfer Learning

Date of Ingest

2024-10-14T11:14:54-05:00

Abstract

Subseasonal climate forecasting is a new and rapidly evolving field that aims to bridge the gap between weather forecasting and long-term climate projections. Machine learning models used in this field are typically trained on historical observations, which have a small sample size due to the short recent observational period. To address this challenge, researchers have explored the use of simulated climate model output to supplement training data. By applying transfer learning, models can utilize both datasets with the goal of obtaining more skillful predictions. In preparation for transfer learning, this project conditions two variants of an Autoencoder model trained on observational climate variables for predicting temperature anomalies over the continental United States. Both models are constrained to only use climate variables that are present in the CESM2 Subseasonal to Seasonal (S2S) Hindcast dataset. The skillfulness of these models is evaluated using the Cosine Similarity across both spatial and temporal dimensions. Results show that the predictive skill of both models varies across different regions in the United States, depending on both the architecture of the models and the availability of covariates. Additionally, predictive skill varies depending on the season that we are predicting. These results demonstrate the importance of carefully selecting the architecture and covariates when setting up machine learning models in preparation for transfer learning.

Type of Resource

text

Owning Collections