Preliminary investigation of machine learning approaches to improve projections of future climate in Hawai‘i

Detailed, reliable projections of future changes in climate are needed by Hawai‘i’s resource managers, such as water utilities managers, land managers, conservation organizations, and decision makers. However, global climate models (or “general circulation models”), which are currently the most commonly used tool for projecting future climate variations, are known for representing large-scale climate patterns and are not ideal for simulating climate processes at small spatial scales, such as those relevant to Hawaiʻi’s climate. Traditionally, the technique of “downscaling” is used to bridge this gap between global climate models and local climate conditions. Due to the lack of downscaled datasets for Hawaiʻi, however, there are conflicting projections for future climate in the region, leading scientists to turn to a relatively new method for these projections, called “machine learning” (i.e. a tool for data analysis that automates analytical modeling).

Open forest view of taller trees and lower ferns
The continuation of lush native forests like this one in Kalopa depend upon what precipitation patterns the future may bring. (Photo: Alan Cressler, USGS)

Machine learning is a promising tool that remains relatively unexplored within the climate adaptation community’s development of climate projections. Machine learning approaches offer a range of alternatives, never previously tested in Hawai‘i, for developing statistical models at scales applicable for land managers to make decisions. Hawaiʻi’s complex weather and climate are heavily affected by interactions between wind and the rough terrain of Hawai‘i’s mountainous islands, resulting in extreme differences from place to place in rainfall, cloudiness, temperature, wind speed, and other factors. These complex situations present in Hawaiʻi provide an ideal opportunity to test the utility of machine learning approaches in teasing out patterns otherwise obscured by traditional global climate modeling efforts.

In this project, researchers will test innovative new machine learning techniques for use in developing useful future rainfall predictions in Hawai‘i. The results of this project will be an assessment of the viability of these new approaches to develop at-scale climate projections for Hawai‘i, thus guiding future downscaling efforts. Future work based on this analysis will add significantly to the available projections, providing resource managers and decision makers with more information with which to develop scenarios of changing water resource availability.





Tom Giambelluca
Professor of Geography, UH Mānoa


Yusuke Hatanaka
Department of Computer and Data Science, UH Mānoa


Victoria Keener
Research Fellow, PacRISA, East-West Center, UH Mānoa