Professor Im Jeong-ho's Team Predicts Thaw Concentration Using UNET Deep Learning
Average Prediction Error Below 6%... Forecast Up to 1 Year
An artificial intelligence (AI) model capable of predicting changes in Arctic sea ice up to one year in advance has been developed.
Providing mid- to long-term forecast information, it is expected to aid in the development of Arctic shipping routes and marine resource exploration.
A research team led by Professor Lim Jeong-ho of the Department of Earth Environmental and Urban Engineering at UNIST (President Park Jong-rae) developed an AI model that can predict Arctic sea ice concentration one year ahead with an error accuracy within 6%. Sea ice concentration refers to the proportion of an area covered by ice within a unit area.
The research team developed this AI model by using UNET to learn the complex relationships between past Arctic sea ice concentration change patterns and key climate factors such as air temperature, sea temperature, solar radiation, and wind. UNET is a deep learning algorithm that enables AI to learn relationships between image data such as satellite images.
The developed model showed high accuracy in mid- to long-term forecasts. When evaluating accuracy by comparing the AI model’s predicted values with past actual sea ice concentration values, the average prediction error was below 6% for 3-month, 6-month, and 12-month forecasts. Existing models showed increasing average prediction errors as the forecast period lengthened.
Moreover, this model demonstrated stable predictive performance even in situations where sea ice rapidly decreased unusually. In cases like the summers of 2007 and 2012, when sea ice melted sharply, existing models recorded an average prediction error of 17.35%, whereas the developed AI model recorded an average prediction error of 7.07%, reducing the average prediction error by more than half.
The research team also identified climate factors that play important roles in mid- to long-term sea ice concentration forecasts. Analysis of differences in UNET model prediction results revealed that solar radiation and wind were major variables at the edges of sea ice where the ice thickness is thin.
Professor Lim Jeong-ho said, “This study overcomes the limitations of existing physics-based models and clarifies the complex effects of various environmental factors on Arctic sea ice changes,” adding, “It will help in developing Arctic shipping routes, exploring marine resources, and establishing climate change response policies.”
This research was published online on December 11 in the international journal Remote Sensing of Environment and was conducted with support from the Korea Polar Research Institute and the Ministry of Oceans and Fisheries.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
