Type |
Article |
Date |
2015-10 |
Language |
English |
Author(s) |
He Guelton Liyun1, Fablet Ronan2, Chapron Bertrand1, Tournadre Jean1 |
Affiliation(s) |
1 : IFREMER, Lab Oceanog Space, F-29280 Plouzane, France. 2 : Telecom Bretagne, SC, F-29280 Brest, France. |
Source |
Ieee Journal Of Selected Topics In Applied Earth Observations And Remote Sensing (1939-1404) (Ieee-inst Electrical Electronics Engineers Inc), 2015-10 , Vol. 8 , N. 10 , P. 4742-4750 |
DOI |
10.1109/JSTARS.2015.2496503 |
WOS© Times Cited |
7 |
Keyword(s) |
Coastal wind, downscaling, high resolution (HR), machine learning, support vector regression (SVR) |
Abstract |
The availability of sea surface wind conditions with a high-resolution (HR) space-time sampling is a critical issue for a wide range of applications. Currently, no observation systems nor model forecasts provide relevant information with a high sampling rate in both space and time. Synthetic aperture radar (SAR) satellite systems deliver HR sea surface fields, with a spatial resolution below 0.01., but they are also characterized by a large revisit time up 7 to 10 days for temperate zones. Meanwhile, operational model predictions typically involve a high temporal resolution (e.g., every 6 h), but also a low spatial resolution (0.5 degrees). With a view to leve-raging both data sources, we investigate statistical downscaling schemes. In this study, a new model based on a machine learning method, namely support vector regression (SVR), is built to reconstruct HR sea surface wind fields from low-resolution operational model forecasts. The considered case study off Norway demonstrates the relevance of the proposed SVR model. It outperforms state-of-the-art approaches [namely, linear, analog, and empirical orthogonal function (EOF) downscaling models] in terms of mean square error. It also realistically reproduces complex space-time variabilities of the observed SAR wind fields. We further discuss the SVR model as a generalization of the popular linear and analog models. |
Full Text |
File |
Pages |
Size |
Access |
|
9 |
10 MB |
Access on demand |
Author's final draft |
10 |
10 MB |
Open access |
|