Satellite altimeter wind speed estimation and tropical cyclone characterization using machine learning

Abstract

Sea surface wind speed (Uio) is one of the vital variables for tropical cyclone analysis in providing accurate wind intensity information to the warning center. However, rough sea state condition, has caused the U10 observations by buoy to become unreliable. Although satellite altimeter can measure U10, the operational Gourrion algorithm was designed for normal sea state conditions. Extreme oceanatmospheric interaction worsen by the rain contamination on the altimeter signal has impaired the quality of the derived U10, hence putting low attention in tropical cyclone study. This operational U10 product which only incorporates the backscatter and the surface wave height at Ku-band as principal parameters is insufficient to emulate a complex cyclone environment. Though higher U10 regime saturated at 20 ms-1 and heavy rainy conditions have reduced the U10 accuracy, other ocean-related parameters are worth considering. Therefore, this study was aimed to analyse the altimeter ocean-related parameters and thus estimate high accuracy U10 for tropical cyclone wind characterization. This study established a relationship between parameters response from Joint Altimetry Satellite Oceanography Network (Jason)-2 and Jason-3; and the coincident U10 from Meteorological Operational (MetOp)-A and MetOp-B scatterometers in 350 tropical cyclones captured between 2015 and 2018 globally. Quantitative assessment on the quality of altimeter C-band parameters and other simultaneously observed radiometric ocean parameters namely brightness temperatures at 18.7, 23.8, and 34.0 GHz, water vapor content, and liquid water content related to extreme U10 were presented. Correlation of C-band parameters to U10 outperformed that of the Ku-band counterpart by at least 29% and the inclusion of radiometric parameters contributed to a significant error reduction of about 48%. New and high accuracy U10 models were developed using Multiple Linear Regression and machine learning techniques namely Artificial Neural Network, Support Vector Machine, and Gaussian Process Regression. The Gaussian Process Regression with all parameters considered was proved to be the best model that could estimate U10 up to 35 ± 1 ms-1 with the improvement of 35% and 75% inside the rain and at the higher U10 regime respectively. The study clearly presented the tropical cyclone wind characters that could now be objectively estimated. The uncertainty of the derived maximum sustained wind speed intensity could be reduced to 70% compared to that of operational U10. The storm center location, eye width, radius of inner and outer circle relatively at 50- knot and 30-knot respectively were distinguishable and well agreed to the reported tropical cyclone best-track. This study successfully established the fundamental analysis on the performance of altimetry and radiometry parameters acquired by Jason mission and integrate them to represent the tropical cyclone environment. The finescale altimeter along-track resolution of extracted tropical cyclone wind characters is exclusively demonstrated and has become a vital complement to the optical satellite image observation.