Application of Artificial Neural Networks for Sea-Surface Wind-Speed Retrieval From IRS-P4 (MSMR) Brightness Temperature

Indian Remote Sensing Satellite Multifrequency Scanning Microwave Radiometer (MSMR)-measured brightness temperatures (T_B) in 6.6-, 10.65-, 18-, and 21-GHz channels with dual polarizations were utilized to retrieve sea-surface wind speed (SSWS). A concurrent and collocated database was constructed on MSMR T_B- and deep-sea (DS)-buoy-recorded wind speeds for the period of June 1999-July 2001 over the north Indian Ocean. A radial-basis-function-based artificial-neural-network (ANN) algorithm was developed to estimate SSWS from MSMR T_B values. Multiple ANNs were generated by the systematic variation of the architecture of input- and hidden-layer nodes. The performance of the most successful algorithm was evaluated based on statistical summary and network performance. The accuracy of the ANN-based wind-speed algorithm was compared with DS-buoy observations, and the result was then compared with the output of the regression analysis between buoy- and MSMR operational-global-retrieval-algorithm (OGRA)-derived SSWS values. On the average, 84% (92%) of ANN-estimated MSMR SSWS observations are within ±2 m/s ( ±3 m/s) when compared with DS-buoy observations. The correlation and root mean square error of 0.80 and 1.79 m/s, respectively, for ANN-predicted SSWS values are much better than that obtained from OGRA. The performance of the ANN algorithm was also evaluated during a super cyclone (October 1999) over the Bay of Bengal. The ANN algorithm could capture the high cyclonic winds, and the values match reasonably well with Special Sensor Microwave/Imager and SeaWinds Scatterometer (QuikSCAT) operational wind products.