Wind Speed Estimation From CYGNSS Using Artificial Neural Networks

In this article, a retrieval algorithm based on the use of an artificial neural network (ANN) is proposed for wind speed estimations from cyclone global navigation satellite system (CYGNSS). The delay/Doppler map average and the leading edge slope observables, derived from CYGNSS delay/Doppler maps, are used as inputs to the network, along with geographical, geometry, and hardware antenna information. The derivation of the optimal number of hidden layers and neurons is obtained using statistical metrics of agreement between the CYGNSS data and the wind matchups obtained from modelled winds output by the wavewatch 3 (WW3) model. A cumulative distribution function (CDF) matching step is applied to the network outputs, to impose that the CDF of the retrievals matches that of the matchups. The resulting wind speeds are unbiased with respect to WW3 modeled winds, and deliver a global root mean square (RMS) difference (RMSD) of 1.51 m/s, over a dynamic range of wind speeds up to 32 m/s. The obtained RMSD is the lowest among those seen in literature for wind speed retrievals from CYGNSS. A comparison is carried out between the winds retrieved from the ANN approach and those derived using the fully developed sea approach, which represent the CYGNSS baseline wind product. The comparison highlights that the ANN approach outperforms the baseline approach for both low and high wind speeds and removes most of the geographical biases between baseline winds and WW3 winds seen in monthly maps of wind speeds. The ANN approach could well be applied to the entire CYGNSS dataset to generate an enhanced wind speed product