A neural network as a nonlinear transfer function model for retrieving surface wind speeds from the special sensor microwave imager

TLDR

The study discusses the limitations of neural network algorithms in the context of surface wind speed retrieval from SSM/I data. The authors aim to develop a single extended-range neural network to model the SSM/I transfer function for surface wind speed retrievals and to introduce a cloud liquid water–based moisture criterion to improve performance. The method employs a single extended-range neural network trained on SSM/I data, incorporating a cloud liquid water–based moisture retrieval criterion and optionally 85 GHz brightness temperatures. When applied to standard SSM/I datasets, the SER NN achieved a bias of 0.05 m/s and an rms of 1.65 m/s, matching the accuracy of separate neural networks for clear and cloudy conditions; using the cloud liquid water criterion reduced bias to 0.03 m/s, rms to 1.58 m/s, and rejected only 2 % of data while recovering 40 % of previously discarded cases, with an additional ~10 % accuracy gain for cloudy scenes when 85 GHz brightness temperatures were included.

Abstract

A single, extended‐range neural network (SER NN) has been developed to model the transfer function for special sensor microwave imager (SSM/I) surface wind speed retrievals. Applied to data sets used in previous SSM/I wind speed retrieval studies, this algorithm yields a bias of 0.05 m/s and an rms difference of 1.65 m/s, compared to buoy observations. The accuracy of the SER NN for clear (low moisture) and cloudy (higher moisture/light rain) conditions equals the accuracy of NNs trained separately for each of these cases. A new moisture retrieval criterion based on a single, physically interpretable parameter, cloud liquid water, is proposed in conjunction with the SER NN. Using this retrieval criterion, (1) a moisture retrieval threshold for cloud liquid water of 0.5 kg/m 2 was estimated, and (2) 40% of the data rejected by previous rain flags could be recovered. When the SER NN was trained using this retrieval criterion, a bias of 0.03 m/s and an rms value of 1.58 m/s were obtained and only 2% of the data were rejected. Also, a slight improvement in retrieval accuracy for cloudy conditions was achieved (∼10%) by including SSM/I brightness temperatures at 85 GHz. Finally, the limitations of NN algorithms are discussed in light of the present application.