A New Method for Deriving Ocean Surface Specific Humidity and Air Temperature: An Artificial Neural Network Approach

Abstract

A new methodology for deriving monthly averages of surface specific humidity (Qa) and air temperature (Ta) is described. Two main aspects characterize the new approach. First, remotely sensed parameters, total precipitable water (W), and sea surface temperature (SST) are used to derive Qa and Ta. Second, artificial neural networks (ANN) are employed to find transfer functions relating the input (W, SST) and output (Qa and Ta) parameters. Input data consist of nearly six years (January 1988?November 1993) of monthly averages of total precipitable water from Special Sensor Microwave/Imager data and sea surface temperature analysis from the National Centers for Environmental Prediction. Surface marine observations of Qa and Ta are used to develop and evaluate the new methodology. The performance of the algorithm is measured with surface marine observations not used in the development phase. Higher seasonally dependent discrepancies between Qa and Ta derived from the new method and in situ data are observed in regions such as the Kuroshio and Gulf Stream currents. After removal of systematic biases, the new method indicates that the combination of W and SST as input parameters and the ANN algorithm provides an interesting alternative for deriving monthly averaged surface parameters. The global mean bias in Qa is 0.010 ± 0.23 g kg?1 over most oceanic areas, whereas root-mean-square (rms) differences are 0.77 ± 0.39 g kg?1. Likewise, the global mean bias and rms in Ta are on the order of ?7.3 ? 10?5 ± 0.27°C and 0.72 ± 0.38°C, respectively.