Neural Networks for Modelling Wind Vectors

Abstract

The ERS-1 satellite was launched in 1991. It carries a scatterometer with three antennae which measure the reflected radar power from the surface of the Earth. This backscatter is due to the reflection of the micro-wave radar beam from small ripples on the surface of the ocean which are generated by instantaneous winds. The resulting measurement triplet can be used to infer wind vectors. The implementation of a forward model which maps wind vectors to radar backscatter is addressed here, applying techniques from the field of neural networks. An empirical approach is adopted here. The neural networks are trained with wind data from the European Centre for Medium-Range Weather Forecasting in which high wind speeds occur. The poor quality of the models obtained demonstrates that the noise in this input data cannot be neglected. A Bayesian framework is then adopted to account for this noise. Compared to existing reference models, the fit of the model in target space is improved, especially at high wind speeds which are of greatest interest for meteorological studies. Although the inversion of the model is not implemented, its potential accuracy is higher than existing models.