A machine learning based forward operator for visible and near-infrared satellite images

Abstract:

Visible and near-infrared satellite observations from geostationary satellite provide high-resolution information on clouds and aerosols. To use this information for data assimilation and for the evaluation of numerical weather prediction models, sufficiently fast and accurate forward operators are required, which compute synthetic satellite images from the model state. At these wavelengths, scattering is important, which makes radiative transfer methods complicated and computationally expensive. Only recently, a fast method for visible images has become available, which is based on a compressed eight-dimensional reflectance look-up table computed with a much slower standard radiative transfer method. While the usefulness of this operator has been demonstrated, it is restricted to clouds and non-absorbing channels. Taking aerosols and additional effects like absorption by water vapor into account with the same approach is not very practical, as it would require additional dimensions and thus a strongly increased table size. Here we report on using feed-forward neural networks as an alternative to the look-up table and demonstrate that it is possible to achieve a comparable speed and accuracy. The amount of training data and the memory required by the operator can be reduced by orders of magnitude, which makes additional input dimensions feasible. Moreover, the neural-network approach has an advantage for variational or hybrid data assimilation systems: Tangent-linear and adjoint versions of the neural network inference code can easily be derived for arbitrary network structures and do not have to be changed when the network is trained with different data.