A Neural Network Technique for Retrieving Atmospheric Species from Microwave Limb Sounders
Atmospheric radiometry requires solving an inversion problem due to the indirect nature of the measurements. Traditional iterative methods for non-linear inversions are computationally intensive and faster algorithms are desirable. This work contributes by developing a technique for fast inversion of microwave limb sounding observations of the atmosphere. The technique uses multilayer perceptrons, a type of neural network, to perform a non-linear regression between the spectra measured by the sensor and the atmospheric species to be retrieved. There is theoretical grounding to propose this inversion method, but its implementation is far from trivial and the practical realization can have a large impact on the solutions obtained.
The retrieval algorithm is implemented by training a set of multilayer perceptrons with a set of species realizations and the corresponding spectra. The objective is to provide a model of the conditional distribution of the species state for each given measurement. The set is generated by first randomly disturbing a set of mean species states and then running a forward model on the generated species realizations. This is followed by an eigenvector expansion technique that reduces the dimensionality of the spectral space. Then the topology of the multilayer perceptrons is decided, followed by a training that finds the most probable weights for the given training set.
The retrieval algorithm has been practically tested by inverting simulated and measured spectra from the sub-millimetre limb sounder Odin-SMR. First retrievals from simulated data showed that the new technique compared well with an iterative approach based on optimal estimation, but doing much faster inversions. First tests on real data confirmed the previous results, but also showed the sensitivity of the technique to spectral artefacts. The operational iterative inversions are much slower but seem to have an advantage here. Nevertheless, the proposed algorithm is a promising technique for fast inversion of well characterised limb sounder observations.