The use of principal component analysis for the assimilation of high‐resolution infrared sounder observations for numerical weather prediction

Abstract

Methodologies are discussed for the efficient representation of observations from high-resolution infrared sounders for the purposes of assimilation into numerical weather prediction models. The use of principal component analysis is explored and it is noted that, while the available information in the observations is stored efficiently, the non-locality of the Jacobians that arise may cause practical problems in an operational assimilation system. Reconstructing radiance spectra from the principal components appears to be a more realistic approach in the near term.

Statistical retrieval of atmospheric profiles with deep convolutional neural networks

“…Alternatively, one could use a different feature extraction method, for instance Independent Component Analysis (ICA) [32] where the uncorrelated and statistically independent variates maximize a measure of non-Gaussianity such as negentropy in all original variables. Taking into account the noise estimate when designing PCA for infrared sounders has been shown to be important [10,11]. In [33] we applied a minimum noise fraction (MNF) transformation that simultaneously minimizes the noise fraction or equivalently maximize the signal-to-noise ratio (given a noise model) in all original variables and takes into account the information contained in the spatial neighbors.…”

“…Although this result may appear counterintuitive since compression implies reduction on the amount of information in the images, a certain level of compression is actually beneficial because: 1) compression removes information but also noise, and it could be useful to remove the components with low signal-to-noise ratio (SNR); and 2) spatial compression introduces information about the neighbouring pixels in an indirect yet simple way. Including the noise estimate in the design of PCA of infrared sounders has been considered before [10], and actually it is currently implemented in the IASI pipeline [11]. However the inclusion of the spatial information, while it is important [12], has obtained less attention.…”

Inversion for atmospheric thermodynamical parameters of IASI data in the principal components space

“…While it is true that we cannot build up a bijection betweenỹ F andỹ, it is true that this kind of mapping could exist between a subset of the elements ofỹ F and the vector c τ , because M τ < M. As suggested by Collard et al (2010), since the M τ PC scores are linearly independent, there could exist a subset of filtered radiances, which is itself linearly independent and, therefore, the covariance matrix S u could have an inverse. Although there may exist a subset of filtered radiances with M ρ = M τ , which admits an inverse for S u , in general this is not expected for any choice of the filtered radiances.…”

“…We also address the problem of using filtered radiances within a given inverse scheme and how and when this usage becomes equivalent to the direct inversion within the PC space. The use of filtered radiances for the assimilation of high-resolution infrared sounder observations has been recently analyzed and described in Collard et al (2010). The emphasis of our paper is rather on the direct assimilation of PC scores and therefore it contains complementary information.…”

Implementation of reconstructed radiance of IASI observations in WRF data assimilation system

“…Reconstructed radiance [5] from the raw spectrum is proposed to represent the IASI observations more efficiently. The method has received a great deal of attention as a mean for assimilation of high-resolution infrared data.…”