Dimension reduction of satellite observations for remote sensing. Part 2: Illustration using hyperspectral microwave observations

In the first article of this series, the two classical strategies to reduce satellite data dimension (i.e. compression and channel selection) were presented, together with the introduction of a new method, the so‐called ‘bottleneck channels’ (BC). BC are a compromise between the two classical approaches and can benefit from the advantages of both. In this article, the three methodologies are tested using experiments on a synthetic dataset corresponding to a hyperspectral conceptual instrument in the microwave, for frequencies up to 500 GHz. As expected, principal component analysis (PCA) based methods are best to compress data, but their components lack the physical interpretability of real channels. Channel selection methods preserve this physical meaning but require a much larger number of channels in order to use the redundancy of information to reduce instrumental noise. The new BC method appears to be a good compromise. It can be seen as a PCA compression method where the components are constrained to be instrument channels, facilitating their understanding, inversion or assimilation. BC allows for an easy calibration of data based on radiative transfer simulations and also alleviates the mixing problem of the PCA technique, where various physical variabilities (e.g. temperature, humidity, clouds) can be mixed in the same extracted components. Furthermore, the BC compression rate is equivalent to that of PCA‐based methods even with a limited number of BC.