Machine learning spectral clustering techniques: Application to Jovian clouds from Juno/JIRAM and JWST/NIRSpec

We present a new method, based on a joint application of a principal component analysis (PCA) and Gaussian mixture models (GMM), to automatically find similar groups of spectra in a collection. We applied the method (condensed in the public code chopper.py) to archival Jupiter spectral data in the 2–5 µm range collected by NASA Juno/JIRAM in its first perijove passage (August 2016) and to mosaics of the great red spot (GRS) acquired by JWST/NIRSpec (July 2022). Using JIRAM data analyzed in previous work, we show that using a PCA+GMM clustering can increase the efficiency of the retrieval stage without any loss of accuracy in terms of the retrieved parameters. We show that a PCA+GMM approach is able to automatically identify spectra of known regions of interest (e.g., belts, zones, GRS) belonging to different clusters. The application of the method to the NIRSpec data leads to detection of substructures inside the GRS, which appears to be composed of an outer halo characterized by low reflectivity and an inner brighter main oval. By applying these techniques to JIRAM data, we were able to identify the same substructure. We remark that these new structures have not been seen before at visible wavelengths. In both cases, the spectra belonging to the inner oval have solar and thermal signals comparable to those belonging to the halo, but they present broadened 2.73 µm solar-reflected peaks. Performing forward simulations with the NEMESIS radiative transfer suite, we propose that the broadening may be caused by differences in the vertical extension of the main cloud layer. This finding is consistent with recent 3D fluid dynamics simulations.