Identification of H2O ice at lunar south polar region based on deep learning

The presence of frozen volatiles (especially H 2 O ice) has been proposed in the permanently shadowed regions (PSRs) near the poles of the Moon, based on various remote measurements including the visible and near-infrared (VNIR) spectroscopy. Compared with the middle- and low-latitude areas, the VNIR spectral signals in the PSRs are noisy due to poor solar illumination. Coupled with the lunar regolith coverage and mixing effects, the available VNIR spectral characteristics for the identification of H 2 O ice in the PSRs are limited. Deep learning models, as emerging techniques in lunar exploration, are able to learn spectral features and patterns, and discover complex spectral patterns and nonlinear relationships from large datasets, enabling them applicable on lunar hyperspectral remote sensing data and H 2 O-ice identification task. Here we present H 2 O ice identification results by a deep learning-based model named one-dimensional convolutional autoencoder. During the model application, there are intrinsic differences between the remote sensing spectra obtained by the orbital spectrometers and the laboratory spectra acquired by state-of-the-art instruments. To address the challenges of limited data and the difficulty of matching laboratory and remote sensing spectra, we introduce self-supervised learning method to achieve pixel-level identification and mapping of H 2 O ice in the lunar south polar region. Our model is applied to the level 2 reflectance data of Moon Mineralogy Mapper. The spectra of the identified H 2 O ice-bearing pixels were extracted to perform dual validation using spectral angle mapping and peak clustering methods, further confirming the identification of most pixels containing H 2 O ice. The spectral characteristics of H 2 O ice in the lunar south polar region related to the crystal structure, grain size, and mixing effect of H 2 O ice are also discussed. H 2 O ice in the lunar south polar region tends to exist in the form of smaller particles (~70 μm in size), while the weak/absent 2-μm absorption indicate the existence of unusually large particles. Crystalline ice is the main phase responsible for the identified spectra of ice-bearing surface however the possibility of amorphous H2O ice beneath optically sensed depth cannot be ruled out. • This work applied artificial intelligence for the first time to the exploration of lunar H2O ice. • A deep learning-based workflow is developed to realize pixel level identification of H2O ice. • H2O ice in the lunar south polar region tends to exist in the form of smaller particles. • Crystalline ice is the main phase responsible for the identified spectra of ice-bearing surface.