The model is the message: Lightweight convolutional autoencoders applied to noisy imaging data for planetary science and astrobiology

TL;DR

This paper demonstrates that lightweight convolutional autoencoders can effectively reconstruct highly noisy planetary images, offering a useful tool for data reduction and analysis in challenging observational conditions.

Contribution

It introduces a simple convolutional autoencoder model tailored for noisy planetary imaging data, highlighting its effectiveness in reconstructing images with extensive random noise.

Findings

Multi-color image reconstruction is effective with 90% areal noise coverage.

Autoencoders can recover images obscured by low illumination or sensor degradation.

Latent space representations may be more useful than raw data for scientific analysis.

Abstract

The application of convolutional autoencoder deep learning to imaging data for planetary science and astrobiological use is briefly reviewed and explored with a focus on the need to understand algorithmic rationale, process, and results when machine learning is utilized. Successful autoencoders train to build a model that captures the features of data in a dimensionally reduced form (the latent representation) that can then be used to recreate the original input. One application is the reconstruction of incomplete or noisy data. Here a baseline, lightweight convolutional autoencoder is used to examine the utility for planetary image reconstruction or inpainting in situations where there is destructive random noise (i.e., either luminance noise with zero returned data in some image pixels, or color noise with random additive levels across pixel channels). It is shown that, in certain use cases, multi-color image reconstruction can be usefully applied even with extensive random destructive noise with 90% areal coverage and higher. This capability is discussed in the context of intentional masking to reduce data bandwidth, or situations with low-illumination levels and other factors that obscure image data (e.g., sensor degradation or atmospheric conditions). It is further suggested that for some scientific use cases the model latent space and representations have more utility than large raw imaging datasets.