Publication

Abstract :

The increasing number of objects in Earth’s orbit is raising serious safety concerns for space assets and missions. This paper presents development, analysis and selection of a suitable deep learning framework among U-Net and convolutional autoencoder, for the identification of space debris and other irregular structures in astronomical imagery. A publicly available multi-class visible spectrum astronomical dataset is preprocessed using a two-stage enhancement pipeline: CLAHE to amplify faint features, and bilateral filtering to dwindle noise while retaining sharp edges. The convolutional autoencoder, trained exclusively on synthetic visible spectrum images, learns latent representations of normal space observations. Anomaly scores derived from reconstruction errors are used to detect deviations from these learned patterns. Cross-domain generalization is validated using real James Webb Space Telescope infrared images, preprocessed via asinh scaling, and false-colour mapping to unify input representation. The spectrum-agnostic deep learning approach for anomaly detection is analyzed for different images under visible and infrared domains for its satisfactory performance.