Contourlet-Guided Latent Diffusion for Super-Resolution of Sentinel-2 Imagery

Abstract. High-spatial-resolution satellite imagery is important for en-
vironmentalmonitoring,forest-disturbanceassessment,andprecisionagri-
culture. However, widely available sensors such as Sentinel-2 provide
only moderate spatial resolution, limiting the representation of fine spa-
tial structures, textures, and directional landscape patterns. Recent ad-
vances in generative artificial intelligence, particularly diffusion models,
have demonstrated strong performance in image restoration and super-
resolution [1,7]. Nevertheless, preserving geometrically meaningful edges
anddirectionalstructuresremainschallenginginremote-sensingimagery.
Thisworkproposesasuper-resolutionframeworkthatintegratesthecon-
tourlet transform [4] with a latent diffusion model [5]. A spatial encoder
extracts local image features, while a parallel contourlet branch cap-
tures multi-scale directional information describing edges, contours, and
textures. The two feature representations are fused into a structural con-
ditioning signal that guides iterative denoising in latent space. A decoder
then reconstructs the super-resolved image.
The model is trained using the SEN2NAIP dataset [6], in which high-
resolution NAIP imagery is used to generate Sentinel-2-like inputs for
supervised learning. Preliminary evaluation shows that the proposed
method achieves 30.12 dB PSNR and 0.843 SSIM, outperforming bicu-
bic interpolation, SR3, and an unguided latent-diffusion baseline. Quan-
titative and visual comparisons indicate improved preservation of fine
boundaries, textures, and directional structures. Ongoing work focuses
on validation using real ground-truth data and geographically diverse
Sentinel-2 scenes.

Keywords: Remote sensing· Sentinel-2· super-resolution· latent dif-
fusion models· contourlet transform· generative artificial intelligence·
SEN2NAIP