Remote sensing super-resolution stops chasing scale factors: SpatialSR rebuilds at absolute resolution, reaching 66.7 mIoU on Vaihingen
Synopsis
Arbitrary-scale super-resolution for remote sensing images has mostly been conditioned on the relative upscaling factor between low- and high-resolution images, leaving out absolute spatial resolution in meters per pixel, a physical prior under which the same factor can correspond to very different texture distributions. SpatialSR instead conditions reconstruction on a target spatial resolution: it estimates the low-resolution image's absolute resolution to derive the required factor, draws resolution-specific texture style from a pretrained resolution-to-image diffusion model, and folds the diffusion process into a single-step training constraint. On Vaihingen at a 0.1 m/pixel target from 0.5 m/pixel sources, its segmentation mIoU is 66.7, above the listed comparison methods.
Interpretation
The target of remote sensing super-resolution shifts from a relative upscaling factor to an absolute spatial resolution, so reconstructed images are required to reach a specified meters-per-pixel value. Earlier arbitrary-scale remote sensing super-resolution methods conditioned on the relative scale factor between low- and high-resolution images and did not treat absolute spatial resolution itself as a reconstruction prior. Texture distributions differ at the same upscaling factor: 4 m/pixel to 1 m/pixel and 0.4 m/pixel to 0.1 m/pixel are both four-times super-resolution, yet their high-resolution texture distributions differ substantially, as do their high-frequency energy distributions in the frequency domain.
The pipeline first estimates the low-resolution image's own spatial resolution and then determines how much to upscale. The spatial resolution estimation module SREM takes candidate patches with a sliding window, selects subsets by random, Laplacian and Haar-wavelet sampling, concatenates them along channels, and predicts resolution with a network built from ResNet blocks and VisionMamba blocks. On the MSRUAV validation set, within-domain UAV testing gives an estimation mean absolute error of 0.0312 with a coefficient of determination of 0.94; transferring to satellite UCMD imagery gives a mean absolute error of 0.0487 and a coefficient of determination of 0.88.
The diffusion model only supplies resolution-related texture style and enters training as a single-step constraint, with no iterative sampling at inference. A pretrained Stable Diffusion v1.5 is fine-tuned on a resolution-to-image task, and its intermediate denoising features are precomputed into a lookup table keyed by resolution, with intermediate resolutions obtained by linear interpolation between the two nearest keys. The trajectory consistency loss aligns reconstructed images to the posterior mean of the DDPM denoising trajectory; for four-times super-resolution on 128×128 low-resolution inputs, single-step inference takes 0.0074 s with 1.76 M parameters and 51.96 G FLOPs.
Perspective
The demonstrated scope is arbitrary-scale super-resolution of remote sensing images and its downstream semantic segmentation and scene classification: synthetic settings use resampled images as references, real-world settings process source images directly, and downstream models are applied zero-shot with official weights and no additional fine-tuning.
The method struggles with target spatial resolutions beyond the predefined range, and resolution estimation accuracy degrades in cross-sensor scenarios because of distribution shift; the resolution style lookup table is discrete, so intermediate resolutions are approximated by linear interpolation.