Technical Summary: Mind the Gap: Quantifying the Domain Gap in Cross-Sensor Diffusion Super-Resolution

Problem Statement

Remote sensing applications, such as wildfire monitoring and resource management, face a critical spatial resolution dilemma. Freely available missions like Sentinel-2 offer high temporal frequency but moderate resolution (10 m/pixel), whereas commercial systems like PlanetScope provide higher resolution (3 m/pixel) at a cost. Super-Resolution (SR) aims to bridge this gap by reconstructing high-resolution (HR) images from low-resolution (LR) observations.

However, a fundamental challenge persists: the absence of true paired LR–HR observations from different sensors. Consequently, most SR models are trained on synthetically degraded data (e.g., bicubic downsampling of HR images). This creates a significant domain gap when these models are applied to real cross-sensor imagery. The synthetic training data fails to capture the complex spectral, radiometric, and geometric properties of real sensors, leading to unstable performance and potential degradation of image quality when deployed on actual satellite data. While diffusion-based models have achieved state-of-the-art results in natural image SR, their generalization capabilities in this specific cross-sensor remote sensing context remain largely unquantified.

Methodology

Dataset Construction

The authors constructed a large-scale, geometrically and temporally aligned dataset comprising 175 regions from the USA, pairing Sentinel-2 (10 m) and PlanetScope (3 m) imagery.

• Alignment: Pairs were acquired via PlanetHub with sub-pixel geometric alignment.
• Temporal Proximity: Pairs were filtered to minimize scene evolution, with the majority falling within a ±15-day window.
• Spectral Alignment: To ensure physical validity, six overlapping spectral bands (Coastal Blue, Blue, Green, Red, Red Edge, NIR) were selected from both sensors, avoiding the "hallucination" of spectral information.
• Experimental Configurations: The study utilized three distinct setups:
  1. Case I (Synthetic Baseline): Training and testing on synthetically degraded PlanetScope data (PlanetLR → PlanetHR).
  2. Case II (Domain Gap Quantification): Training on synthetic PlanetScope data but testing on real Sentinel-2 data (SentinelLR → PlanetHR).
  3. Case III (Direct Mapping): Training and testing directly on real Sentinel-2 to PlanetScope pairs (SentinelLR → PlanetHR).

Benchmarked Models

Five state-of-the-art diffusion-based SR architectures were evaluated under identical conditions using a unified U-Net backbone (30M parameters) to isolate the effects of the diffusion formulation:

1. DDPM: Standard denoising diffusion probabilistic model.
2. Flow Matching (FM): A deterministic generative paradigm learning a vector field.
3. Image-to-Image Schrödinger Bridge (I2SB): Frames translation as a stochastic bridge between distributions.
4. ResShift: Reparameterizes the Markov chain to shift residuals between HR and LR images.
5. UniDB: Generalizes diffusion bridges using Stochastic Optimal Control.

Evaluation Metrics

The study employed standard metrics (PSNR, SSIM, LPIPS) and introduced a novel metric:

• LPIPSSat: A domain-adapted perceptual metric based on a ResNet-50 backbone pretrained in a self-supervised manner on Sentinel-2 RGB data. Unlike standard LPIPS (trained on ImageNet), LPIPSSat is tailored to the spectral and textural characteristics of Earth observation imagery.

Downstream Task

To assess practical utility, the SR-enhanced images were evaluated on a burned-area change delineation task using a U-Net trained on the CaBuAr dataset. This tested whether SR models preserved task-relevant information or introduced artifacts that hindered segmentation.

Key Results

Quantitative Performance

• Synthetic Success, Real Failure (Case I vs. Case II): Models trained on synthetic data (Case I) achieved high scores (PSNR > 45 dB). However, when applied to real cross-sensor data (Case II), performance collapsed catastrophically. PSNR dropped to ~16.7 dB, and perceptual scores indicated that models were actively degrading the input by imposing synthetic priors (e.g., bicubic-like sharpness) that did not match the physical reality of Sentinel-2 data.
• Limited Adaptation (Case III): Training directly on real cross-sensor data (Case III) allowed for partial recovery of performance. Perceptual metrics improved significantly over Case II, but results remained far below the synthetic baseline (Case I). High variance in the real-world data (due to illumination and sensor differences) hindered consistent optimization across architectures.
• Metric Sensitivity: LPIPSSat proved more sensitive to spectral and structural discrepancies than standard LPIPS, better reflecting the failures in real-world scenarios.

Downstream Task Impact

• Synthetic Models (Case II): When used as a preprocessing step for the burned-area task, synthetic-trained models provided modest but consistent improvements over the baseline, likely because they enhanced visual features without introducing complex domain-specific artifacts.
• Real-World Models (Case III): Models trained on real data introduced "destructive artifacts" and hallucinations that degraded the downstream segmentation performance. The dual objective of learning both super-resolution and domain adaptation simultaneously proved unstable, leading to outputs that were less useful for the specific task than the original low-resolution inputs.

Key Contributions

1. Systematic Quantification: The first systematic study quantifying the synthetic-to-real generalization gap in diffusion-based SR for remote sensing, demonstrating that synthetic training does not prepare models for real cross-sensor challenges.
2. LPIPSSat Metric: Introduction of a novel, domain-adapted perceptual metric based on Sentinel-2 self-supervised features, offering a more faithful evaluation of Earth observation imagery than generic metrics.
3. Large-Scale Dataset: Construction of a rigorously aligned, large-scale paired dataset of Sentinel-2 and PlanetScope imagery, enabling rigorous cross-sensor SR evaluation.

Significance and Claims

The paper argues that the current reliance on synthetic training for satellite SR is a "key limitation" of the field. While synthetic training offers a stable baseline, it fails completely when transferred to real cross-sensor scenarios. Conversely, training directly on real data introduces optimization difficulties and entangled objectives (SR + domain adaptation) that result in unstable, artifact-ridden outputs.

The authors conclude that the field faces a dilemma: synthetic training is stable but limited in utility, while real-world training is necessary for adaptation but currently too unstable for practical deployment. They posit that future research must focus on disentangling super-resolution from domain adaptation, suggesting that current architectures cannot fully adapt to the physical and radiometric diversity of real-world sensors without new methodological approaches. The work serves as a foundation for developing robust cross-sensor generalization methods rather than simply improving model architectures.