Rethinking Pan-sharpening: A New Training Process for Full-Resolution Generalization

TL;DR

This work shifts pan-sharpening research from architecture-centric scaling to a practical, process-level paradigm: all-in-one training of a single, compact model on multiple datasets (WV2, WV3, GF2). The PanTiny framework embodies this paradigm with a lightweight single-encoder architecture, a simple yet effective fusion module, and a robust composite loss that substantially improves full-resolution generalization (QNR) across datasets. Empirically, all-in-one training yields universal FR gains for diverse models, while PanTiny achieves state-of-the-art performance with far lower parameter counts and FLOPs than large, specialized methods. The paper also demonstrates the importance of loss design and simple architectural choices for robustness, reproducibility, and deployability in real-world pan-sharpening deployments.

Abstract

The field of pan-sharpening has recently seen a trend towards increasingly large and complex models, often trained on single, specific satellite datasets. This one-dataset, one-model approach leads to high computational overhead and impractical deployment. More critically, it overlooks a core challenge: poor generalization from reduced-resolution (RR) training to real-world full-resolution (FR) data. In response to this issue, we challenge this paradigm. We introduce a multiple-in-one training strategy, where a single, compact model is trained simultaneously on three distinct satellite datasets (WV2, WV3, and GF2). Our experiments show the primary benefit of this unified strategy is a significant and universal boost in FR generalization (QNR) across all tested models, directly addressing this overlooked problem. This paradigm also inherently solves the one-model-per-dataset challenge, and we support it with a highly reproducible, dependency-free codebase for true usability. Finally, we propose PanTiny, a lightweight framework designed specifically for this new, robust paradigm. We demonstrate it achieves a superior performance-to-efficiency balance, proving that principled, simple and robust design is more effective than brute-force scaling in this practical setting. Our work advocates for a community-wide shift towards creating efficient, deployable, and truly generalizable models for pan-sharpening. The code is open-sourced at https://github.com/Zirconium233/PanTiny.

Figures (7)

• Figure 1: Our proposed PanTiny framework enables training a single, unified model on multiple datasets (WV2, WV3, GF2) simultaneously. This all-in-one approach achieves SOTA performance while maintaining a significantly smaller model size and lower computational cost compared to methods that require separate, specialized models for each dataset.
• Figure 2: The overall architecture of our proposed PanTiny framework. It consists of a single lightweight convolutional encoder for the MS input, a simple yet effective fusion module to integrate PAN information, a body of standard Transformer blocks for deep feature interaction, and a final convolutional layer for refinement.
• Figure 3: Quality comparison across SOTA methods on WV3 dataset. Refer to supplementary materials for more results.
• Figure 4: Performance trajectory of the ablation study. This figure illustrates the performance changes of different models under the "all-in-one" training paradigm.
• Figure 5: Visual comparison on the WorldView-2 (WV2) dataset.