ArtFusion: An Arbitrary-Reference Spatiotemporal Fusion Model for Seamless Remote Sensing Image Reconstruction

Abstract

Spatiotemporal fusion (STF) represents a vital solution for continuous high-resolution Earth observation. Despite the increasing abundance of historical satellite archives, most existing STF methods are constrained by a limited number of auxiliary reference images (typically one or two), leading to underutilized multitemporal information and reduced reconstruction reliability, particularly when References are compromised by cloud contamination or abrupt land-surface variations. To address these limitations, this study proposes an arbitrary-reference STF (ArtFusion) model. The architecture incorporates an efficient contrast-aware hybrid block (ECHB) for deep feature extraction, an explicit temporal information encoder (ETIE) to utilize acquisition metadata, and a multihead cross-reference attention fusion (MCAF) module designed to facilitate the integration of an arbitrary number of reference images. A comprehensive evaluation across 24 experimental cases in three representative study regions demonstrates that ArtFusion consistently outperforms four state-of-the-art (SOTA) benchmarks [multilevel feature fusion with generative adversarial network (MLFF-GAN), C-ROBOT, RealFusion, and frequency-selected differential fusion transformer (FSDFormer)]. Compared with the best-performing baseline among the four competing methods, ArtFusion achieves average improvements of 6.8% in spectral accuracy [root-mean-square error (RMSE)] and 15.3% in spatial accuracy [improved edge difference metric (iEDGE)]. Notably, high reliability is maintained across four challenging scenarios: rapid phenological changes, drastic morphological variations, highly heterogeneous landscapes, and frequent cloud contamination, while also demonstrating strong cross-regional transferability. Despite its superior performance, ArtFusion has an extremely compact structure with 0.19 million trainable parameters, only 2.2% of MLFF-GAN’s parameters. This work demonstrates the potential of leveraging multiple reference images to push the boundaries of STF accuracy, rather than merely increasing model complexity. This flexible multireference fusion scheme provides a promising pathway for robust, large-scale Earth observation in cloudy and dynamically changing landscapes. The source code is available at: https://github.com/Andy-cumt/ArtFusion-STF