FSDFormer: A Frequency-Selected Differential Fusion Transformer for Remote Sensing Image Spatiotemporal Fusion

Spatiotemporal fusion (STF) plays a critical role in remote sensing and Earth observation by integrating high spatial resolution details with high temporal consistency, serving as an essential tool for environmental monitoring and postdisaster assessment. Recently, numerous STF methods, especially deep learning-based approaches, have been proposed and shown promising results. However, most of them struggle with effectively capturing complex spatiotemporal interactions as well as balancing computational efficiency and interpretability. To address these challenges, we propose an innovative STF framework, frequency-selected differential fusion Transformer (FSDFormer), which comprehensively processes multiscale spatiotemporal data while maintaining linear computational complexity. FSDFormer integrates three key components: the spatial enhanced agent Transformer (SEAT) block, the frequency-adaptive filtering network (FAFN), and the spatiotemporal differential fusion (SDF) block. The SEAT block employs a novel attention mechanism to reduce computational complexity, while the FAFN enhances feature extraction by introducing frequency-domain filtering in the forward network. Grounded in a spatiotemporal dependency-aware fusion strategy, the SDF block leverages a two-stage mechanism to excavate and fuse temporal and spatial information sequentially. Extensive experiments show that FSDFormer delivers superior performance on three STF datasets, outperforming state-of-the-art (SOTA) methods in accuracy, robustness, and generalizability. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Sichen-Lu/FSDFormer</uri>