arXiv:2611.23987v1 Announce Type: new Abstract: Low Earth Orbit (LEO) satellite communications are essential for achieving ubiquitous global connectivity. However, the high mobility of LEO satellites induces massive Doppler shifts and fast time-varying fading, severely impairing signal synchronization. Accurate Carrier Frequency Offset (CFO) estimation is critical for receiver synchronization. Nevertheless, conventional estimators and deep learning approaches often degrade significantly under highly dynamic, low-SNR conditions. To address these challenges, this paper proposes a novel Convolution Augmented Transformer architecture for robust, blind CFO estimation in LEO communication systems. Initially, to overcome the inherent limitations of standard linear embedding layers in capturing nonlinear phase rotations, we construct a multi-domain input feature that encompasses in-phase/quadrature (I/Q), polar, and statistical autocorrelation properties. Subsequently, a Convolutional Pre-Denoising (CPD) module, integrating multi-scale dilated convolutions and an attention mechanism, is deployed at the front end to adaptively suppress localized noise and extract robust spatiotemporal representations. These refined latent features are then processed by a Transformer encoder to explicitly track global, long-range phase dependencies. Extensive simulation results demonstrate that the proposed data-driven approach significantly outperforms conventional estimation algorithms, achieving enhanced estimation accuracy and superior resilience in highly dynamic LEO environments.