A Spatio-temporal Residual Attention Network with SNR-adaptive Weighting for MIMO-OFDM Channel Estimation

Abstract

This paper addresses channel estimation in MIMO-OFDM systems under time-varying Rician fading and spatially correlated propagation conditions, where reliable coherent detection requires accurate channel state information. Although least-squares (LS) estimation is simple and computationally efficient, it remains highly sensitive to noise. In addition, conventional deep learning-based estimators may suffer from tracking bias when temporal channel variations are not properly handled. To overcome these imitations, this paper proposes SMART-CE, a spatio-temporal residual attention network with SNR-adaptive weighting. The proposed approach refines a sequence of LS estimates by extracting spatial-frequency features, exploiting temporal channel correlation, and adaptively balancing historical and current channel information according to the operating SNR. The final channel estimate is obtained through a residual correction of the current LS estimate. Simulation results on an 8 x 8 MIMO-OFDM system show that SMART-CE generally achieves lower NMSE and BER than CNN, residual CNN, residual-attention CNN, recursive least-squares (RLS), and normalized least-mean-square (NLMS) baselines, with the most pronounced gains observed in the low-to-medium SNR regime.