Underwater acoustic channels exhibit typical doubly selective time-frequency characteristics. Severe multipath propagation, Doppler effects, and large delay spreads give rise to pronounced time variation and sparse structural features, making accurate channel tracking and parameter estimation relatively difficult. To improve the transmission reliability of underwater acoustic communication systems under non-Gaussian noise and cluster-sparse channel conditions, this paper first establishes a Polar-coded underwater acoustic OFDM system. To address the problem of robust estimation of cluster-sparse time-varying underwater acoustic OFDM channels under limited pilot and non-Gaussian heavy-tailed noise conditions, a Decision-Aided Cluster-aware Subspace Pursuit (DA-CSP) channel estimation method is proposed. The method establishes cluster center priors from the first-symbol full-pilot estimate, constrains the sparse search space of subsequent symbols to local delay intervals via a cluster diffusion window, and introduces cluster-aggregated correlation to suppress anomalous tap misselection caused by heavy-tailed noise. When the signal-to-noise ratio meets a reliability threshold, high-confidence data decisions are used to construct virtual pilots, forming an extended observation matrix to enhance the observability of channel taps within the cluster window, thereby realizing a two-stage estimation procedure of pilot pre-estimation and data-aided refinement. Compared with conventional least-squares and MBSE methods, DA-CSP compresses the matrix inversion dimension from the full channel length L to the cluster window width W (W<