In this paper, we investigate the performance of adaptive modulation(AM) orthogonal frequency division multiplexing(OFDM) system in underwater acoustic(UWA) communications. The aim is to solve the problem of large feedback overhead for channel state information(CSI) in every subcarrier. A novel CSI feedback scheme is proposed based on the theory of compressed sensing(CS). We propose a feedback from the receiver that only feedback the sparse channel parameters. Additionally, prediction of the channel state is proposed every several symbols to realize the AM in practice. We describe a linear channel prediction algorithm which is used in adaptive transmission. This system has been tested in the real underwater acoustic channel. The linear channel prediction makes the AM transmission techniques more feasible for acoustic channel communications. The simulation and experiment show that significant improvements can be obtained both in bit error rate(BER) and throughput in the AM scheme compared with the fixed Quadrature Phase Shift Keying(QPSK) modulation scheme. Moreover, the performance with standard CS outperforms the Discrete Cosine Transform(DCT) method. In this paper, we investigate the performance of adaptive modulation (AM) orthogonal frequency division multiplexing (OFDM) system in underwater acoustic (UWA) communications. The aim is to solve the problem of large feedback overhead for channel state information (CSI) in every A novel CSI feedback scheme is proposed based on the theory of compressed sensing (CS). We propose a feedback from the receiver that only feedback the sparse channel parameters. This system has been tested in the real underwater acoustic channel. The linear channel prediction makes the AM transmission techniques more feasible for acoustic channel communications. The simulation and experiment show that significant improvements can be obtained both in bit error rate (BER) and throughput in the AM scheme compar ed with the fixed Quadrature Phase Shift Keying (QPSK) modulation scheme. Moreover, the performance with standard CS outperforms the Discrete Cosine Transform (DCT) method.