Abstract: To eliminate effect brought from the noise in underwater OFDM system for sparse multi-path channel estimation, iterative least square adaptive channel estimation method based on threshold detection is proposed. According to least-square criterion, error squares are weighted and summed by introducing weighted factor and a iterative equation is derived to execute adaptive channel estimation. The computation complexity of iterative equation is low and it can not bring out big matrix calculation. Based on the solution of LS estimation, the most significant paths of channel are detected using predetermined threshold and the non-significant paths of time channel impulse response are set zero, reducing the effect of noise for channel estimation and increasing estimation precision. At the end, the effect of variant training symbol lengths for MSE performance is discussed, at the same time, the performance simulation of the proposed method is implemented on condition that the training symbol length, threshold, and so on are prescribed and the simulation result is analyzed.