Description: For the first time, we present the application of an unstructured triangular grid to the Finite-Volume Community Ocean Model using the ensemble Kalman filter scheme, to assimilate coastal acoustic tomography (CAT) data. The fine horizontal and vertical current field structures around the island inside the observation region were both reproduced well. The assimilated depth-averaged velocities had better agreement with the independent acoustic Doppler current profiler (ADCP) data than the velocities obtained by inversion and simulation. The root mean square difference (RMSD) between depth-averaged current velocities obtained by data assimilation and those obtained by ADCPs was 0.07 m s −1 , which was less than the corresponding difference obtained by inversion and simulation (0.12 m s −1 and 0.17 m s −1 , respectively). The assimilated vertical layer velocities also exhibited better agreement with ADCP than the velocities obtained by simulation. RMSDs between assimilated and ADCP data in vertical layers ranged from 0.02 to 0.14 m s −1 , while RMSDs between simulation and ADCP data ranged from 0.08 to 0.27 m s −1 . These results indicate that assimilation had the highest accuracy. Sensitivity experiments involving the elimination of sound transmission lines showed that missing data had less impact on assimilation than on inversion. Sensitivity experiments involving the elimination of CAT stations showed that the assimilation with four CAT stations was the relatively economical and reasonable procedure in this experiment. These results indicate that, compared with inversion and simulation, data assimilation of CAT data with an unstructured triangular grid is more effective in reconstructing the current field.