This paper examines the recognition of underwater targets by the method of the poles and presents the experimental results. The underwater targets used in experiment are hollow cylinders made from different materials, and with different lengths and thicknesses.The features of the underwater target are extracted in the following three steps. First, the impulse response function of the target is identified in the frequency domain by using the short pulse as the incident signal. Second, the method of the so-called "convolution-modified recursive least-squares" is employed to estimate the parameters of the ARMA model of the target. Third, the poles are calculated in the frequency domain. The poles are adopted as the features of the target and considered to be a constant quantity. The computer simulation shows that the anti-noise ability of the method of the pole is 12dB improved compared with the AR model.In the experiment, the matched- filtering classifier is applied to classify the underwater target. The experimental results has demonstrated that 100% correct recognition rate can be achieved for signal-only cases. With white noise added to the echoes, 100% correct recognition rate is still achieved when signal-to-noise ratio is as low as 8dB, and 87.5% correct recognition rate is achieved for S/N = 5dB.