Moving target classification is an important issue in wireless sensors. The wild environment makes it a difficult problem for the acoustic signals. In this paper, a new classification method for moving targets in the wild is proposed based on microphone array and linear sparse auto-encoder (LSAE). First, the acoustic signals of moving targets are enhanced by delay-and-sum (DS) beamformer in the narrowband way for the simplicity. The enhancing effects are given a detailed analysis. Then, a spatial feature named noise likelihood (NLH) is presented to further resist the interferences and noise widely existing in the wild. The NLH has a good ability to distinguish between the moving targets and noise. Moreover, to make full use of both the signals beamformed and the NLH, a classification network combining the LSAE layers to learn their representations by self-taught learning and the softmax layer for the classification is built. Experiments show that not only the representations learned by the LSAE layers are robust and much distinguishable but also the proposed method achieves a much better classification performance in comparison with the baseline classifiers for moving targets in the wild.