The phase-sensitive optical time-domain reflectometer systemissensitive to ambient environmentsuch as sound waves, air flow and transient high-frequency noise. Thus the practical signal-to-noise-ratio is always low, the real intrusion signals are always severely obscured and difficult to be detected, as a consequence, false alarm occurs frequently. To solve the problem, a signal processing schemeusing feed-forward neural network together with singular spectrum feature extraction algorithm was presented for disturbance detection. The main points of the method were mapping the slidinglongitudinal time series to a sequence of multi-dimensional lagged vectors which form trajectory matrix and computing the singular value decompositionof the trajectorymatrix. This method can effectively eliminate sound waves and transient high-frequency acoustic noise, etc. Performance results showthat with breeze interference, the correct detection rate is 90% and false alarm rate is less than 2% at a range of 14 km.