To detect and recognize passive fish acoustic signal from marine noisy environment, sparse decomposition is present to realize the endpoint detection by coherent ratio feature. During training，the algorithm extracts feature of fish and wave acoustic as test object under different signal-to-noise(SNR), feature from noisy signal segment is extracted in the testing to classify with test object, finally it is realized by threshold endpoint detection method. Experiment shows that effective signal segments can be detected more accurate by this algorithm than by the power spectrum feature algorithm under low SNR.