The second order statistics of the noise seen in the recent continuous line acquisition data from multimeasurement streamers vary in frequency, time and space domains. Incorrect estimation of such noise statistics can lead to poor reconstruction of the wavefield by the generalized matching pursuit (GMP). A new noise power estimation technique based on a Bayesian recursive framework is proposed. This algorithm can not only update the noise power estimation when the signal absence is presumed, but also track dynamically to the realistic levels of non-stationary noise. Results using a 3D testdata shows improved noise estimation performance over the original method that assumes the noise statistics is stationary within any shot gather. This improvement is achieved by reducing the leaked turn noise into the GMP output.