In this paper, we propose a new method to designing a time-frequency (TF) binary mask for speech enhancement in low SNR environments. The mask is related to the reconstructed speech signal using a robust time-varying filtering (RTVF) algorithm. The key to the RTVF algorithm is the instantaneous frequency (IF) estimation of the encoded speech signal. Considering the characteristic of the speech spectrogram, we particularly utilize mathematical morphology to achieve F estimation instead of estimating the peak of the TF distribution directly. Based on an initial F information, the reconstructed speech signal with less noise is obtained via RTVF. According to the energy distribution of the reconstructed signal`s TF representation, we predict the TF binary mask which selectively retains or discards TF points. Experimental results show that the proposed method achieves satisfactory performance in low SNR environments, and a comparison with other conventional speech enhancement algorithms is also given.