Robust principal component analysis (RPCA), a novel method for speech enhancement (SE), is expected to decompose the spectrogram of a noisy speech into a low-rank matrix and a sparse matrix, which contain noise components and speech components, respectively. However, some speech components, which are not so variable in different time frames, are possible to be decomposed into a low-rank matrix as noise mistakenly. To address this problem, a novel SE method based on spectrogram-rearranged RPCA (SRPCA) is proposed for a sparse matrix with better decomposition for all speech components in white noise environments. For further improvement under coloured noises corruption, the multi-band method is introduced for SRPCA to be applied in all bands individually. Accordingly, a time-domain enhanced speech is reconstructed from the processed sparse matrix. Numerical experiments show the effectiveness of the proposed method.