Multispectral images can carry more data information to represent color than common three channel images, which causes problems in storage space and communication. In order to solve the above problems, researchers propose to use an interim connection space (ICS). Multispectral data is compressed into the ICS before storage and transmission, spectral data is reconstructed from the ICS when needed, and the interim connection space determines the effect of the transition. Derhak et al. ［JIST, 50: 53-63, 2006］ proposed a 6 dimension ICS called LabPQR. First three dimensions of this space for a given spectral reflectance r are the tristimulus values XYZ (denoted by a column vector t) under a specified viewing condition (represented by a weighting table matrix H). The rest three dimensions is the combination coefficients, denoted by a column vector tPQR, for the metameric black rb under the first three main unit and orthogonal basis vectors, denoted as a matrix B, for the metameric black space, funded using principal component analysis. Here, the spectral decomposition gives the metameric black rb based on the compressed tristimulus value vector t, i. e., rb=r-Mt, where the mapping matrix M is the well-known “R-matrix”. The metameric black space consists of all metameric black rb from the spectral image or an independent training reflectance dataset. The reconstructed reflectance rp is simply given by Mt+BtPQR。 In this paper, a new ICS is proposed and is named MLabPQR. The difference between MLabPQR and LabPQR is the choice of the mapping matrix M. For the proposed MLabPQR, the matrix M was chosen as the “Wiener estimation matrix”. The “Wiener estimation matrix” does not only depend on the viewing condition matrix H but also depends on the training reflectance dataset. Therefore, the choice of the Wiener estimation matrix can keep the main spectral information for the spectral image, which, we hope, can improve the spectral and colorimetric accuracies for the reconstruction. The proposed ICS was tested using the NCS reflectance dataset and a spectral image, and compared with other ICSs such as LabPQR, LabRGB, XYZLMS and LabW2P in terms of spectral accuracy measures (root mean square error (RMSE) and goodness of fit coefficient (GFC)) and colorimetric accuracy measure (CIELAB colour difference). All ICSs were trained using an independent Munsell reflectance and test datasets. Comparison results showed that our proposed ICS out performed all other ICSs in terms of both spectral and colorimetric accuracy measures. Hence, the proposed ICS is expected to find applications in spectral image compression and cross media reproduction.