A state-of-the-art deep learning framework, HyperReconNet, recover an spectral image from its compressed measurements. However, HyperReconNet does not take the sensing matrix into a account on the training. We propose a residual modbased convolutional neural network to address this limitation.