There is no uniform frame for evaluating a parametric model. A new criterion named by RIA for the model selection is presented which synthetically considers the approximation precision, sparsity of parameters and the residual information. Here the residual information with Gauss distribution is measured by relativity of the residual. According to a simple thought that the whole system is comprised of model information and residual information after modeling the data. This new criterion is transformed into a measurement form for model information. And its reasonability is demonstrated here by theoretic analysis with the application in space data processing.