Source number estimation is an important problem in signal processing. The performances of traditional methods in this field are severely degraded under some extreme conditions, such as relatively short observed signal length, low signal-to-noise ratio (SNR), or arbitrary noise covariance structure. Recently, machine learning has been used for source number estimation with different data formats as input data due to the advantages of not requiring subjective parameter settings and being data-driven. However, it also has shortcomings such as fluctuating performance and requiring more training samples. In view of this problem, based on the statistical characteristics of the eigenvalues of the covariance matrix, this letter proposes two preprocessing schemes to further manifest the input features of the network model. The input features transformed by the proposed preprocessing methods have specific distributions, which can improve the discriminability between features, thereby enhancing the prediction ability of the network. Simulation results verify that the proposed preprocessing schemes are capable of improving the estimation performance while reducing the fluctuation of the performance.