The mechanisms of atrial fibrillation (AF) are not completely understood and a precise analysis of the atrial activity (AA) signal in electrocardiogram (ECG) recordings is necessary for a better understanding. Blind source separation (BSS) techniques as principal and independent component analysis (PCA and ICA), and block term decomposition (BTD), have proven useful in extracting the AA source from ECG recordings. However, the automated selection of the AA source among the other sources is still an issue. In this scenario, the present work proposed two contributions: i) the normalized mean square error of the TQ segment (NMSE-TQ) as a new feature to quantify the AA content of a source, and; ii) an automated classification of AA and non-AA sources using 3 machine learning algorithms, linear and quadratic discriminant analysis (LDA and QDA), and support vector machine (SVM). The 3 BSS techniques mentioned above were applied in 12 short AF ECGs from 12 different patients (BTD was applied several times, since its performance depends on the initialization of its computation algorithm), generating 767 sources that were visually labeled as 340 AA sources and 427 non-AA sources. Three parameters were extracted and used as features for classification: spectral concentration (SC) in %, kurtosis of the signal in frequency domain and NMSE-TQ. The accurate performance of LDA, QDA and SVM were 82.63%, 85.14%, and 90.63%, respectively, overcoming the techniques of the state-of-the-art, which the most accurate one presented 78.53% of accuracy. The NMSE-TQ alone presented 78.01% of accuracy in selecting the AA source. It was also observed that the mean and standard deviation (m +/- std) of SC, kurtosis and NMSE-TQ for the AA sources was 62.7 +/- 15.2, 136.5 +/- 4.5 and 1.7 +/- 3.7, while for the non-AA sources was 39.7 +/- 15.6, 42.4 +/- 30.7 and 32.6 +/- 73.1, respectively.