Atrial Fibrillation (AF) is the most common cardiac arrhythmia, linked to increased risk of stroke, heart failure, and mortality poses a serious health challenge. This work proposes a noninvasive method to characterize Body Surface Potential Maps (BSPMs) of varying complexity using the Wavelet Scattering Transform (WST). The goal is to distinguish between normal sinus rhythm, AF with fibrosis, and AF with multiple rotors through a classification framework. The processing pipeline includes dimensionality reduction via PCA, robust time-frequency feature extraction using WST, and classification with a Random Forest model. The best results were achieved on dataset with noise injection of 10 dB SNR, reaching 99.6\% accuracy. Misclassification analysis indicated that spectral overlap, particularly in cases with interacting rotors, can hinder class separation. These results support the potential of WST-based BSPM analysis for noninvasive AF mechanism characterization.