Estimating the number of concurrent sound sources is a fundamental step for spatial audio processing, yet it remains challenging in highly reverberant environments. While recent deep learning methods utilize spherical microphone arrays, they predominantly rely on magnitude-only spatial power maps, discarding the phase information of the incident sound field. We propose SGCT (Spherical GCN-Transformer), an end-to-end architecture that estimates source counts directly from complex-valued Steered Response Functional (SRF) maps. The model projects the SRF onto a HEALPix-discretized sphere, using a Graph Convolutional Network (GCN) to capture spatial geometry and a Transformer encoder to model temporal dynamics across frames. Experimental results on simulated reverberant scenes demonstrate that retaining phase information and leveraging spherical graph convolutions yields a macro F1 score of 0.919, significantly outperforming classical estimators and magnitude-only baselines, particularly for high source counts.