The emergence and establishment of SARS-CoV-2 variants of interest (VOI) and variants of concern (VOC) highlight the importance of genomic surveillance. We propose a statistical learning strategy (SLS) for identifying and spatiotemporally tracking potentially relevant Spike protein mutations. We analyzed 167,893 Spike protein sequences from US COVID-19 cases (excluding 21,391 sequences from VOI/VOC strains) deposited at GISAID from January 19, 2020 to March 15, 2021. Alignment against the reference Spike protein sequence led to the identification of viral residue variants (VRVs), i.e., residues harboring a substitution compared to the reference strain. Next, generalized additive models were applied to model VRV temporal dynamics, to identify VRVs with significant and substantial dynamics (false discovery rate q-value <0.01; maximum VRV proportion > 10% on at least one day). Unsupervised learning was then applied to hierarchically organize VRVs by spatiotemporal patterns and identify VRV-haplotypes. Finally, homology modelling was performed to gain insight into potential impact of VRVs on Spike protein structure. We identified 90 VRVs, 71 of which have not previously been observed in a VOI/VOC, and 35 of which have emerged recently and are durably present. Our analysis identifies 17 VRVs ~91 days earlier than their first corresponding VOI/VOC publication. Unsupervised learning revealed eight VRV-haplotypes of 4 VRVs or more, suggesting two emerging strains (B1.1.222 and B.1.234). Structural modeling supported potential functional impact of the D1118H and L452R mutations. The SLS approach equally monitors all Spike residues over time, independently of existing phylogenic classifications, and is complementary to existing genomic surveillance methods.