Surveillance data used by epidemic alert systems are typically fully aggregated in space at the national level. However, epidemics may be spatially heterogeneous, undergoing distinct dynamics in distinct regions of the surveillance area. We unveiled this in retrospective analyses by classifying incidence time series. We used Pearson correlation to quantify the similarity between local time series and then classified them using modularity maximization. The surveillance area was thus divided into regions with different incidence patterns. We analyzed 31 years (1985-2016) of data on influenza-like illness from the French Sentinelles system and found spatial heterogeneity in 19 of 31 influenza seasons. However, distinct epidemic regions could be identified only 4-5 weeks after a nationwide alert. The impact of spatial heterogeneity on influenza epidemiology was complex. First, when a nationwide alert was triggered, 32%-41% of the administrative regions of France were experiencing an epidemic, while the others were not. Second, the nationwide alert was timely for the whole surveillance area, but subsequently regions experienced distinct epidemic dynamics. Third, the epidemic dynamics were homogeneous in space. Spatial heterogeneity analyses can provide information on the timing of the peak and end of the epidemic, in various regions, for use in tailoring disease monitoring and control.