Extreme precipitation events have significant environmental and societal impacts. Frequency analysis of such extremes is challenging due to the high spatial and temporal variability of rainfall and sparse gauge coverage. Conventional regionalization methods address this by grouping gauges into homogeneous regions and applying the index flood method, which assumes stations in the same region share a common distribution scaled by local indices. Although widely used, this method faces challenges in clustering, as grouping by geographic proximity alone may overlook important statistical similarities in rainfall extremes. In this study, we introduce a network-based framework to define pooling regions for regional frequency analysis of daily annual maximum rainfall in the region of Thessaly, Greece. First, we build a weighted adjacency matrix representing the Euclidean distances between all pairs of gauges using four widely available physical covariates: latitude, longitude, elevation, and mean annual precipitation. The decision of whether two nodes should be connected is determined by a genetic algorithm, which optimizes a cost function designed to capture global network characteristics. The network communities are identified using the Walktrap algorithm, which analyzes short random walks that tend to remain within the same community. For comparison, we construct a second network in which connections are based on both spatial proximity (latitude, longitude) and statistical similarity expressed with L-moments. Our comparison shows that the physically based method produces larger regions that are more heterogeneous. On the contrary, the statistically informed method identifies smaller, more fragmented homogeneous regions. Notably, the physical clustering fails to achieve an optimal partition, meaning it does not group all stations into homogeneous regions. These differences emphasize the need to understand why rainfall extremes behave differently statistically than their physical locations suggest. This understanding will lead to more robust estimates of extreme rainfall.