Regional flood frequency analysis (RFFA) has been adopted to obtain design flood quantiles in ungauged locations. However, advancement in identifying homogeneous regions-particularly in integrating river network types into RFFA, is lacking. Therefore, we explored the performance of RFFA in estimating flood quantile changes by considering river network information; different river network types were used to define homogenous basins. An ensemble artificial neural network (EANN) and generalized additive model (GAM) were applied to explore regional flood quantile estimates. To perform RFFA, 95 hydrometric stations in arid and semiarid areas of the United States were selected. Among the 95 river networks, 46, 33, and 16 were classified as dendritic, pinnate, and trellis networks, respectively. Better estimates of flood quantiles from the RFFA models were determined and verified using statistical indices in the river network types than those in random networks (non-homogenous basins). Although some regional estimates derived from the GAM model in random networks showed poor performance, most regional estimates from the GAM model provided better accuracy than those of the EANN model when river network types were used as homogenous regions. For instance, the GAM model improved the NASH index by 14.76 and 0.93% in 10-and 50-year quantiles based on dendritic networks over the EANN model, respectively. These results suggest that river network types can be applied to define homogenous basins to develop RFFA procedures, offering better design flood estimates for water-related disasters in arid and semi-arid regions.