The lack of up-to-date, easily available flood risk data has slowed down Canadian attempts to mitigate flood damage, notwithstanding the substantial historical losses associated with frequent and devastating flood episodes. To address this problem, we present a novel flood risk estimation framework. The pipeline is extremely modular and consists of an advanced geographical pre-processing system, a probabilistic deep learning based marginal regional regression model, an advanced hydraulic model, and a web-interface based SaaS platform for zonal risk estimation and visualization. The model is implemented on entire North America covering almost 14,000 network groups and 1.9 million catchments. There are 2 flavors of the model coarse-resolution (30m) and high-resolution (1m where HRDEM is available). The geographic pre-processing system include creation of topology of the river and catchment network, DEM conditioning, LU/LC processing, for the hydraulic model. It also pre-processes the location of the station, soil parameters, topographic variables, climate parameters for the input to the hydrologic model. The deep learning based probabilistic regional regression models relates an almost 130 catchment specific and upstream aggregated co-variates with the distribution of the annual maximum discharge of almost 16,000 stations. This model is further used to infer the distribution and in turn 15 quantiles/return periods (2-1500) of the discharge at all the ungauged locations. These discharge return periods are then used as input to the volumetric manning’s equation based hydraulic model which estimates the corresponding flood maps for each catchment. The testing of the flood maps against observations and govt. studies indicated good performance.