The execution of hydraulic models at large spatial scales has yielded a step change in our understanding of flood risk. Yet their necessary simplification through the use of coarsened terrain data results in an artificially smooth digital elevation model with diminished representation of flood defense structures. Current approaches in dealing with this, if anything is done at all, involve either employing incomplete inventories of flood defense information or making largely unsubstantiated assumptions about defense locations and standards based on socioeconomic data. Here, we introduce a novel solution for application at scale. The geomorphometric characteristics of defense structures are sampled, and these are fed into a probabilistic algorithm to identify hydraulically relevant features in the source digital elevation model. The elevation of these features is then preserved during the grid coarsening process. The method was shown to compare favorably to surveyed U.S. levee crest heights. When incorporated into a continental‐scale hydrodynamic model based on LISFLOOD‐FP and compared to local flood models in Iowa (USA), median correspondence was 69% for high‐frequency floods and 80% for low‐frequency floods, approaching the error inherent in quantifying extreme flows. However, improvements versus a model with no defenses were muted, and risk‐based deviations between the local and continental models were large. When simulating an event on the Po River (Italy), built and tested with higher quality data, the method outperformed both undefended and even engineering‐grade models. As such, particularly when employed alongside model components of commensurate quality, the method here generates improved‐accuracy simulations of flood inundation. Plain Language Summary Traditional flood risk assessments are carried out using computer models built with local data, but their spatial coverage is impaired by how expensive and time‐consuming they are to produce. Recent advances in data availability, understanding of necessary physical process representation, and computational capacity have enabled hydraulic models of the entire globe to be built in an automated fashion at a fraction of the financial and human cost. However, their accuracy can be significantly impaired by a lack of information on flood defenses. As the model is built, elevation data are coarsened to reduce the number of calculations required to simulate flooding over such wide areas. This results in flood defense structures being smoothed out of the terrain information used in the model. Publicly available defense inventories are of insufficient coverage to ameliorate this issue. In this paper, a method is presented, which automatically detects levee‐like features in high‐resolution elevation data and accurately represents their heights during this necessary coarsening process. Simulating flood inundation over this “defended” topography results in high correspondence between local models and observations for test cases in the United States and Italy, with improvements particularly felt where a lack of defense information is the dominant source of error. Key Points Flood defense representation is presently poor in large‐scale flood models, impairing their ability to map flood hazard accurately A new method is presented, which automatically identifies hydraulic structures in terrain data and accurately preserves their elevations Hydraulic simulations where a lack of defense data is the dominant error show significant improvements in skill when incorporating this method