Reducing flood risk through improved disaster planning and risk management requires accurate and reliable estimates of flood damages. Models can provide such information by calculating the costs of flooding to exposed assets, such as buildings within a community. Computational or data constraints often lead to the construction of such models from coarse aggregated data, the effect of which is poorly understood. Through the application of a novel spatial segregation framework, we are able to show mathematically that aggregating flood grids through averaging will always introduce a systematic error in a particular direction in partially inundated regions. By applying this framework to a case study we spatially attribute these errors and demonstrate how the exposure of buildings can be an order of magnitude more sensitive to these errors than uninhabited regions. This work provides insight into, and recommendations for, upscaling grids used by flood risk models. Further, we demonstrate a positive dependence of systematic error magnitude on scale coarseness, suggesting coarse models be used with caution and greater attention be paid to issues of scale.