The work shown is a solution to the limitations of long-term use for surrogate machine learning (ML) models on dynamic domains for the purpose of improved flood warning systems. The solution includes three major components, including a single model for combined flood control structural operations and inundation mapping; the modularization of a single model to minimize the computational cost for future domain updates; and a model input bootstrapping method to leverage observations while minimizing biases in the resulting surrogate ML training dataset. Together, these solutions are tested with various ML architectures to prove viability and highlight the final hurdles for implementation.