Hydrodynamic models are an essential tool for studying the movement of water and other materials across the Earth surface. However, the possible questions which models can address remain limited by practical constraints on model size and resolution, particularly in fluvial and coastal environments in which hydrodynamically-relevant landscape features are topologically complex and span a wide range of spatial scales. The rise in popularity of unstructured meshes has helped address this problem by allowing mesh resolution to vary spatially, and many models support local refinement of the mesh using breaklines or internal regions-of-interest. However, there remains no standardized, objective, or easily reproducible method to define or implement internal features between different users. The present study aims to address whether remote sensing information can be used to fill in that gap, by embedding information about hydrological connectivity and landscape structure directly into an unstructured mesh. We present a fully-automated image processing methodology for preserving dynamically-active connected features in the unstructured 2D shallow-water model ANUGA, while reducing computational demand in other less active areas of the domain. The Unstructured Mesh Refinement Method (UMRM) works by converting a binary input raster into a collection of closed, simple polygons which can be used to internally refine the model mesh, meanwhile preserving landscape connectivity and enforcing model-related constraints. The UMRM and ANUGA are both fully open-source and agnostic regarding the source of remote sensing data used as input, which can include optical, radar, and topographic datasets. We demonstrate the use of the UMRM workflow by applying it to a large-scale model of the Wax Lake and Atchafalaya Delta distributary system in coastal Louisiana. Our model mesh is refined using a long-term time-series of optical Planet imagery, a short-term time-series of interferometric SAR measurements of water level change, and lidar-derived topography data. We compare the results of the connectivity-preserving mesh (CPM) to results from an unrefined mesh using a uniform mesh resolution, and find that the UMRM decreased the number of mesh elements, simulation time, and output data size by around a third, without any loss in model accuracy when compared to in-situ and remotely-sensed water level measurements. To our knowledge, this study is the first to use non-topographic remote sensing data to constrain the mesh structure of a hydrodynamic model, and results from our test application suggest that doing so can result in noteworthy reductions in computational demand.