Critical hydrology related algorithms to trace the path of surface water flows, including flow accumulation, stream order, watershed delineation, and runoff simulation, can be difficult to execute for large aerial extents at fine spatial and temporal resolutions. Libraries like GDAL that use multi-threaded CPU-based implementations running on a single host may be slow, and distributed infrastructures like Google Earth Engine may not support the kind of computational primitives required by these algorithms. We have developed a GPU-accelerated framework that re-engineers these four algorithms and is able to process areas as large as river basins of 250,000 km2 on commodity GPU workstations. We express these algorithms in terms of easily parallelizable primitives of pixel independent (PI) and short-pixel (SP) operations, and iterative primitives of long-pixel (LP) operations. Each algorithm uses a different mix of the primitives which helps us ensure that the implementation is generic. We show that our implementation of these algorithms produces accurate outputs and is able to achieve significant performance benefits over alternative methods. Being able to execute the algorithms on a commodity GPU workstation paves the path to use on-prem infrastructure to produce national-scale outputs, and collaborate to pool multiple national-scale outputs together for global-scale analysis.

Note: the talk at the conference will also combine the “STACD: STAC Extension with DAGs for Geospatial Data and Algorithm Management” paper into one slot.