Advancing Cyclone Tracking with HIMPACT: High-Resolution Multilevel Python-Based Algorithm for Cyclones’ Centroid Tracking

Convection-permitting simulations resolve the deep convective cells that organise Mediterranean tropical-like cyclones. They also generate localised pressure minima that can capture a conventional cyclone tracker and pull it away from the synoptic-scale centre. We introduce High-Resolution Multilevel Python-Based Algorithm for Cyclones’ Centroid Tracking (HIMPACT), an open-source Python algorithm developed by the corresponding author within the CETEMPS framework, that stabilises cyclone-centre identification by combining three elements: a multi-level geopotential analysis restricted to the 800–950 hPa layer, a percentile-based threshold that isolates the vortex core from convective perturbations, and a convex-hull centroid that depends on the geometry of a percentile-defined core rather than on a single extreme grid point, so that an isolated convective pressure deficit cannot displace the estimate by more than a fraction of the core radius. HIMPACT was evaluated in four tracking experiments across three Mediterranean cyclones at grid spacings from 2 to 28 km using WRF, ICON-DREAM and ERA5, while MPAS was additionally used to test portability and computational scaling on an unstructured Voronoi mesh. Across the three experiments in which the driving data resolve a coherent lower-tropospheric cyclone structure, the best five-level configurations reduce root-mean-square displacement errors by approximately 16–48% relative to the corresponding single-level configurations. Activating the absolute minimum alongside the centroid more than doubles the error variance when the pressure field is multi-modal. The 800–950 hPa window avoids both surface extrapolation artefacts below 950 hPa and mid-tropospheric steering signatures above 800 hPa. A counterexample with an extratropical storm exposes a data-quality threshold: when the driving dataset does not resolve a vertically coherent cyclone structure, the multi-level weighted mean diverges, and single-level tracking becomes the safer choice. HIMPACT is model-agnostic, requires no format conversion, and runs on a single CPU core at approximately 9.8–41.3 s per time step for the recommended five-level configuration across the tested back-ends; substantially larger costs occur for high-level-count MPAS configurations.