Applicant
Prof. Dr. Ralph Kühn
Molecular Zoology, Chair of Zoology
TUM School of Life Sciences
Technical University of Munich (TUM)
Scientific Project Manager
Paul Kasko
Molecular Zoology, Chair of Zoology
TUM School of Life Sciences
Technical University of Munich (TUM)
Project Overview
Movement and redistribution provide observable signatures of how mobile entities interact with heterogeneous and changing environments. Modern tracking and Earth Observation technologies now make it possible to combine high-resolution trajectories with equally detailed information on vegetation, climate, land use, and other spatially and temporally varying conditions. Developed primarily for wildlife telemetry and spatial ecology, hrHSA addresses the broader problem of inferring how environmental structure, individual variation and changing conditions shape observed movement and space use. This creates substantial computational demands: high-resolution analyses may require billions of spatial queries, repeated statistical inference, simulation and prediction across environmental datasets that exceed the memory of individual compute nodes.
Within the KONWIHR project, we developed hrHSA (high-resolution Habitat Selection Analysis), an open-source Python framework that integrates movement trajectories with spatially and temporally resolved environmental data and provides a unified workflow from data preparation to statistical inference, validation, prediction and simulation. hrHSA supports Resource Selection Functions (RSF), Step Selection Functions (SSF) and integrated Step Selection Functions (iSSF), including frequentist and hierarchical Bayesian approaches. Integration with the terrabyte High Performance Data Analytics platform (DLR & LRZ) enables direct coupling of trajectory analyses with large Earth Observation datasets. Its computational concepts apply more broadly wherever large sets of spatial observations are related to heterogeneous environmental fields.
A central objective was to make these analyses scalable without sacrificing statistical rigour or reproducibility. Systematic profiling on CoolMUC-4 identified three dominant computational kernels: sparse sampling from large raster datasets, statistical inference, and dense spatial prediction. Each required a distinct optimisation strategy. hrHSA therefore combines storage-aware spatial access, vectorised likelihood formulations, bounded out-of-core processing, reduced memory allocation and workload-specific parallel execution. Rather than simply increasing CPU core counts, process/thread geometry and task granularity were matched to the computational structure of each workload.
The resulting implementation scales well beyond conventional workstation analyses. On CoolMUC-4, hrHSA handled one billion spatial queries against a 768-GiB raster stack and processed a 900-GiB logical raster workload within a 224-GiB node-memory allocation, demonstrating efficient out-of-core execution. Optimised RSF, SSF and iSSF workflows achieved end-to-end speed-ups of approximately 1.27–1.50× relative to the reference implementation.
hrHSA is a publicly available, extensively documented and reusable research framework that makes high-resolution habitat-selection analysis reproducible and accessible while retaining the controls required for HPC-scale execution. It also provides a computational foundation for future developments such as heterogeneous CPU/GPU computing, accelerator-native statistical inference and deep-learning approaches that combine learned environmental representations with mechanistic movement models. KONWIHR support has thus created both an immediately usable research resource and a platform for the next generation of data-intensive movement ecology.
Further information
[1] hrHSA documentation
[2] Analytical workflow
[3] Performance and scaling
