FeatureZ: A General Framework for Feature-Preserving Compression via Pointwise Bounds and Star Classification

Abstract

Geometric and topological features, such as isosurfaces, quantiles, merge trees, and Morse-Smale complexes, are central to the analysis and visualization of scientific data across diverse domains, including medical imaging, climate science, materials science, and astronomy. However, most lossy compressors for scientific data provide only pointwise error guarantees and do not preserve derived features. Existing feature-preserving compressors are often difficult to develop and typically tailored to a single feature. In this paper, we introduce FeatureZ, a lossy compression framework for structured volumetric scalar fields that can preserve a wide class of geometric and topological features. In particular, FeatureZ frames feature-preserving compression as preserving pointwise upper and lower bounds together with a consistent classification of each point's star (i.e., its incident cells) in the underlying structured mesh. Although not all notions of feature preservation fit this framework, many features and topological descriptors targeted by existing compressors do. FeatureZ provides an efficient implementation of feature-preserving compression under this formulation. In particular, FeatureZ operates as an augmentation layer that refines the output of an existing lossy compressor. It first applies quantization to enforce pointwise bounds, and then employs an iterative procedure to ensure consistent star classification. We demonstrate that FeatureZ preserves diverse features during compression with minimal overhead, achieving compression ratios comparable to or better than methods specialized for individual features.