Reclassifying Fusion

An Empirically Grounded Alternative to the JDL Level Hierarchy

Abstract

The multisensor data fusion community has spent over three decades cycling through numbered "level" schemes, from the original JDL model (1985) through the Revised JDL (Steinberg, Bowman, White, 1999), the DFIG extensions (Blasch et al., 2002, 2005), Revision II (Llinas et al., 2004), and Steinberg's ontological refinement (2022, 2023), producing at least ten distinct naming conventions that remain in concurrent use with no formal deprecation of any predecessor. The result is a field where "Level 2" can mean situation assessment, feature fusion, or decision fusion depending on which scheme an author happens to adopt. Drawing on an empirical catalog of 1,647 fusion methods across four volumes, spanning nearly five decades of literature from the late 1970s through 2026 and thousands of primary references, I identify seven root causes of the naming confusion (pipeline implication, inconsistent counting, semantic overloading of level numbers, orthogonality of the Dasarathy input output model, informality of the low/ medium/high scheme, version fragmentation across revisions, and conflation of estimation with control). I then propose a two axis EntityType.Operation naming convention grounded in what the catalog actually reveals: methods group naturally by the type of entity whose state they estimate or control (Signal, Object, Relational, Predictive, Control, Human) and by the mathematical operation they perform (Detection, Estimation, Classification, Combination, Optimization). The convention is descriptive rather than ordinal, self documenting, free of implied processing sequence, tolerant of methods that straddle traditional level boundaries, and directly mappable to real system architectures. A full crosswalk table maps the proposed convention back to JDL, DFIG, Dasarathy, and the three tier low/medium/high scheme, demonstrating that EntityType.Operation subsumes rather than replaces the prior work.