Cross-scene and cross-clutter generalization of task-based compression results

Characterize the performance of the two-sided task-based radar compression criterion and the resulting codec frontiers across substantially broader scenes, clutter regimes, and regions not represented in training.

Background

The primary experiment uses one Sentinel-1 agricultural/exurban scene and the independent replication uses one AFRL Gotcha pass, channel, and vehicle. Although the qualitative concentration-versus-false-alarm mechanism replicates, the study does not establish broad statistical generalization.

The learned codec is evaluated with a within-scene split, so its reported behavior measures within-scene rather than broad geographic or clutter-regime generalization. Additional Sentinel-1 scenes had been acquired but not evaluated at the time of writing.