Task-Based Evaluation of Raw Radar Data Compression: A Pre-Registered Study of Where Classical Codecs Fail to Preserve Target Detection, and Why

Abstract

Synthetic aperture radar (SAR) systems collect raw I/Q echo data at rates that exceed downlink capacity; fielded systems compress onboard with block-adaptive quantization (BAQ/FDBAQ). Proposed replacements, including learned compression, are typically evaluated with image-quality metrics (PSNR, SSIM, SQNR); none measures whether the data still supports its operational use. We introduce a pre-registered, task-based evaluation methodology for raw radar compression: codecs are scored after SAR focusing against a two-sided detection criterion -- a CFAR detection-agreement floor and a false-alarm budget at matched threshold -- using frozen task models trained once on uncompressed data. On Sentinel-1 stripmap Level-0 data the evaluation reproduces the operating point of the fielded FDBAQ codec (3-bit BAQ sustains utility at 6.12 bits per complex sample), with a previously reported reconstruction-lattice artifact tested for and ruled out. No classical configuration tested sustains utility below 4.86 bits per complex sample (roughly 13:1) on this scene, and each failure has an identifiable mechanism: raw echoes offer no transform-coding gain, while energy concentration after focusing delivers the predicted detection gain (Pd 0.888 vs 0.691 at 2 bits per complex sample) but converts it into false alarms through wavelet ringing; detection probability alone would have certified a codec producing roughly 53 times the tolerable false-alarm count. Applying the same protocol to AFRL Gotcha GMTI data (airborne, never compressed onboard) replicates both findings independently (frontier at 7.99 bits per complex sample). We release the evaluation harness, a unitary dechirp/focus transform with verified round-trip invertibility (relative error ~1e-7), and the complete pre-registration trail, including one retracted overclaim and a bounded negative result for a small learned codec, reported as a lower bound.