Assessing the Reliability of Micro-Doppler Spectrogram Classification from the DIAT-μSAT JPEG Archive

Abstract

Reliable AI-assisted interpretation of radar-derived micro-Doppler images requires evaluation protocols that separate image-archive regularities from physical generalization. We audited 4849 DIAT-μSAT JPEG spectrograms and evaluated partition sensitivity, image controls, and representation-randomized training. Raw radar recordings and acquisition metadata were unavailable. A post hoc feature audit found within-class continuity in filename order and a measurable shift between early training and late test blocks in all six classes, using both frozen ResNet18 and HOG features; the source of this order structure remains unknown. Frozen ResNet18 balanced accuracy decreased from 97.71% under a random split to 88.64% under the class-stratified, late-index S3 split. Boundary-restricted images retained class information but did not identify a non-physical artifact or prove shortcut learning. Native JPEG file attributes were separately class-associated but were not supplied to the image classifiers. In a confirmatory long-budget comparison with ten paired optimization seeds and matched checkpoint selection, representation-randomized F2 exceeded the ColorJitter-plus-RandomGrayscale A3WC comparator by 1.69 percentage points in mean seed-wise worst-case balanced accuracy over the protocol-defined K0/K1/K3/K4 rendering set (seed-bootstrap 95% interval +0.45 to +2.78; exact two-sided sign-flip p = 0.0352). F2 did not improve the held-out K2 rendering stress test. The results establish an image-level archive-order evaluation stress and a limited finite-rendering training benefit, not acquisition-independent radar recognition.