RaDOD-Net: A Data-Contracted CNN–LSTM–Attention Architecture for Radar Enhancement and Airport Obstacle Detection

Abstract

Radar learning studies often place denoising, temporal feature extraction, attention, and target detection inside one pipeline while leaving the underlying radar product insufficiently specified. This omission is consequential: a polar intensity scan, a range–Doppler map, a range–azimuth map, and a raw complex radar cube do not support the same transformations, supervision, or claims. This paper reconstructs RaDOD-Net as a testable, product-aware architecture rather than as a report of unverified performance. The proposed system combines a shared convolutional encoder, patchwise causal LSTM states, local multi-head attention, a residual reconstruction head, and a calibrated detection head. Its central contribution is a radar data contract that binds every model component and metric to explicit axes, units, calibration, labels, temporal alignment, and provenance. A staged validation protocol separates signal restoration, target preservation, detection operating characteristics, domain transfer, and operational readiness. Oxford Radar RobotCar is treated as an urban polar-radar source that can support representation and transfer experiments after new annotation, but not as raw range–Doppler evidence or an airport benchmark. No numerical superiority is claimed. Instead, the paper provides tensor-level interfaces, leakage-resistant splits, matched baselines, uncertainty reporting, ablations, resource accounting, and evidence gates needed to test the architecture reproducibly. The result is a falsifiable research design for determining whether joint enhancement and detection improves weak-target surveillance without erasing target energy or inflating false alarms.