MorphoNeural CFAR: Interpretable Deep Refinement of Morphological Detection for Radar Range-Doppler Maps

Constant false alarm rate (CFAR) detection degrades severely in nonuniform clutter, which is common in radar range-Doppler (RD) maps. This letter presents a morphological CFAR detector (MNWTH) that replaces statistical background modeling with a ring-shaped morphological opening to obtain structure-preserving background estimates. A refinement CNN is further introduced in the logit domain to adaptively suppress residual clutter while preserving interpretability. Experiments on real bistatic radar data show that the refined MNWTH improves the AUC from 0.7496 to 0.8263 and achieves 2-3× higher partial AUC under low false-alarm rates compared with CA-/GOCA-/SOCA-CFAR and median CFAR. The proposed framework offers a simple, robust, and computationally efficient solution for radar small-target detection in remote sensing applications.