Physics-Guided Learnable Wavelet Framework for Micro-Doppler Characterization of Low-Altitude Radar Targets

Abstract

Reliable micro-Doppler characterization is relevant to low-altitude radar sensing, where measured echoes may contain weak returns. Fixed-parameter time–frequency methods can produce smeared or fragmented time–frequency representations (TFRs), which can affect weak-component observability and Doppler-ridge traceability. This article presents WConV-TF, a physics-guided learnable wavelet front end for radar micro-motion analysis. WConV-TF converts preprocessed target-range-cell echoes into micro-Doppler TFRs by embedding parameterized Morlet bases in a learnable complex wavelet convolution layer. The normalized center frequencies and bandwidths are optimized from data while retaining a time–frequency interpretation. A multiscale residual module then refines the magnitude-domain TFR. The method is evaluated using simulated data, measured frequency-modulated continuous-wave (FMCW) radar data, and measured S-band array radar data. Across four measured target categories, WConV-TF reduces Rényi entropy by approximately 0.4%–46.3% relative to the strongest competing method for each category and reduces the controlled-rotor mean absolute localization error (MALE) by approximately 66.9% relative to the best baseline. On the common four-category sequence-level split, it improves downstream classification accuracy by 4.66 percentage points over short-time Fourier transform (STFT). The results across simulated signals, two measured radar systems, and four target types indicate applicability to low-altitude radar signal analysis under the evaluated protocols.