DTFIRNet: Deep learning-based Radar Perception for Urban Low-Altitude Wireless Networks

In urban low-altitude wireless network (LAWN) scenarios, radar signals used for aerial drone monitoring are often corrupted by building occlusions and noise, impairing reliable perception. To address this, we propose Deep Time-Frequency Inverse Reconstruction Network (DTFIRNet), a deep learning framework for robust radar perception under occluded conditions. DTFIRNet integrates an enhanced UNet with residual connections and spatio-temporal attention, and a trainable iSTFT module guided by STFT-derived time-frequency representations (TFRs), enabling robust signal restoration under occlusion. Experimental results show that DTFIRNet significantly outperforms existing methods in waveform recovery and delay estimation, achieving an average range error of 0.611~{m}. We also construct and release a dataset of corrupted LFM radar echoes via IEEE DataPort (https://dx.doi.org/10.21227/qw6h-ns61), supporting further research in radar-based LAWN sensing.