Sub-terahertz (sub-THz) radar is used for emerging applications of sensing owing to its high bandwidth and fine range resolution, but its low signal-to-noise ratio (SNR), waveform distortion, phase instability, impulsive interference, and non-stationary clutter can significantly affect target detection. Thus, this paper proposes a lightweight convolutional neural network (CNN) based target detection method that operates in the spectrogram domain in degraded and statistically mismatched sub-THz radar environments. The proposed detector is based on log-spectrograms of Short-Time Fourier Transform and uses a distributed learning approach based on time-frequency features distributed in space and observed over time, rather than relying on a particular model of target waveform or a fixed clutter distribution. In a controlled simulation framework, a stochastic clutter variation, impulsive interference, spectral spreading, phase perturbations, waveform distortion, and low SNR conditions are included. In order to assess the system beyond the nominal training distribution, the testing environment is also modified using well-controlled variations of the target strength, the characteristics of the waveform, the phase behavior, and the clutter statistics. In addition, the detector is compared against the matched-filter, spectral cell-averaging constant false-alarm rate, and a compact learning-based baseline detector, using ROC-AUC and probability of detection at low false-alarm operating conditions. The results show that the proposed CNN can achieve a high ability to discriminate targets under different degraded and shifted sensing conditions, and the performance degradation can be observed under strong target-strength reduction and highly distorted sensing conditions. An assessment using computation reveals that the compact architecture has 60 609 parameters and the FP32 model size is about 0.2312 MB, with a single-sample inference latency of about 21.8 ms on the measured evaluation platform. The obtained results show that the spectrogram domain feature learning method is compact and under low-SNR, cluttered, and domain-shifted sensing conditions, which is an experimentally supported method for target detection in sub-THz radar systems.