The sensitivity of underwater propagation models to acoustic and environmental variability increases with the signal frequency; therefore, realizing accurate acoustic propagation predictions is difficult. Owing to this mismatch between the model and actual scenarios, achieving high-frequency source localization using model-based methods is generally difficult. To address this issue, we propose a deep learning approach trained on real data. In this study, we focused on depth estimation. Several 18-layer residual neural networks were trained on a normalized log-scaled spectrogram that was measured using a single hydrophone. The algorithm was evaluated using measured data transmitted from the linear frequency modulation chirp probe (11-31 kHz) in the shallow-water acoustic variability experiment 2015. The signal was received through two vertical line arrays (VLAs). The proposed method was applied to all 16 sensors of the VLA to determine the estimation performance with respect to the receiver depth. Furthermore, frequency-difference matched field processing was applied to the experimental data for comparison. The results indicate that ResNet can determine complicated features of high-frequency signals and predict depths, regardless of the receiver depth, while exhibiting robust environmental and positional variability.