Abstract Bryde’s whales form a major coastal aggregation in the Beibu Gulf, China. Using 1 year of continuous island‐based seismic recordings from Xieyang Island, we established a large labeled data set of coastal Bryde’s whale calls with more than 1.7 million samples. A convolutional neural network trained on these data achieved 99.7% accuracy in binary classification of seismic spectrograms as whale call or non‐call. Analysis of the detected calls revealed ultralow frequencies down to 5 Hz and rhythmic repetition with variable interpulse intervals. Vocal activity persisted from January to July, with a preliminary suggestion of reduced calling during early afternoon hours. These results demonstrate the effectiveness of onshore seismometers and deep learning for continuous, low‐cost monitoring of coastal baleen whales.