Abstract: Spike-interference is one of the most common disturbances in 1-second geomagnetic sampling data, and its proper correction directly affects the quality of 1-second geomagnetic observations. Conventional methods, such as the median-average approach, require extensive manual intervention and show limited adaptability when dealing with interference lasting longer than one second. To improve the accuracy of spike-interference correction, this study formulates the task as a machine-learning regression problem and proposes an XGBoost-based model for correcting spike disturbances in 1-second geomagnetic sampling data. Using raw Z-component observational data from the automated geomagnetic station at the Manzhouli Seismic Observatory between January 2024 and April 2025, 17720 samples of magnetic-field variations—capturing and surrounding spike-interference events—were extracted to construct the dataset for model training. The model achieved the following performance metrics on the test set: a mean squared error of 0.010971 nT², a root mean squared error of 0.104742 nT, a mean absolute error of 0.074217 nT, and a coefficient of determination of 0.999852. The results demonstrate the excellent performance of the XGBoost model in correcting spike-interference in 1-second geomagnetic sampling data, providing a robust and high-precision method for automated correction of such disturbances.