Digital elevation models (DEMs) play a vital role in terrain analysis, disaster warning, and environmental monitoring. However, existing ground point extraction methods from point clouds often suffer from high misclassification rates in complex geomorphological settings, significantly compromising DEM accuracy-especially in regions like the Chinese Loess Plateau. To effectively tackle this challenge, we introduce seed-point-guided adaptive ground misclassification detection and correction (SPAG-DC), a novel method based on airborne laser scanning (ALS) sensor data. SPAGDC employs dynamic region growing to extract a reliable ground point core, uses elevation gradients to adaptively construct grids for initial seed points, and introduces a consistency set algorithm to purify these seed points by removing nonground artifacts. High-quality seed points then guide terrain surface fitting using thin-plate splines (TPSs), enabling accurate detection and correction of misclassified points based on distance thresholds. Experiments were conducted based on the existing classical and deep learning (DL)- based filtering methods, with evaluations performed across four dimensions: class II error rate, internal and external conformity, and reference DEM deviation. To further evaluate the generalization ability of the proposed algorithm, the open-source dataset OpenGF was employed for comparative experiments. Results show that SPAG-DC effectively reduces misclassifications, with most elevation deviations between the generated DEMs and ground truth falling within +/- 0.1 m across multiple test areas. The method demonstrates strong spatial consistency and robustness in DEM modeling. Overall, SPAG-DC offers an effective and reliable solution for generating high-precision DEMs in the challenging topography of the Loess Plateau.