Change detection-based potential landslides in three prone areas of South Korea

Geological, morphological, and meteorological conditions have made South Korea prone to landslides. This study proposes a change detection method to detect landslides in the Yecheon, Yeongju, and Jeongseon areas of South Korea. We compared the performances of support vector machine (SVM), maximum likelihood (ML), and random tree (RT) algorithms for detecting potential landslides. The process begins by classifying PlanetScope images taken before and after reported landslides that occurred during the 2023 rainy season. Our evaluation showed that SVM model outperformed the other two models, achieving mean precision, recall, and F1-score values of 60.92%, 80.15%, and 68.24%, respectively; RT and ML algorithms had accuracy metrics that were lower by 2%–8%. Approximately 369 potential landslides were detected in the Yecheon area, 412 in the Yongju area, and 108 in the Jeongseon area based on the SVM data. The proposed method enables rapid and effective generation of a potential landslide map, offering valuable insights for the development of mitigation measures and prevention policies.

地质、地貌与气象条件的综合影响，使得韩国山体滑坡灾害风险居高不下。本研究针对韩国醴川（Yecheon）、荣州（Yeongju）与旌善（Jeongseon）三地，提出一种用于潜在滑坡检测的变化检测方法，并对比了支持向量机（SVM）、最大似然（ML）与随机树（RT）三种算法在该任务中的性能表现。实验流程首先对2023年雨季发生的已报道滑坡事件的前后时序PlanetScope影像开展分类工作。评估结果显示，支持向量机（SVM）模型的性能优于其余两种对比模型，其平均精确率、召回率与F1值分别达到60.92%、80.15%与68.24%；随机树（RT）与最大似然（ML）算法的精度指标较该模型低2%至8%。基于SVM模型的检测结果，醴川地区共检出约369处潜在滑坡，荣州地区412处，旌善地区108处。本研究所提方法可快速高效地生成潜在滑坡空间分布图，为滑坡减灾措施制定与防灾政策优化提供了极具价值的参考依据。