Scholarlink

<p>There is a substantial gap in global research concerning the temporal prediction of landslides. Advances in machine learning within the geosciences present an opportunity for disaster-prone countries, such as the Philippines, to leverage limited resources in developing effective landslide prediction models. In this study, we present a landslide prediction system for a mountain highway in Benguet, Philippines, that relies on a minimal dataset consisting of rainfall measurements and documented landslide occurrences. To enhance the accuracy of the prediction models, a spatial component was incorporated by subdividing the landslide inventory into datasets based on three distinct lithologic domains. These datasets were balanced and input into five machine learning algorithms: Gaussian Naïve Bayes (GNB), Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF). Among these, the RF models demonstrated the highest performance, with the best model achieving an Area Under the Receiver Operating Characteristic Curve (AUROC) of 82.7%, a True Positive Rate (TPR) of 85.4%, and a True Negative Rate (TNR) of 80%. Although all RF models are reliable (AUROC ≥ 0.7, TPR ≥ 0.7, and TNR ≥ 0.7), certain models trained using lithologically- constrained datasets show better performance than those trained on the unconstrained dataset. This finding suggests that grouping landslide events based on lithologic domains prior to model development can enhance temporal landslide prediction models. This approach can be incorporated into early warning systems, as it enables spatially-informed temporal prediction of landslides.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lithologically-constrained, machine learning-based temporal landslide prediction models using rainfall time series for the Benguet First Engineering District, Philippines

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Jamila B. Abuda,
  • Ricarido M. Saturay Jr,
  • Sandra G. Catane,
  • Ivy Guevarra

摘要

There is a substantial gap in global research concerning the temporal prediction of landslides. Advances in machine learning within the geosciences present an opportunity for disaster-prone countries, such as the Philippines, to leverage limited resources in developing effective landslide prediction models. In this study, we present a landslide prediction system for a mountain highway in Benguet, Philippines, that relies on a minimal dataset consisting of rainfall measurements and documented landslide occurrences. To enhance the accuracy of the prediction models, a spatial component was incorporated by subdividing the landslide inventory into datasets based on three distinct lithologic domains. These datasets were balanced and input into five machine learning algorithms: Gaussian Naïve Bayes (GNB), Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF). Among these, the RF models demonstrated the highest performance, with the best model achieving an Area Under the Receiver Operating Characteristic Curve (AUROC) of 82.7%, a True Positive Rate (TPR) of 85.4%, and a True Negative Rate (TNR) of 80%. Although all RF models are reliable (AUROC ≥ 0.7, TPR ≥ 0.7, and TNR ≥ 0.7), certain models trained using lithologically- constrained datasets show better performance than those trained on the unconstrained dataset. This finding suggests that grouping landslide events based on lithologic domains prior to model development can enhance temporal landslide prediction models. This approach can be incorporated into early warning systems, as it enables spatially-informed temporal prediction of landslides.