Scholarlink

<p>To effectively manage and utilize groundwater resources, assessing groundwater potential is essential. However, evaluation precision is limited by uncertainties in groundwater potential mapping due to the number of negative samples (randomly selected non-spring locations outside existing spring buffer zones) and modeling method constraints. This study focused on Yiliang County, Yunnan Province, China. Fifteen thematic layers were selected from five categories: topography and geomorphology, hydrography, geology, human activities, and environment. Four datasets with positive-to-negative sample ratios (ratio of known spring locations to non-spring locations) of 1:0.5, 1:1, 1:2, and 1:3 were constructed to explore the impact of negative sample quantity. Two classic machine learning models—support vector machine (SVM) and random forest (RF)—were compared with a novel multi-grained cascade forest (GCF) model. The computationally efficient dung beetle optimizer (DBO) was employed for parameter optimization, forming three hybrid models: SVM-DBO, RF-DBO, and GCF-DBO. Model performance was evaluated using five metrics: raincloud plots, area under the receiver operating characteristic curve (AUC), accuracy (ACC), true positive rate (TPR), and true negative rate (TNR). Results indicate that increasing negative samples reduces prediction errors for non-spring locations but increases errors for spring identification. The 1:1 ratio achieved optimal performance, with all models exceeding 70% accuracy for both spring and non-spring predictions. The GCF-DBO model demonstrated superior performance (ACC = 0.795, AUC = 0.845). These findings highlight the critical role of sample balance and model selection in enhancing groundwater potential predictions. This research provides actionable insights for sustainable groundwater resource management.</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

Optimizing groundwater potential assessment: uncertainty reduction through sample balancing and enhanced hybrid modeling

  • 期刊论文

  • 发表日期：

    2025年2月26日

查看全文
  • Rui Liu,
  • Juncheng Gou,
  • Jialiang Han,
  • Kai Cao,
  • Xin Pan,
  • Decheng Wang,
  • Liangshuai Wei

摘要

To effectively manage and utilize groundwater resources, assessing groundwater potential is essential. However, evaluation precision is limited by uncertainties in groundwater potential mapping due to the number of negative samples (randomly selected non-spring locations outside existing spring buffer zones) and modeling method constraints. This study focused on Yiliang County, Yunnan Province, China. Fifteen thematic layers were selected from five categories: topography and geomorphology, hydrography, geology, human activities, and environment. Four datasets with positive-to-negative sample ratios (ratio of known spring locations to non-spring locations) of 1:0.5, 1:1, 1:2, and 1:3 were constructed to explore the impact of negative sample quantity. Two classic machine learning models—support vector machine (SVM) and random forest (RF)—were compared with a novel multi-grained cascade forest (GCF) model. The computationally efficient dung beetle optimizer (DBO) was employed for parameter optimization, forming three hybrid models: SVM-DBO, RF-DBO, and GCF-DBO. Model performance was evaluated using five metrics: raincloud plots, area under the receiver operating characteristic curve (AUC), accuracy (ACC), true positive rate (TPR), and true negative rate (TNR). Results indicate that increasing negative samples reduces prediction errors for non-spring locations but increases errors for spring identification. The 1:1 ratio achieved optimal performance, with all models exceeding 70% accuracy for both spring and non-spring predictions. The GCF-DBO model demonstrated superior performance (ACC = 0.795, AUC = 0.845). These findings highlight the critical role of sample balance and model selection in enhancing groundwater potential predictions. This research provides actionable insights for sustainable groundwater resource management.