Scholarlink

<p>The classification of Hydrologic Soil Groups (HSG) is a major stepping stone for rainfall runoff modeling process; however, existing HSG systems are not available at local spatial resolution. In this study, we developed a locally consistent, mapped dataset identifying HSG from 29 characteristics by combining remote sensing, field investigation, and advanced deep learning (DL) techniques. The soil category mapping was produced using four DL algorithms, namely, Deep Neural Networks (DNN), Multi-Layer Perceptrons (MLP), Multi-Layer Perceptrons weight, and DeepBoost. From 202 soil inventory locations, hydrologic soil group classes were determined and then randomly divided into ratios of 70/30% for training and validation. Four common statistical modeling methods including overall accuracy, area under the curve, confusion matrix, and Matthew’s correlation coefficient (MCC) were used to evaluate the effectiveness of model’s predictions. Our findings revealed that changing the feature selection and k values significantly impact the models’ performance. We observed that the models progressively performed better with low dimensionality data and smaller k values. The optimal fold number k = 3 provides the best classification accuracy with DeepBoost, MLP, and MLP-weight. The relevant hydraulic conductivity parameter is the most important driving factor for hydrologic soil classification in the study area. The HSG-D corresponding to high runoff potential was the most prevalent, covering approximately the total area, with values of 89.92%, 88.62%, 88.99%, and 100% across all the models. Furthermore, soil texture types such as clay, silt loam, silty clay loam, silty clay, loam, clay loam, sandy clay loam, and sandy clay were the most prevalent textures predicted by the models, which is consistent with field observations. As multiclassification performance measures accuracy, AUC, and MCC, the MLP-weight exhibited the best results, followed by the MLP and DeepBoost, showing their high ability to handle imbalanced data, nonlinear features, and multiclass classification problems.</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

Deep learning-based aided spatial mapping of local scale hydrologic soil groups in the Upper Oum Er-Rbia Basin, Morocco

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月26日

查看全文
  • Elhousna Faouzi,
  • Abdelkrim Arioua,
  • Mohammed Hssaisoune,
  • Ismaguil Hanadé Houmma,
  • Yassine Ait Brahim,
  • Karaoui Ismail,
  • Arfan Arshad,
  • Lhoussaine Bouchaou

摘要

The classification of Hydrologic Soil Groups (HSG) is a major stepping stone for rainfall runoff modeling process; however, existing HSG systems are not available at local spatial resolution. In this study, we developed a locally consistent, mapped dataset identifying HSG from 29 characteristics by combining remote sensing, field investigation, and advanced deep learning (DL) techniques. The soil category mapping was produced using four DL algorithms, namely, Deep Neural Networks (DNN), Multi-Layer Perceptrons (MLP), Multi-Layer Perceptrons weight, and DeepBoost. From 202 soil inventory locations, hydrologic soil group classes were determined and then randomly divided into ratios of 70/30% for training and validation. Four common statistical modeling methods including overall accuracy, area under the curve, confusion matrix, and Matthew’s correlation coefficient (MCC) were used to evaluate the effectiveness of model’s predictions. Our findings revealed that changing the feature selection and k values significantly impact the models’ performance. We observed that the models progressively performed better with low dimensionality data and smaller k values. The optimal fold number k = 3 provides the best classification accuracy with DeepBoost, MLP, and MLP-weight. The relevant hydraulic conductivity parameter is the most important driving factor for hydrologic soil classification in the study area. The HSG-D corresponding to high runoff potential was the most prevalent, covering approximately the total area, with values of 89.92%, 88.62%, 88.99%, and 100% across all the models. Furthermore, soil texture types such as clay, silt loam, silty clay loam, silty clay, loam, clay loam, sandy clay loam, and sandy clay were the most prevalent textures predicted by the models, which is consistent with field observations. As multiclassification performance measures accuracy, AUC, and MCC, the MLP-weight exhibited the best results, followed by the MLP and DeepBoost, showing their high ability to handle imbalanced data, nonlinear features, and multiclass classification problems.