Scholarlink

<p>In rockfall susceptibility mapping, the reliability of predictive models depends heavily on the quality and structure of input data, particularly the rockfall inventory that forms the foundation of any analysis. Despite its central role, one critical aspect is often overlooked: the geometric representation of this inventory. Many studies apply different formats, points, polylines, or polygons, without questioning their methodological appropriateness or impact on model performance. This study addresses this issue by assessing how inventory geometry influences model accuracy, interpretability, and efficiency. Using the multilayer perceptron (MLP) technique, we compare three inventory types on a geologically complex limestone ridge in northern Morocco. Model evaluations rely on standard indicators, such as AUC, confusion matrices, and model-reality overlays, ensuring a robust assessment of predictive capabilities while accounting for geological and geomorphological processes in rockfall source areas. Results show that the spatial representation of rockfall events has a direct and substantial impact on susceptibility mapping outcomes. The point-based model requires a more complex neural structure (14 nodes) and yields modest performance (AUC = 0.718, recall = 0.71, F1 = 0.83), tending to overgeneralize susceptibility even in unsuitable areas. It emphasizes structural factors but overlooks environmental dynamics such as erosion. In contrast, models based on polyline and polygon inventories provide higher accuracy and spatial coherence. The polyline model better reflects hydro-morphological processes, while the polygon inventory offers the most integrated approach, capturing both triggering factors and long-term accumulation of displaced material. With simpler neural networks (8 nodes, matching the theoretical optimum for this dataset) and superior metrics (AUC = 0.932, recall = 0.97, F1 = 0.98), the polygon-based model proves the most effective. This study demonstrates that inventory geometry is not merely a technical choice but a strategic parameter that shapes model behavior and the quality of risk assessment. It offers a methodological contribution to improve susceptibility mapping and supports more accurate, reliable decisions in natural hazard management and land-use planning.</p> Graphical Abstract <p></p> <p>This work presents an in-depth comparative analysis of Rockfall Susceptibility Mapping (RSM) using the Multi-Layer Perceptron (MLP) model. The study addresses a critical and often overlooked methodological question: the impact of the geometric representation of the rockfall inventory (point, polyline, and polygon) on the predictive model’s performance and interpretability. The analysis is conducted at the level of a geologically complex area in northern Morocco, namely the Dorsale calcaire. The data used in this analysis include topographic and geological maps, satellite images, and a high-resolution digital terrain model (DEM). Several conditioning factors (slope, curvature, aspect, elevation, lithology, and others) have been integrated as explanatory variables. Each inventory geometry datum is divided into two parts: training (70%) and testing (30%), with the same spatial distribution. The latter are subjected to the multilayer neural network model via specific neural architectures. The results obtained highlight the differential contribution of variables according to the geometry used during the analysis and in the various resulting susceptibility maps. In this study, the polygon appears to be the most reliable geometry for modeling rockfall susceptibility, while the point performs the least well. This confirms that the geometry of the rockfall inventory has a decisive impact on the quality and reliability of Rockfall Susceptibility Models (RSM), leading to differential variable contributions and distinct optimal network configurations. Far from being a neutral technical choice, the inventory’s representation format profoundly influences the ranking of predisposing factors and the model’s ability to reflect geomorphological reality. This conclusion is fundamental to attempts to standardize methodologies for assessing rockfall susceptibility mapping or to choose the appropriate inventory geometry according to the phenomenon studied, the data used, and the method employed.</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

Unlocking Accuracy: Inventory Geometry as a Missing Link in Machine Learning-Based Rockfall Susceptibility Mapping (Case Study: Calcareous Dorsal, Rif, Morocco)

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月8日

查看全文
  • Youssef El Miloudi,
  • Younes El Kharim,
  • Rachid El Hamdouni

摘要

In rockfall susceptibility mapping, the reliability of predictive models depends heavily on the quality and structure of input data, particularly the rockfall inventory that forms the foundation of any analysis. Despite its central role, one critical aspect is often overlooked: the geometric representation of this inventory. Many studies apply different formats, points, polylines, or polygons, without questioning their methodological appropriateness or impact on model performance. This study addresses this issue by assessing how inventory geometry influences model accuracy, interpretability, and efficiency. Using the multilayer perceptron (MLP) technique, we compare three inventory types on a geologically complex limestone ridge in northern Morocco. Model evaluations rely on standard indicators, such as AUC, confusion matrices, and model-reality overlays, ensuring a robust assessment of predictive capabilities while accounting for geological and geomorphological processes in rockfall source areas. Results show that the spatial representation of rockfall events has a direct and substantial impact on susceptibility mapping outcomes. The point-based model requires a more complex neural structure (14 nodes) and yields modest performance (AUC = 0.718, recall = 0.71, F1 = 0.83), tending to overgeneralize susceptibility even in unsuitable areas. It emphasizes structural factors but overlooks environmental dynamics such as erosion. In contrast, models based on polyline and polygon inventories provide higher accuracy and spatial coherence. The polyline model better reflects hydro-morphological processes, while the polygon inventory offers the most integrated approach, capturing both triggering factors and long-term accumulation of displaced material. With simpler neural networks (8 nodes, matching the theoretical optimum for this dataset) and superior metrics (AUC = 0.932, recall = 0.97, F1 = 0.98), the polygon-based model proves the most effective. This study demonstrates that inventory geometry is not merely a technical choice but a strategic parameter that shapes model behavior and the quality of risk assessment. It offers a methodological contribution to improve susceptibility mapping and supports more accurate, reliable decisions in natural hazard management and land-use planning.

Graphical Abstract

This work presents an in-depth comparative analysis of Rockfall Susceptibility Mapping (RSM) using the Multi-Layer Perceptron (MLP) model. The study addresses a critical and often overlooked methodological question: the impact of the geometric representation of the rockfall inventory (point, polyline, and polygon) on the predictive model’s performance and interpretability. The analysis is conducted at the level of a geologically complex area in northern Morocco, namely the Dorsale calcaire. The data used in this analysis include topographic and geological maps, satellite images, and a high-resolution digital terrain model (DEM). Several conditioning factors (slope, curvature, aspect, elevation, lithology, and others) have been integrated as explanatory variables. Each inventory geometry datum is divided into two parts: training (70%) and testing (30%), with the same spatial distribution. The latter are subjected to the multilayer neural network model via specific neural architectures. The results obtained highlight the differential contribution of variables according to the geometry used during the analysis and in the various resulting susceptibility maps. In this study, the polygon appears to be the most reliable geometry for modeling rockfall susceptibility, while the point performs the least well. This confirms that the geometry of the rockfall inventory has a decisive impact on the quality and reliability of Rockfall Susceptibility Models (RSM), leading to differential variable contributions and distinct optimal network configurations. Far from being a neutral technical choice, the inventory’s representation format profoundly influences the ranking of predisposing factors and the model’s ability to reflect geomorphological reality. This conclusion is fundamental to attempts to standardize methodologies for assessing rockfall susceptibility mapping or to choose the appropriate inventory geometry according to the phenomenon studied, the data used, and the method employed.