Scholarlink

<p>Accurate bathymetric mapping is crucial for marine navigation, ecosystem conservation, and offshore development, yet over 80% of the ocean floor remains unmapped at high resolution. Traditional methods like multibeam sonar are costly and time-consuming, while satellite-derived bathymetry struggles in deep or turbid waters. Machine learning (ML) offers a promising alternative, but existing approaches often fail to account for spatial autocorrelation, suffer from poor generalization, and lack systematic evaluation across diverse marine environments. This study presents a spatially aware ML framework for bathymetric prediction, addressing key limitations through: (1) geographically coherent dataset splitting to preserve spatial continuity, (2) adaptive normalization (Z-score vs. Robust scaling) tailored to seabed complexity, and (3) feature engineering incorporating spatial coordinates and temporal trends (lagged depths, rolling statistics). We evaluate the framework across six Australian transects spanning shallow shelves, slopes, and abyssal plains, using Random Forest (RF). Results demonstrate that sequential splitting reduces spatial overfitting, with test errors increasing by &lt; 15% compared to training. Z-score scaling excels in Gaussian-like terrains (e.g., Sydney Offshore: Root Mean Square Error (RMSE) = 4.2&#xa0;m, R<sup>2</sup> = 0.986), while Robust scaling outperforms in outlier-prone regions (e.g., Perth-Melbourne: 23% lower Mean Absolute Error (MAE)). Spatial coordinates dominate feature importance (58–72%), though temporal features enhance predictions in gradual slopes. Errors concentrate in steep gradients (&gt; 15°) and ultra-deep zones (&gt; 2000&#xa0;m), highlighting challenges in complex terrains. Our framework advances scalable, interpretable bathymetric modeling, aligning with global initiatives like the Seabed 2030 Project. This work establishes a foundation for high-resolution, spatially coherent ocean floor mapping in the era of big data.</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

Explainable machine learning for bathymetric mapping: adaptive normalization and feature engineering in complex seabed terrains

  • 期刊论文

  • 发表日期：

    2025年11月13日

查看全文
  • Ahmet Durap

摘要

Accurate bathymetric mapping is crucial for marine navigation, ecosystem conservation, and offshore development, yet over 80% of the ocean floor remains unmapped at high resolution. Traditional methods like multibeam sonar are costly and time-consuming, while satellite-derived bathymetry struggles in deep or turbid waters. Machine learning (ML) offers a promising alternative, but existing approaches often fail to account for spatial autocorrelation, suffer from poor generalization, and lack systematic evaluation across diverse marine environments. This study presents a spatially aware ML framework for bathymetric prediction, addressing key limitations through: (1) geographically coherent dataset splitting to preserve spatial continuity, (2) adaptive normalization (Z-score vs. Robust scaling) tailored to seabed complexity, and (3) feature engineering incorporating spatial coordinates and temporal trends (lagged depths, rolling statistics). We evaluate the framework across six Australian transects spanning shallow shelves, slopes, and abyssal plains, using Random Forest (RF). Results demonstrate that sequential splitting reduces spatial overfitting, with test errors increasing by < 15% compared to training. Z-score scaling excels in Gaussian-like terrains (e.g., Sydney Offshore: Root Mean Square Error (RMSE) = 4.2 m, R2 = 0.986), while Robust scaling outperforms in outlier-prone regions (e.g., Perth-Melbourne: 23% lower Mean Absolute Error (MAE)). Spatial coordinates dominate feature importance (58–72%), though temporal features enhance predictions in gradual slopes. Errors concentrate in steep gradients (> 15°) and ultra-deep zones (> 2000 m), highlighting challenges in complex terrains. Our framework advances scalable, interpretable bathymetric modeling, aligning with global initiatives like the Seabed 2030 Project. This work establishes a foundation for high-resolution, spatially coherent ocean floor mapping in the era of big data.