Scholarlink

<p>Accurate water depth estimation is vital for effective lake management, yet traditional surveys are costly, labour-intensive, and spatially limited. This study estimated the bathymetric of Lake Tana, Ethiopia’s largest lake, by integrating Landsat 8 Level 2 imagery with in-situ measurements. Six models were applied, including five machine learning (ML) algorithms: random forest (RF), support vector machine (SVM), artificial neural network (ANN), gradient boosting machine (GBM), and gaussian process regression (GPR)), along with one multiple linear regression (MLR) model, to capture linear and nonlinear relationships between spectral reflectance and water depth. From 12,410 depth measurements (120&#xa0;m intervals, 5&#xa0;km block spacing), 8,612 were used for training and testing with an 80:20 split. Five spectral predictors were employed: the blue, green, and red bands, along with two logarithmic band ratios (blue/red and green/red), with hyperparameters were optimized through cross-validation. Among the models, RF achieved the highest accuracy, with training root mean square error (RMSE), mean relative error (MRE), and Nash-Sutcliffe Efficiency (NSE) values of 0.29&#xa0;m, 2.95%, and 0.99, and testing values of 0.43&#xa0;m, 4.6%, and 0.97, respectively. Other ML models showed moderate performance, while MLR had the lowest accuracy, with training RMSE, MRE, and NSE values of 0.87&#xa0;m, 9.4% and 0.94 and testing values of 0.89&#xa0;m, 9.6%, and 0.94, respectively. Spatial depth maps revealed model dependent under- and overestimation pattern; RF achieved the highest statistical accuracy but showed lower performance in capturing both shallow and deep zones, whereas models with relatively lower statistical performance, such as MLR better captured these zones. To address these discrepancies and leverage each algorithm’s strengths, a weighted ensemble approach was applied, balancing biases and improving spatial consistency, achieving intermediate performance (RMSE = 0.71&#xa0;m, MRE = 7.73%, and NSE = 0.96). Overall, this integrated approach offers a cost-effective and scalable approach for bathymetric mapping, supporting sediment management, water resource planning, navigation safety, and ecosystem conservation in Lake Tana and similar water bodies.</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

Leveraging remote sensing and machine learning for efficient bathymetric surveying in Lake Tana, Ethiopia

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Samuel Berihun Kassa,
  • Atsushi Tsunekawa,
  • Nigussie Haregeweyn,
  • Mitsuru Tsubo,
  • Ayele Almaw Fenta,
  • Dagnachew Aklog,
  • Mulatu Liyew Berihun,
  • Dagnenet Sultan,
  • Takeshi Abe,
  • Taye Minichil Meshesha,
  • Ashebir Sewale Belay,
  • Demesew A. Mhiret,
  • Yoseph Buta Hailu,
  • Zena Tessema Terefe

摘要

Accurate water depth estimation is vital for effective lake management, yet traditional surveys are costly, labour-intensive, and spatially limited. This study estimated the bathymetric of Lake Tana, Ethiopia’s largest lake, by integrating Landsat 8 Level 2 imagery with in-situ measurements. Six models were applied, including five machine learning (ML) algorithms: random forest (RF), support vector machine (SVM), artificial neural network (ANN), gradient boosting machine (GBM), and gaussian process regression (GPR)), along with one multiple linear regression (MLR) model, to capture linear and nonlinear relationships between spectral reflectance and water depth. From 12,410 depth measurements (120 m intervals, 5 km block spacing), 8,612 were used for training and testing with an 80:20 split. Five spectral predictors were employed: the blue, green, and red bands, along with two logarithmic band ratios (blue/red and green/red), with hyperparameters were optimized through cross-validation. Among the models, RF achieved the highest accuracy, with training root mean square error (RMSE), mean relative error (MRE), and Nash-Sutcliffe Efficiency (NSE) values of 0.29 m, 2.95%, and 0.99, and testing values of 0.43 m, 4.6%, and 0.97, respectively. Other ML models showed moderate performance, while MLR had the lowest accuracy, with training RMSE, MRE, and NSE values of 0.87 m, 9.4% and 0.94 and testing values of 0.89 m, 9.6%, and 0.94, respectively. Spatial depth maps revealed model dependent under- and overestimation pattern; RF achieved the highest statistical accuracy but showed lower performance in capturing both shallow and deep zones, whereas models with relatively lower statistical performance, such as MLR better captured these zones. To address these discrepancies and leverage each algorithm’s strengths, a weighted ensemble approach was applied, balancing biases and improving spatial consistency, achieving intermediate performance (RMSE = 0.71 m, MRE = 7.73%, and NSE = 0.96). Overall, this integrated approach offers a cost-effective and scalable approach for bathymetric mapping, supporting sediment management, water resource planning, navigation safety, and ecosystem conservation in Lake Tana and similar water bodies.