Evaluation of Loss Functions (Standard, Robust, and Data-Distribution-Based) in the XGBoost Model for Suspended Sediment Simulation and Analysis of the Effect of Logarithmic Transformation on Its Performance (Case Study: Mashin Hydrometry Station, Ro | ||
| مهندسی و مدیریت آبخیز | ||
| Articles in Press, Accepted Manuscript, Available Online from 09 August 2026 | ||
| Document Type: Research Paper | ||
| DOI: 10.22092/ijwmse.2026.372606.2165 | ||
| Author | ||
| Mahmoudreza Tabatabaei* | ||
| Member of scientific boAssistant Professor, Soil Conservation and Watershed Management Research Institute, Agricultural Research, Education and Extension Organization (AREEO), Tehran, Iran | ||
| Abstract | ||
| Extended abstract Introduction Understanding the precise dynamics of suspended sediment in high-erosion rivers, particularly in arid and semi-arid regions such as the Roodzard River basin, is essential for sustainable water resource management and the optimal operation of dam reservoirs. The influx of large sediment loads not only reduces water quality and deteriorates engineering structures but also alters the natural river morphology, entailing significant economic and environmental costs. Given the difficulties of direct measurement methods and the complexities of physical modeling, data-driven approaches especially machine learning algorithms, including boosting methods have emerged as effective and efficient alternatives due to their ability to identify nonlinear patterns and complex hydrological interactions. In the present study, to accurately simulate and predict the suspended sediment load in the Roodzard River (Mashin hydrometric station), the XGBoost model, one of the most powerful decision-tree-based algorithms, was employed. The main innovation of this research is summarized in a dual approach: first, a comprehensive evaluation of a set of loss functions including classical functions (MSE and MAE), robust functions (Pseudo-Huber), and functions based on the statistical distribution of the data (Poisson, Gamma, and Tweedie) to identify the optimal loss function in terms of statistical compatibility with the skewed and heteroscedastic nature of sediment data; and second, a systematic analysis of the effect of logarithmic transformation, as a common preprocessing method, on the performance of the aforementioned models. The ultimate goal of this research is to enhance prediction accuracy and effectively manage outlier data when confronted with the hydroclimatic heterogeneities prevailing in the study area. | ||
| Keywords | ||
| Evaluation metrics; Logarithmic transformation; Loss functions; River modeling; Roodzard River; Suspended sediment; XGBoost | ||
|
Statistics Article View: 69 |
||
Journal management system. designed by sinaweb