Assessing machine learning-based feature selection methods for estimating daily reference evapotranspiration using meteorological data

Mostafa Sadeghzadeh

doi:10.1007/s00704-026-06568-2

Assessing machine learning-based feature selection methods for estimating daily reference evapotranspiration using meteorological data

2026-09-11

0

PRE

AI

DOI：10.1007/s00704-026-06568-2

Original

Original request for help

Save

Abstract

En 中文

This study presents a novel hybrid framework, integrating seven advanced filter-based and embedded feature selection (FS) techniques with an Extreme Gradient Boosting (XGBoost) model, to enhance reference evapotranspiration (ETo) estimation. XGBoost was selected because of its proven efficacy with tabular meteorological datasets, built-in regularization that mitigates overfitting, native handling of non-linear interactions, and computational efficiency on moderate-sized station records. Pearson, minimum redundancy-maximum relevance (mRMR), Relief algorithm version F (ReliefF), SHapley Additive exPlanations (SHAP), least absolute shrinkage and selection operator (LASSO), Fisher, and Stability methods were employed as FS methods. Data from two distinct climatic conditions, namely the arid and humid areas, were used to assess the applied methods. The framework systematically performed a feature-based domain adaptation procedure through evaluating the relevance/redundancy of key meteorological variables, and minimized the input space features, which helps decrease model complexity and computational costs, e.g.R² ≥ 0.980, RMSE ≤ 0.344 mm d⁻¹ and MAE ≤ 0.270 mm d⁻¹ in the arid region and R² ≥ 0.982, RMSE ≤ 0.188 mm d⁻¹ and MAE ≤ 0.142 mm d⁻¹ in the humid region. It enhances the model’s flexibility and provides modeling tools exclusively suitable for data-scarce conditions, when reliable data are limited. Two modeling stages were defined, where the three most important selected features by each technique were used as inputs to XGBoost in the first stage, while the second stage utilized the two most important features as input variables. Bootstrap, effect size, and decision boundary analyses further evaluated the models’ capabilities. ReliefF and the LASSO methods provided the lowest error indices for the arid region using the two most important features. Pearson, SHAP, Fisher, and Stability methods provided the best result for the humid region with the same approach. The results revealed that, when relying on the suitable FS method, the combined XGBoost-FS model presented promising results for ETo estimation for both climatic conditions (R² up to 0.982, RMSE as low as 0.188 mm d⁻¹). The XGBoost models were compared with Random Forest and a feed-forward Artificial Neural Network, which confirmed the superiority of the XGBoost model.

Assessing machine learning-based feature selection methods for estimating daily reference evapotranspiration using meteorological data

Abstract

Journal

Organization

Cited Papers