بازسازی داده‌های گمشده اقلیمی با استفاده از ترکیب روش بازسازی چندگانه با معادلات زنجیره‌ای (MICE) و مدل‌های تقویتی یادگیری ماشین در حوضه آبریز دریاچه ارومیه

نوع مقاله : مقاله پژوهشی

نویسندگان

• محمد شایان‌نژاد
• محمد جمالی
• سعید اسلامیان

گروه علوم و مهندسی آب، دانشکده کشاورزی، دانشگاه صنعتی اصفهان، اصفهان، ایران

چکیده

کلیدواژه‌ها

• الگوریتم MICE
• بازسازی داده‌های گمشده
• حوضه آبریز دریاچه ارومیه
• متغیرهای اقلیمی
• یادگیری تقویتی

موضوعات

• هواشناسی

عنوان مقاله [English]

Reconstruction of Missing Climatic Data Using Combination of Multiple Imputation by Chained Equations (MICE) and Boosting-Based Machine Learning Approaches in the Urmia Lake Basin

نویسندگان [English]

• Mohammad Shayannejad
• Mohammad Jamali
• Saeid Eslamian

کلیدواژه‌ها [English]

• Boosting learning
• Climatic variables
• MICE algorithm
• Missing data reconstruction
• Urmia Lake Basin

Introduction

Accurate and continuous climate data are essential for climatological analysis, hydrological modeling, and water resources management, particularly in regions facing rapid climatic fluctuations. However, meteorological records often contain missing values due to sensor malfunctions, network interruptions, and human errors. If these missing data are not properly reconstructed, subsequent analyses such as drought assessment, evapotranspiration modeling, and climate change projections may be biased and unreliable (Afrifa-Yamoah et al., 2020; Hersbach et al., 2020).

Traditional gap-filling techniques such as linear regression, spatial interpolation, and principal component analysis have been widely used, yet they often fail to accurately capture nonlinear and interdependent relationships among climatic variables (Matinzadeh et al., 2013). In recent years, machine learning methods—especially boosting-based algorithms such as Gradient Boosting (GBR), Extreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM)—have shown superior performance in modeling complex, nonlinear datasets (Alejo-Sanchez et al., 2025).

The present study aims to evaluate and compare the performance of four gap-filling models—MICE, MICE–GBR, MICE–XGB, and MICE–LGBM—for reconstructing missing climate data across six meteorological stations in the Urmia Lake Basin, northwestern Iran. By integrating the multivariate iterative chained equations (MICE) approach with ensemble boosting algorithms, this research investigates whether hybrid learning frameworks can significantly improve reconstruction accuracy and reduce systematic bias in climatic datasets.

Method

The study area includes six representative synoptic stations—Tabriz, Bonab, Urmia, Maragheh, Saqez, and Sarab—with elevations ranging from 1315 to 1682 m. Daily time series of key climatic variables, including temperature (Tmean, Tmin, Tmax), relative humidity (RH), sea-level pressure (SLP), vapor pressure (SVP), cloudiness (CLD), and evapotranspiration (ET), were analyzed. Four gap-filling approaches were developed and implemented in Python using the scikit-learn and XGBoost/LightGBM libraries:

MICE (baseline model): a chained regression-based multiple imputation method;

MICE–GBR: MICE integrated with Gradient Boosting Regression;

MICE–XGB: MICE coupled with Extreme Gradient Boosting;

MICE–LGBM: MICE combined with LightGBM.

The models were evaluated using four widely adopted statistical indices:

Coefficient of Determination (R²) for accuracy and explained variance;

Normalized Root Mean Square Error (NRMSE) for relative reconstruction error;

Percent Bias (|PBIAS|) for assessing systematic deviation;

Kling–Gupta Efficiency (KGE) for combined evaluation of bias, correlation, and variability.

In addition, the computational runtime of each model was recorded to assess efficiency and potential trade-offs between accuracy and speed.

Results

The comparative analysis demonstrated that the hybrid boosting-based models substantially outperformed the basic MICE model in reconstructing missing climate data. The MICE–XGB model achieved the highest accuracy across most stations and variables, with mean values of R² > 0.90 and KGE > 0.92. The MICE–LGBM model followed closely, offering comparable performance with slightly lower computational cost. The lowest reconstruction errors were obtained for temperature and pressure-related variables, which exhibit smoother temporal patterns and stronger inter-variable correlations. In contrast, cloudiness (CLD) and evapotranspiration (ET) showed higher error values due to their nonlinear and discontinuous nature. These findings align with Li et al. (2021) and Badrzadeh et al. (2022), who reported similar challenges in reconstructing cloud and radiation data.

Across all models and stations, the absolute percent bias (|PBIAS|) remained below 0.025%, confirming the absence of systematic bias and the robustness of the reconstruction framework. Moreover, spatial evaluation revealed consistent model performance across stations with varying topography and elevation, highlighting the generalizability of the hybrid MICE–boosting methods.

Regarding computational efficiency, the MICE–XGB model, despite being slightly slower than MICE and MICE–LGBM, achieved a favorable balance between accuracy and runtime. Its runtime was less than half that of standard deep learning approaches reported in similar studies, indicating the computational practicality of boosting-based models for large-scale climatic applications.

Conclusions

This study demonstrated that integrating the MICE framework with boosting algorithms such as XGB, GBR, and LGBM significantly enhances the accuracy and reliability of climate data reconstruction. Among the evaluated models, MICE–XGB provided the most consistent and accurate results, particularly for temperature and pressure variables. The extremely low |PBIAS| and high KGE values indicate excellent agreement between reconstructed and observed data, confirming the suitability of these hybrid methods for climatological and hydrological modeling. From a practical perspective, the findings highlight the potential of ensemble-based machine learning approaches in addressing missing data challenges in meteorological datasets—especially in basins such as Urmia Lake, where data discontinuity poses serious limitations for environmental analysis. The balance between computational efficiency and predictive precision makes these hybrid models ideal candidates for operational climate monitoring systems and regional reanalysis datasets. Future work should focus on extending this framework to spatiotemporal imputation of gridded datasets, integrating remote sensing and reanalysis data, and exploring deep hybrid networks (e.g., MICE–CatBoost or MICE–LSTM) for improved temporal pattern reconstruction.

Funding

The authors received no specific funding for this work.

Authorship contribution

Conceptualization, M.J. and M.Sh.; methodology, M.J. and M.Sh.; software, M.Sh.; validation, M.J., M.Sh. and S.E.; formal analysis, M.J.; investigation, M.J. and M.Sh.; resources, M.J. and S.E.; data curation, M.J.; writing—original draft preparation, M.J.; writing—review and editing, M.Sh. and S.E. All authors have read and agreed to the published version of the manuscript.

Declaration of Generative AI and AI-assisted technologies in the writing process

The authors declare that no generative AI or AI-assisted technologies were used in the writing, analysis, or preparation of this manuscript. The authors take full responsibility for the content of this publication.

Data availability statement

Data available on request from the authors

Acknowledgements

We acknowledge the Iran Meteorological Organization for providing the historical data. The authors thank the anonymous reviewers for their valuable comments and suggestions.

Ethical considerations

The authors avoided data fabrication, falsification, plagiarism, and misconduct.

Conflict of interest

The authors declare no conflict of interest.