Evaluation of Statistical and Machine Learning-Based Bias Correction Techniques for CMIP6 Daily Air Temperature over Bangladesh
Manuscript Under Preparation , 2026
Abstract
Background: The Coupled Model Intercomparison Project Phase 6 (CMIP6) provides standardized climate simulations for assessing historical variability and future climate change. In Bangladesh, reliable temperature simulations are important for climate-sensitive sectors such as health, agriculture, and water resources. However, CMIP6 models contain systematic biases that affect the representation of daily temperatures and extremes, limiting their direct use in regional impact assessments. Bias correction is therefore needed to reduce these discrepancies and improve the representation of observed temperature characteristics.
Objectives: This study aimed to evaluate and compare statistical and machine learning (ML)-based approaches for bias correction of CMIP6 daily air temperature over Bangladesh. Specifically, the study aimed to (i) assess the performance and consistency of ERA5 and ERA5-Land temperature reanalysis datasets against observations and identify an appropriate reference dataset for bias correction; (ii) evaluate the performance of traditional statistical bias correction methods for correcting CMIP6 daily maximum and minimum temperature; (iii)assess the effectiveness of Light Gradient Boosting (LightGBM) as an ML-based bias correction approach; and (iv) compare statistical and ML-based bias correction methods using overall, spatial, seasonal, and percentile-wise evaluation, with particular emphasis on their ability to reproduce temperature extremes.
Data and Methods: First, ERA5 and ERA5-Land daily temperature reanalysis datasets were evaluated against observations from Bangladesh Meteorological Department (BMD) stations to identify the most appropriate gridded reference dataset for bias correction. Based on the comparatively better agreement with station observations, ERA5 was selected as the reference dataset. Subsequently, daily maximum and minimum temperature simulations from four CMIP6 models were bias-corrected using three statistical Quantile Mapping (QM) methods, namely Empirical Quantile Mapping (EQM), Quantile Delta Mapping (QDM), and Scaled Distribution Mapping (SDM), and one ML-based method, LightGBM. The resulting bias-corrected simulations were then evaluated across four dimensions: overall, spatial, seasonal, and percentile-wise, to assess their ability to reduce systematic biases and reproduce the observed characteristics of daily temperature, including temperature extremes.
Results: The results showed substantial errors and systematic biases in the raw CMIP6 simulations, with T_max RMSE ranging from 3.67 to 5.14°C and moderate correlations with ERA5; T_min also showed notable errors and systematic biases. All bias correction methods substantially reduced these discrepancies for both T_max and T_min. LightGBM generally outperformed the traditional QM methods; for example, for INM-CM5-0, it reduced T_max RMSE from 5.14°C to 2.27°C (55.8% reduction) and achieved lower MAE and higher correlation than the QM methods. Spatially, LightGBM produced more uniform error reduction across Bangladesh, with RMSE generally limited to 2.0–2.5°C for T_max and 1.5–2.0°C for T_min, whereas QM retained larger residual errors (3°C–4°C) in the northern regions. Seasonal evaluation showed that QM provided more spatially consistent corrections with residual biases closer to zero, while LightGBM generally achieved lower seasonal RMSE and MAE. LightGBM also performed better in representing temperature extremes. This was especially clear for the highest temperature values. Raw T_max RMSE exceeded 10°C in this range, while QM reduced it to 4–6.5°C and LightGBM generally reduced it to below 4°C.
Conclusion: Overall, the findings indicate that LightGBM provided more effective bias correction than traditional QM methods, particularly for reducing spatial error and representing temperature extremes. These results show the potential of ML-based bias correction to improve the reliability of CMIP6 daily temperature simulations over Bangladesh and support their use in climate impact assessments.
