Improving the Indian Monsoon Data Assimilation and Analysis regional reanalysis maximum and minimum temperatures over India using <scp>machine‐learning</scp> techniques

Abstract To complement forecasts from numerical weather prediction (NWP) models and to better understand historical weather patterns, reanalysis datasets like IMDAA (Indian Monsoon Data Assimilation and Analysis) are widely used. However, these datasets carry inherent biases from the underlying NWP models. Bias correction techniques can significantly improve their reliability. This study applies multivariate bias correction to location‐specific IMDAA 2‐m maximum ( T max ) and 2‐m minimum ( T min ) temperatures using five machine‐learning (ML) methods: Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGB) and Convolutional Neural Network (CNN). Overfitting analysis indicates that RF, MLR and CNN show strong generalization, while SVM shows strong overfitting at specific stations and XGB demonstrates moderate overfitting at many locations. Accuracy analysis of the bias correction models shows that mean error (ME) and root‐mean‐squared error (RMSE) are significantly reduced for all stations. XGB is optimal for T max , maintaining low bias and variability, while RF shows the best performance for T min , especially in coastal and peninsular regions. Categorical verification is carried out using the Equitable Threat Score (ETS), Heidke Skill Score (HSS), and Peirce Skill Score (PSS) for T max (&gt;35, 38 and 40°C) and T min (&gt;28, 30, and 35°C), which confirms statistically significant improvements (95% confidence intervals, CI) for lower T max thresholds (&gt;35°C). Maximum improvement is seen in T min where the scores change from negative to positive for all thresholds and regions. Finally, Relative Economic Value (REV) is computed which demonstrates that bias correction increases forecast value, particularly at lower thresholds. Results indicate that bias correction significantly enhances skill of the reanalysis dataset, with T max showing 20%–50% improvement (˜80%–90% along coasts) and T min improving by 70%–90%. However, the magnitude of improvement varies by region and threshold, therefore a region‐specific bias correction strategy is recommended.