<p>Climate variability is vital for effective climate adaptation and risk management. This study investigates the temperature variations during the Boro season in Bangladesh and evaluates the performance of multiple machine learning models for predicting both maximum and minimum temperatures across diverse ecosystems. In this study, we employed several machines learning models and the model performance was evaluated using metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Squared Error (RMSE). The results revealed that CatBoost model consistently outperformed other models for Barind and Haor and SVM outperformed for Coastal region, achieving the lowest error metrics across both maximum and minimum temperature predictions. However, The Diebold–Mariano test revealed that linear, DT, and KNN models performed similarly but significantly worse than advanced algorithms, while ensemble methods (RF, GBM, XGBoost, CatBoost) showed no significant differences, indicating robust performance; neural models (CNN, LSTM) yielded mixed results, sometimes aligning with ensembles and sometimes differing significantly. Spatial analysis identified high-risk areas with extreme temperature conditions, particularly in regions like Rajshahi, Natore, and Pabna. These findings emphasize the need for region-specific temperature prediction models and targeted climate adaptation strategies in Bangladesh’s diverse ecosystems. The results highlight the importance of localized predictive models and the need for targeted climate adaptation strategies in the face of temperature extremes across Bangladesh’s diverse ecosystems.</p>