Regional climate variability and rice yields prediction in Bangladesh using machine learning and deep learning models

This study compared conventional regression, machine-learning, and deep-learning models for Aman and Boro rice across the Barind, Coastal, and Riverine regions of Bangladesh using data from 1970 to 2024. Climate records from the Bangladesh Meteorological Department (BMD) and rice-yield data from the Bangladesh Bureau of Statistics (BBS) were matched by location and date; 1287 records were retained for each season before region-specific modeling. Predictors included maximum and minimum temperature, rainfall, relative humidity, and sunshine duration. Nine models were evaluated using cross-validation-based hyperparameter selection and multiple metrics, including MAE, RMSE, MAPE, R 2 , and prediction bias; MAE and RMSE are reported in t ha⁻ 1 . Correlation analysis showed consistently negative associations of temperature with rice production ( r = − 0.13 to − 0.28), while rainfall ( r = 0.03–0.21) and relative humidity ( r = 0.07–0.32) showed positive but region-dependent associations; sunshine relationships were weak. TFT achieved the lowest RMSE in four of six region–season combinations: Coastal Aman (0.22), Barind Boro (0.30), Riverine Aman (0.16), and Riverine Boro (0.31 t ha⁻ 1 ). For Coastal Aman, TFT achieved MAE = 0.20 t ha⁻ 1 and R 2 = 0.958; for Riverine Aman, MAE = 0.14 t ha⁻ 1 and R 2 = 0.773. GBM performed best for Coastal Boro (RMSE = 0.64 t ha⁻ 1 ; R 2 = 0.308), while CatBoost and GBM showed the lowest RMSE for Barind Aman (0.54 t ha⁻ 1 ). The study provides a region- and season-specific comparison of statistical, machine-learning, and deep-learning approaches, supporting targeted forecasting and adaptation priorities for drought-prone Barind, salinity-affected Coastal, and flood-prone Riverine areas.