Understanding the relationship between large-scale atmospheric circulation patterns, climate memory, and drought variability is critical for improving drought forecasting. This study aims to predict drought conditions in Khuzestan Province by integrating climate teleconnection indices and machine learning methods—specifically the Random Forest (RF) algorithm—while evaluating the impact of climate memory across various timescales.

Monthly precipitation and air temperature data from 13 synoptic meteorological stations across Khuzestan Province were obtained from the Iran Meteorological Organization for the period 1982 to 2024. These data were used to calculate the Standardized Precipitation Evapotranspiration Index (SPEI) at four timescales (1, 3, 6, and 12-month periods) to represent short-term to long-term drought variability. To investigate the role of large-scale climate drivers, 37 climate variables—comprising teleconnection indices and synoptic ocean-atmosphere components—were incorporated as predictors into the Random Forest (RF) model, employing two time-lag scenarios: Lag1-3 and Lag1-6. Before implementing the RF model, a feature selection process was conducted to identify the most suitable input variables. The Variance Inflation Factor (VIF) was used to mitigate multicollinearity among predictors, ensuring that only the most influential and independent variables were retained. The dataset was then partitioned into two subsets: 80% for model training and 20% for testing and validation. Model performance was evaluated using the coefficient of determination (R2), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE).

The results demonstrated that prediction accuracy improves significantly with increasing SPEI timescales. The evaluation metrics (R², MAE, and RMSE) consistently confirmed that the best performance was achieved at the medium- and long-term drought scales (SPEI-6 and SPEI-12), with an R² of 0.426. A comparative analysis of two input lag scenarios (Lag1-3 and Lag1-6) revealed that while performance differences were negligible at short timescales, the Lag1-6 scenario exhibited higher predictive accuracy for medium- and long-term droughts, underscoring the significance of climate memory exceeding three months in the region. Furthermore, the RF model successfully reproduced the dominant spatial patterns of drought across Khuzestan Province with reasonable accuracy. Spatial distribution maps, derived from observed and predicted SPEI values using the Inverse Distance Weighting (IDW) interpolation method, indicate that the northern and northeastern parts of the province (including Safiabad, Dezfol, and Izeh) generally experience relatively more humid conditions. In contrast, the central and southwestern regions—particularly the Ahvaz–Bostan–Abadan axis—are characterized by lower SPEI values and more severe drought conditions. However, the RF model tended to dampen the range of variability during extreme events, leading to a slight underestimation of very severe drought periods. Additionally, the findings revealed that drought drivers vary depending on the timescale. At shorter timescales, atmospheric dynamical fluctuations, such as wind speed anomalies, play a more significant role. At longer timescales, large-scale teleconnection indices, specifically the El Niño–Southern Oscillation (ENSO) and DMI, become the dominant factors explaining drought fluctuations. Overall, this study demonstrates that integrating large-scale climate drivers, climate memory, and advanced machine learning provides an efficient framework for drought forecasting in Khuzestan Province, supporting improved water resource management and climate risk mitigation in southwestern Iran.