Utilizing Logistic Regression and Random Forest to Model Meteorological Drought Persistence Across Seasonal Transitions

This study aims to assess the persistence and temporal dynamics of inter-seasonal meteorological drought by evaluating the predictive performance of Random Forest (RF) and Logistic Regression (LR) models. Utilizing the Standardized Precipitation Index (SPI) at a three-month timescale (SPI-3) as a drought indicator, we analyze the temporal evolution of drought conditions across individual meteorological stations. The primary objective is to determine the efficacy of these models in forecasting drought persistence from one season to the next, thereby enhancing the understanding of drought dynamics and improving predictive capabilities for localized drought monitoring. A range of statistical metrics is employed to evaluate model performance, with sensitivity playing a crucial role in detecting persistent drought events. Random Forest (RF) generally outperformed Logistic Regression (LR) in most cases, particularly in terms of sensitivity. RF showed higher sensitivity across multiple stations, including Bahawalnagar (1.0000 vs. 0.9048 for LR), Bahawalpur (0.8125 vs. 0.6875), and Mianwali (0.7857 vs. 0.7143), ensuring that more drought cases were correctly identified. Additionally, RF achieved higher Kappa values and accuracy in most stations, such as Bahawalpur (Kappa: 0.3618 vs. 0.2966 for LR) and Mianwali (0.5245 vs. 0.4056), indicating better overall classification agreement. Given its superior performance in detecting drought persistence, RF emerges as the more effective model. The findings of this study emphasize the importance of using reliable statistical models to analyze drought persistence and its temporal variations at localized scales. RF proves to be the more effective model for detecting drought persistence, primarily due to its higher sensitivity in identifying persistent drought cases. These insights are particularly valuable for policymakers and drought management authorities in enhancing localized drought preparedness strategies. By leveraging RF for accurate drought detection, decision-makers can develop more adaptive and data-driven drought management frameworks. This approach can aid in efficient resource allocation, agricultural planning, and water management, especially in climate-sensitive regions where precise drought monitoring is crucial for mitigation efforts. Strengthening statistical modeling through advanced techniques will further enhance drought risk management and resilience-building strategies at both regional and national levels.