Multi-Output Random Forest Model for Spatial Drought Prediction

Abstract

In regions with limited meteorological monitoring systems, spatial drought modeling is of importance for efficient water resource management. This study recommends an alternative drought modeling strategy for Standardized Precipitation Evapotranspiration Index (SPEI) prediction at multiple target stations using data from neighboring stations. The Multi-Output Random Forest (MORF) model is implemented in this study to consider the spatial correlations among stations for the simultaneous prediction of SPEI for multiple stations instead of training independent models for each station. The efficiency of MORF is further compared to Multi-Output Support Vector Regression (MOSVR) and three baselines; a single-output RF, a monthly climatology model, and a persistence model. In addition to statistical performance criteria, drought characteristics are evaluated using intensity-duration-frequency analysis for three temporal scales (SPEI-3, SPEI-6, and SPEI-12). Results demonstrate that MORF outperformed MOSVR and RF in approximating observed drought intensity, duration, and frequency under moderate, severe, and extreme drought scenarios. Furthermore, spatial analysis reveals that MORF accurately captured the seasonal evolution of drought conditions including onset and recovery phases. The remarkable success of MORF in contrast to MOSVR and three traditional baselines can be explained by its ability to detect nonlinear and complex interactions of drought condition among various neighboring stations. This study emphasizes the promise of multi-output machine learning algorithms for drought monitoring in water resource management and climate adaptation planning in data-scarce regions.