Bayesian-optimized Support Vector Regression for Monthly Drought Estimation Using Rainfall and Large-scale Climate Teleconnections

This study developed a Bayesian-optimized support vector regression (SVR) framework for estimating one-month Standardized Precipitation Index (SPI-1) conditions at Kesinga station in the Tel River Basin, Odisha, India. Monthly rainfall for January 1981-December 2025 was obtained from the NASA POWER platform. Large-scale climate covariates comprised the Southern Oscillation Index (SOI), Nin o 3.4 sea-surface-temperature anomaly, Dipole Mode Index (DMI) and Pacific Decadal Oscillation (PDO), while sine and cosine month terms represented annual seasonality. The 540- month record was separated chronologically into 70% training (January 1981-June 2012) and 30% testing (July 2012- December 2025). A conventional radial-basis-function SVR was compared with an Optuna-optimized SVR, for which the kernel and hyperparameters were selected over 100 trials using a tree-structured Parzen estimator and five-fold expanding-window time-series cross-validation. The original SVR produced test RMSE, MAE, R², NSE, correlation, bias and index-of-agreement values of 0.461, 0.347, 0.777, 0.777, 0.884, -0.053 and 0.932, respectively. The optimized SVR improved these values to 0.391, 0.302, 0.840, 0.840, 0.920, -0.056 and 0.958. Optimization reduced RMSE by 15.15% and MAE by 12.93%, while increasing R² and NSE by 8.03%. Permutation analysis showed that rainfall and seasonal harmonics were the dominant predictors, whereas the teleconnection indices provided smaller but non-zero incremental contributions. Because contemporaneous rainfall was used to estimate same-month SPI-1, the framework is interpreted as drought-status estimation and monitoring rather than advance forecasting.