Long-Term Drought Analysis and Forecasting Using Hybrid Wavelet Denoise Random Forest Models with SPI, Z-Score, and China Z-Index

Abstract

Accurately forecasting drought events using data-driven models is essential for mitigating the global impacts of climate change. In this study, drought in two districts of Gujarat, namely Aravalli and Gandhinagar, is assessed to characterize dry, normal, and wet conditions using the standardized precipitation index (SPI) with 3, 6, and 12-month lead times (SPI-3, SPI-6, and SPI-12), along with the China Z-index and Z-score using historical long time series data from 1960 to 2019. Furthermore, the prediction of SPI (SPI-3, SPI-6, and SPI-12) events in the Aravalli and Gandhinagar districts has been assessed using machine-learning models, including the random forest (RF) model and hybrid model such as the Wavelet Denoise Random Forest (WDRF) model. The results of the Forecasted models have been analyzed using performance indicators such as RMSE, MAE, R2, and NSE. The study results show that, over the 60-year analysis period, the Aravalli district experienced 28 dry years, with the remaining 32 years being normal or wet. Similarly, Gandhinagar district experienced 26 dry years, with the remaining 34 years being normal or wet. The forecast results of this study indicate that the coupled WDRF model has shown significantly improved prediction accuracy compared to the standard RF model. Additionally, the performance indicators show that the WDRF model attained the highest percentage improvement in prediction accuracy for SPI-3 (3-month SPI), followed by SPI-6 (6-month SPI), and finally SPI-12 (12-month SPI), compared to the RF model in both the Aravalli and Gandhinagar districts. Further comparison of the overall performance SPI-12 shows higher prediction accuracy in both the RF and WDRF models. © 2025 Elsevier B.V., All rights reserved.