A Hybrid Deep Learning Method Based on Savitzky–Golay Filter Using the Standardized Precipitation Index for Drought Prediction

Drought is one of the significant phenomena arising from the variability and climate change in recent decades. In this study, trend dynamics in monthly total rainfall time series measured at the uMkhanyakude district were analyzed using the Mann–Kendall (MK) test, Modified Mann–Kendall (MMK) test, and innovative trend analysis (ITA). Forecasting drought is crucial for preparing and implementing a mitigation plan. This study utilized a hybrid predictive model that combined the temporal convolutional network (TCN) and long short-term memory (LSTM) with a Savitzky-Golay filter (i.e., SG-TCN-LSTM) to forecast drought using the Standard Precipitation Index (SPI). Daily rainfall data from six stations at the uMkhanyakude district in KwaZulu-Natal province, South Africa, from 1980 to 2023 were used in this study. SPI data of 6-, 9-, and 12-month periods were then calculated using the rainfall data. In terms of trend analysis of the monthly total rainfall, the MK and MMK tests detected at Mkuze Game Reserve, Pongolapoort Dam, Hlabisa Mbazwana, False Bay Park, and Ingwavuma Manguzi a statistically significant decreasing trend with negative z-scores of -2.634 & -3.005; -3.392 & -3.925; -3.068 & -3.846; -6.749 & -6.088; and -1.979 & -2.268, respectively, while Riverview shows a statistically significant increasing trend in rainfall with positive z-scores of 2.295 and 4.657. The ITA results are consistent with the MK and MMK results. The SPI forecasting results show that the SG-TCN-LSTM had the highest prediction accuracy of the models at all SPI timescales. For example, at Riverview station, the Root Mean Square Error (RMSE) values of the SG-TCN-LSTM hybrid model are 0.218, 0.102, and 0.049 for SPI-6, SPI-9, and SPI-12, respectively. The R² for this hybrid model is 0.950, 0.979, and 0.992 for SPI-6, SPI-9, and SPI-12, respectively. These results highlight not only the spatial variability in rainfall trends across the district but also demonstrate the potential efficacy of the hybrid approach as a reliable and accurate tool for drought prediction. The incorporation of advanced deep learning techniques with signal smoothing enhances the stability of forecasts, thereby offering valuable insights for proactive drought risk management and the development of targeted adaptation strategies in regions susceptible to drought.