Hybrid machine learning framework for drought forecasting: integration of LSBoost, RQNN, CNN-RNN, CNN-LSTM and advanced data decomposition

Drought is a natural disaster with serious impacts on agriculture, water resources management and ecosystem balance. Ongoing global warming and climate change exacerbate drought cases. Therefore, proper tools are essential for long-term drought monitoring and efficient drought forecasts. The Palmer Drought Severity Index (PDSI) is a reliable indicator developed to monitor long-term drought trends and is widely used to define agricultural and hydrological droughts. In this study, drought predictions were performed with the use of PDSI data for the period January 1958–December 2023 for Kayseri meteorological station of Türkiye. Four different machine learning models Least squares boosting, Regression Quantile Neural Network (RQNN), Convolutional Neural Network—Recurrent Neural Network (CNN-RNN) and Convolutional Neural Network—Long Short-Term Memory (CNN-LSTM) were combined with Empirical mode decomposition (EMD), Maximal overlap discrete wavelet transform (MODWT), Tunable Q-factor wavelet transform (TQWT), Empirical wavelet transform (EWT) and Variational Mode Decomposition (VMD) data decomposition techniques. Then, PDSI prediction performance of independent and hybrid models was assessed in detail. Data from January 1958 to March 2007 was used in the training phase of the models, while data from April 2007 to December 2023 was used in the testing phase. The t-1, t-2, and t-3 lagged values of the PDSI were set as input variables and the t value was set as the output variable. Among the evaluated models, EWT-RQNN yielded relatively low error values and high values for the considered performance criteria. The TQWT-RQNN and EWT-CNN-LSTM models also attracted attention with their superior performance. Present findings revealed that machine learning models hybridized with appropriate data decomposition techniques significantly improved PDSI prediction performance. Kruskal–Wallis H-test results revealed no significant difference between predicted and observed PDSI values and confirmed that the developed models generally produced successful predictions. It was concluded based on present findings that not only robust model architectures, but also effective pre-processing methods play a critical role in predicting drought-like complex and non-linear processes.