Application of Machine Learning Models for Short-term Drought Analysis Based on Streamflow Drought Index

This study investigates the drought condition based on streamflow drought index (SDI) using various machine learning (ML) techniques. The ML models include Multiple Linear Regression, Artificial Neural Networks, K-Nearest Neighbors (KNN), and eXtreme Gradient Boosting Regressor (XGBR). The SDI-based drought analysis is conducted at 3, 6, 9, and 12 months at two stations in the Drava River considering six different lead times. Furthermore, the reliability of ML-based estimations is explored. Overall, the obtained results demonstrated that performances of the ML models vary for each case scenario. Moreover, the optimal choice of lead times varies across different SDIs, with 3 for the 3-month SDI, 4 for the 6- and 9-month SDIs, and 6 for the 12-month SDI. It can be concluded that with the increase of SDI month, the optimal lead time also enhances. Furthermore, the reliability analysis reveales that while KNN models tend to overfit, XGBR models provided a proper balance between the training and testing reliability, making it a desirable choice for SDI prediction. Additionally, the confidence percentage (CP) analysis indicated a surge in CP with an increase in the SDI month, demonstrating the significant role of the number of SDI months. Therefore, this study highlights the importance of selecting appropriate lead times, SDI months, and ML models to improve predictive performance and reliability in short-term drought forecasting.