Scale-Dependent Gradient Boosting Algorithms for SPEI Drought Prediction

The drought monitoring and forecasting are essential for effective water resources management and decreasing climate risks because of increasing climatic variability. In order to simulate the 12-month Standardized Precipitation Evapotranspiration Index (SPEI-12), this paper evaluates the appropriateness and the comparative performance of gradient boosting-based machine learning models namely; Gradient Boosting Regressor, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost). A rigorous evaluation methodology is adopted to ensure scientific accuracy and applicability of the operation where statistical goodness of fit measures, hydrological efficiency measures, diagnostics of errors, bias measures, test of significance, and accuracy of threshold-based drought classification are all undertaken. According to the results, the learning capacity of all gradient boosting models is high in the course of the training, and R² and NSE values are between 0.98 and 0.99, which suggests that the variability of SPEI-12 is depicted well. LightGBM and CatBoost outperformed the other approaches in both R², NSE, and KGE values and lower RMSE and bias in the testing stage, therefore, the models were the most predictable and applicable. It is interesting to note that LightGBM is generally accurate and efficient, whilst CatBoost is more resistant to outliers, which is demonstrated by lower average relative error. LightGBM is the most superior approach when compared to other model with evaluation metrics (R² of 0.87, NSE of 0.86, KGE of 0.83, and the lowest RMSE of 0.37). Evaluation using the threshold indicates the operational strength of the proposed framework, and all models were highly accurate in detecting moderate and severe situations of drought. In 67.23% of the test cases the model correctly forecasted an event of drought at a tolerance of 10% which rose to 90.64% at a tolerance of 100 percent which is corroborated by the fact that it is a realistic model that can be useful in an operational drought early warning system. Models were most effective under intense drought conditions with a high degree of accuracy of over 90 percent at the 100 percent mark, which means that it is reliably applicable in detecting severe drought conditions that are necessary in emergency response planning. The model performance was strongly validated by means of the rigorous statistical analysis using various statistical metrics which included: R² NSE, KGE, RMSE, P-Bias, and F-statistics. This multimeric method ensured comprehensive evaluation that can be used in operation in different climatic regions. In general, the findings indicate that machine learning models based on the gradient boosting are a valid and useful approach to predict the drought index over the long run. This paper demonstrates the unique advantages of boosting techniques in the long-term drought index (SPEI-12) modelling and the importance of selecting and validating the model with numerous statistical measures. The proposed approach holds tremendous potential in improving risk assessment for drought monitoring.