Global-scale streamflow simulation framework used in the paper of "Abrupt surface water decline during 2023–2024 record warming"

Summary: This repository contains the data and source code for a global-scale streamflow simulation framework utilizing an ensemble of machine learning models. The methodology addresses the challenge of hydrological variability by implementing a climate-zone-specific training approach. Results from the Step 3 script can be extrapolated globally to simulate monthly streamflow across all HydroSHEDS basins in the subsequent step. Key technical features: Specialized ensembles: The framework employs five machine learning architectures: Random Forest, Gradient Boosting, SVM, Neural Networks, and Gaussian Process Regression. Climate segmentation: Basins are categorized using an Aridity Index (AI) threshold of 0.65. Asymmetric weighting: To mitigate common overestimation issues in arid hydrology, dry-region models are trained with a 2x penalty for overestimation errors. Bayesian Model Averaging (BMA): Individual model outputs are integrated using climate-zone-specific BMA weights to produce a robust final simulation. Dataset contents: Input data: Compiled monthly training features, observed monthly streamflow for 150 basins, optimized hyperparameters, and geospatial shapefiles. Output data: Fully trained model objects, calibrated annual time series, and performance evaluation visualizations. Source code: A comprehensive MATLAB script for data imputation, model training, and metric calculation (NSE, KGE, PBIAS) (see step3_ML_DL_fulltraining_inputs_outputs_rep2.m). Temporal Coverage: 1991 – 2024.