Teleconnection‐Informed Machine Learning for Seasonal Streamflow Characterisation in the Vu Gia–Thu Bon River Basin: A SHAP ‐Based Interpretability Framework

Seasonal streamflow characterisation in monsoon‐driven tropical catchments remains a major challenge for regional water resource management, particularly where extreme hydrological regimes are co‐modulated by multiple large‐scale ocean–atmosphere teleconnections. The Vu Gia–Thu Bon River Basin in Central Vietnam exemplifies this challenge: a 41‐year record (1979–2019) reveals a coefficient of variation of 0.95 at the principal gauging station, with pronounced inter‐annual variability shaped by compound ENSO, PDO, and IOD forcing. While teleconnection‐informed machine learning approaches are increasingly adopted for streamflow estimation, most regional studies treat climate indices as independent covariates and provide limited model‐level interpretation of how compound climate information is represented in streamflow estimates. This study develops a regionally tailored, interpretable machine learning framework for the VGTB Basin that addresses these limitations through three sequential contributions. Training‐period lag‐correlation analysis was used to identify candidate lag windows for ENSO, IOD, and PDO, while Morlet wavelet coherence was applied to the full record only as descriptive multi‐scale hydroclimatic context. Three machine learning models (XGBoost, SVR, and ElasticNet) were developed using a 41‐predictor feature matrix comprising lagged rainfall, antecedent discharge, seasonal harmonics, teleconnection indices, and compound ENSO × PDO interaction terms. These models were evaluated on an independent 12‐year test set using leakage‐controlled, temporally blocked cross‐validation. XGBoost achieved NSE = 0.856 at Nong Son; at the regulated Thanh My station, overall NSE was 0.711, comprising 0.873 before the Dak Mi 4 inter‐basin diversion and 0.495 in 2016–2019 after it. Controlled ablation analysis indicated that this performance is attributable almost entirely to local hydrometeorological predictors, as teleconnection indices and their compound interaction terms did not improve out‐of‐sample skill. SHAP compound‐phase analysis and observed‐discharge stratification demonstrated that La Niña × PDO− months were associated with significantly higher wet‐season discharge compared to El Niño × PDO+ months, with a 45.8% increase observed at Nong Son. However, a formal interaction test indicated that ENSO and PDO effects combine in an approximately additive manner at the monthly scale. These findings support the use of teleconnection phases as descriptive hydroclimatic context, although controlled ablation analysis reveals that they do not enhance out‐of‐sample monthly streamflow estimation beyond what is achieved using local predictors.