Integrating XGBoost and SHAP to uncover feature contributions for river network selection across different patterns

• Integrating XGBoost and SHAP explains river network selection across varying patterns. • Macro-, meso-, and micro-level analyses reveal feature roles in the selection. • Identifies three feature types: universal, pattern-sensitive, redundant. • Pattern-sensitive features preserve specific drainage patterns. River network selection is essential for generating multiscale river network datasets used in earth observation (EO) applications, such as hydrological analysis and terrain modeling. However, how individual river features influence selection decisions, especially across different drainage patterns, remains unclear. To address this gap, this study introduces an explainable artificial intelligence framework that integrates eXtreme Gradient Boosting (XGBoost) with SHapley Additive exPlanations (SHAP) to uncover feature contributions for river network selection across different patterns. First, river strokes were constructed as analytical units and multiple geometric, topological, and hydrological features were extracted. Then, XGBoost-based selection models were trained for the parallel and rectangular drainage networks. Finally, the macro-, meso-, and micro-level SHAP analysis was conducted over the trained selection models. The experimental results reveal that features fall into three categories for the two types of patterns considered here: universal features (represented by upstream cumulative area, Horton code) that dominate selection regardless of pattern; pattern-sensitive features (represented by confluence angle, river proximity distance) whose influence varies with drainage patterns; and low-contribution features with negligible contribution. These findings explain why certain rivers are retained or removed across different drainage patterns, providing explainable insights to support automated, pattern-preserving generation of multiscale river networks.