Abstract: Accurate flood forecasting plays a critical role in flood prevention,disaster mitigation,and sustainable water resources management.This study introduces VD-LSTM-Bootstrap,an integrated machine learning framework,to enhance forecasting reliability through two methodological advancements: the integration of flood process vectorization at the LSTM input layer for characterizing runoff variation patterns and the implementation of Bootstrap resampling at the output layer for uncertainty quantification.The model was applied to the Jingle and Lushi River basins in China,utilizing 50 and 20 historical flood events,respectively,divided into training and validation sets at a ratio of 7∶3.The results indicate that the VD-LSTM model achieves superior performance compared to standard LSTM model,with NSE values consistently above 0.8,peak flow relative errors remaining below 15%,and significantly reduced root mean square error and bias metrics.The enhanced hydrograph fitting capability effectively addresses both the underestimation of flood peaks and time lag issues commonly encountered in conventional LSTM approaches.Furthermore,the VD-LSTM-Bootstrap model generates reliable confidence intervals,exhibiting coverage rates of 90.1%,85.5%,and 80.3% in Jingle Basin and 91.7%,86.2%,and 81.6% in Lushi Basin,systematically surpassing their respective theoretical confidence levels.This methodology establishes a novel framework for combining process-based hydrological analysis with data-driven modeling techniques.