Abstract: The convergence time of small and medium-sized watersheds in mountainous areas is short, and the monitoring station network is sparse. Sudden flash floods often cannot be warned in time, and the problem of underreporting is prominent. Taking the Guanshan River Basin in Shiyan City, Hubei Province as an example, a flash flood warning method based on the Entity Sensing Long Short Term Memory Network (EA-LSTM) was constructed by integrating hourly ground observations from 2010 to 2025 with the Global Precipitation Monitoring Program (GPM), the High Resolution Land Reanalysis Dataset (ERA5-Land), and the Medium Resolution Imaging Spectroradiometer (MODIS) multi-source remote sensing data. Unlike most studies, this study uses a persistent baseline Q (t+k) =Q (t) as a mandatory control and event level warning indicators as the core evaluation dimension. In an independent test set (January 2022 to January 2025, 26 343 hours, 23 flood events, 8 sub robustness experiments), this method showed an improvement of 0.072 in critical success index (CSI) compared to the sustainability baseline during the (t+6) hour foresight period, a relative decrease of 38.4% in false alarm rate (FAR), and an increase of 3.4 percentage points in hit rate (POD); The median of the Continuity Skills Score (PSS) for the entire foreseeable period is negative, indicating that this method is not superior to the Continuity Baseline in terms of continuous fitting metrics, and its value is reflected in the event level alert dimension of business attention. The ablation experiment showed that time convolutional encoding mainly compresses the variance of the long foresight Nash Sacliffe Efficiency Coefficient (NSE), while EA gating raises the median of the long foresight NSE to a positive value. The research results indicate that in high autocorrelation traffic sequences, NSE is not the most suitable representative indicator for flash flood warning capability, and business evaluation should mainly focus on CSI, POD, and FAR.