Coastal flooding has increased significantly in recent decades which is further exacerbated by warming and is expected to be more intense in the future. The increasing trend has already affected at least 40 % of the United States population residing in the coastal region. The recent occurrence of compound flooding had a devastating impact and loss of human life. For instance, hurricane Isaac in 2012 caused about $2.35 billion in total property damage with several fatalities. Therefore, monitoring and predicting coastal floods are crucial for mitigation efforts. Along the backwater zone, the flood impact is significantly exacerbated due to the occurrence of multivariate flooding like fluvial flood (FF), storm surge (SS), and Low-frequency surge (LFS). Here, we used a Long Short-Term Memory (LSTM) data assimilation approach to predict fluvial flood, storm surge, and low-frequency surge over backwater zones of Susquehanna and Delaware River basins (SBR and DRB). LSTM trained on a range of high-resolution meteorological conditions shows an accuracy of 0.96, 0.63, and 0.72 Nash–Sutcliffe efficiency respectively, and F1-Score of 0.92,0.93, and 0.80 respectively, in FF, SS, and LFS for multivariate flood prediction. We find a significant improvement in predicting seasonally variable multivariate flood events by integrating upstream datasets for learning in LSTM. Further, we were able to attribute from our LSTM model that flood events are majorly driven by changes in snowmelt runoff, precipitation, and water level. Our findings demonstrate the potential of LSTM for accurately predicting complex multivariate floods in backwater zones.