Study region: The study focuses on the Tarim River Basin in the hyper arid regions, where flooding events have become increasingly frequent and severe due to climate change. The desert-oasis transition zones are highly vulnerable because of their limited vegetation cover, low soil permeability, and strong hydrological variability, which together complicate flood monitoring and management. Study focus: Comprehensive flood simulations that simultaneously capture flood extent and streamflow dynamics at high spatiotemporal resolution remain scarce in arid environments. To address this gap, we propose an interpretable deep learning framework for full-process flood modeling. The framework integrates daily 30 m Seamless Data Cube (SDC) remote sensing data with deep learning-based hydrological models. A U-shaped network (UNet) is used to extract daily flood extents, while hybrid and pure deep learning models simulate daily streamflow under data-scarce conditions. The integration of these models enables a consistent representation of flood processes from surface inundation to river discharge. New hydrological insights for the region: Validation with Landsat imagery confirms that SDC-derived flood maps achieve an average bias below 5 %, while the streamflow simulations produce median Kling-Gupta Efficiency (KGE) and Nash-Sutcliffe Efficiency (NSE) values exceeding 0.8. The proposed framework successfully captures both the spatial and temporal dynamics of floods in arid regions. Furthermore, interpretability analysis reveals that accelerated snowmelt is the dominant driver of recent flood events. This study demonstrates a transferable and data-efficient approach for improving flood modeling and monitoring across arid regions worldwide.