The paper introduces FloodCast, a large-scale flood modeling and forecasting framework that combines multi-satellite observations and hydrodynamic modeling. The framework aims to address the limitations of traditional hydrodynamic models, which rely on fixed-resolution grids and parameters, leading to high computational costs and inaccuracies in flood predictions. FloodCast is designed to be fast, stable, accurate, resolution-invariant, and geometry-adaptive, making it suitable for large-scale applications. The multi-satellite observation module includes a real-time unsupervised change detection method and a rainfall processing and analysis tool to leverage multi-satellite data for flood prediction. The hydrodynamic modeling module introduces GeoPINS, a geometry-adaptive physics-informed neural solver that combines the advantages of physics-informed neural networks (PINNs) and neural operators. GeoPINS is trained without the need for extensive training data and is designed to handle complex river geometries and different spatial and temporal resolutions. The paper evaluates the performance of FloodCast using the 2022 Pakistan flood as a benchmark dataset. The model is validated in three dimensions—flood inundation range, depth, and spatiotemporal downscaling—using SAR-based flood data, traditional hydrodynamic benchmarks, and optical remote sensing images. The results show that the sequence-to-sequence GeoPINS model outperforms traditional hydrodynamics in terms of prediction accuracy, with smaller prediction errors and higher precision in predicting spatially varying water depths. Key contributions of the paper include the development of FloodCast, the proposal of GeoPINS, and the establishment of a benchmark dataset for flood prediction methods. The effectiveness of the proposed methods is demonstrated through comprehensive evaluations, showing that FloodCast can provide high-precision, large-scale flood modeling with an average MAPE of 14.93% and an average MAE of 0.0610m for 14-day water depth predictions.The paper introduces FloodCast, a large-scale flood modeling and forecasting framework that combines multi-satellite observations and hydrodynamic modeling. The framework aims to address the limitations of traditional hydrodynamic models, which rely on fixed-resolution grids and parameters, leading to high computational costs and inaccuracies in flood predictions. FloodCast is designed to be fast, stable, accurate, resolution-invariant, and geometry-adaptive, making it suitable for large-scale applications. The multi-satellite observation module includes a real-time unsupervised change detection method and a rainfall processing and analysis tool to leverage multi-satellite data for flood prediction. The hydrodynamic modeling module introduces GeoPINS, a geometry-adaptive physics-informed neural solver that combines the advantages of physics-informed neural networks (PINNs) and neural operators. GeoPINS is trained without the need for extensive training data and is designed to handle complex river geometries and different spatial and temporal resolutions. The paper evaluates the performance of FloodCast using the 2022 Pakistan flood as a benchmark dataset. The model is validated in three dimensions—flood inundation range, depth, and spatiotemporal downscaling—using SAR-based flood data, traditional hydrodynamic benchmarks, and optical remote sensing images. The results show that the sequence-to-sequence GeoPINS model outperforms traditional hydrodynamics in terms of prediction accuracy, with smaller prediction errors and higher precision in predicting spatially varying water depths. Key contributions of the paper include the development of FloodCast, the proposal of GeoPINS, and the establishment of a benchmark dataset for flood prediction methods. The effectiveness of the proposed methods is demonstrated through comprehensive evaluations, showing that FloodCast can provide high-precision, large-scale flood modeling with an average MAPE of 14.93% and an average MAE of 0.0610m for 14-day water depth predictions.