Flooding has been the world most frequent natural disaster over the past decades. Near realtime flood mapping is critical for mitigating the devastating impacts of flooding (e.g., life/economic losses). The United Nations (UN) Sustainable Development Goal 11 (2015-2030) is calling for global efforts to monitor large-scale floods in near realtime for protecting the vulnerable communities. Remote sensing (RS) has the unique capability of earth observation with applications in global disaster response. Recent advances in micro-satellite constellations are offering opportunities for daily satellite image acquisition with a high spatial resolution. Such a high temporal resolution (i.e., daily) is the key to the success of RS for near realtime global flood monitoring. Moreover, the large volume of RS imagery further unleashes the vast potential of data-driven machine learning (ML) approaches for accurate flood mapping. Recent developments in deep learning, a sub-field of ML, have demonstrated the superiority of ML in flood mapping given massive RS data and corresponding labels. Unfortunately, training labels are very scarce in the domain of disaster response. This dissertation explores ML models for flood mapping without massive training labels by introducing the prior knowledge in RS data including the spatial, temporal, and spectral information: First, this dissertation proposes the PSNet to model flood mapping as bitemporal change detection by leveraging the temporal information of RS data. With only a few hand labels, the model can map floods with an over 0.85 F1 score for various heterogeneous flooded areas. Second, to further accelerate flood mapping, a self-supervised bitemporal change detection model, Patch-CVA, is developed to detect changed areas without hand labels. Further post-processing removes most false alarms by introducing the spatial and spectral information of floodwaters. Third, while the Patch-CVA has achieved good performance, its underlying assumption (i.e., most changes are flooded areas) may not hold in certain environments (e.g., crop growth). This dissertation proposes a weak label generation framework by fusing the spatial, temporal, and spectral information of geospatial data. A large weakly labeled dataset is obtained to train data-driven ML models for pixel-wise flood mapping with an over 0.9 F1 score.