An Unsupervised Surface Water Un-Mixing Method Using Landsat and Modis Images for Rapid Inundation Observation

Surface water has been monitored by using remote sensing techniques for decades. Nevertheless, remotely sensed images inevitably compromise temporal and spatial resolution, which limits the urgent application in sudden flooding response. Images with a high spatial resolution cannot be acquired and processed frequently to monitor the fast-evolving flood events that could change dramatically within a few days. To address such a challenge, this paper presents a method to un-mix surface water from MODIS derived water indices, which could be generated daily. First, before flooding, the water indices are calculated from satellite sensed reflectance, and water fraction is calculated from classified Landsat images. Second, the relationship between water fraction and water indices is learned by using a regression tree ensemble method. Third, after flooding, water indices are obtained from the new MODIS images, and water fraction is retrieved by applying the learned regression tree ensemble. By spatially un-mixing coarse but daily accessible surface water observations, this proposed method facilitates near-real-time emergency responses to flooding and inundation monitoring.