Extrapolating Satellite-Based Flood Masks by One-Class Classification - A Test Case in Houston

Flood masks are among the most common remote sensing products, used for rapid crisis information and as input for hydraulic and impact models. Despite the high relevance of such products, vegetated and urban areas are still unreliably mapped and are sometimes even excluded from analysis. The information content of synthetic aperture radar (SAR) images is limited in these areas due to the side-looking imaging geometry of radar sensors and complex interactions of the microwave signal with trees and urban structures. Classification from SAR data can only be optimized to reduce false positives, but cannot avoid false negatives in areas that are essentially unobservable to the sensor, for example, due to radar shadows, layover, speckle and other effects. We therefore propose to treat satellite-based flood masks as intermediate products with true positives, and unlabeled cells instead of negatives. This corresponds to the input of a positive-unlabeled (PU) learning one-class classifier (OCC). Assuming that flood extent is at least partially explainable by topography, we present a novel procedure to estimate the true extent of the flood, given the initial mask, by using the satellite-based products as input to a PU OCC algorithm learned on topographic features. Additional rainfall data and distance to buildings had only minor effect on the models in our experiments. All three of the tested initial flood masks were considerably improved by the presented procedure, with obtainable increases in the overall κ score ranging from 0.2 for a high quality initial mask to 0.7 in the best case for a standard emergency response product. An assessment of κ for vegetated and urban areas separately shows that the performance in urban areas is still better when learning from a high quality initial mask.