Levin, PhillipDavies, Ian Phillip2020-10-262020-10-262020-10-262020Davies_washington_0250O_22099.pdfhttp://hdl.handle.net/1773/46475Thesis (Master's)--University of Washington, 2020Maps of flood inundation derived from satellite imagery during and after a flood event are critical tools for disaster management. Their utility, however, is limited by optically thick cloud cover that obscures many spaceborne sensors. This study explores a data-driven method to predict flooding in cloud-obscured pixels. For an obscured image, models were trained on visible pixels using 30-meter flood conditioning features and then used to predict flooding on the cloud-covered pixels of that same image. Logistic regression, random forest, and neural networks were evaluated. To obtain prediction uncertainty estimates, a Bayesian neural network using Monte Carlo dropout was trained and compared to Logistic regression confidence intervals. Logistic regression and neural networks averaged 96% accuracy and 86% AUC, but poor recall of <35%. The Bayesian neural network provided useful measures of uncertainty that tracked well with prediction errors. Finer resolution data and more input features may improve this method.application/pdfen-USCC BYBayesian neural networksdata-drivenfloodingGoogle Earth Engineremote sensingTensorFlowGeographyForestryThesis