Probabilistic flood forecast model based on remote sensing information

Abstract

This study aims to explore the possibility of employing remote sensing images to build a probabilistic flood extent forecasting model. This model is constructed and tested in two study areas: New Orleans and Miami. Images that recorded flooding events are first performed with segmentation method Seed Region Growing, and segmented images are classified by Maximum Likelihood classifier. Area detected as water subtracting the permanent water area is the detected flood extent. In total there are nineteen images being processed. The flood detection result is validated by flooded locations from NOAA flood reports and the news, and the accuracy is at 70.4%. The detection result, with flood conditioning factors which include precipitation, sea level, elevation, drainage capacity and distance to the water area, is the input to the probabilistic forecasting model. All inputs are standardised to a common grid system and every cell in that grid system contains a set of data. Two kinds of model structures are proposed and both models are trained with logistic regression and probit regression, both of which are the members of the Generalised Linear Model(GLM). The first kind of model structure is only tested in the New Orleans study area and the second kind of model structure is examined at both study areas. The precision of the first model structure is at 20% with a kappa value at zero. For the second model structure, over-sampling method SMOTE is used to increase the number of data points of the class 'flooded'. The highest precision of the second model structure at New Orleans is 12.6% and at Miami 23.6%, and the highest cohen's kappa values are 0.127 and 0.131 for New Orleans and Miami respectively. The first model structure actually failed in building a success model at a large portion of the study area due to limited records. For the second kind of model structure, most variables are linked to the flooding by the model correctly. The precipitation has a positive relation with flooding, especially when time effect is considered. Elevation reduces the probability of flooding. At the Miami study area where no sea dike exists the sea level has a strong positive relation with flooding. Drainage capacity used in the New Orleans study area does not show an influence on flooding, which requires modelling the intricate drainage system more accurately. In Miami study area, when the study area is confined to the seaside and Miami Beach area, the accuracy of prediction is improved, which informs that land-use type is crucial to be considered in the input. This study innovative collected information from several remote sensing images of different flood events and applied the information to build a probabilistic model, which shows that the information provided by mages could link flooding conditioning factors with flooding. It is recommended to incorporate the remote sensing technique in the flood extent forecast model in the future.