Flooding extent cartography with Landsat TM imagery and regularized kernel Fisher's discriminant analysis

Abstract

In this paper the combined use of the regularized kernel Fisher's discriminant analysis classifier (kFDA) with Landsat TM multispectral imagery is explored for flooded area cartography purposes. This classifier provides an efficient and regularized solution for the non-linear delineation of pixels corresponding to flooded surface. The flood mapping issue is tackled from both uni- and multi-temporal classification perspectives: the former recasts the problem as a classical image classification procedure – with class water as target; the latter considers the extraction of flooded area as a change detection problem – in which only the non-permanent standing water is considered as flood. As a case study is used a Landsat TM dataset of the James River in South Dakota (USA), a region that experienced a heterogeneous flooding in spring 2011. Findings from our analysis suggest that precisely delineating the exceeding water extent requires a non-linear classifier applied in a multi-temporal setting.