SEMI-AUTOMATIC TEXTURE SEGMENTATION OF REMOTELY SENSED IMAGERY FOR LANDSLIDE HAZARD ASSESSMENT

Javier Hervás (1) and Paul L. Rosin (2)

1. Joint Research Centre, Space Applications Institute, 21020 Ispra (Va), Italy
2. Brunel University, Dept. of Information Systems, Uxbridge, Middlesex, UK

Image segmentation implies the division of the image into regions of similar attribute, where texture can be taken as such attribute. On remotely sensed imagery of landslide-prone areas texture is often the expression of specific landforms and/or vegetation disruption patterns encompassing landsliding. In this paper a semi-automatic texture segmentation approach for recognising unstable slopes is presented. The method entails three main stages: i) selection of training samples; ii) texture transformation based on the image texture spectrum; iii) thresholding of the texture image into discrete landslide hazard zones. This technique has been applied to both air-borne and satellite-borne panchromatic and multispectral digital image sets at resolutions ranging from 3.5 m to 30 m, including ATM, IRS-1C Pan, SPOT Pan and Landsat TM imagery of a semi-arid Mediterranean area. On multispectral imagery, discriminant analysis is further undertaken for selecting the best texture-based segmented bands. Adequacy of spatial resolution and spectral coverage for discriminating landslide-related texture on imagery is also discussed.