Segment Optimisation for Object-based Landslide Detection

Advances in remote sensing technology and image analysis systems have led to an increase in automatic feature extraction technique for several novel applications. Object-oriented analysis (OOA) of high resolution remote sensing data is one such technique, wherein objects/segments are the image primitives that form the basis for automatic feature extraction, and thus have critical influence on the accuracy of the subsequent processing. In this study we assess how the fractal net evolution segmentation approach, embedded in Definiens Developer, can be optimised for automatic detection of landslides. Landslides are important natural mass wasting processes, and their fast detection using high resolution satellite and elevation data can inform disaster mitigation strategies. In recent work we showed how spectral, contextual and morphometric information can be successfully used to detect and characterise landslides of different types. However, given the complex shape and size of landslides, as well as their spectral heterogeneity, segmentation with a single set of parameters is not possible and remains a weakness in our method. Here we assess multiple segmentation strategies to derive an appropriate segmentation for various classes to be identified in subsequent rule-based classification. Segments were subjected to spatial autocorrelation analysis using Moran’s I index and intrasegment variance analysis to optimise the parameter selection for multiresolution segmentation. Instead of one object level pertaining to a single parameter, multiple object levels with a set of optimum parameters were used to detect landslides of variable sizes. We test our approaches with Resourcesat-1 LISS-IV multispectral imagery (5.8 m) and a 10 m DEM derived from 2.5 m stereoscopic Cartosat-1 data for two landslide prone areas in the High Indian Himalayas. * Corresponding author.