Object-based classification of multi-sensor optical imagery to generate terrain surface roughness information for input to wind risk simulation

Geoscience Australia is conducting a series of national risk assessments for a range of natural hazards such as severe winds. The impact of severe wind varies considerably between equivalent structures located at different sites due to local roughness of the upwind terrain, shielding provided by upwind structures and topographic factors. Terrain surface roughness information is a critical spatial input to generate wind multipliers. It is generally the first spatial field to be evaluated, as it is utilised in both the generation of the terrain and topographic wind multiplier. Landsat imagery was employed to generate a terrain surface roughness product for six major metropolitan areas across Australia. It was necessary to investigate the applicability of multi-sensor approaches to generate a regional/national terrain surface roughness map based on the Australian/New Zealand wind loading standard (AS/NZS 1170.2). This paper discusses the methodology that developed a procedure to derive terrain surface roughness from various multi-source satellite images. MODIS, Landsat, and IKONOS imagery were acquired (from 12 September - 26 November 2002) covering a significant portion of the New South Wales, Australia. An object-based image segmentation and classification technique was tested for seven bands of MODIS, six bands of Landsat Thematic Mapper, and four bands of IKONOS. Eleven terrain categories were identified using this technique which achieved classification accuracies of 79% and 93% over metropolitan (Sydney) and rural/urban areas respectively. It was revealed that the object-based image classification enhances the quality of the terrain product compared to traditional spectral- based maximum likelihood classification methods. To further improve the derivation of terrain roughness classification results, an integrated textural-spectral analysis merged Synthetic Aperture Radar and optical datasets provided in a study by [1]. A comparison with results derived from textu- ral-spectral classification showed considerable improvement over the results from earlier classification techniques.