Classification of Urban Structures by Combining Spectral and Spatial Information of Satellite Images

Abstract

Some experimental studies on the classification accuracy of urban structures were conducted using local density average (LDA) and local dentity variation (LDV) as the combined features of spectral and spatial information which replace conventional pixel-wise spectral information, for the purpose of verifying new possibility of data analysis by high resolution satellite data such as Landsat/TM and Spot/HRV. Airborne MSS data were used for simulating pseudo-satellite data of various ground resolutions from 5 meters to 80 meters. By the airborne MSS simulation experiment, the feature LDA or LDA combined with LDV was verified to give higher classification accuracy then the pixel-wise spectral feature, if ground resolution is higher than 30 meters. In addition, the resolution of 20 meters or 30 meters was proved to result in the highest accuracy with the feature LDA combined with LDV. By the experiment using actual satellite data, Landsat/TM and MSS data resulted in the reasonable accuracy which was predicted from the airborne MSS simulation. However, Spot/HRV data showed lower accuracy about 20 percent than the airborne MSS simulation accuracy, which is considered to be due to the spectral variation caused by offnadir observation from satellite. These experiments suggest new data utilization for urban analysis with high resolution TM or HRV data.