Influence of Data Source and Training Size on Impervious Surface Areas Classification Using VHR Satellite and Aerial Imagery Through an Object-Based Approach

Abstract

Two very-high-resolution (VHR) satellite images from the GeoEye-1 and WorldView-2 sensors have been used in order to extract impervious surface areas (ISAs) over a Mediterranean coastal area of Almeria (Spain) through an object-based image analysis (OBIA). Different feature sets (basic multispectral information, relative spectral indices, and texture indices based on local variance) were used to feed a support vector machine (SVM) classifier in order to determine the most suitable information for ISAs classification. The classification results coming from both satellite images were compared to each other and also against those provided by a previous similar work carried out on an archival orthoimage. An estimation of the most appropriate number of training samples was performed for each data source by a sampling size reduction procedure. The accuracy assessment of the classification results showed that texture based on local variance was a valuable feature to improve ISA classification accuracy. When texture based on variance was included, the classification accuracy results provided by the archival orthoimage experiment (overall accuracy: 88.1% and KHAT: 0.760) were similar to those obtained from the VHR-satellite images (overall accuracy: 90.4% and 89.7%, KHAT: 0.806 and 0.792 for GeoEye-1 and WorldView-2, respectively). Finally, the influence of the data source and training size on ISA classification accuracy was also proved.