Object-based classification of IKONOS data for vegetation mapping in central Japan

Abstract

Vegetation mapping using IKONOS data was implemented at a countryside study area in central Japan, where small patches of various plant communities are mixed together in a complicated mosaic pattern. Pixel-based and object-based classifications using only spectral features were implemented and their accuracies were compared. In addition, the object-based classification was also performed on a combination of spectral and textural features, with a stepwise regression model used in the discriminate analysis to select the most relevant features. Classifications were implemented at four levels, the highest of which used seven vegetation categories. The object-based classification proved more accurate than the pixel-based classification. In addition, the addition of textural features generated significant improvements in accuracy. The overall classification accuracy and Kappa coefficients at the highest level were 52.8% and 0.373 for the pixel-based classification; 58.9% and 0.458 for the object-based with spectral features only; and 65.0% and 0.542 for the object-based with additional features. Some problems with misclassification remained, but the overall results demonstrate that object-based classification of very high resolution satellite images using additional features is a practical tool for vegetation mapping in Japan.