Texture features for land cover change detection at 250 m resolution-an application of machine learning feature subset selection

Texture measures have been widely studied in addition to spectral features to characterize land cover type. Texture measures can be used to update land cover maps and enhance classification accuracy. Most previous studies have focussed only on fine resolution data such as Landsat, SPOT and Synthetic Aperture Radar imagery. Texture analysis at coarser resolution (250 m) has not been investigated because data are only beginning to become available with the launch of MODIS (Moderate Resolution Imaging Spectroradiometer). To meet the need of global change research, the MODIS instrument is planned to provide global coverage at 250 m resolution in the red and infrared bands. Early warning of human-induced land cover changes, such as urbanization and deforestation, is expected from MODIS. This paper examines whether texture analysis at this scale would be useful for detecting changes. There are at least two issues in using texture analysis: 1) which texture measures should be used, and 2) what window size is most appropriate for capturing textures. In this study we test a large number of texture measures and window sizes on MODIS data simulated from Landsat data using the data mining technique of feature selection