Sensitivity of mixture modelling to end‐member selection

Abstract

This paper applies five different image‐based end‐member selection methods for use in a linear mixture model applied to Along Track Scanning Radiometer 2 (ATSR‐2) data. The study area is a 10 000 km2 site based in central Finland in a landscape that is dominated by forests. Observed forest proportions are derived from the Finnish Forest Research Institute forest inventory data, a multi‐source inventory based on interpolation between a systematic grid of field plots applying Landsat Thematic Mapper (TM) data at a resolution of 25 m. A stepwise multiple linear regression between the ATSR‐2 bands and the observed forest proportion data for the study site results in a standard error of 20.1% between observed forest proportions and predicted forest proportions and an R 2 value of 45.7%. This implies a relatively weak linear relation between the data bands and the observed forest proportions. The end‐member selection methods used within this study are based on using a principal components analysis (PCA) approach and using ancillary data to determine pure pixels within the image data. The accuracy of the modelled proportions estimates are evaluated using the rms error of the modelled forest proportions to the observed forest proportions. The PCA‐based end‐member model has the lowest rms error of 23.75%. The high‐pixel method has the highest rms error of 63.28% and also has the greatest tendency to return pixel estimates to either 100% forest or 0% forest. The proportions estimates from all of the models (when excluding pure pixels) are systematically related and highly correlated.