This paper is concerned with subpixel modelling of land cover estimation. A new method is presented to estimate the abundance of different cover types within individual pixels. Several linear and non-linear multivariate calibration techniques are compared with respect to their ability to establish a relation between pixel values and fractions of ground cover. The method is demonstrated using Landsat-TM imagery and data from Dutch heathlands. By using an optimization procedure for matching field data to image data, a solution was found for the problem of positioning errors in the training set formation.