Nonparametric and Probabilistic Classification of Agricultural Crops Using Multitemporal Images

In this paper, a new approach for classification of multitemporal satellite data sets, combining multispectral and change detection techniques is proposed. The algorithm is based on the nearest neighbor method and derived in order to optimize the average probability for correct classification, i.e. each class is equally important. The new algorithm was applied to a study area where satellite images (SPOT and Landsat TM) from different seasons were used. It showed that using five seasonal images can substantially improve the classification accuracy compared to using a single image. As a large scale application, the approach was applied to the River Dalälven drainage basin. As the distributions for different classes are highly overlapping it is not possible to get satisfactory accuracy at pixel level. Instead it is necessary to introduce a new concept, pixel-wise probabilistic classifiers. The pixel-wise vectors of probabilities can be used to judge how reliable a traditional classification is and to derive measures of the uncertainty (entropy) for the individual pixels. The probabilistic classifier gives also unbiased area estimates over arbitrary areas. It has been tested on two test sites of arable land with different characteristics