The computer-assisted landcover classification of large areas is a common task in remote sensing. Depending from the individual question and the available inputdata, classes reach different levels of separability. For higher overallaccuracies of classification, additional features like vegetation indices or texture features can be produced through datatransformation. The data set for this thesis are four chanel (near Infrared, red, green, blue) digital aerial images of the bavarian alps with in five resolutions (20 cm, 100 cm, 100 cm focal median, 200 cm, 500 cm). The influence of different aspects of spectral and textural features to the classificationresult, the impact of featureselection, amount and size of trainingsamples and the transferability of the classificator to bordering testsites are investigated. For classification an object based approach using Random Forest classifier was chosen. Six classes were classified: coniferous trees, broad-leafed trees, mountain pine (Pinus mugo), grasland, rock and shadows. Objects containing more than two classes were left out of the systematic test. The highest overallaccuracy of the independend testdata using the spectral information from all four channels near infrared, red, green and blue reached 77.4 %. Using all features a result of 85.5 %, through featureselection of the 300 most important of all features, rated by their MDA importance, 86.1 % was reached. Using spectral features and texture features produced with the graylevel co-occurence matrix approach and the discrete stationary wavelet transform, overallaccuracies of 84.2 % of the independend testdataset were reached. The highest Producer- and Useraccuracies were reached for the classes rock and shadow, the lowest for the class broad-leafed trees. Missclassification occured most frequently between the classes coniferous trees, borad-leafed trees and mountain pine.