An automated method for sample data selection for Landsat TM and ETM+ images is presented in this paper. Training data is sampled with the use of reflected electromagnetic radiation in separate frequency bands and their combinations. Data is selected for four land cover types of interest, differentiated by land use – water bodies, wetlands, agricultural land, and forests. The proposed method can be used when reference data is lacking or incomplete. For a quality check of the prepared sampling data k-nearest neighbour’s algorithm was used. A high accuracy of classification result was acquired, demonstrated by the results of the experiment section of the study.