Data mining applied for land cover classification using Landsat 8

This paper is committed to investigate the spectral attributes extracted from Landsat 8 image bands and topographic attributes derived from TOPODATA, meant to discriminate land cover classes by means of decision trees, a technique in the scope of data mining. Statistical measures of samples corresponding to 12 land cover classes collected in Rio de Janeiro city were calculated from a database composed of 18 layers, from which four decision trees were generated. The results showed that the mean and median were the most relevant statistical attributes. As to spectral attributes, Band 1 is worth of mention, which has been selected to classify water classes, besides discriminating vegetation and non-vegetation classes. Regarding vegetation indices, the data mining algorithm exclusively relied on the Simple Ratio Index in all trees to the detriment of the NDVI. Slope has been employed in three decision trees to separate rock outcrop from low-height vegetation. On the other hand, radiance has not been used in any of the four decision trees. Considering the ever-increasing volume of remotely sensed data currently available, it ought to be acknowledged that data mining represents a crucial solution to efficiently extract information from large databases in a short time.

Keywords:
semantic networks; images classification; decision trees