ABSTRACT
The article presents preliminary results of object based classification of multispectral middle resolution MODIS satellite images. In order to use information about spectral diversity of land cover classes, two MODIS Terra images, registered in April and August 2009, were analysed simultaneously. As the study area, a polygon of 85 267 km2 situated in central Poland was selected. Within its borders all typical classes of Polish land cover can be found. The object oriented classification was performed using red and infrared bands of both images. Additionally, channel PC 2 obtained as an outcome of the Principal Component Analysis, was incorporated into the classification as one of the main features for discrimination of land cover classes. The first segmentation is prepared only for classification of water bodies. Next segmentation of the rest of the scene is done, and after that the object domain is divided into two parts characterized by high and low values of PC 2 channel. Using consecutive processes inside the first part of the object domain, the objects were classified as: urban areas, forest and agriculture areas. Grasslands and other agriculture areas are recognized in the second part of the object domain. The applied classification rules, based on threshold values of the bands R, IR, PC 2 and NDVI index, allow detection of six basic land cover classes: water, dense urban areas, spread urban areas, forest, agricultural area and grassland. The accuracy assessment of the final classification was done using the CORINE CLC2006 datasets. Before determining the error matrix, the classification image was re-projected from Sinusoidal into the 1992 coordinate system while the vector data base CLC2006 was exported to the raster format. The overall accuracy was estimated at the level of 78%.