Pixel- versus object-based classification of forest and agricultural areas from multiresolution satellite images

Managing of natural resources including agriculture and forestry is a very important subject for governments and decision makers. Up-to-date, accurate, and timely geospatial information about natural resources is needed in the management process. Remote sensing technology plays a significant role in the production of this geospatial information. Compared to terrestrial work, the analysis of larger areas with remote sensing techniques can be done on a shorter timescale and at lower cost. Image classification in remote sensing is one of the most popular methods used for the detection of forest and agricultural areas. However, the accuracy of classification changes according to the source and reference data, the classification method, and the producer's knowledge and experience. In this research, the identification of forests and agricultural areas was studied in terms of both their geometry and attribution using different classification methods and different source data. Landsat, Aster, and RapidEye images, which have different spatial and spectral resolution, were used as the source data. Pixel- and object-based classification algorithms were also tested. Classification accuracy results were evaluated at 300 stratified random pixels. It was found that the best overall accuracy was obtained from Aster imagery with object-based classification using the nearest neighbor method. The results also showed that spatial resolution is important for discrimination of classes and spectral resolution is important for definition of features, and confirmed the well-established paradigm of remote sensing that there are no perfect source data or method of classification for all situations.