Object-oriented classification in association with accumulated reflectance and data mining tools

Abstract:

The objective of this work was to use the accumulated reflectance technique and data mining application, followed by object-oriented classification, in images of Operational Land Imager (OLI) sensor, Landsat 8, for the classification of native vegetation and agricultural coverage of Cerrado. Four reflectance images were used for the discrimination of six classes - agriculture, livestock, wetland, savannah, forest, and grassland -, for classification of Parque Nacional das Emas and surrounding areas in the state of Goiás, Brazil. The images were segmented for the extraction of sample spectral attributes and application of attribute combinations (mean + mode, all attributes) on data mining. The Weka software was used to construct the decision trees. This methodology indicated that the differentiation among targets increased from the temporal accumulation of the reflectance in all bands and classes, and that the optimal image was that of the sum of the four dates. The classification based on the attribute associations mean + mode showed no restraints in the decision rules processing, unlike the association of all attributes. The mean + mode classification showed a satisfactory accuracy (global accuracy, 69%; Kappa, 58%; and TAU, 63%). The integration of these techniques shows potential to differentiate native and anthropogenic vegetation in the Cerrado.

Index terms:
object-based analysis; multitemporal analysis; anthropization; supervised classification; mining; remote sensing