Optimizing land cover classification accuracy for change detection, a combined pixel-based and object-based approach in a mountainous area in Mexico

Abstract

Inventories of past and present land cover changes form the basis of future conservation and landscape management strategies. Modern classification techniques can be applied to more efficiently extract information from traditional remote-sensing sources. Landsat ETM+ images of a mountainous area in Mexico form the input for a combined object-based and pixel-based land cover classification. The land cover categories with the highest individual classification accuracies determined based on these two methods are extracted and merged into combined land cover classifications. In total, seven common land cover categories were recognized and merged into single combined best-classification layers. A comparison of the overall classification accuracies for 1999 and 2006 of the pixel-based (0.74 and 0.81), object-based (0.77 and 0.71) and combined (0.88 and 0.87) classifications shows that the combination method produces the best results. These combined classifications then form the input for a change detection analysis between the two dates by applying post-classification, object-based change analysis using image differencing. It is concluded that the combined classification method together with the object-based change detection analysis leads to an improved classification accuracy and land cover change detection. This approach has the potential to be applied to land cover change analyses in similar mountainous areas using medium-resolution imagery.