This paper shows the work developed with three Advanced Wide Field Sensor (AWiFS) scenes from three intra-annual time periods within a research project of the Remote Sensing Unit (GDR) from the Portuguese Geographic Institute (IGP). The goal was to produce a land cover map using an approach that combines the perfield classification and the object oriented image analysis. First, the AWiFS scenes were segmented to produce objects to be used as classification units. Second, per-pixel classification was performed with decision trees. The training observation collection consisted in a deterministically identification of single pixels spread across the study area concerning a 15-class nomenclature. Third, each object was classified with the modal class provided by the previous per-pixel classification. Four, the combined pixel/object classification was generalized to a minimum mapping unit of 10 ha with an in-house developed software. Finally, the land cover map was evaluated based on a probabilistic accuracy assessment. The methodology was tested on a study area located in Tagus and Sado watersheds. This study area was selected due to its characteristics, such as land cover variety and landscape heterogeneity, that are representative of the Portuguese territory. The land cover map produced with the developed approach has an overall accuracy of 66.9%.
|Title of host publication||Anais do Simpósio Brasileiro de Sensoriamento Remoto|
|Publication status||Published - 1 Jan 2009|
|Event||XIV Simpósio Brasileiro de Sensoriamento Remoto - |
Duration: 1 Jan 2009 → …
|Conference||XIV Simpósio Brasileiro de Sensoriamento Remoto|
|Period||1/01/09 → …|