AUTOMATIC SELECTION OF TRAINING AREAS USING EXISTING LAND COVER MAPS

摘要

Supervised image classification requires a setrnof sampling units for each class. Traditionalrnsampling is costly because it has to be donernmanually by image interpreters. In this paperrnwe present an automatic approach forrndefining training samples that only needs anrnold land cover map for the same area withrndifferent technical attributes. This map isrnfirstly reclassified into the nomenclature ofrnthe map one wants to produce. Then a randomrnstratified sampling is executed using the basernmap classes as strata. The sample is thenrnautomatically edited in order to eliminaternmislabelled sampling units. The methodologyrnwas tested with MERIS data of ContinentalrnPortugal. Although the overall accuracy isrnsmaller compared with the traditional methodrn(76% against 80%), the training sampling costrnwas greatly reduced. The results suggest thatrnthe proposed methodology is promising sincernit allows reducing the cost of samplingrnwithout excessive loss of thematic accuracy.