Improving Image classification Accuracy: a Method to Incorporate Uncertainty in the Selection of Training Sample Set

Luisa M S Goncalves1, Cidalia C Fonte2, Hugo Carrao3 and Mario Caetano3
1.Polytechnic Institute of Leiria Department of Civil Engineering, Portugal Institute for Systems and Computers Engineering at Coimbra (INESC Coimbra) Coimbra, Portugal
2.Institute for Systems and Computers Engineering at Coimbra (INESC Coimbra) Coimbra, Portugal Department of Mathematics - University of Coimbra Coimbra, Portugal
3.Remote Sensing Unit (RSU) Portuguese Geographic Institute (IGP) Lisboa, Portugal Research Centre for Statistics and Information Management (CEGI), Institute for Statistics and Information Management (ISEGI),Universidade Nova de Lisboa, Campus de Campolide Lisboa, Portugal; 2.; 3. {hugo.carrao, mario.caetano}

Abstract: The automatic production of land cover maps using multispectral remote sensing images requires the use of learning classifiers for mapping the imagery data into a set of discrete classes. A group of classifiers commonly used are the supervised classifiers. The first stage of a supervised classification consists on the identification of training areas in the satellite image for each class, which are then used as descriptors of the spectral characteristics of the different classes. The classification results are therefore influenced by the sample pixels selected as training sets. This paper proposes an automatic method to assist the selection of training samples for mapping land cover from satellite images with the aid of ancillary information, namely elder or contemporaneous maps with lower spatial resolution, the Normalized Difference Vegetation Index and information provided by the classification uncertainty. It is shown that more accurate outputs may be derived with this methodology and some conclusions are drawn.

Keywords: training samples, soft classification, measures of uncertainty, classification accuracy

Concalves.pdf970.48 KB