Application of a Spectral Angular Mapper on the multispectral Daedalus images improved classification quality of the indicators of the minefields

: In a frame of the project » ; ; Space and airborne Mined Area Reduction Tools – SMART” (European Commission, IST-2000-25044), was used set of multispectral images acquired by scanner Daedalus (DLR, Oberpfaffenhofen, Germany). These images were classified with different methods at the pixel level (RMA, ULB – Brussels, Belgium) and at the region level (ULB – Brussels, Belgium). The representative set of the training and validation patches containing the ground truth data was provided and used. The relevant classes in the project are related to the likelihood of the landmine presence (indicators of mine presence – IMP) and to the likelihood of the landmine absence (the indicators of mine absence IMA), and are not ordinary land cover and land use classes. These classes were defined by iterative research that finished by approved list of IMP and IMA, that depend on the context. The detection of several important IMP and IMA was not possible without use of the multi-band polarymetric synthetic aperture radar data (E-SAR, DLR). The goal of the current work was to improve classification quality of IMA if only multispectral (Daedalus) images are available. In the paper we report about improvement of the IMA detection by supervised classification methods (Mahalanobis, Maximum likelyhood, Minimum distance to mean) if the information obtained by the Spectral Angular Mapping (SAM) method and a priori knowledge about dimensions and shapes of ther fields were fuzed. The most important omission errors of IMA were significantly reduced, and the application of SAM method was approved as useful for the considered problem.