Multiscale classification by using optimum Path-Forest
Nowadays, the main problems in region recognition of remote sensing images are: (1) the dependence of the classification methodson the segmentation quality; and (2) the selection of representativesamples for training. The major challenge is that the samplesindicated by the user are not always enough to define the best segmentation scale.Furthermore, the indication of samples can be costly, since it often requires to visit studied places in loco.The objective of this research project is to develop aninteractive multiscale classification approach that allows segmentation andclassification refinement according to user indications. The segmentation-dependence problem will be addressed by using techniques that rely on multiple scales instead ofonly one segmentation result. The selection of representative samples, in turn, willbe supported by the development of new approaches based on active learning with user interactions.The proposed method will be validated in three applications associated withdistinct research areas found at Institute of Computing, Unicamp: (1) phenologicalpattern recongnition; (2) agricultural region classification by usingmultisensor data; and (3) illegal region identification in aerial images.