Development of Advanced Semi-Supervised Kernel Methods for Classification of Hyperspectral Remote Sensing Images

HYPERKERNEL DIT-PRJ-05-139

Status NOT active project
DISI role Coordinator
Project type Research Project
Dimension International
Acquisition date 2005-12-16
Start date 2006-05-19
End date 2009-06-22

Project details

Project astract Nowadays, the new generation of remote sensing satellites offer many possibilities for monitoring the Earth surface (e.g., for the analysis of climate changes, urbanization, forest fires, coastal areas, water quality, crop fields, detection of contaminants and pollution elements, etc). Despite the good capabilities of multispectral sensors (which have a relatively small number of spectral channels) to address these problems, in the last years the development of hyperspectral sensors has offered improved performance for the detection and classification of the Earth land-cover classes. In particular, the information contained in hyperspectral images allows the characterization, identification, and classification of the land-covers with improved accuracy and robustness. However, several critical problems should be considered in classification of hyperspectral data, among which: (i) the high number of spectral channels, (ii) the spatial variability of the spectral signature, (iii) the high cost of true sample labeling, and (iv) the quality of data. In particular, the high number of spectral channels and low number of labeled training samples pose the problem of the curse of dimensionality (i.e., the Hughes phenomenon) and, as a consequence, result in the risk of overfitting the training data. These problems have been recently alleviated by the introduction of kernel-based classifiers (e.g., support vector machines). However, further improvement in the classification accuracy could be achieved by taking into account the information provided by the high number of unlabeled samples in the image. In recent years a considerable attention has been devoted to semi-supervised learning, which differs from traditional supervised learning as it exploits unlabeled data. The main objective and motivation driving this project is to extend the use of kernel-based methods to the challenging semi-supervised classification framework, with the aim of defining advanced classification techniques for hyperspectral remote sensing images.
Keywords kernel Based Methods, SVM, Machine Learning, Remote Sensing, Hyperspectral Data
Fundings 26000 €
Partners
  • DIT - UniTN
  • University of Valencia

DISI Sub-project details

Project astract The main objective and motivation driving this project is to extend the use of kernel-based methods to the challenging semi-supervised classification framework, with the aim of defining advanced classification techniques for hyperspectral remote sensing images.
Fundings 16000 €
Manager Lorenzo Bruzzone