Mine and vehicle detection in hyperspectral image data: waveband selection

Abstract

Hyperspectral (HS) data contains spectral response information that provides detailed descriptions of an object. These new sensor data are useful in automatic target recognition applications. However, such high-dimensional data introduces problems due to the curse of dimensionality, the need to reduce the number of features (λ responses) used to accommodate realistic small training set sizes, and the need to employ discriminatory features and still achieve good generalization (comparable training and test set performance). HS sensors produce high-dimensional data; this is characterized by a training set size (N i ) per class that is less than the number of input features (N F ). A new highdimensional generalized discriminant (HDGD) feature extraction algorithm and a new modified branch and bound (MBB) feature selection algorithm are described and compared to other feature reduction methods for two HS target detection applications (mine and vehicle detection). Both space and spectral parameters are adapted. A new blobcoloring hit-miss transform is introduced.

Feature selection from high-dimensional hyperspectral and polarimetric data for target detection

“…When the number of original features is manageable (28 in this case), the modified BB is preferred as it guarantees to find optimal solutions in an efficient manner. However, if the number of original features is large (e.g., 50), BB based algorithms are slow [11]. In that case, suboptimal solutions are needed such as FFS.…”

“…Branch and bound algorithm is efficient compared to exhaustive search algorithms and has been shown useful in practical applications [11]. However, when applied to high dimensionality problems (e.g., select 10 features out of 100 original features), BB or its modified versions are slow.…”

“…J is then evaluated at the nodes at a given level of the tree; if J < B for a given node, then J need not be evaluated at nodes under that node (J decreases as we proceed down the tree, since fewer features are then used, i.e., more features are omitted). To further reduce the searching space, some modifications include [11]: efficient evaluation of paths through the tree with no branches, starting searches at levels 1/4 and 1/2 of the way down the tree, ordering the features (nodes) by which are best using forward selection, and calculation of the initial bound B using the m best features selected by forward selection. At the upper levels of the search tree, we do not expect J < B, since only one or two features (the best ones by forward selection) are omitted.…”

“…In these cases, a two-step search method is preferred: the first step reduces the feature size (filtering) and the second step selects the best subset of features from the remaining features. In [12], a forward selection is first applied to reduce the original feature size to about 30; a modified branch and bound algorithm [11] then selects the best subset of features from the 30 features. In this paper, we propose to use floating forward selection to select an initial subset of 28 features (more are possible, as the number of subsequent calculations is not excessive).…”

“…The BB algorithm works for many practical problems for which an exhaustive search would be impractical or impossible. More efficient BB algorithms [10,11] are recently proposed and are shown to outperform the original BB algorithm by reducing redundant J evaluations in a BB algorithm. A two step feature selection algorithm with sequential forward selection as a filter and branch and bound algorithm as the final feature selector is also proposed for high dimensional feature selection applications in which even the most efficient BB algorithms alone are not practical because of the huge amount of computations [12].…”