We present a parallel image classification approach, referred to as the parallel positive Boolean function (PPBF), to multisource remote sensing images. PPBF is originally from the positive Boolean function (PBF) classifier scheme. The PBF multiclassifier is developed from a stack filter to classify specific classes of land covers. In order to enhance the efficiency of PBF, we propose PPBF to reduce the execution time using parallel computing techniques. PPBF fully utilizes the significant parallelism embedded in PBF to create a set of PBF stack filters on each parallel node based on different classes of land uses. It is implemented by combining the message-passing interface library and the open multiprocessing (OpenMP) application programing interface in a hybrid mode. The experimental results demonstrate that PPBF significantly reduces the computational loads of PBF classification.
In hyperspectral imagery, greedy modular eigenspace (GME) was developed by clustering highly correlated bands into a smaller subset based on the greedy algorithm. Unfortunately, GME is hard to find the optimal set by greedy scheme except by exhaustive iteration. The long execution time has been the major drawback in practice. Accordingly, finding the optimal (or near-optimal) solution is very expensive. Instead of adopting the band-subset-selection paradigm underlying this approach, we introduce a simulated annealing band selection (SABS) approach, which takes sets of non-correlated bands for high-dimensional remote sensing images based on a heuristic optimization algorithm, to overcome this disadvantage. It utilizes the inherent separability of different classes embedded in high-dimensional data sets to reduce dimensionality and formulate the optimal or near-optimal GME feature. Our proposed SABS scheme has a number of merits. Unlike traditional principal component analysis, it avoids the bias problems that arise from transforming the information into linear combinations of bands. SABS can not only speed up the procedure to simultaneously select the most significant features according to the simulated annealing optimization scheme to find GME sets, but also further extend the convergence abilities in the solution space based on simulated annealing method to reach the global optimal or near-optimal solution and escape from local minima. The effectiveness of the proposed SABS is evaluated by NASA MODIS/ASTER (MASTER) airborne simulator data sets and airborne synthetic aperture radar images for land cover classification during the Pacrim II campaign. The performance of our proposed SABS is validated by supervised k-nearest neighbor classifier. The experimental results show that SABS is an effective technique of band subset selection and can be used as an alternative to the existing dimensionality reduction method.
Satellite remote sensing images can be interpreted to provide important information of large-scale natural resources, such
as lands, oceans, mountains, rivers, forests and minerals for Earth observations. Recent advances of remote sensing
technologies have improved the availability of satellite imagery in a wide range of applications including high
dimensional remote sensing data sets (e.g. high spectral and high spatial resolution images). The information of high
dimensional remote sensing images obtained by state-of-the-art sensor technologies can be identified more accurately
than images acquired by conventional remote sensing techniques. However, due to its large volume of image data, it
requires a huge amount of storages and computing time. In response, the computational complexity of data processing
for high dimensional remote sensing data analysis will increase. Consequently, this paper proposes a novel classification
algorithm based on semi-matroid structure, known as the parallel k-dimensional tree semi-matroid (PKTSM)
classification, which adopts a new hybrid parallel approach to deal with high dimensional data sets. It is implemented by
combining the message passing interface (MPI) library, the open multi-processing (OpenMP) application programming
interface and the compute unified device architecture (CUDA) of graphics processing units (GPU) in a hybrid mode. The
effectiveness of the proposed PKTSM is evaluated by using MODIS/ASTER airborne simulator (MASTER) images and
airborne synthetic aperture radar (AIRSAR) images for land cover classification during the Pacrim II campaign. The
experimental results demonstrated that the proposed hybrid PKTSM can significantly improve the performance in terms
of both computational speed-up and classification accuracy.
For hyperspectral imagery, simulated annealing (SA) and greedy modular eigenspaces (GME) have been successfully
developed to cluster highly correlated hyperspectral bands into a smaller subset of band modules. This paper introduces a
novel band selection technique of combining these two approaches, called the SA and GME band selection (SGBS), for
hyperspectral imagery. The SGBS selects sets of non-correlated hyperspectral bands for hyperspectral images based on
heuristic and greedy algorithms, utilizes the inherent separability of different classes in hyperspectral images to reduce
dimensionality, and further generates a unique clustered eigenspace (CE) feature set effectively. The proposed SGBS
features can 1) avoid the bias problems of transforming the information into linear combinations of bands as does the
traditional principal components analysis, 2) evince improved discriminatory properties, crucial to subsequent classification
compared with conventional band selection techniques, 3) provide a fast procedure to simultaneously select the most
significant features by merging SA and GME schemes, and 4) select each band by a simple logical operation, called the
CE feature scale uniformity transformation (CE/FSUT), to include different classes into the most common feature modules
of the hyperspectral bands. The performance of the proposed SGBS method was evaluated by MODIS/ASTER airborne
simulator (MASTER) images for land cover classification during the Pacific Rim II campaign. Encouraging experimental
results showed that the proposed SGBS approach is effective and can be used as an alternative to the existing band selection
algorithms.
For hyperspectral imagery, greedy modular eigenspaces (GME) has been developed by clustering highly correlated hyperspectral bands into a smaller subset of band modules based on greedy algorithm. Instead of greedy paradigm as adopted in GME approach, this paper introduces a simulated annealing band selection (SABS) approach for hyperspectral imagery. SABS selects sets of non-correlated hyperspectral bands for hyperspectral images based on simulated annealing (SA) algorithm while utilizing the inherent separability of different classes in hyperspectral images to reduce dimensionality and further to effectively generate a unique simulated annealing module eigenspace (SAME) feature. The proposed SABS features: (1) avoiding the bias problems of transforming the information into linear combinations of bands as does the traditional principal components analysis (PCA); (2) selecting each band by a simple logical operation, call SAME feature scale uniformity transformation (SAME/FSUT), to include different classes into the most common feature clustered subset of bands; (3) providing a fast procedure to simultaneously select the most significant features according to SA scheme. The experimental results show that our proposed SABS approach is effective and can be used as an alternative to the existing band selection algorithms.
Multispectral sensors are still widely used in satellite remote sensing. They usually have spectral bands less than ten
channels. The problem for so few channels is that it can not directly solve linear mixture model by least square unmixing
for subpixel target detection. In order for least square approach to be effective, the number of bands must be greater than
or equal to that of signatures to be classified, i.e., the number of equations should be no less than the number of
unknowns. This ensures that there are sufficient dimensions to accommodate orthogonal projections resulting from the
individual signatures. It is known as band number constraint (BNC). Such constraint is not an issue for hyperspectral
images since they generally have hundreds of bands, however, this may not be true for multispectral images where the
number of signatures to be classified might be greater than the number of bands. In order to relax this constraint, we
present two signature reduction methods to reduce the number of unknowns, based on signature selection and signature
fusion. A SPOT image scene will be used for experiment to demonstrate the performance.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.