By Topic

Neuroinspired Architecture for Robust Classifier Fusion of Multisensor Imagery

Sign In

Cookies must be enabled to login.After enabling cookies , please use refresh or reload or ctrl+f5 on the browser for the login options.

Formats Non-Member Member
$33 $13
Learn how you can qualify for the best price for this item!
Become an IEEE Member or Subscribe to
IEEE Xplore for exclusive pricing!
close button

puzzle piece

IEEE membership options for an individual and IEEE Xplore subscriptions for an organization offer the most affordable access to essential journal articles, conference papers, standards, eBooks, and eLearning courses.

Learn more about:

IEEE membership

IEEE Xplore subscriptions

1 Author(s)
Bogdanov, A.V. ; Ruhr- Univ. Bochum, Bochum

Two new algorithms for robust and fault-tolerant classifier combination are presented. The attractor dynamics (AD) algorithm models some properties of sensory integration in the central nervous system and is based on the application of the dynamical systems for classifier fusion. The classifier masking (CM) algorithm is a nonneural version of the AD algorithm based on finding intersecting classifier intervals. Both of the proposed algorithms employ the idea of consensus among individual classifiers. The individual classifiers have been trained using resampled feature sets. They fuse the information from advanced synthetic aperture radar, medium resolution imaging spectrometer, and advanced along track scanning radiometer envisat satellite sensors for the improved sea ice classification. The results of our experiments show that training and combing the individual classifier outputs in a multiple classifier system significantly improve the robustness and the fault tolerance of the classification system as compared to the single classifier combining all sources of information. The robustness of the single classifier has been largely reduced in cases of single sensor failures (87.9 % in normal conditions versus 64.8% and 66.1% for two artificially corrupted data sets), whereas the CM algorithm is more tolerant to the sensor and preprocessing errors (86.4% in normal conditions versus 78.9% and 73.6% for two artificially corrupted data sets). The performance of the CM algorithm is superior to those of the simple multiple classifier combination strategies based on classifier averaging and majority voting (78.9% versus 70.9% and 69.5%, respectively) because the AD and CM algorithms are able to discard the corrupted classifier outputs based on classifier agreement and, in fact, represent hybrid approaches combining the properties of classifier averaging and classifier selection methods.

Published in:

Geoscience and Remote Sensing, IEEE Transactions on  (Volume:46 ,  Issue: 5 )