MISR abstract

Garay, M.J., Mazzoni, D.M., Davies, R., DeCoste, D.M., and Braverman, A. J. (2003). A Hybrid Global MISR Cloud Mask using Support Vector Machines and Active Learning. Eos Trans. AGU, 84(46), Fall Meet. Suppl. 2003, Abstract # H11F-0913

The Multiangle Imaging SpectroRadiometer (MISR) onboard NASA's Terra EOS satellite provides unique sensing capabilities that promise potentially much better global cloud identification. A number of algorithms have been developed and implemented for detecting clouds in MISR data, some of which use MISR's unique multiangle sensing capability. All of these techniques are firmly grounded in the physics of remote sensing, but the accuracy of each method is highly dependent on different specific conditions. This presents a unique opportunity for soft computing methods. We are investigating techniques that use Support Vector Machines (SVMs) to combine the raw MISR data and the output of existing MISR cloud mask algorithms into a new and more robust global cloud mask. One of the main challenges in training a SVM (or any other supervised classifier) is that it is very expensive and time consuming to collect training data. To address this problem we have incorporated and are continuing to refine the relatively new technique known as active learning, in which the algorithm queries the human expert to supply training labels in regions that would be most beneficial for improving the model. We have developed an interactive application which utilizes SVMs and active learning to allow a scientist to quickly train a classifier for MISR data. In addition, we have performed a number of small-scale case studies and a global sampling study which compare the accuracy of the existing MISR cloud mask algorithms to our best SVM models.

[Full text not yet available]