An operational MISR pixel classifier using support vector machines

The Multi-angle Imaging SpectroRadiometer (MISR) data products now include a scene classification for each 1.1-km, pixel that was developed using Support Vector Machines (SVMs), a cutting-edge machine learning technique for supervised classification. Using a combination of spectral, angular, and texture features, each pixel is classified as land, water, cloud, aerosol, or snowlice, with the aerosol class further divided into smoke, dust, and other aerosols. The classifier was trained by MISR scientists who labeled hundreds of scenes using a custom interactive tool that showed them the results of the training in real time, making the process significantly faster. Preliminary validation shows that the accuracy of the classifier is approximately 81% globally at the 1.1-km pixel level. Applications of this classifier include global studies of cloud and aerosol distribution, as well as data mining applications such as searching for smoke plumes. This is one of the largest and most ambitious operational uses of machine learning techniques for a remote-sensing instrument, and the success of this system will hopefully lead to further use of this approach. (c) 2006 Elsevier Inc. All rights reserved.