We will describe a number of applications that have been developed to use Support Vector Machines (SVMs) to classify imagery from the Multiangle Imaging SpectroRadiometer (MISR) on the Terra satellite. SVMs are a type of supervised learning algorithm, in the same category as Artificial Neural Networks, Decision Trees, and Naïve Bayesian Classifiers.

The MISR instrument has been operational for over four years, acquiring over 50 Terabytes of high quality radiance data in nine view angles and four spectral bands at a resolution of just over 1.1 x 1.1 km, with a subset of that data also available at a resolution of 275 x 275 m. One of the primary objectives of the MISR mission is to improve our understanding of the role of clouds and aerosols in the earth's global energy balance. The three applications we will discuss are the development of a global cloud mask, a global cirrus cloud detector, and a global aerosol classifier and plume detector.

Previously, neural networks have been used in attempts to learn cloud detection models. Typically, these models have been applied only to a small range of scenes or within a geographically limited region. To the best of our knowledge, no model has attempted to work with global satellite data, such as that obtained from MISR. With such a large dataset, of course, artificial intelligence methods can be usefully applied to aid researchers in understanding the data, as well as for the development of models that can characterize the data globally, such as estimating the average amount of global cloud cover.

In these investigations we have used SVMs to take advantage of spectral, spatial and the unique angular information from MISR's suite of nine cameras. Recent studies have shown that SVMs are particularly adept at generalizing and they are sometimes able to achieve higher accuracy than other supervised learning algorithms on standard classification benchmark tests. As with any other supervised classifier, one of the main challenges in training SVMs is that it is very expensive and time consuming to collect training data. To address this problem we have developed an interactive application made possible by a fast SVM trainer and a Graphical User Interface (GUI) designed to make it very easy to see the implications of classification labels.

We will demonstrate the accuracy of the SVM approach as related to each of the three applications above, especially in comparison to traditional single angle, nadir-looking, threshold-based retrievals.