A MISR cloud-type classifier using reduced Support Vector Machines

We are developing a pixel-level cloud-type classifier for the Multi-angle Imaging SpectroRadiometer (MISR), an instrument used to study clouds and aerosols from NASA’s Terra satellite. To augment MISR’s existing high-level products (including cloud masks, cloud heights, and aerosol optical depth retrievals), our cloudtype classifier labels each 1.1-km pixel as clear, or as belonging to one of several types of cloud. In the past, similar classifiers have been developed for other remote-sensing instruments using various machine learning techniques, such as artificial neural networks. However, support vector machines (SVMs) are not typically used, in part because the computational cost of evaluating new examples with an SVM can be much higher. Our novel approach to achieving high classification accuracy within the computational requirements of the operational MISR processing system involves training a very large multi-class SVM using thousands of training points and then applying cutting-edge reduced-set techniques to yield a computationally manageable number of support vectors. The resulting product will help provide new insights for those constructing cloud climatologies, modeling radiative transfer through clouds, and studying the effects of clouds on climate, in addition to demonstrating the effectiveness of using SVMs in a production science