Cloud detection and characterization using topological data analysis

Abstract

The presence of cirrus clouds introduces complex heating and cooling effects on the atmosphere and can also interfere with remote sensing from satellite-based sensors or from high-altitude aircraft. Detection of cirrus clouds thus provides an opportunity for atmospheric correction to introduce accurate compensation to images of the earth’s surface. Previous work on detection and characterization of cirrus clouds have been based on observing spectral signatures on a spectral channel with significant water absorption, or calculation of radiant intensity ratios over a water band to a reference spectral channel. Our proposed approach is based on applying computational homology to characterize the topological properties of cirrus clouds. We utilize an application called JPLEX to study the persistent homology of multi-dimensional simplicial complexes built from available hyperspectral or multispectral data. The technique has been successfully applied to discriminate subtle features in high dimensional noisy data sets. Previous examples include anomaly detection in hyperspectral images. The analysis makes use of the entire multidimensional data set (not just one or a combination of spectral bands) which may offer advantages in discriminating among various cloud types in a scene, as well as determining other characteristics of cirrus clouds such as altitude and thickness. Our initial computational experiment with an AVIRIS scene has demonstrated that JPLEX is able to discriminate between cumulus and cirrus clouds.