Cloud field segmentation via multiscale convexity analysis

Abstract

Cloud fields retrieved from remotely sensed satellite data resemble functions depicting spectral values at each spatial position (x,y). Segmenting such cloud fields through a simple thresholding technique may not provide any structurally significant information about each segmented category. An approach based on the use of multiscale convexity analysis to derive structurally significant regions from cloud fields is addressed in this paper. This analysis requires (1) the generation of cloud fields at coarser resolutions and (2) the construction of convex hulls of cloud fields, at corresponding resolutions by employing multiscale morphologic opening transformation and half-plane closings with certain logical operations. The three basic parameters required from these generated multiscale phenomena in order to accomplish the structure-based segmentation include (1) the areas of multiscale cloud fields, (2) the areas of corresponding convex hulls, and (3) the estimation of convexity measures at corresponding resolutions by employing the areas of cloud fields and areas of corresponding convex hulls. These convexity measures computed for multiscale cloud fields are plotted as a function of the resolution imposed owing to multiscale opening to derive a causal relationship. The scaling exponents derived from these graphical plots are taken as the basis for (1) determining the transition zones between the regimes and (2) segmenting the cloud fields into morphologically significant regions. We demonstrated this approach on two different cloud fields retrieved from Moderate Resolution Imaging Spectroradiometer (MODIS) data. The segmented regions from these cloud fields possess different degrees of spatial complexities. As many macroscale and microscale atmospheric fields are classified according to spatial variability indexes, the framework proposed here would supplement those existing atmospheric field classification methodologies.