A Hybrid Classification Algorithm for Segmenting Clouds from Multispectral Thermal Images

摘要

This paper reports on a novel approach to atmospheric cloud segmentation from a space based multispectral pushbroom satellite system. The algorithm consists of a hybrid-classification system in the sense that supervised and unsupervised networks are used in conjunction. The approach uses cloud and ground signatures as derived from 15 scenes which contain an assortment of climate, land, and cloud variations. In total 27 prototypes are used for the supervised portion of the classification. For performance evaluation, a series of numerical comparisons to human derived cloud borders were performed. A set of 40 scenes was selected to represent various climate zones with different land cover from around the world. These images contained several classes of varying cloud coverage (0% - 100%) and cloud types. A classification rate of 90% was found when using one short wave infrared (IR) band and one mid-wave IR band. A sensitivity of 75% and specificity of 90% were reported as well. A noise analysis was performed by adding Gaussian noise to the bands at SNRs ranging from 0dB to -30dB. Using two bands the algorithm produced a 66% classification rate at -30dB and a 75% rate at 0dB.