Abstract

A multistep algorithm to detect cloudy objects in multispectral images is presented. Clustering spatial pixels by the k-means method and applying spectral criteria of cloudy/clear sky to fragments of obtained clusters are carried out in each step of the algorithm. One cloudy object is found in one step. Results of testing the algorithm on images from a sensor HYPERION (199 non-zero spectral bands in a 425 nm – 2400 nm interval under high spatial resolution of 30 m) are given. Images with discontinuous cloud cover above different surfaces (ocean, vegetation, desert, town, snow) are considered. An alternative method, in which the same spectral criteria are applied to each pixel, is also used in testing. Cloud masks obtained by both algorithms are compared. Mean spectra of obtained cloudy objects are given. The presented algorithm finds 1-3 cloudy objects corresponding to the brightness distribution in RGB images. Using the alternative algorithm (without preliminary clustering) leads to detection errors on the cloud edges. Three quality parameters are offered. The ratio of dispersion of "cloudy" spectra to dispersion of "clear" spectra is found to be most informative. This ratio should be much less than 1 when using a good cloudy mask.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call