Automatic Cloud Cover Assessment for SPOT Images
C.F.Chen, Y.S.Juang, and A.J.Chen
Center for space and Remote sensing Research
National Central University
Chung-Li, Taiwan, R.O.C.
Tel : 886-3-4227151-7624, Fax: 886-3-4254908
E-mail:cfchen@csrsr.ncu.edu.tw
Abstract
The cloud cover of the earth resources satellite images, e.g Landsat and SPOT, has always been the great concern to the ground receiving station and the users. For the ground receiving stations, the cloud cover is recorded as being one of the major meta data in the images archive, while for the users, the cloud cover frequently represents the unwanted information of the images. In this study, an automatic algorithm is developed to assess SPOT image's cloud cover. The algorithm consists of two major stages; threshold segmentation, and post processing . in the first stage, a threshold-based filter is designed to find out the pixels that normally have greater spectral responses in the images . in the second stage, the post processing is used recover the remaining cloud pixels that have been excluded from the initial segmentation stage. Te proposed algorithm has been compared with the manual method. The results indicate that the proposed method is able to reach a comparable level qualitatively and quantitatively with the manual assessment.
Introduction
The first resources satellite receiving station in Taiwan has started receiving and processing data since 1993. a tremendous amount of remote sensing data has been received and widely used many applications in this country since then. For the images obtained from the optical sensors, e.g, Landsat and SPOT, the cloud cover has always been a great concern to the users. In order to help the users to identify the candidate images, the receiving station has developed an images catalog system which offers the quick-look images and their cloud cover percentage. At present, the cloud cover identification and percentage assessment are manually analyzed on the monitor screen. The routine practice shows that the manual identification is a labor-intensive and time -consuming task for the analyst. This study attempts to develop an automatic method to detect cloud location and calculate the cloud percentage in SPOT quick, look images. The method consists of two major stages: threshold segmentation, and post processing. In ht first stage, because the cloud has relatively high spectral response than other features in the image, a threshold-based filter using the spectral statistics ( mean and standard deviation ) of IR, Red, and Green bands of SPOT images is designed to divide the cloud pixels form other features ( Lillesand and kifer , 1995). The aim of the first stages is to make certain of the pixels identified are exclusively the cloud. In the second stage, the post processing is used to recover the remaining clouds excluded from the initial segmentation stage, and eliminate the misidentified cloud pixels. The procedures include the region growing (Rosenfeld and Kak, 1982) and the shape recognition (Pratt, 1991). The region growing is used to find more cloud pixels form the cloud boundary identified during the first stage. The shape recognition is aimed to eliminate highly reflective landscapes (e.g. rivers ) that the grouped as the cloud pixels in the region growing step. The detail method will be addressed in next section. The following section will present the testing results. Some conclusion remarks are given in the final section.
