328 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 5, NO. 3, JULY 2008 Wavelet-Based Technique to Extract Convective Clouds From Infrared Satellite Images Bhupendra A. Raut, Student Member, IEEE, R. N. Karekar, and Dileep M. Puranik Abstract—Extraction of all cumuliform clouds from infrared satellite images is important for cloud studies. Existing methods have focused on extracting only the cumulonimbus clouds. Over monsoon Asia, warm cumulus and cumulus congestus clouds are a large fraction of total cumuliform clouds and are covered by cirrus. An extraction method based only on brightness tempera- tures (BT) is not sufficient for the detection of these cumuliform clouds. In this letter, a new cloud extraction technique based on spatial characteristics of the convective clouds is presented. The à trous wavelet transform (ATWT) is shown here to successfully extract both shallow and deep convective clouds. The depression in BT caused by the cold cloud tops corresponds to negative wavelet components found by the ATWT. Index Terms—À trous wavelet transform (ATWT), infrared (IR) satellite data, monsoon, warm convective clouds. I. I NTRODUCTION D URING the southwest monsoon season, a tropical east- erly jet (TEJ) persists in the upper troposphere over the Indian region. The TEJ blows the anvils of cumulonimbus to great distances, covering a very large area of the region with cirrus. The cirrus obscures the shallow convective clouds that make a major contribution to rainfall. To study this cloud sce- nario over the Indian (and Asian) region during the southwest monsoon, infrared (IR) satellite data are preferred because of the day as well as nighttime availability. Good spatial and temporal resolution of the present-day geostationary sensors is an added advantage. Although all cumuliform clouds are important and although cumulonimbus fraction in total cumuliform cloud cover is very small [1], the focus has been on extracting cumulonimbus. The methods to do so include the use of brightness temperature (BT) thresholds [2], gradients [3], and the use of water vapor channel in addition to IR [4]. In the monsoon region, the cirrus cover meets both the criteria of low BT [2] and significant gradients [3]. The result is misclassification of cirrus as cumuliform and the often missing cumulus and cumulus congestus. The objective of this letter is to present a robust method of extraction of all cumuliform clouds in IR satellite data in the presence of cirrus and other clouds. The dataset is from Meteosat-5 satellite with an image resolution of 5 km. Manuscript received August 6, 2007; revised January 3, 2008. This work was supported by the Indian Space Research Organization and was conducted at the University of Pune. The authors are with the Department of Atmospheric and Space Sci- ences, University of Pune, Pune-411 007, India (e-mail: baraut@ieee.org; dileepmp@unipune.ernet.in). Digital Object Identifier 10.1109/LGRS.2008.916072 Fig. 1. Schematic of clouds illustrating the concept of depression in BT (dBT) with clouds of different scales. Meteosat-5 is positioned at 63 ◦ E. The region covered by the data is 4 ◦ N to 33 ◦ N, 65 ◦ E to 94 ◦ E. II. EXTRACTION SCHEME Fig. 1 shows the schematic model of small cumuli (5 km size, 3–5 km height), cumulus congestus (10–15 km size, 5–6 km height), the rare cumulonimbus (20–40 km size, 10–16 km height), and extensive cirrus (hundreds of kilometers in size). Their BTs decrease with an increase in height. To extract these cumuliform clouds, the selected method should separate clouds including cirrus by size or scale. The extracted quantity would be the depression of BT (dBT) from the background. Wavelet transforms (WTs) are a class of tools that perform scale analysis. At every stage, the image scale doubles, and ob- jects of up to double in size are successively revealed. However, typical WTs shrink (i.e., decimate) the images at larger spatial scales. Hence, in the cloud separation study, useful information would be lost should the extracted cloud feature change its size or shape. Hence, a nondecimating WT is required. The à trous wavelet transform (ATWT) [5], [6] is a nondecimating, dyadic algorithm in which the image retains its size at all stages. The ATWT allows for multiscale analysis. Translation invariance is maintained. Since information is not lost in deci- mation, the wavelet components are separately interpretable as radiance changes at each scale. The optimal scaling function for analysis is the B-3 spline [7]. The convolution mask used here is hB3(l)= {0.0625, 0.25, 0.375, 0.25, 0.0625}. Puranik and Karekar [8] used the ATWT for separating structures in thunderstorms in advanced microwave sounding unit-B data. To explain the working of the ATWT algorithm, a one- dimensional signal is used in Fig. 2, which shows one such 1545-598X/$25.00 © 2008 IEEE