AUTOMATED PRODUCTION OF CLOUD-FREE AND CLOUD 
SHADOW-FREE IMAGE MOSAICS FROM CLOUDY SATELLITE IMAGERY 
Min LI, Soo Chin LIEW and Leong Keong KWOH 
Centre for Remote Imaging, Sensing and Processing, National University of Singapore 
BLK SOCI Level 2, Lower Kent Ridge Road, Singapore 119260 - (crslimin, crslsc, ersklk)@nus.edu.sg 
KEY WORDS: Mosaic, Feature, Detection, Algorithms, High Resolution, IKONOS, SPOT, Imagery 
ABSTRACT: 
The humid tropical region is always under partial or complete cloud covers. As a result, optical remote sensing images of this region 
always encounter the problem of cloud covers and associated shadows. In this paper, an operational system for producing cloud-free 
and cloud shadow-free image mosaics from cloudy optical satellite imagery is presented. The inputs are several cloudy images of the 
same area acquired by the IKONOS or SPOT satellites. By mosaicking the cloud-free and cloud shadow-free areas in the set of 
images, a reasonably cloud-free and cloud shadow-free composite scene can be made. This technique is especially valuable in 
tropical regions with persistent and extensive cloud cover. 
l. INTRODUCTION 
Cloud cover is a big problem in optical remote sensing of the 
earth surfaces, especially over the humid tropical regions. This 
problem can usually be solved by producing a cloud-free 
mosaic from several multi-date images acquired over the same 
area of interest. In this method, an image containing the least 
cloud covers is taken as the base image. The cloudy areas in the 
image are masked out, and then filled in by cloud-free areas 
from other images acquired at different time. It is equivalent to 
the manual "cut-and-paste" method. The cloud-masking process 
can be automated by intensity-thresholding to discriminate the 
bright cloudy areas from cloud-free areas. However, simple 
thresholds cannot handle thin clouds and cloud shadows, and 
often confuse bright land surfaces as clouds. 
In this paper, we present an automated procedure for producing 
cloud-free and cloud shadow-free image mosaic from cloudy 
optical imagery, that is able to overcome the pitfalls 
encountered by the simple thresholding method. This method 
works for both multispectral and panchromatic images. In this 
procedure, the pixels are classified into clouds, vegetation, 
buildings or bare soil based on the pixel intensity, colour, size 
and shape features. Cloud shadows are automatically located 
from the knowledge of the imaging geometry and the intensity 
gradients at cloud edges. Each pixel/patch in each of the images 
is then ranked according to some predefined ranking criteria. 
The highest ranked pixels/patches are preferably used to 
compose the mosaic. 
2. DESCRIPTION OF THE ALGORITHM 
Figure 1 shows a schematic diagram of the system for 
operational production of cloud-free and cloud shadow-free 
mosaics from optical satellite imagery. 
2.1 Input Images 
The inputs to the system are multispectral/panchromatic images 
of the same region acquired within a specified time interval. 
The images are co-registered before being fed into the system. 
754 
2.2 Balancing of Grey Level 
The brightness of pixels at the same location from two different 
scenes will be slightly different due to the atmospheric effects, 
sun angles and sensor look angles during acquisition. This 
disparity is especially prominent in low-albedo vegetated areas. 
Therefore, it is necessary to balance the intensity of the patches 
so as to minimize the variation. An image from the set of input 
images is chosen as the reference image. The pixel values of all 
other images in the same set are adjusted according to 
Pa Ep t (5-0) (1) 
where P is the output pixel value, S is the input pixel value, 
P is the mean pixel value of an overlap area around the 
mask patch being processed from the reference image, Æ is the 
mean pixel value of the same overlap area from the image to be 
balanced. Please note that the grey-level balancing procedure 
must be applied to each band. 
2.3 Cloud and Cloud-Shadow Masking 
Initial cloud and cloud shadow masks are produced using 
simple intensity thresholds. However, bright pixels of bare soil 
or building may be confused with cloud pixels. Such confusions 
are resolved by making use of size, shape and colour 
information of the bright pixel clusters. Clouds that need to be 
masked out are much larger than individual buildings. Man- 
made features such as buildings and bare soil normally have 
simple geometric shapes. An automatic method has been 
developed not only to calculate the size of bright patches but 
also to detect the lines, simple shapes and colour of the bright 
land surface in order to eliminate improper masking of these 
buildings and bare soil as clouds by the initial intensity 
thresholds. We employ a technique based on a geometric model, 
solar illumination direction and sensor viewing direction, as 
well as the intensity gradient to automatically predict the 
approximate location of cloud shadows near to the cloud edges. 
Inte 
Th 
thr 
sol 
ini 
arc 
2.4 
In 
sti 
le 
da 
us 
pi: 
int 
pi 
ac 
In 
in 
th 
th 
or 
an 
or 
of 
an