Figure 8.4; Boundaries (black pixels) of all different groups: threshold 0.2K.
The assum ption of the spatial uniform ity o f SST m ust take into account some dynam ical processes occurring in the upper part of ocean such as ed dies, frontal regions and coastal upweUing. C ontinental-shelf tidal fronts have tem perature variations of a few ten ths o f a degree over length scales of l Okm. M esoscale eddies, however, often have tem perature anomalies of ±2iC on space scales of 50 to 200k m. Coastal upwelhng regions have tem perature variations of up to 5 K over distances of 50fc7n [RobS5, p p .227-228]. If we assume that SST anomahes show linear spatial variations on given scales, th e tem perature variation of O . l K over a 1k m length scale [ 5 K over 5 0km) can be the m ax im um spatial SST variation for A TSR im ages. The threshold of 0. 2 K was determ ined by taking into account the radiom etric noise of A T SR discussed in Section 2.2.1 as well as more abrupt SST variations. The threshold of 0. 2 K was shown to be suitable by m anual observation of the 18 detected cloud im ages as described in Section 8.4.
The spatial variabihty test detects the groups consisting of less than 10 pixels. These groups correspond to the targets w ith highly variable tem pera ture spatially. This test is similar to the conventional spatial coherence test which calculated the STD of the 3 x 3 pixel area around a given pixel. This conventional test, however, gives high standard deviation for a com pletely cloudy-free pixel adjacent to a cloud edge pixel. Therefore, it degrades the spatial resolution of the detected cloud image. Although a modified coherence
8 .3 Al g o r i t h m 1 2 6
test was proposed by Thiermann and Ruprecbt [TR92] showing more sensi tivity than the conventional approach, it is also subject to the same problem around cloud edges. The proposed m ethod using region segmentation, how ever, differentiates cloud-free pixels from the edge pixels of clouds, and it does not misinterpret cloud-free pixels. In addition, the calculation of the standard deviation requires tens of multiplications and additions for a single pixel. The region segmentation technique requires only two comparisons for a pixel and some housekeeping operations in th e grouping procedure and one comparison with the threshold of 10 pixels for each group. This reduces the computation time dramatically. In addition, this segmentation is directly used to d e te rm in e
the IR threshold between sea and clouds.
The IR threshold determ ination begins with detecting sea groups. The tem perature of clear sea is shown to be spatially uniform by the spatial variabil ity test. The groups with more than 1000 contiguous pixels, called hig groups^
were chosen as sea group candidates. Note th at low and uniform stratus clouds can also form big groups.
The sea groups were distinguished from the uniform cloud groups by examining neighbourhood regions. The boundaries of sea groups touch cloud edges only. Since the region segmentation thresholds were chosen such th at cloud-free pixels were contiguous, one sea group cannot touch another sea group. On the other hand, the uniform cloud groups are surrounded by sea and/or lower, warmer clouds. It is very unlikely th at a big uniform cloud group is completely surrounded by higher and colder clouds. In an extreme case such as when an image with the area of 512^m X h l2 k m is fully cloudy, there must
be a significant number of partially cloudy pixels. These partially cloudy pixels provide warmer neighbours to big uniform cloud groups, and therefore, they prevent a big uniform cloud group from being misinterpreted as a clear sea group in the proposed algorithm.
Therefore, the sea groups touch only cloud groups which have a lower tem perature. Since the big cloud groups touch the underlying sea groups, underlying cloud groups, or higher cloud groups, the big groups which do not touch the groups with higher tem perature are identified as sea groups.
In practice, pixels th a t are subject to error have to be considered. For example, a sea group can have four boundary pixels which touch a single error
8 .3 Al g o r i t h m 1 2 7
pixel with a higher tem perature. In order to solve this problem, the number of boundary pixels, Nb, adjacent to a pixel with a higher tem perature, Nh,
and also those adjacent to a lower tem perature, Ni, were counted as a part of the big groups. Since this procedure counts the number of the boundary pixels of the big groups, it adds only a small amount of computation. If a large enough num ber of pixels touch the groups with higher or lower tem perature to eliminate the effects of the error pixels, the big group is determined to touch the group with higher or lower tem perature. In other words.
if if
^ > 0.05 ^ > 0-05
then touches the warmer groups,
then touches the colder groups. (8.1)
This algorithm can, therefore, work over ocean with small and warm islands as fax as the circumferences of islands are much shorter (less than 5 percent) than th at of the big sea group which contains the islands.
After the selection of the sea groups, the Tni-nimn-m BT of the sea groups are found. It becomes th e threshold of the IR gross test. The block diagram of the overall algorithm is illustrated in Figure 8.5.
Region Segmentation j Spatial Variability Test Big Groups Determined
Threshold Determination IR Gross Cloud Test Distinguish Sea Groups from
Uniform Cloud Groups
Figure 8.5: Block diagram of the proposed cloud detection algorithm. If the image does not have some big cloud-free groups, i.e. almost whole
8 . 3 Al g o r i t h m 128
area in the image is covered by clonds, it is impossible to determine the thresh olds from the image. In th at case, the algorithm can use pre-determined thresholds for detecting cloud-free pixels which do not form big groups or the algorithm can determine th at the whole image is cloudy.
The problem of partially filled cloudy pixels is addressed here for the conventional spatial coherence test and for the proposed spatial variability test. Let To denote the maximum BT difference of two adjacent cloud-free pixels. T+2To T+To T T+To T T-To T T-To T-2To I I T+AT T T I T T T (a) (b) T+To T-To T T+AT T (c) (d)
Figure 8.6: Examples of the 3x3 areas: (a)(c) cloud-free area with the maxi mum BT variation, (b)(d) one partially cloudy pixel. (a)(b) is for the conven tional spatial coherence test and (c)(d) is for the proposed spatial variability test.
Figure 8.6 (a) shows a 3x3 cloud-free area with the maximum standard deviation, S T D = The threshold on the STD in the conventional spatial coherence test is, therefore, A partially cloudy pixel w ith a very small BT decrement A T (A T -C T ) as shown in Figure 8.6 (b), can be
8.4 Results and Discussion 129
detected only if
S T D ~ > — T„ A T |> 2v/3T„. ( 8 .2 )
9 9
As shown in Figure 8.6 (c), the BT difference threshold used in the proposed spatial variability test is To in order to connect tw o adjacent cloud- free pixels with the m axim um B T difference of To. A partially fiUed cloud pixel with a B T decrem ent A T as shown in Figure 8.6 (d ), can be detected if I A T 1> To. The proposed spatial variabihty algorithm can, therefore, detect partially cloudy pixels w ith much smaller fraction (To) than the conventional spatial coherence test (2\/3 T o ). This means that the proposed algorithm has a better capabihty of distinguishing partially filled cloudy pixels from clear sea pixels.
8.4
R esults and D iscussion
% V
( b )
Figure 8.7: D etected cloud im ages by the proposed algorithm: (a) IR gross cloud test, (b) spatial variability test. W hite pixels are cloudy.
Figure 8.7 (a) shows the detected cloud images by the IR gross cloud test and th e spatial variability test of the proposed algorithm is shown in Figure 8.7 (b). Compared to the result of the EO DC cloud detection algorithm illustrated in Figure 8.2, the proposed IR gross cloud test detected more cloud pixels by
8 .4 Re s u l t s a n d Di s c u s s i o n 1 3 0
virtue of the more accurate threshold. It can be seen th at uniform cloud groups of more than 10 pixels in size passed the spatial variability test but they were d ^ ^ ^ e d by the IR gross cloud test.
5000 Post-determined threshold 4000
2
3000J
2000 lOOO 250.0 260.0 270-0 280.0 290.0 300.0Figure 8.8: Histogram of the original i ^ i g ^ ^ d IR gross cloud test thresholds (pre-determined and post-determined).
The accuracy of the autom atic IR threshold determination is illustrated in Figure 8.8. The proposed algorithm determined an accurate threshold near the narrow peak of clear sea pixels in the histogram while the pre-determined threshold allowed many warm and/or partially cloudy pixels to pass the IR gross cloud test. By using the accurately determined IR threshold, the pro posed algorithm detected 14% more cloudy pixels from this image than the EODC algorithm.
The proposed algorithm was applied to 18 images with various types of clouds. They are ATSR IR ( llfim ) images over ocean. They were taken from October 1991 to February 1992. The image location is between 6.8°N/17.9®W and 15.4°S/20.9®W, which is a limited area over the tropical Atlantic. Since the contrast between uniform low cloud and sea surface is generally greater at mid-latitudes than at the tropics due to the smaller optical thickness and the lower amount of water vapour in the m id-latitude atmosphere [FC94], the
8 ,4 Re s u l t s a n d Di s c u s s i o n 1 3 1
chosen examples are considered to be more difficult cases. Figures 8.9 and 8.10 shows each image with its histogram, the threshold determined by the proposed algorithm (thick line) and EODC pre-determined threshold (thin line).
As shown in Figure 8.9 and 8.10, the IR thresholds determined by the proposed algorithm are much closer to clear sea peaks in BT histograms than than the pre-fixed thresholds. This means th at the thresholds determined by the proposed algorithm are likely to be more accurate than the pre-fixed thresholds.
There is an exceptional case. Figure 8.10 (m). This image is mostly covered by clouds and no big groups were found from the spatial variability test. The pre-determined thresholds by the EODC algorithm were too low to detect any cloudy pixels from the image shown in Figure 8.9 (a).
The accuracy of IR thresholds was assessed also by manual observation. The procedure of the m anual assessment is as follows. An original BT image is displayed on two adjacent canvases. The cloud mask which was determined by the IR gross cloud test with the thresholds determined by the proposed technique is overlaid on the original image displayed on the left canvas. The grey-level of the cloud mask corresponds to the mean tem perature of cloud- free pixels. The cloud mask determined by the EODC pre-fixed thresholds is overlaid on the right canvas in the same way. The amount of undetected cloudy pixels can then be measured visually by the amount of grey-level variation in the canvases. In order to help the visual assessment, the grey-level contrast in the canvases is controlled within various grey-level ranges by changing a colormap of the canvases. The result of this procedure is shown in Figure 8.11. The image shown in Figure 8.11 is the one in Figure 8.9 (i).
It can be seen from Figure 8.11 th at a number of warm and/or par tially filled cloudy pixels passed the EODC IR gross cloud test with pre-fixed thresholds while almost all cloudy pixels were detected by the proposed IR threshold determination technique. The poor IR threshold of the EODC algo rithm detected only a small num ber of high and cold clouds. A small grey-level variation in the left canvas is due to either the BT variation of clear sea or some partially filled cloudy pixels. These partially filled cloudy pixels which passed the IR gross cloud test were detected by the spatial variability test.
Figure 8.12 is an example of the same observation technique as th at