A Simple Metod for the Cloud Detection over Land Using Daytime AVHRR Data
2. Detection of the cloud-free pixels
The AVHRR onboard NOAA-series has five (or found) channels at 1.1 km spatial resolution ranging from the visible to infrared spectrum (0.58-0.68, 0.725-1.10, 3.55-3.93, 10.3-11.3, 11.5-12.5 mm) (Planet, 1988). Data from these channels are calibrated in reflectance (channel 1,2 ) and brightness temperature (channel 3,4,5) using the given method in Planet (1988). So, they can discriminate the target area through two ways at least, the reflectance difference and the brightness temperature difference among targets. In general, cloud has a relatively higher reflectance and lower temperature than those of land and sea surface. The numerous previous works to identify the cloud from satellite remote sensing data employed this simple logics(Reynolds and Vonder Haar, 1977; Schiffer and Rossow, 1983; Chou et al., 1986; Seze and Debois, 1987; Derrien et atl., 1993; Saunders and Kriebel, 1993; Suh and park, 1993).
In this algorithm, five threshold tests were applied to the day time AVHRR data for the detection of cloud-free pixels. In this process, there were some possibility of incorrectly identifying the cloud-free pixels as cloudy ones. But this was the safest way to ensoure no escape of cloud contaminated pixels from detection as pointed out by Saunders and Kriebel (1988), because the goal of this algorithm was to find the purely cloud-free pixels for the accurate calculation of NDVI.
The overall process of this algorithm are composed of 7 steps (Fig. 1). This algorithm's principle is very similar to those of Saunders and Kriebel (1988) and Derrien
et al., (1993). Main differences of this algorithm from the others are in the way to set up cloud-free reference data.
Figure1. General description of the tests used in the cloud detection scheme (p,r,1,2,4 and 5 mean pixel, reference, channel 1, channel 2, channel 4 and channel 5, respectively)
Step 1 : To set up a clear sky reference data
Its aim was to set a reference data to be used in the next threshold tests. First, highness temperature (channel 4) composition (HBTC) for the period of 15-day was processed. After that the representative clear sky brightness temperature (CS-T4) for the each sub-area(size : 20 x 20 AVHRR pixels) was calculated by neglecting the lowest 10% brightness temperature to avoid the permanent cloudy pixels, and selecting and averaging the next lowest 40% brightness temperature. Second, the same processes were applied to channel 1 data (DX_R1). Finally, upper composition processes were applied to the channel 1 and 4 data for the period of 1 month(for the completely cloud free data) and calculated the LSD of each channel.
Step 2: To estimate dynamic thresholds
The final thresholds of each sub-area were determined as a combination of CS_T4(CS_R1), the cloud free land surface temperature (AD_T4) and reflectance (AD-r1) of the analysis data, and temperature (reflectance) difference (Diff_T4/Diff_R1) between land surface and cloud top. The temperature (reflectance) values (Diff_T4/Diff_R1) were adopted from Derrien et al., (1993) and Rossow (1988), respectively. The final thresholds of each sub-area(FT_T4(sa)FT_r1(sa)) were determined through the combination of three elements:
FT_T4(sa) = (CST_4(sa) + AD_T4)/2 - Diff_T4(10K)
FT_R1(sa) = (CSR_4(sa) + AD_R1)/2 - Diff_R4(7%)
Step 3: Brightness temperature test
The purpose of this test was to detect the low temperature pixels(high or middle-level cloud) If the brightness temperature of a pixel was lower than the FT_T4, then the pixel was masked as a cloudy one.
Step 4: Reflectance test
It aimed to detect the cloud(low-level cloud) characterized by high reflectivity and high temperature. If reflectance of a pixel was higher than the FT_R1 then the pixel was masked as a cloudy one.
Step 5: Spatial coherence test
Spatial coherence test was applied to the channel 2 data because the magnitude of LSD of monthly composted channel 2 data was smallest among them. In this process, the partly-contaminated pixels by cloud edge or small cumulus were targeted. It LSD of a pixel larger than that of reference data by 2.0 then the pixels was classified as a cloud-contaminated one.
Step 6: Split-window temperature difference test
The split-window channel temperature difference (T4-T5) can be used to detect semi-transparent cirrus cloud due to the different emissivities of the cloud at the two wavelength (Inoque 1985, 1987); Saunders and Kribel(1988); Derrien et at., (1993)). We used the look-up table developed by Saunders and Kriebel (1988) to detect the semi-transparent cirrus cloud. But we did not reflect the spectral difference in
emissivity according to the land cover type.
Step 7: Inverse relation threshold test
The channel 1(0.58-0.68mm) reflectivity of bare soil is higher than that of vegetation or forest, and the temperature of bare soil is higher than that of vegetation or forest in general. When a pixel had a little higher reflectivity and colder temperature than neighborhood ones had, it was classified as a cloud contaminated one.
