Study Area and Materials

The forest area in the upstream basin of the Shihmen reservoir is selected as the study area (Figure 4-1). One meteorological station, GaoYi station (21C080), is selected to represent the forest area upstream from the Shihmen reservoir. The precipitation records of the station, which have been kept for more than 30 years, are collected.

The acquisition dates of the SPOT images and relative information are given in Table 4-1 and the six SPOT false-color images are given in Figure 4-2. For SPOT images, there are four spectral bands: green (0.51−0.59 μm), red (0.61−0.68 μm), near infrared (0.79−0.89 μm), and short wave infrared (1.58−1.75 μm), and a 20 meter spatial resolution. Every SPOT image is geometrically rectified, and the error of rectification is less than 10 meters. Other satellite images, with coarser spatial resolution (1.1 km at nadir) are also collected. Several AVHRR images taken in 2002 and 2004 are obtained from the Center for Space and Remote Sensing Research, National Central University. Thirty-nine images with less cloud covered are selected in year 2002, and 30 images for the year 2004; the dates of the selected images are listed in Table 4-2. AVHRR data include five spectral bands—red (0.58−0.68 μm), near infrared (0.725−1.1 μm), mid-wave infrared (3.55−3.93 μm), and two thermal channels (10.3−11.3 μm and 11.5−12.5 μm)—and provide information about vegetation and surface temperature.

The AVHRR data are preprocessed using WinChips software, which coordinates

Figure 4-1 Northern Taiwan and the upstream basin of the Shihmen reservoir.

Table 4-1 SPOT images used in this study.

Physical gain value Minimum radiance (W^/m²×μm×sr) Sensor Date

Band 1 Band 2 Band 3 Band 1 Band 2 Band 3 SPOT-4 1999/05/11 1.935 2.28786 2.42568 26.17 11.80 10.31 SPOT-4 2000/05/09 1.467 1.83253 0.876* 31.36 18.01 22.83 SPOT-4 2001/05/25 1.4085 1.78991 1.93605 43.31 24.02 9.30 SPOT-4 2002/05/29 1.3545 1.76272 1.2735 31.91 18.15 7.85 SPOT-4 2003/05/07 1.3545 1.76272 1.2735 36.18 15.32 6.28 SPOT-2 2004/05/11 1.43772 1.2662 1.15374 70.94 18.95 17.33

*: Extremely low physical gain value due to different mode setting when the image was acquired.

Table 4-2 Selected AVHRR images with less cloud cover.

Dates of AVHRR image

2002 2004

2002-01-04-0511* 2002-07-22-0416 2004-01-03-0555 2004-07-06-0425 2002-01-06-0450 2002-07-24-0533 2004-01-04-0544 2004-07-24-0558 2002-01-10-0547 2002-08-19-0410 2004-01-10-0437 2004-07-25-0546 2002-02-17-0540 2002-08-27-0422 2004-02-10-0526 2004-08-01-0429 2002-02-22-0446 2002-08-28-0411 2004-02-14-0442 2004-08-03-0544 2002-02-26-0405 2002-08-29-0401 2004-02-22-0451 2004-11-20-0506 2002-03-07-0546 2002-09-17-0530 2004-03-11-0448 2004-11-29-0503 2002-03-08-0357 2002-09-18-0519 2004-03-22-0603 2004-12-02-0608 2002-04-06-0521 2002-10-14-0533 2004-04-11-0536 2004-12-09-0450 2002-04-08-0459 2002-10-15-0522 2004-04-21-0523 2004-12-11-0605 2002-04-14-0534 2002-10-16-0511 2004-05-03-0448 2004-12-12-0554 2002-04-15-0523 2002-11-05-0450 2004-05-11-0638 2004-12-13-0543 2002-04-22-0409 2002-11-08-0418 2004-05-16-0540 2004-12-16-0509 2002-04-23-0537 2002-11-09-0408 2004-06-02-0548 2004-12-29-0600 2002-05-02-0401 2002-12-04-0430 2004-06-10-0557

2002-05-10-0552 2002-12-05-0420 2004-06-28-0553 2002-06-19-0515 2002-12-13-0431

2002-06-21-0453 2002-12-14-0559 2002-06-24-0601 2002-12-15-0548

2002-12-17-0526

*: The last four digits represent the Greenwich Mean Time (GMT) of image acquired.

1999 2000

2001 2002

2003 2004 Figure 4-2 False-color representations of SPOT images from 1999 to 2004.

the images into a longitude and latitude system. The WinChips output data are albedos for two visible channels and brightness temperatures for three infrared channels.

However, the largest geometric registration error of the output images may be as large as 10 km. Therefore, manually geometric registration of each AVHRR image is conducted by selecting several ground control points (GCPs) around the boundary of Taiwan and controlling the registration error to within 1.1 km.

Besides the registration problem, the other difficulty for land resource monitoring is the presence of cloud cover in the study area. Cloud screening plays an important role in series satellite images applications. Therefore, a standard and efficient cloud screening procedure is necessary. Saunders and Kriebel (1988) applied a radiance band ratio, channel 2 divided by channel 1, and determined a threshold to identify cloud pixels in an AVHRR image. Welch (1998) mentioned that the spatial variability of the spectral signal increases when cloud cover present in an image. He utilized the GLCM (Gley Level Co-occurrence Matrix) method to assess the spatial texture of cloud. The variables of the GLCM, including angular second moment (ASM), contrast (CON), and entropy (ENT), are commonly used in texture analysis. The other method for cloud screening is based on the physical characteristics of clouds. However, prior knowledge is needed for the physical-based method.

The cloud screening procedure (see Figure 4-3) proposed in this study combines the reflective spectral features of cloud and the spatial texture characteristics. We assume that when cloud presents in a pixel, the sensor-received reflective radiance and the spatial contrast will increase. We exclude the feature of low cloud top temperature because the cloud top temperature may vary with seasons and types of cloud. The thresholds of channel 1 and channel 2, albedos equal to 15%, are first applied. If there is a pixel for which the albedos of channel 1 (a₁) and channel 2 (a₂) are both larger

than 15%, the pixel is probability covered by clouds; otherwise, the pixel will be assigned to a land surface. The next step is to calculate the band ratio (also called Q value) by the radiances received by channel 1 and channel 2, which can be expressed as (Saunders and Kriebel, 1988)

1 2

I

Q= I (4-1)

where I₁, I₂ are the received radiances of channel 1 and channel 2. The relationship between radiance and albedo is expressed as

( )

where the subscript represents the number of the channel (i =1,2). a is the albedo of _i channel 1 or channel 2 in percentage, and F is the extra-terrestrial solar irradiance _i in channel 1 or channel 2. The values of F₁ and F₂ are 133.2 and 243.1 (W/m2), respectively. The thresholds of Q value given by Saunders and Kriebel, 0.75~1.6, are modified to 1.0~2.5 which is more suitable for application in Taiwan. If a pixel’s Q value is within the suggested threshold, the pixel will be assigned to cloud covered;

otherwise, the pixel will be considered to represent land surface. After the above two steps, the pixels covered by thick cloud are identified. Thus, the pixels covered by thin cloud and near the boundary of the cloud mass are not filtered. To resolve this problem the contrast texture (CON) is applied which is sensitive to the boundary of cloud mass and some kinds of thin cloud. Therefore, the third step is to calculate the contrast texture, which is defined as

( ) ∑ ( )

respectively, which are determined from a 3 × 3 window (Figure 4-4). The g is the gray level and M is the maximum gray level presented in an image. The CON texture is calculated by the albedo of channel 2 (near infrared band) in this study. The maximum CON value may be larger than 30000 and varies with the images. We normalized the CON value by the maximum CON value (CON_max), which is expressed as

CONmax

CON_N = CON (4-4)

where the CON represents the normalized CON value. The selected threshold _N value of the CON equals 0.23. This value is given somewhat arbitrarily, but is _N capable of screening thin clouds and the boundary area of cloud mass (Figure 4-5).