Keywords : AVHRR NDVI, TWO, SZA

Abstract
Detection and monitoring of land cover change in global scale provides an important input to global change studies. Time series AVHRR data have potential to extract large land cover change areas. To pursue this potential, undesirable effects/components in time series AVHRR data should be removed. These undesirable effects can be divided into the following two types. The first one is high temporal frequency effects such as cloud effects and signal noises in data transmission. The other one is low temporal frequency effects such as the effect by solar zenith angle(SZA). This paper describes the preprocessing method to remove these effects for the use of NOAA/NASA Pathfinder AVHRR Land Data Set. In order to remove high temporal frequency effects, "Temporal Window Operation(TWO) method" was developed. This method is applicable for time series 10-day composite NDVI data. The investigation of the effect by SZA to NDVI reveals that the SZA less than 60 degree has little effect and that this range of SZA is recommended for the use of AVHRR NDVI data.

Introduction
The data used for this study are NOAA/NASA Pathfinder AVHRR Land(PAL) Data Set 10-day composite NDVI data. The PAL data are 8 km resampled data from Global Area Coverage(GAC) data with nominal resolution of 4 km. The period of the data authors used for this study is for 14 years from August, 1981 to September, 1994. Though these data have potential for global land cover change monitoring, the PAL NDVI data have undesirable components to be removed. In this study, authors divided these components into two categories. One is high temporal frequency effects such as cloud effects and signal noises in data transmission. The other is low temporal frequency effects such as the effect by solar zenith angle(SZA). In order to remove high temporal frequency effects, authors refined the Temporal Window Operation(TWO) method(Park 1998). In order to investigate low temporal frequency effects, homogeneous land cover area were selected and mode NDVI values of the area were extracted and analyzed because mode value is free from the effect of misregistration.

Removal of high temporal frequency effects
- Temporal Window Operation(TWO) method -

The TWO method consists of the following three steps
(Step 1) Search and the removal of high value noise (preprocessing of the TWO method)
AVHRR channel 1 and 2 data may have high or low value noises at the time of transmission and reception when the angle of an antenna becomes low. Though NDVI value becomes high or low due to this reason, only high NDVI noises remains through Maximum Value Composite(MVC) processing. Therefore the problem is how to eliminate high value noises. The proposed method is the combination of temporal threshold and spatial threshold.

The temporal threshold is the 115% of the maximum NDVI from three consecutive NDVIs before and after the examined 10-day composite NDVI. If the examined NDVI is larger than 115% value, it is considered as a noise.

The spatial threshold is calculated from mean(M) and standard deviation(SD) of the surrounding 24 NDVI values in 5 by 5 matrix region with the examined pixel as a center. The threshold value is M + 1.5*SD. If the examined NDVI is larger than the threshold, it is considered as a noise.

The threshold values were decided empirically.

(Step 2) Temporal Window Operation(TWO) method
The TWO method is based on the assumption that the real NDVI free from noises changes smoothly in a year with one or two peaks(local maximums) and bottoms(local minimums). In this case, NDVI changes monotonously between a peak and a bottom. The TWO method is carried out in a moving temporal windows with a predetermined temporal period(window size). Longer the window size is, there are more chance to have a real NDVI free from noises in the window. However, a window size longer than a period between a peak and a bottom has no meaning because monotonous NDVI change occurs only within this period. Figure 1 shows the NDVI value conversion by TWO method. The TWO method is carried out for time series 10-day NDVI data as follows.

1. Find the NDVI value at the start(first) point in a temporal window.
2. Search the nearest temporal point with larger NDVI than the start point in a temporal window.

If there is such a point, this point will be the next start point. If there is not such a point, the point with maximum NDVI in a window will be the next start point.

3. NDVI at points between the former start point and a new start point are converted by linear interpolation of two NDVIs of these consecutive start points.
4. iteration to 2.

Figure 1. Temporal Window Operation (TWO) STEP2