Removal of Noise by Wavelet Method to Generate High Quality Temporal Data of Terrestrial MODIS Products

Time-series terrestrial parameters derived from NOAA/AVHRR, SPOT/VEGETATION, TERRA, or AQUA/MODIS data, such as Normalized Difference Vegetation Index (NDVI), Leaf Index Area (LAI), and Albedo, have been extensively applied to global climate change. However, the noise impedes these data from being further analyzed and used. In this paper, a wavelet-based method is used to remove the contaminated data from time-series observations, which can effectively maintain the temporal pattern and approximate the “true” signals. The method is composed of two steps: (a), timeseries values are linearly interpolated with the help of quality flags and the blue band, and (b), time series are decomposed into different scales and the highest correlation among several adjacent scales is used, which is more robust and objective than the threshold-based method. Our objective was to reduce noise in MODIS NDVI, LAI, and Albedo timeseries data and to compare this technique with the BISE algorithm, Fourier-based fitting method, and the SavitzkyGolay filter method. The results indicate that our newly developed method enhances the ability to remove noise in all three time-series data products. Introduction Time-series data for some land surface parameters, such as Normalized Difference Vegetation Index (NDVI), Leaf Area Index (LAI), and albedo, have been successfully used a wide range of fields. Many analysis methods have been developed from NDVI time-series data to (a) detect land-cover changes (Zhan et al., 2002; Friedl et al., 2002; Roy et al., 2002), (b) derive biophysical parameters for other models (Sellers et al., 1994; Moody and Johnson, 2001; Lu et al., 2003), and (c) monitor vegetation dynamics (Sakamoto et al., 2005; Beck et al., 2006). Observing the change of LAI in time and space plays a significant part in understanding and modeling the Removal of Noise by Wavelet Method to Generate High Quality Temporal Data of Terrestrial MODIS Products Xiaoliang Lu, Ronggao Liu, Jiyuan Liu, and Shunlin Liang land surface processes in the entire climate system (Running et al., 1988; Potter et al., 1993; Chase et al., 1996). Studying albedo time series plays a central role in global energy budget and climate forcing issues (Dirmeyer and Shukla, 1994; Dickinson, 1995; Roesch et al., 2002). After National Aeronautics and Space Administration (NASA) launched the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor aboard both the Terra and Aqua satellites, respectively, in December 1999 and May 2002, researchers were given an unprecedented way to get a variety of time-series data. However, these time-series data inevitably contain disturbances caused by cloud presence (Gutman, 1991), atmospheric variability (Huete and Liu, 1994), and aerosol scattering (Xiao et al., 2003). Noise degrades data and hinders analysis. To reduce noise, the Maximum Value Composite (MVC) method (Holben et al., 1986) is usually composited to get a higher percentage of clear-sky data. However, if the composite period is long, the land surface does not remain static; and if it is too short, the atmospheric disturbance cannot be removed effectively, especially in cloudy regions. For example, there exist many low quality pixels in 8or 16-day composite MODIS products (Moody et al., 2005). Several methods, based on interpolation of time series data, have been proposed to remove such noise and to reconstruct high-quality NDVI time-series data. These methods can be generally categorized into two general types. The first methods include removing noise in the time domain, such as the best index slope extraction (BISE) algorithm (Viovy et al., 1992), the asymmetric Gaussian function fitting approach (Jonsson and Eklundh, 2002), the weighted least-squares linear regression approach (Swets et al., 1999), the Savitzky-Golay filter approach (Chen et al., 2005), and the ecosystem-dependent temporal interpolation technique (Moody et al., 2005). The second type includes noise-removal methods in the frequency domain, such as Fourier-based fitting methods (Sellers et al., 1994; Roerink et al., 2000). Each of the approaches has advantages. Before time-series data can be utilized in advanced research applications, one of these data smoothing approaches must be applied. The BISE algorithm has been used to classify vegetation and forest types (Xiao et al., 2002). The Fourier-based fitting approach has been employed to derive terrestrial biophysical parameters (Moody and Johnson, 2001), and to classify land-cover types (Anders, 1994). Asymmetric Gaussian PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Oc t obe r 2007 1129 Xiaoliang Lu is with the Department of Earth & Atmospheric Sciences, Purdue University, CIVIL 550 Stadium Mall Drive, West Lafayette, IN 47907, and formerly with the Institute of Geographical Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China ([email protected]). Ronggao Liu, and Jiyuan Liu are with the Institute of Geographical Sciences and Natural Resources Research, Chinese Academy of Sciences, No.11A, Datun Road, Chaoyang District, Beijing, 100101 China. Shunlin Liang is with the Department of Geography, University of Maryland, 2181 LeFrak Hall, College Park, MD 20742. Photogrammetric Engineering & Remote Sensing Vol. 73, No. 10, October 2007, pp. 1129–1139. 0099-1112/07/7310–1129/$3.00/0 © 2007 American Society for Photogrammetry and Remote Sensing PMSRS-02.qxd 9/14/07 11:18 PM Page 1129