Improved method for noise removal from hyperspectral vegetation spectrum using second generation wavelets

Many of vegetation studies make use of the vegetation reflectance spectra acquired by hyperspectral remote sensing technique. However, the hyperspectral vegetation spectra are highly noisy and the presence of noise affects the results of the spectral discrimination between vegetation species. Moreover, in hyperspectral studies it is common to perform analysis based on derivatives of the spectrum. This method is very sensitive to noise; therefore, noise removal is essential before performing derivative analysis. Also, to relate the cover reflectance to image reflectance in hyperspectral remote sensing imagery, noise-free field spectra are essential. Compared to traditional smoothing methods, wavelet transform shows promising results in denoising area. This thesis applied different types of wavelet transforms including discrete wavelet transform (DWT), lifting wavelet transform (LWT) which is the basis of the second generation wavelets, stationary wavelet transform (SWT),and the proposed method that is based on a combination of LWT and SWT methods that is called stationary lifting wavelet transform (SLWT). The objective of this research is to propose an accurate and stable method based on SLWT for noise removal from hyperspectral vegetation spectrum. The proposed method takes into account the characteristics of the vegetation reflectance spectrum and its results are compared with other three wavelet methods that are DWT, LWT, and SWT. These wavelet techniques were examined on a synthetic vegetation spectrum which is created by PROSPECT leaf model (a model of leaf optical properties spectra) and on several real-world vegetation spectra. To assess the effects of denoising several indicators including root mean square error (RMSE), signal-to-noise ratio (SNR), correlation coefficient and visual evaluation methods were employed. The experimental results showed that compared to other wavelet methods, the proposed method produced highly accurate statistical results. The best denoising results were acquired by applying Haar mother wavelet by making 13% improvement for SNR and by giving an RMSE as low as 0.0002 and correlation coefficient value of almost one. The visual evaluation showed that the proposed method preserves the absorption features and inflection points, as well as the wavelength positioning of local minima and maxima. Furthermore, the following novel results are concluded from this thesis: the proposed method is level-independent and narrow downs the choice of mother wavelet to a few low-order mother wavelets; as a result, it highly lowers the complexity of the denoising process. Unlike other wavelet-based methods, the proposed method gives reliable and predictable statistical results therefore is a stable method for noise removal from hyperspectral vegetation spectrum.

Preview PDF FK 2014 40R.pdf Download (1MB) | Preview

Actions (login required)

View Item