De-noising spectral signatures from shallow water bodies for water quality determination purposes
ISSN der Zeitschrift
Anforderungen an die Agrarinformatik durch Globalisierung und Klimaveränderung
Regular Research Papers
Gesellschaft für Informatik e.V.
Spectral signatures collected by means of field spectrometer from shallow water bodies are noise contaminated from the atmosphere or the sensor itself. Existing spectral analysis methods are based on the variation within the data; therefore, they are very sensitive to noise effects. Noise can obscure important features such as peaks, valleys, or peak widths, or make calculation of signal features such as slopes, areas and peak widths difficult. The filter approach should maintain the sharpest absorption/reflectance features in the original signal. The level of the noise highly depends on the atmospheric conditions during data acquisition such as clouds and wind. Each spectral signature even when taken from the same water body but on another date is processed and filtered according to noise level.. The application of the methods is based on the criteria that the selected model must smooth out high frequency noise while maintaining the smallest features that could be associated with biophysical attributes of the water (absorption troughs and reflectance peaks). We found that the higher polynomial or wavelet orders do not provide more optical information than the lower ones; simple models were selected accordingly. For normal weather conditions, it is best to use the DWT filter with sym5 or sym8 Symlet wavelet depending on noise level. Symlet wavelet seeks to preserve shapes of reflectance peaks and essentially performs a local polynomial regression to determine the smoothed value for each data point. This method is superior to Adjacent Averaging because it tends to preserve features such as peak height and width, which are usually 'washed out' by Adjacent Averaging. At the same time, in case of windy weather, the best filter was application of Savitzky-Golay filter with lower frame size (e.g. 17) and subsequent application of Discrete Wavelet Transformation with sym5 wavelet. In all filters a polynomial degree of 3 preserved best the shape of the spectra and has been used for all data.