Comparison of selected noise reduction techniques for MODIS daily NDVI: An empirical analysis on corn and soybean

Remote sensing derived NDVI data is fundamental to crop monitoring and crop yield estimation research. The usefulness of NDVI relies on reducing noise caused by varying atmospheric conditions such as cloud, haze, and dust as well as by sensor viewing geometry. Different techniques have been applied for noise reduction of composite NDVI products. However, the noise reduction for daily NDVI remains a challenge compare to multi-day composite products as the noise is already reduced to a certain degree in composite products. To address this issue, this research conducts experiments on reducing the noise of MODIS daily NDVI by selected techniques and applying a unique two level filtering on two crops: the first-level filter to take best possible NDVI values and then the second-level filter to smooth the curve by interpolating values selected in the first filter. The Best Index Slope Extraction (BISE) and running average are selected as first level filter. Five other techniques such as first Fourier transformation, Savitzky-Golay filter, asymmetric Gaussian function, double sigmoidal function and double logistic function fitting are selected as second level filter. The performances of noise reduction techniques are evaluated based on the correlation coefficient between crop NDVI and crop yield. The result demonstrates that the overall performance of Best Index Slope Extraction (BISE) as first filter is better than running average. The combination of BISE and Savitzky-Golay filter (SavGol) revealed better performance over other techniques in MODIS daily NDVI noise reduction based on the coefficient of determination (R², 0.86 for corn and 0.80 for soybean) between area under NDVI curve and crop yield. The noise-reduced NDVI profile generated through this filter combination can explain 86% variability in corn yield and 80% variability in soybean yield.