Is time series smoothing function necessary for crop mapping? — Evidence from spectral angle mapper after empirical analysis

Abstract

Time series smoothing functions have been frequently applied to fit multi-temporal vegetation index for better extraction of plant seasonal/growing parameters. Questions are raised that whether the smoothing is necessary for crop mapping. Four time series smoothing functions, namely, HANTS, Savitzky-Golay (S-G), double logistics and asymmetric Gaussian, were used to smooth 23 MODIS 16-days composite NDVI images in one year. The effectiveness were compared through visual check, correlation coefficient R, root mean square error (RMSE), and local signal noise ratio (SNR). The best smoothing time series NDVI images, along with the original time series images, were then used to map corn and soybeans by spectral angle mapper (SAM) method and their mapping accuracies were compared. Comparison of smoothing results showed that S-G fitted data got the strongest correlation coefficient R, the lowest RMSE and lower local SNR. Comparison of mapping results further showed that time smoothing function does not improve the classification accuracy obviously with the same training sample and same temporal bands. The whole analysis indicates that it is the mapping method that matters more than time series smoothing function for classification precision.