Noise detection method for time-series vegetation index derived from remote sensing images

Abstract

A noise detection method for time-series vegetation index (TSVI) derived from remote sensing images. Firstly, unit root test is used to classify observation values of each pixel into a stationary series or a non-stationary series; for the non-stationary, an appropriate mathematical model is used to model discrete TSVI, then differences between actual observation values and prediction values of the model are calculated and recorded as a deviation. As the deviation has removed seasonal components, the non-stationary series is transformed into a stationary series. For a stationary series or deviation data, noise detection is performed based on the assumption that observation values are distributed within a certain range around mean values; then model fitting and noise detection are iteratively carried out with remained observation values—until the iterations reached the maximum number or no noise is detected at one iteration. The time series is then converted back to image space to obtain a noise mask and optimized. The present invention can obtain an accurate noise mask and improve reliability of land surface-related applications.