Vegetation Dynamics from Denoised NDVI Using Empirical Mode Decomposition

A novel approach to study vegetation dynamics is introduced, using the Empirical Mode Decomposition (EMD) to analyze NDVI time series. The NDVI time series which is nonlinear and nonstationary can be decomposed by EMD into components called intrinsic mode functions (IMFs), based on inherent temporal scales. The highest frequency component which has been found to represent noise is subtracted from the original NDVI series; thus smoothing the noisy signal. The different key features describing vegetation phenology have been extracted by analyzing the noise free signal. The lowest frequency component (last IMF) is the trend in the NDVI series. The trend in the series has been identified finding the Sen’s slope of last IMF, and the non-parametric seasonal Mann–Kendall test has been used to confirm the significance of the observed trend. The method has been applied on per–pixel basis to the SPOT Vegetation NDVI product covering Northeast India and surrounding regions for the time span of 1998–2009. Results show that the method has performed well in identifying the pixel clusters with significant trends. Hotspot regions with severe vegetation degeneration have been identified, and the relationship of the observed trends with the expected causative variables such as land use and land cover, topographic relief, and anthropogenic causes has been explored. The spatial locations of these critical regions closely matches with the findings of the previous studies carried out locally in the region, mainly indicating the shifting cultivation practice to be the main cause for land cover change.