Wavelets transform and Linear Spectral Mixture Model applied to MODIS time series for land cover change analysis

This work presents a methodology that uses fraction images derived from Linear Spectral Mixture Model and wavelets transform from MODIS time-series for land cover change analysis. Our approach uses MODIS/Terra surface reflectance images acquired from 2000 to 2006 time period. For this study, a test site was selected in the Mato Grosso State, Brazilian Amazonia, encompassing several landscape types as tropical forest, savanna, transitional forest, regrowth, deforested areas, croplands and pasture. The samples of land cover classes were collected during four field campaigns (2003, 2004, 2005, and 2006) to be used as ground truth. The linear spectral mixture model was applied to the MODIS surface reflectance images of RED, NIR and MIR spectral bands. This model generated the vegetation, shade, and soil fraction images. In the next step, the Meyer orthogonal Discrete Wavelets Transform was used for filtering the time-series of MODIS fraction images. The filtered signal was reconstructed excluding high frequencies for each pixel in the fraction images (soil, vegetation, and shade) of the time-series. This procedure allows to observe the original signal without clouds and other noises. The accumulated precipitation data were used for dynamic phenological analysis, which showed the temporal lags between wet season and vegetation growing stages. The results show that wavelets transform can provide a gain in multitemporal analysis and visualization on inter-annual fraction images variability patterns.