Spatio-Temporal Change Detection from Multidimensional Arrays: detecting deforestation from MODIS time series

Growing availability of long-term satellite imagery enables change modeling with advanced spatio-temporal statistical methods. Multidimensional arrays naturally match the structure of spatio-temporal satellite data and can provide a clean modeling process for complex spatio-temporal analysis over large datasets. Our study case illustrates the detection of breakpoints in MODIS imagery time series for land cover change in the Brazilian Amazon using the BFAST (Breaks For Additive Season and Trend) change detection framework. BFAST includes an Empirical Fluctuation Process (EFP) to alarm the change and a change point time locating process. We extend the EFP to account for the spatial autocorrelation between spatial neighbors and assess the effects of spatial correlation when applying BFAST on satellite image time series. In addition, we evaluate how sensitive EFP is to the assumption that its time series residuals are temporally uncorrelated, by modelling it as an autoregressive process. We use arrays as a unified data structure for the modeling process, R to execute the analysis, and an array database management system to scale computation. Our results point to BFAST as a robust approach against mild temporal and spatial correlation, to the use of arrays to ease the modeling process of spatio-temporal change, and towards communicable and scalable analysis.