Distance metric-based forest cover change detection using MODIS time series

Abstract More than 12 years of global observations are now available from NASA's Moderate Resolution Imaging Spectroradiometer (MODIS). As this time series grows, the MODIS archive provides new opportunities for identification and characterization of land cover at regional to global spatial scales and interannual to decadal temporal scales. In particular, the high temporal frequency of MODIS provides a rich basis for monitoring land cover dynamics. At the same time, the relatively coarse spatial resolution of MODIS (250–500 m) presents significant challenges for land cover change studies. In this paper, we present a distance metric-based change detection method for identifying changed pixels at annual time steps using 500 m MODIS time series data. The approach we describe uses distance metrics to measure (1) the similarity between a pixel's annual time series to annual time series for pixels of the same land cover class and (2) the similarity between annual time series from different years at the same pixel. Pre-processing, including gap-filling, smoothing and temporal subsetting of MODIS 500 m Nadir BRDF-adjusted Reflectance (NBAR) time series is essential to the success of our method. We evaluated our approach using three case studies. We first explored the ability of our method to detect change in temperate and boreal forest training sites in North America and Eurasia. We applied our method to map regional forest change in the Pacific Northwest region of the United States, and in tropical forests of the Xingu River Basin in Mato Grosso, Brazil. Results from these case studies show that the method successfully identified pixels affected by logging and fire disturbance in temperate and boreal forest sites. Change detection results in the Pacific Northwest compared well with a Landsat-based disturbance map, yielding a producer's accuracy of 85%. Assessment of change detection results for the Xingu River Basin demonstrated that detection accuracy improves as the fraction of deforestation within a MODIS pixel increases, but that relatively small changes in forest cover were still detectable from MODIS. Annually, over 80% of pixels with >20% deforested area were correctly identified and the timing of change showed good agreement with reference data. Errors of commission were largely associated with pixels located at the edges of disturbance events and inadequate characterization of land cover changes unrelated to deforestation in the reference data. Although our case studies focused on forests, this method is not specific to detection of forest cover change and has the potential to be applied to other types of land cover change including urban and agricultural expansion and intensification.