Wednesday, 17 December 2014
Nagendra Singh1 and Ranga Raju Vatsavai1,2, (1)Oak Ridge National Laboratory, Oak Ridge, TN, United States, (2)North Carolina State University at Raleigh, Center for Geospatial Analytics, Raleigh, NC, United States
MODIS data has been extensively used to study vegetation and crop phenological characteristics and monitoring. Recently, land cover change detection techniques have been proposed that identify changes in normalized difference vegetation index (NDVI) time series data collected by applying time series change detection techniques but these studies do not address most of the challenges associated with the land cover change detection problem. In this study, we present a novel un-supervised change detection method based on computing similarity between two given time series images (MODIS NDVI bi-weekly time series images). The widely used K-Means clustering is not very effective in modeling temporal correlations and crop phenology in particular. In addition clusters have to be manually labeled in order to identify changes. Our proposed research addresses these two limitations. Basic methodology is based on premise that if there is no change between two years, then the corresponding time series (and phenology) should be very “similar.” There are several measures for estimating the similarity between two give time series. Here we have adopted the widely used “Fréchet distance” as measure of similarity between two time series. Unlike regular clustering based time series, where year-wise time series are clustered separately, our algorithm builds a single cluster model for a given pair of year-wise time series. Using this single cluster model, each year-wise time series are then clustered into a given number of groups. Therefore, a direct comparison can be made between any two given clustered images to identify changes. We test our approach using data for years 2006 and 2011 and evaluating it against NLCD and USDA class labels. Results, challenges, and future research directions are discussed in details.