Using Time Series Segmentation for Deriving Vegetation Phenology Indices from MODIS NDVI Data Varun Chandola ∗ , Dafeng Hui † , Lianhong Gu ‡ , Budhendra Bhaduri ∗ and Ranga Raju Vatsavai ∗ ∗ Geographic Information Science & Technology Oak Ridge National Laboratory Oak Ridge, TN 37831-6017 Email: chandolav,vatsavairr,bhaduribl@ornl.gov † Department of Biological Sciences Tennessee State University Nashville, TN 37209 Email: dhui@tnstate.edu ‡ Environmental Sciences Division Oak Ridge National Laboratory Oak Ridge, TN 37831-6017 Email: lianhong-gu@ornl.gov Abstract—Characterizing vegetation phenology is a highly significant problem, due to its importance in regulating ecosys- tem carbon cycling, interacting with climate changes, and decision-making of croplands managements. While ground based sensors, such as the AmeriFlux sensors, can provide measurements at high temporal resolution (every hour) and can be used to accurately calculate vegetation phenology indices, they are limited to only a few sites. Remote sensing data, such as the Normalized Difference Vegetation Index (NDVI), collected using the MODerate Resolution Imaging Spectroradiometer (MODIS), can provide global coverage, though at a much coarser temporal resolution (16 days). In this study we use data mining based time series segmentation methods to derive phenology indices from NDVI data, and compare it with the phenology indices derived from the AmeriFlux data using a widely used model fitting approach. Results show a significant correlation (as high as 0.60) between the indices derived from these two different data sources. This study demonstrates that data driven methods can be effectively employed to provide realistic estimates of vegetation phenology indices using periodic time series data and has the potential to be used at large spatial scales and for long-term remote sensing data. Keywords-vegetation phenology; time series; segmentation I. I NTRODUCTION Vegetation (or plant) phenology is the study of the timing of seasonal cycles of vegetation activity, such as onset of greening in the spring, timing of the maximum of the growing season, leaf senescence, vegetation dormancy, and the total length of growing season [1], [2]. Many such events are recurring plant life cycle states that are initiated by environmental factors and are sensitive to climatic variation and change [3], [4]. Thus, phenological studies can be used to evaluate the effects of climate change. Vegetation phenological stages of crops are also important indicators in agricultural production, management, planning and decision- making [2], [5]. In addition, vegetation phenology is impor- tant for predicting ecosystem carbon, nitrogen, and water fluxes [6], [7], as the seasonal and interannual variation of phenology have been linked to net primary production estimation, crop yields, and water supply [8], [9]. Variability in growing season length is likely to have a direct impact on the ecosystem carbon and water balances [9], [10]. Char- acterization of vegetation phenology at site, regional, and national scales has been recognized as important for many scientific and practical applications [2]. Accurate assessment of phenological events, therefore, becomes increasingly vital for investigating vegetation-climate interactions, quantifying ecosystem fluxes and croplands managements [11]. 0 50 100 150 200 250 300 350 400 2000 3000 4000 5000 6000 7000 8000 9000 10000 Days NDVI Maximum NDVI Rate of Green up Start of Season (Green up) Duration of Growth End of Season Rate of Senescence Figure 1. Phenological Characteristics from NDVI Time Series for a Cropland Site (Bondville, USA, 2006) Vegetation phenology can be assessed using different approaches at scales from small plots to large spatial 2010 IEEE International Conference on Data Mining Workshops 978-0-7695-4257-7/10 $26.00 © 2010 IEEE DOI 10.1109/ICDMW.2010.143 202 2010 IEEE International Conference on Data Mining Workshops 978-0-7695-4257-7/10 $26.00 © 2010 IEEE DOI 10.1109/ICDMW.2010.143 202