Mapping short-rotation plantations at regional scale using MODIS time series: Case of eucalypt plantations in Brazil Guerric le Maire a, ⁎, Stéphane Dupuy b , Yann Nouvellon a,c , Rodolfo Araujo Loos d , Rodrigo Hakamada e a Cirad, UMR Eco&Sols, Montpellier, France b Cirad, UMR TETIS, Montpellier, France c Atmospheric Sciences Department, USP, IAG, São Paulo, Brazil d Technology Center, Fibria Celulose S.A., Aracruz, ES, Brazil e International Paper, SP 340 road, km 171, Mogi Guaçu SP, Brazil abstract article info Article history: Received 3 December 2013 Received in revised form 28 April 2014 Accepted 27 May 2014 Available online xxxx Keywords: Fast-growing plantations Eucalypt MOD13Q1 Vegetation indices Landsat Bounding Envelope Time series pattern analysis Pattern recognition Subsequence matching Mining time series data Short-rotation plantations are extending worldwide due to the increased demand for pulp and wood. Reliable es- timations of recent expansion of short-rotation plantation areas and associated land use changes are a prerequisite to assess their environmental impact on regional carbon and water cycles, and on climate. A binary classification methodology using MODerate resolution Imaging Spectroradiometer (MODIS) 16-day 250 m NDVI time series was developed and applied to classify Eucalyptus plantations across Brazil. The identification of Eucalyptus planta- tions specific patterns in the time series was based on the calculation of matching functions between the NDVI time series and a ~2 years long reference time series. Among the seven tested matching functions, the bounding enve- lope was the most successful. This method was robust to residual noise on the NDVI time series, and a threshold coefficient for the binary classification was adjusted using an omission-commission criteria. With this method, it was possible to detect any presence of Eucalyptus between 2003 and 2009 at monthly time-steps, including the periods of bare soils between two rotations that are typically 6–7 years long. The dates of first afforestation, of clear-cut at the end of a rotation, and of re-planting at the beginning of a new rotation were retrieved from the NDVI time series with a precision of ~66 days. The final almost continuous tri-dimensional map (space and time) was validated with three different datasets, from local to regional data. All three datasets gave similarly high global accuracy statistics, but a global underestimation of Eucalyptus areas compared to large scales census was observed. Discrepancies and way to improve the Eucalyptus area estimates were discussed in this study. The developed methodology could be applied to other short-rotation tree plantations. © 2014 Elsevier Inc. All rights reserved. 1. Introduction Tracking the land uses and land cover changes at a regional scale is of critical importance to analyze the modifications of global biogeochemical cycles and the impacts of environmental policies. Several global land cover maps have been produced from classification of remote sensing data (MODIS land cover product, USGS-IGBP, UMD, GLC2000, GlobCover, etc.). The classification algorithms were an ensemble supervised decision trees, e.g. for MODIS MCD12Q1 product (Friedl, Sulla-Menashe, Tan, Schneider, Ramankutty, Sibley and Huang, 2010), unsupervised classifi- cation followed by post-classification refinement, e.g. USGS-IGBP prod- uct (Loveland, Reed, Brown, Ohlen, Zhu, Yang and Merchant, 2000), clustered supervised and unsupervised classification, e.g. GlobCover (Bontemps, Defourny, Van Bogaert, Arino, Kalogirou and Ramos Perez, 2011). Such global maps obviously have a small number of classes and have a coarse spatial resolution, and are therefore of limited interest to monitor the area covered by specific crops or plantations. In parallel to the development of these global maps, researchers have used the same satellite image resources to produce maps of crop classes at farm or land- scape levels in order to assess regionally and annually the land use changes of the main crops (e.g. Arvor, Jonathan, Meirelles, Dubreuil, & Durieux, 2011; Brown, Kastens, Coutinho, Victoria, & Bishop, 2013; Epiphanio, Formaggio, Rudorff, Maeda, & Luiz, 2010; Galford, Melillo, Mustard, Cerri, & Cerri, 2010; Wardlow, Egbert, & Kastens, 2007). All these studies have shown the potential of satellite image series to classify different crops and cropping systems, and therefore to assess the conse- quences of agricultural practices on land use changes. Indeed, the knowl- edge of the crop, forest or grassland phenology, together with their spectral signature, makes it possible to greatly improve the precision of the determination of subclasses. As a consequence, it is difficult to get a unified methodology and many different methods have been used to clas- sify coarse resolution satellite image time series for the production of crop maps, each method depending on the objective of the study and of the crop type under consideration (García-Mora, Mas, & Hinkley, 2011). Remote Sensing of Environment 152 (2014) 136–149 ⁎ Corresponding author at: UMR Eco&Sols, 2 place Viala - Bât. 12, 34060 Montpellier cedex 2, France. E-mail address: guerric.le_maire@cirad.fr (G. le Maire). http://dx.doi.org/10.1016/j.rse.2014.05.015 0034-4257/© 2014 Elsevier Inc. All rights reserved. Contents lists available at ScienceDirect Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse