IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 44, NO. 7, JULY 2006 1765 A Sampling Method for the Retrospective Validation of Global Burned Area Products Luigi Boschetti, Member, IEEE, Pietro Alessandro Brivio, Member, IEEE, Hugh D. Eva, Javier Gallego, Andrea Baraldi, and Jean-Marie Grégoire Abstract—This work presents a design-based validation and cali- bration scheme for the Global Burned Area 2000 (GBA2000) prod- ucts. The objective of such a scheme is to assess the margins of un- certainty associated with the burned area products and to estimate calibration coefficients needed to convert burned pixel counts into areal estimates. As the validation of GBA2000 was performed long after 2000, and given the fact that burned areas are a predomi- nantly nonpermanent land cover change, the reference data are obtained from a set of Landsat-7 Enhanced Thematic Mapper Plus high-resolution remotely sensed data. A stratified sampling scheme is presented, specifically designed for the retrospective validation of burned area data; the scheme is based on combining informa- tion from two low-resolution burned area products (GBA2000 it- self and Globscar). The resulting stratification has been applied to the whole global GBA2000 dataset, and preliminary validation re- sults are reported for Africa. The conclusions highlight the limits of a retrospective validation exercise, and summarize some of the open issues in the validation of global burned area maps. Index Terms—Burn scars, low spatial resolution, remote sensing, validation. I. INTRODUCTION I N RECENT years, several studies have attempted to under- stand the influence of greenhouse gases and aerosols pro- duced by anthropogenic activities on the variations of the chem- ical composition of the atmosphere. It is widely recognized that a significant fraction of the total amount of atmospheric emis- sions is due to the biomass burning phenomenon; as a result, numerous studies focused on the estimation of such a contribu- tion, making use of active fires and burned areas detections from remotely sensed (RS) data. Several global fire-related thematic maps, obtained from low- resolution RS data, are available for year 2000, such as Global Burned Area 2000 (GBA2000), developed by the Joint Research Centre—European Commission) [1], Globscar (developed by Manuscript received October 12, 2004; revised January 3, 2006. The work of L. Boschetti was supported in part by the European Commission through a research fellowship. L. Boschetti was with the Joint Research Centre of the European Commission, 20133 Ispra (VA), Italy, and with the Consiglio Nazionale delle Ricerche, Isti- tuto per il Rilevamento Elettromagnetico dell’Ambiente (CNR-IREA), Milano, Italy. He is now with the the Department of Geography, University of Maryland, College Park, MD 20742 USA (e-mail: luigi@hermes.geog.umd.edu). P. A. Brivio is with the Consiglio Nazionale delle Ricerche, Istituto per il Rilevamento Elettromagnetico dell’Ambiente (CNR-IREA), 20133 Milan, Italy (e-mail: brivio.pa@irea.cnr.it). H. D. Eva, J. Gallego, A. Baraldi, and J.-M. Grégoire are with the Joint Re- search Centre of the European Commission (JRC), I-21020 Ispra, Italy (e-mail: hugh.eva@jrc.it; javier.gallego@jrc.it; andrea.baraldi@jrc.it; jean-marie. gregoire@jrc.it). Digital Object Identifier 10.1109/TGRS.2006.874039 the European Space Agency) [2] and World Fire Atlas (WFA), developed by the European Space Agency) [3]. The characteris- tics of these global products derived from low-resolution (1 km at nadir) RS data are summarized in Table I. Focusing on the two burned area products, the most notable difference is that, while Globscar was derived from RS data by applying a single global algorithm, GBA2000 was obtained by adopting ecosystem spe- cific algorithms. Globscar [2] is a burned area product obtained from Euro- pean Remote Sensing 2 satellite (ERS-2) Along Track Scanning Radiometer 2 (ATSR-2) data. The burned areas are detected by means of the combination of two algorithms. The first one [4] makes use of the near infrared (NIR) and thermal infrared (TIR) spectral channels; it adopts an adaptive, window-based approach exploiting the fact that unburned vegetation has higher NIR reflectance and daytime temperature than the other sur- faces. The increase of temperature that occurs over a burned surface during daytime is due to its increased solar radiation absorption and to the absence of evapotranspiration. Because of the adaptive thresholds, the maps produced are affected by systematic commission errors in some ecosystems; as a conse- quence, the algorithm has to be complemented by a second set of rules, based on fixed thresholds. The second algorithm per- forms six different tests, making use of ATSR-2 bands 2, 3, 4, and 6. A pixel is considered as burned only if it is flagged as such by both algorithms. Instead of adopting a single global algorithm for processing the whole dataset, in order to overcome problems such as phe- nology, varying snow cover and flooding or variations in fire intensity, the GBA2000 strategy was to develop a set of algo- rithms, each specific for an ecosystem. A network of scientific institutions was set and each partner institution was responsible for developing and testing an algorithm for their specific region and fire season of interest and expertise. Eight algorithms were developed and tested, and six employed for the final product; a comprehensive overview of all the algorithms can be found in [1]. While design-based validation exercises have been con- ducted for landcover products at a global scale [5], [6], up until now, no global burned area product has been validated, let alone validated with a design-based statistical methodology. A design-based accuracy assessment method is characterized by the selection of reference data via a probability sampling. Such reference data can be subsequently used to compute accuracy metrics, which take into account the sampling probability. Conversely, model-based methods base the inference upon the 0196-2892/$20.00 © 2006 IEEE