National Aeronautics and Space Administration

This paper discusses the process of validating active fire "yes/no" binary fire detection products from moderate- resolution satellite sensors. General concepts and practical issues are illustrated by the validation of the Moderate Resolution Imaging Spectroradiometer (MODIS) active fire product in Siberia. Coincident Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) imagery is used to characterize spatial patterns of flaming at sub-MODIS pixel scale. It is shown that for proper evaluation reference fire observations are needed at the scale of the satellite pixel, as only 60% of the MODIS footprints contain single contiguous clusters of ASTER fire pixels. In Siberia the size of a single ASTER fire cluster within the MODIS footprint that has a 50% probability of being flagged as "fire" is ~60, compared to ~45 in the Brazilian Amazon, whereas previous radiative transfer simulations suggested similar detection probabilities. The lower-than-expected detection rates in Siberia are largely attributable to flaming underneath heavy smoke, which is not detected by the current MODIS algorithm. Pixel-based and cluster-based omission error rates are derived, and it is shown that the probability of flagging as "fire" a MODIS pixel which contains a given number of 30-m ASTER fire pixels is typically 3-5 times lower than detecting a contiguous cluster with the same number of ASTER fire pixels. The procedures described are recommended for a consensus active fire validation protocol, but with the inclusion of multiplatform sensor configurations to complement the near-nadir angular sampling from single-platform observations such as MODIS and ASTER on Terra, the characterization of fires within the satellite pixel, either by partitioning the pixel into flaming, smoldering, and unburned areas and assigning temperature values to each of them [3], [4], or by calculating fire radiative power [5], [6]. In many ways, evaluation of the binary detection and continuous characterization products requires different approaches. In this paper we focus on the validation of the binary active fire detection product only. Experience of the fire remote sensing community in producing global active fire maps and interaction with users has frequently uncovered a misunderstanding about the correct interpretation of satellite-based fire detection data [7]-[9] and to a consequent skepticism about product quality. A need therefore exists for explicit information on product accuracy, which can be achieved through validation. Validation is defined as the process of assessing by independent means the quality of the data products derived from system outputs [10]. Independent information on fires can be obtained either by the direct observation of fires by alternative means or by observing fire effects such as atmospheric emissions or land cover change.