Abstract
Fire is one of the most relevant disturbances affecting terrestrial ecosystems globally, altering vegetation, soil, water, and atmospheric composition. Fire causes a non-permanent land cover change, through the removal of vegetation, the deposition of charcoal and ashes, and the exposure of soil; the temporal persistence of these changes is highly variable, ranging from a few weeks in tropical savannas to years in boreal forests. Global burned area products have been systematically generated in the past 20 years from several coarse spatial resolution (250 m - 1 km) Earth Observation (EO) systems. These products are the main input in global biomass burning atmospheric emission inventories, and in the most recent studies on the role of fire in the global carbon cycle and vegetation dynamics.
Because of the non-permanent nature of burned areas, the algorithms employed for the generation of global burned area products rely on the availability of daily or near-daily observations from coarse resolution EO systems. The high revisit frequency ensures that a sufficient number of cloud-free observations are generally available globally before burned areas disappear, with few exceptions in known locations of persistent cloud cover.
The systematic generation of moderate spatial resolution (10 m - 30 m) burned area products could potentially meet the needs of a variety of fire science and applications communities, and at different scales from global (e.g., pyrogenic carbon emissions estimation) to regional scale (e.g., environmental post-fire assessment and remediation decision support). Algorithms for the generation of moderate resolution burned area maps have been recently prototyped regionally and continentally, and have the potential for global implementation. However, ,moderate resolution sensors have reduced temporal resolution (e.g., 16 days for Landsat) compared to coarse resolution sensors (e.g., ~1 day for MODIS), which could potentially lead to omission errors in ecosystems where the spectral signal associated with burning events disappears quickly, and cloud cover limits the number of valid observations.
My dissertation focuses on estimating the combined effect of the impermanent nature of land cover change typical of burning events and the cloud cover, which reduces the number of valid observations available to detect burns, on global burned area mapping using Landsat data. The dissertation has three objectives. The first objective (Chapter 2) is to estimate the temporal persistence time of the signal associated with burned areas, stratified by ecosystem and land cover type, making use of the global, multiyear MODIS data record. The second objective (Chapter 3) is to evaluate the suitability of the MODIS-derived cloud mask as a proxy for Landsat 7 cloud observations. Finally, the third objective (Chapter 4) is to estimate the potential omission errors in a hypothetical global Landsat burned area product, due to the combined effect of reduced revisit frequency and cloud contamination.
Chapter 2 presents a global analysis of the burned area persistence time defined as the duration of the spectral separability of the burned / unburned areas mapped by the MODIS Global Burned Area Product (MCD64). The separability was computed by analyzing time series of normalized burn ratio (NBR) from nadir BRDF-adjusted MODIS reflectances (MCD43). Results showed that, globally, the median burned area persistence time was estimated as 29 days and 86.6% of the global area, as detected by MODIS, can be detected accurately only for up to 48 days. Furthermore, the results indicated that early and late fires had a shorter persistence time compared to fires burning in the central portion of the fire season. The results, therefore, indicate that the persistence time can be a limiting factor for mapping burned areas using moderate resolution satellite sensors, which have a low temporal resolution (e.g. Landsat 16 days, Sentinel 2A and 2B 10 days each, 5 days when used in combination).
Chapter 3 presents a comparison of Landsat and MODIS cloud data. Landsat 7 Enhanced Thematic Mapper Plus (ETM+) image cloud fractions over land were compared with collocated MODIS cloud fractions, generated by combining the MODIS-Terra global daily cloud mask product (MOD35) with the Landsat 7 ETM+ image footprints and acquisition calendar. The results showed high correlation between the MODIS and Landsat 7 ETM+ cloud fractions (R2 = 0.83), negligible bias (median difference: < 0.01) and low dispersion around the median (inter-quartile range: [-0.02, 0.06]). These results indicated that, globally, the cloud cover detected by MODIS Terra data can be used as a proxy for Landsat 7 ETM+ cloud cover at the Landsat World Reference System (WRS) scale.
Chapter 4 builds on the previous chapters and presents the potential omission error of a hypothetical Landsat global burned area product compared to the MODIS global burned area product. The Landsat omission error was estimated as the amount of burned area detected by MODIS that would not be detected by Landsat 7 because of the combined effect of the impermanent spectral signal associated with burned areas and missing observations due to cloud cover. The simulation was informed by the MODIS global burned area product (MCD64A1), used as fire mask to define the location and timing of burning, and the MODIS-Terra cloud product (MOD35), used to determine the number of post-fire cloud-free observations available following the Landsat 7 acquisition calendar and ground swath footprints. Globally, the resulting omission error was estimated as 19% of the average annual burned area detected by MODIS, with a maximum error over forest land cover (33%) and minimum over shrubland land cover (5%). The results were derived using the acquisition calendar of Landsat 7 only, however, thanks to the aggregation of data from over 15 years of acquisitions, the results can be extended to the other existing Landsat sensors, which are positioned on the same orbit shifted by an 8-days lag, and also to Landsat 9, which is planned to be launched in the same orbit of Landsat 7 by Spring 2021.
The findings of this research have implications for the future development of a global burned area product generated using moderate resolution EO data such as Landsat. The burned area persistence times provide an estimation of the period after the burning date in which burned areas can be mapped reliably and have implications on the length of the rolling periods, used in change detection algorithms to map burned areas. The potential omission error of a Landsat burned area product identified locations and times of the year in which the low revisit frequency of Landsat combined with the occurrence of clouds can have degrading effects on Landsat burned area maps accuracy.