Increasing availability of dense time series of moderate spatial resolution satellite data for mapping global burned areas calls for mapping algorithms designed to easily integrate data at different spatial and temporal resolutions, irrespective of particular grid constraints. In this paper, we describe a novel hybrid approach for global burned area mapping that combines active fire data and time series of surface reflectance using graphs, which provide a flexible and efficient way of extracting spatiotemporal consistent patches. Our approach has three main steps. Firstly, we analyze burn-sensitive vegetation index time series to determine for each location a set of events, which are the dates for which the spectral-temporal signal indicates the possibility of a burn. Secondly, we explore the spatiotemporal distribution of all events and active fires to determine a subset of events with strong evidence of corresponding to burned areas. Those events are used as positive occurrences for training a one-class maximum entropy classifier and obtain, for each candidate event, a likelihood of it actually corresponding to a burn. Finally, we build a graph that combines all previous information, from which we extract spatiotemporal patches of densely connected events. Patches with strong evidence of burning determine the burned area map at any given time period. This research is part of the European Space Agency's Climate Change Initiative (ESA-CCI) and aims ultimately at generating Sentinel-3 daily global burned area products at 500 m spatial resolution. Towards that end, we test our approach with spatially and spectrally similar MODIS gridded surface reflectance (MOD/MYDO9GA), as well as non-gridded active fire (MCD14ML) 2008 data and CCI global land cover maps. Using 105 independent Landsat fire reference perimeters to validate global results, we show that our algorithm applied to MOD/MYDO9GA data (PT-M09) has very similar accuracy (44%) measured by the Dice coefficient compared with MCD64A1 v006 (45%). Moreover, PT-M09 exhibits a higher commission error but a lower omission error than MCD64A1. Due to their coarse resolution, this kind of product cannot capture very small burn areas. The bias relative to the reference burned area indicates that the algorithm presented in the current study underestimates burned area by 14% of the area actually burned according to reference data, which is lower than the underestimation in MCD64A1 (28%). We also analyzed the temporal accuracy of the patch based algorithm and concluded that the average time difference to active fires detection is 3.43 days, which is similar to MCD64A1 (3.75 days). Finally, we performed a sensitivity analysis which shows that total mapped burned areas varies only 3% when the main algorithm threshold that separates "burned" and "unburned" patches varies from quantile 45% to 55%.
原文链接
分享
引用
