Fire and environmental dataset (2003 - 2019) for Phelps et al. (2022, Global Change Biology).
Associated manuscript abstract: Narratives of landscape degradation are often linked to unsustainable fire use by local communities. Madagascar is a case in point: the island is considered globally exceptional, with its remarkable endemic biodiversity seen as threatened by unsustainable anthropogenic fire. Yet, fire regimes on Madagascar have not been empirically characterised or globally contextualised. Here, we apply a comparative approach using MODIS remote sensing data (2003-2019), to determine relationships between Madagascar’s fire regimes and global patterns and trends. We demonstrate that Madagascar’s fire regimes are similar to 88% of tropical burned area, with shared climate and vegetation characteristics. Therefore, rather than a global exception, Madagascar’s fire regimes could usefully be understood as a microcosm of most tropical fire regimes, which contribute to global understanding of fire. We found that landscape-scale fire declined in grassy biomes across the tropics, and at a relatively fast rate on Madagascar. The island’s high tree loss anomalies (1.25 to 4.77x the tropical average) were not explained by any general expansion of grassy biome burning and were centred in forests rather than at forest-savanna boundaries, demonstrating that high rates of forest degradation were not explained by landscape-scale fire escaping from savannas into forests. Associated with forests, landscape-scale fire trends reflected important differences among tropical regions, indicating a need to better understand regional variation in the anthropogenic drivers of change. Unexpectedly, the highest tree loss anomalies on Madagascar were centred in environments without landscape-scale fire, where the role of small-scale fires (<21ha) is unknown. Madagascar’s fire regimes thus contribute two lessons with global implications: first, landscape-scale burning is declining in grassy biomes across the tropics and does not explain high tree loss anomalies on Madagascar. Second, landscape-scale fire is not uniformly associated with forest loss, indicating a need for more socio-ecological context around narratives of tropical fire and ecosystem degradation.

Fire: We defined fire regimes with Global Fire Atlas data (Andela et al. 2019), based on the MODIS dataset MCD64A1 version 6 (Giglio et al. 2018), available from 2003 - 2016. 12 fire characteristics were calculated and hierarchically clustered in order to define Madagascar’s fire regimes (column 4: "orig_Mada_regime"). The 12 fire characteristics (column 12-23) include yearly and monthly variables of average burned area (column 14-15), coefficient of variation (CV) in burned area (column 12-13), average fire number (column 18-19), CV of fire number (column 16-17), average fire size (column 22-23), and CV of fire size (column 20-21). Monthly CV characteristics reflect average annual seasonality, and yearly CV characteristics reflect year-to-year variation across the study period. Average fire season length was also calculated by the number of months burned per year (column 24: "BAseasonLength"). No landscape-scale fire (NLSF) regimes were identified as pixels without burned area (column 5: "NLSF_regime"). Similar fire regimes were identified across the global tropics (column 7-11) using a multivariate environmental similarity surface (MESS) index (Elith et al. 2010; Di Cola et al. 2017). A mutually exclusive subset of similar fire regimes were then selected (where MESS > 24: column 6), in order to make environmental comparisons between similar fire regimes on Madagascar and elsewhere in the global tropics.

Vegetation: Four vegetation characteristics based on NDVI (normalised difference vegetation index) were calculated to compare Madagascar's fire regimes with the global tropics, including monthly and yearly characteristics of average NDVI (column 27-28) and CV of NDVI (column 25-26). NDVI values were calculated from combined MODIS Terra [MOD09GA.006] and Aqua [MYD09GA.006] datasets (Vermote & Wolfe 2015a,b), after removing pixels with cloud cover and water and calculating a 5-day median moving window to reduce noise in the dataset. For the purpose of our study, we use NDVI as a comparative proxy for vegetation (leaf) density between regions and fire regimes, provided that average NDVI values correspond to tree cover (Hansen et al. 2013). In order to compare environments by vegetation type, we identified forests where NDVI values indicated high-stable vegetation (normalised mean NDVI > 80%; normalised annual CV of NDVI < 8%), and grassy biomes where vegetation was associated with open-variable NDVI (normalised mean NDVI < 80% and ≥ 6%; normalised monthly CV of NDVI > 8%). Mean NDVI and CV of NDVI limits – 80% and 8% respectively – were chosen based on a natural break observed in the NDVI values of low-variable fire regimes and no landscape-scale fire (NLSF) regimes (i.e. environments where landscape-scale fire was not detected), which corresponds to previously identified forest limits (Moat & Smith 2007).