全球变化知识资源中心

Monthly burned area (BA) data²² (MCD64A1; 08-2000 onwards) and annual land cover information²⁴ (MCD12C1.51; 2001 onwards) were scaled to the 0.25° resolution of the monthly precipitation data²³ (TRMM 3B43 version 7; 1998 onwards). Our study domain included all African grid cells with annual precipitation rates between 400 and 1,500 mm yr⁻¹ precipitation to focus on savannah regions. We developed a statistical model that aimed to explain variability in annual BA using antecedent precipitation (API) and changes in cropland extent (dCROP) for each grid cell. As input for the statistical model, we removed the mean seasonal cycle from the BA and precipitation data (Supplementary Information). We calculated the annual BA anomaly for each grid cell not on the basis of calendar years but as the sum of the BA anomaly from 5 months before the month of peak burning until 6 months after, to circumvent issues in northern Africa where the peak fire season is in December and January. We used data from August 2000 until June 2013, resulting in 12 annual cycles for all grid cells.

Building on earlier work²⁵, a multiple linear regression model was used to explain the BA anomaly for each grid cell (x, y) and year (t):

where β, b0 and b1 are optimally fitted parameters to minimize the sum of ordinary least squares of the errors (ε). Positive dCROP values indicate conversion of natural vegetation into cropland. The impact of dCROP on annual BA was found to vary mostly with actual cropland extent and mean annual precipitation. To prevent overfitting we clustered grid cells into cropland extent–mean annual precipitation bins, and determined the slope β_i for each bin (i) separately (Fig. 2b). The BA–precipitation response was more variable and the slope b0_{x, y} was therefore optimized for each grid cell individually. We included an intercept that varied between grid cells (b1) as done in ref. 25, representing the initial annual BA. Both explanatory variables are discussed below and in more detail in the Supplementary Information.

Antecedent precipitation is often thought to be the single most important driver of inter-annual variability of BA in savannas^{13, 25}. For each grid cell, the effect of precipitation on annual BA was explored by calculating Pearson’s r between antecedent precipitation and annual BA using averaging periods from 1 up to 24 months. In general, for a given grid cell short averaging periods will be negatively correlated with BA as precipitation shortly before the burning season increases fuel moisture; whereas longer averaging periods will be positively correlated with BA through the process of fuel build-up²⁵. It is therefore possible that precipitation has both a positive and a negative effect on annual burned area for each grid cell. However, we found that both effects are rarely important at the same time (∼3% of the grid cells when p < 0.1; see Supplementary Information). Therefore, we include only one precipitation–BA response variable per grid cell in the model, based on the strongest absolute response (positive or negative). API may therefore be estimated for each grid cell (x, y) and year (t) using:

where T is the averaging period (ranging from 1 to 24 months) for API that led to the highest absolute correlation (Pearson’s r) between the running mean of the monthly precipitation anomaly p′ and the annual BA anomaly and m is the month of maximum burning (Supplementary Fig. 1 and Section 2). As the effect of API may be positive or negative and is based on different T for each grid cell, we decided to optimally fit b0 for each grid cell. Although we used a statistical model and did not investigate underlying processes in more detail, confidence in our model may be derived from the strong correlation between API and the annual BA anomaly and from the clear spatial patterns in optimal T (Supplementary Fig. 1).

Including dCROP in the model is a logical step because the distribution of croplands has a considerable impact on current BA distribution in northern Africa (Fig. 2a), and because strong trends were observed over our relatively short study period. Owing to the limited confidence in inter-annual variation of the land cover product, we used the linear trend derived from the whole 2001 to 2012 study period.

As explained above, the short study period (2001–2012) constrains the number and ways that explanatory variables can be included in the multiple linear regression model, which is further discussed in the Supplementary Information. Trend maps (Fig. 1) are shown for all significance levels, because in general the less significant trends were part of spatial systems of significant and larger trends. Outliers may have affected some of the trends computed, especially trends observed in regions of high inter-annual climate variability, but this effect is thought to be marginal (for further details see Supplementary Information).

To investigate the role of ENSO as a driver of inter-annual variation in BA, we calculated for each pixel Pearson’s r between the annual BA anomaly (as defined above) and the mean MEI index over the same period (from 5 month before the peak burning until 6 months after; Fig. 3). We then computed mean inter-annual variation separately for positively and negatively correlated areas of northern and southern Africa (Supplementary Fig. 3).