Fig. 3. Predictability of African fire by leading time.

The predictability of African fire carbon emission anomalies is estimated using multiple machine learning techniques (MLTs) as a function of lead time. The predictability is represented by the squared correlation coefficient (R²) between the predicted and observed monthly anomalies (n = 60) of the regional average fire carbon emissions across the a northern and b southern African ecoregions. The assessed sets of predictors include previously identified atmospheric and socioeconomic predictors (blue), currently identified oceanic and terrestrial predictors (black), and the combination of all these predictors (red). The assessed models include season-specific models (filled circles), in which the MLTs are built and applied by season, and all-season models (open squares). The circles and squares indicate the mean R² across the 100 ensemble members of the best MLT (see the Methods section), and the vertical lines indicate the range of 10th and 90th percentiles of the 100 ensemble members.