Global burned area decreased in both 2003–2012 and 2013–2022 (GFED5.1), but local and regional trends of one decade did not predict the next decade's trends; a common relative decrease did

Question: does the location of a satellite burned-area trend in one decade predict the trend in the next decade? Data: GFED5.1 annual burned area at 0.25 degrees (2003-2022), the independent ESA FireCCI51 product (2003-2020), and the GFED5.1 near-real-time extension (2023-2025, VIIRS-based beta). Method: we compared linear trends between disjoint windows for 0.25, 1, 2.5, 5 and 10 degree cells and the 14 GFED basis regions. The main test splits each decade into odd and even years. Trends from the two halves of one decade give a reference for full persistence. We also used slope tests with a year-permutation null, 10-degree block bootstrap intervals, and forecasts of the next-window mean. Results: the global total decreased by 0.76 %/yr in 2003-2012 and 1.10 %/yr in 2013-2022 (12.2 % lower decade mean; FireCCI51 13.5 %). Across cells, same-decade trend halves had rank correlations of 0.13-0.34, but cross-decade correlations were 0.02-0.07, about 10-36 % of full persistence (upper 95 % limits 0.24-0.49 at 0.25-5 degrees). With 2003-2011 vs 2012-2020, GFED5.1 and FireCCI51 both gave cross-decade correlations of -0.06 to -0.01. Extrapolation of each unit's earlier trend forecast the next-decade mean worse than no change (skill -1.0 to -1.7 for GFED5.1, -2.3 to -3.3 for FireCCI51). One common relative decrease applied to all units forecast better than no change (skill +0.15 to +0.33 at 0.25-10 degrees, +0.62 for regions). A time-reversed check shows that regression to the mean gives little of this skill at 2.5 degrees or more. The sign of the regional trend changed in 8 of 14 regions. In 2023-2025 the NRT total was 4 % above the 2013-2022 mean, and no tested forecast was clearly better than no change. Conclusion: in 2003-2022, decadal local burned-area trends were mostly multi-year variability; the global decrease was a broad proportional decrease, not a continuation of earlier local trends. Limitations: one pair of decades; both products use MODIS; the 2023-2025 data use a different sensor; we did not test drivers.

burned area · fire · GFED5 · FireCCI51 · trend persistence · forecast skill · global change · remote sensing · regression to the mean · savanna