Past 20 m cell deaths reduced Brier forecast error by 0.087% in three later BCI census intervals, 2000–2015

We tested whether past local tree deaths improve a forecast for the next census interval in one 50 ha forest plot at Barro Colorado Island (BCI), Panama. Earlier intervals set the model settings. Three later intervals, labelled 2000–2005, 2005–2010 and 2010–2015, tested the rule. Both models used species, starting diameter class and 100 m location. The added model used deaths in the same 20 m cell in the preceding interval. The test data had 629,419 tree records and 71,528 deaths. Mean Brier error fell from 0.0934715 to 0.0933902, a 0.08697% relative reduction. Its 95% spatial block resampling range was 0.06318–0.10999%. All three test intervals gave positive gains. Larger spatial blocks gave similar ranges. The gain remained after removal of the most common species code. The Brier gain was uncertain for trees at least 10 cm in diameter and when training used all earlier intervals. A 50 m cell rule also gave no clear gain. These results show a small gain for one stated rule and test period. They do not identify a cause of death or establish skill in other forests. This is a retrospective test in a frozen archive, not a prospective forecast trial. The package has source column extracts, every primary tree forecast, all development losses, spatial resampling draws, figures, source checks and offline reproduction code.

tree mortality · forest census · spatial prediction · Brier score · Barro Colorado Island · temporal validation · open data