Spatially explicit forest mortality forecasts are driven by autocorrelation, not ecological context

Abstract

A warmer and drier climate has increased the severity and frequency of drought and insect-induced forest mortality. Forest mortality is an autocorrelated phenomenon that results from a complex interplay between stressors and forest traits. Remote sensing enables us to measure forest mortality in different ways, including aerial surveys; changes in satellite imagery; and individual dead tree mapping in high-resolution imagery. We evaluated the role of autocorrelation and mortality data sources in forecasting drought and insect-induced forest mortality in the western United States. To achieve these objectives, we compared the performance of gradient-boosted regression models against naive autocorrelation models at continental and local scale. Further, we also evaluated bias and variability among observers in aerial surveys. At continental scale, we compared models trained on aerial surveys with a satellite-derived forest mortality dataset. At local scale, we compared models trained on aerial surveys with models trained on maps of individual dead trees. We found that gradient-boosted regression models slightly outperformed naive models across both scales. We also found that covariates related to autocorrelation were the primary drivers of mortality predictions and that model performance declined when they were excluded. This result was consistent across both scales of analysis and all mortality data sources. Furthermore, we found bias in aerial surveys where sites with prior-year mortality were re-surveyed more often than locations without mortality. Our findings showed that forest mortality forecasts based on remote sensing data are not sufficiently accurate to support forest management. To map future mortality risk, we advocate for scaling plot-level mortality models and incorporating accuracy assessments into aerial survey datasets.