Analysis of the use of spatio-temporal early warning metrics to estimate fire vulnerability resilience in the Amazon rainforest

Organisator Laboratory of Geo-information Science and Remote Sensing

wo 11 april 2018 11:30 tot 12:00

Locatie Lumen, building number 100
Droevendaalsesteeg 3a
6708 PB Wageningen
+31 317 481 700
Zaal/kamer 2

By Elke Hendrix (the Netherlands)

The current Amazon tropical rainforest’s "climate-vegetation" equilibrium is one of the ecosystems at risk of entering an alternative stable state. Heat, drought and tree mortality can cause the current equilibrium as a carbon (C) sink to shift to the "fire dominated tropical savanna" equilibrium, as a C source. The critical transition of the C balance can eventually have a huge impact on the global C cycle. However, it is poorly understood what processes drive the critical transition in ecosystem states. As a result, there is a need to identify early warning signals to help us have a better understanding of critical transition mechanisms. Early warning metrics (EWMs) are commonly used to characterize resilience under stressed regimes, by measuring the properties of time series. Therefore it is important to understand whether critical transitions are preceded by any spatio-temporal ecosystem dynamics. In this study we aimed to evaluate whether a series of EWMs, derived from remote sensing time series (i.e. LST, vegetation indices and VOD), can provide relevant knowledge for better predicting both forest fires and drought vulnerability. To do so, we implemented a baseline random forest (RF) model, constrained by the aforementioned remote sensing variables, which only characterize the contemporary conditions of the ecosystems. Further, we benchmark the baseline model against a RF model constrained both by the current condition and EWMs. It allowed us to better understand the importance of EWMs for predicting fire vulnerability on a local scale and drought stress on a regional scale. We showed that adding EWMs that are consistent with theoretical expectations based on critical slowing down increased the overall accuracy of the predictions by 4.71%, from 68.52% to 73.23%, for the fire vulnerability predictions and by 1.48%, from 85.69% to 87.17%, for the drought stress predictions. Among the different predictive variables tested, spatial autocorrelation derived from LST night datasets was the most important followed by spatial variance derived from LST night datasets. We conclude that the addition of EWMs shows potential for the prediction of fire and related drought stress. But critical note should be made that these additions are relatively small, further research is needed to evaluate whether EWMs can strengthen existing drought stress and related fire risk frameworks. The VOD dataset used in this research was too coarse to use in the RF models, but the time series showed significant signs of critical slowing down for the drought stressed areas. Further research is needed to assess whether a less coarse VOD dataset could increase the accuracy of the predictions for both drought stress and fire vulnerability.

Keywords: Tipping points; Early warning metrics; spatial early warning metrics; forest fires; NDVI; LST; VOD; NBR; critical slowing down; resilience; random forest.