- Helmholtz Centre for Environmental Research
- Image Processing Laboratory (IPL), Universitat de València
Forest mortality is a complex phenomenon because of the interaction of multiple drivers over
a long period. Understanding these interactions and their relevant time scale is important for
forest management. Unlike climate data which are continuous (daily or hourly resolutions),
forest mortality related observations are discrete (e.g. the number of trees, mortality fraction)
with lower frequencies (once/twice a year). Forests also have persistent memory with a large
buffering capacity. All the above-mentioned reasons make the analysis of forest mortality
difficult with the conventional tools. Deep learning is well suited for modelling multivariate
time series with persistent non-linear interactions. In this study, we generate 200,000 years of
hourly climate data using a weather generator (AWE-GEN). We aggregate the hourly data to
daily values and feed it to a process based forest model (FORMIND). The forest model gives
us mortality fractions per year, in line with the forest mortality related observations. For the
method development phase, we use these simulated data. First, we use a variationalautoencoder
to extract climatic features and use that for the prediction of forest mortality. In
the second stage, we do the prediction of forest mortality and feature extraction together and
illustrate the difference between the extracted features. Further different approach to
disentangle the extracted feature are tested. Finally, we present the analysis of performance
versus explainability for different approaches.