Seminar: Praveen Kumar

Tree mortality in Germany has increased due to repeated drought and rising climate variability. Earlier studies have explained many of these patterns, but most rely on limited field observations or fixed site models. In this work I present a data driven method that predicts yearly mortality across Germany by combining remote sensing, soil and forest structure information, and monthly climate data. Mortality labels come from a standing deadwood product based on unmanned aerial vehicle images and Sentinel time series. The model uses transformer based encoders to process static features, time series data, delta information, and yearly climate summaries, and is trained with a spatial cross validation scheme based on semivariance clusters. Results show strong accuracy with reliable uncertainty and highlight the influence of soil fertility, rooting conditions, forest structure, and water related climate stress. The approach provides a scalable way to assess mortality risk under changing climate conditions.