On October 21, 2025, Henning Riecken will present his Innolab results on ” Modeling Tree-Level Beech Leaf Loss in Bavaria Under Climate Stress with Multimodal Deep Learning Architectures ” at 12:00 in seminar room 3, John-Skilton-Str. 4a.
From the abstract: European beech (Fagus sylvatica) is a key species in Central European forests, yet recent years have seen widespread decline in Bavarian forests, primarily driven by drought stress due to changing climate conditions. Since current assessments of leaf loss rely on costly and time-consuming field surveys, remote sensing offers an efficient and comprehensive way to monitor forest health over large areas. This Innolab project explores the potential of a multimodular Long Short-Term Memory (LSTM) model, which integrates high-temporal-resolution climate and Sentinel-2 time series with a Multilayer Perceptron (MLP) processing static environmental data, to predict yearly leaf loss at the single-tree level. The model’s performance was compared to state-of-the-art machine learning approaches to assess the added value of incorporating dynamic vegetation indices and climate time series alongside traditionally used static predictors. Testing such an approach is particularly beneficial for developing early-warning systems and scalable monitoring frameworks, enabling forest managers to detect leaf loss in time to deploy adaptive management strategies.:
Supervisor: Dr. Sarah Schönbrodt-Stitt, EORC
From the abstract: European beech (Fagus sylvatica) is a key species in Central European forests, yet recent years have seen widespread decline in Bavarian forests, primarily driven by drought stress due to changing climate conditions. Since current assessments of leaf loss rely on costly and time-consuming field surveys, remote sensing offers an efficient and comprehensive way to monitor forest health over large areas. This Innolab project explores the potential of a multimodular Long Short-Term Memory (LSTM) model, which integrates high-temporal-resolution climate and Sentinel-2 time series with a Multilayer Perceptron (MLP) processing static environmental data, to predict yearly leaf loss at the single-tree level. The model’s performance was compared to state-of-the-art machine learning approaches to assess the added value of incorporating dynamic vegetation indices and climate time series alongside traditionally used static predictors. Testing such an approach is particularly beneficial for developing early-warning systems and scalable monitoring frameworks, enabling forest managers to detect leaf loss in time to deploy adaptive management strategies.:
Supervisor: Dr. Sarah Schönbrodt-Stitt, EORC
